2024-03-01 00:00:00 CAFIAC FIX

2024-03-11 00:00:00 CAFIAC FIX

2023-05-19 19:00:00 CAFIAC FIX

2023-05-21 19:00:00 CAFIAC FIX

2023-04-21 00:00:01 CAFIAC FIX

2023-04-04 19:09:28 Many people are talking about the Belgian chatbot death, but few seem to familiar with the facts.Here'

2023-04-03 03:02:18 @ceperez ha ha, we certainly do not have AGI and won’t this time next year, moratorium or no.

2023-04-03 03:01:12 @BartWronski I would guess that you haven’t read any of my many critical discussions of Tesla

2023-04-03 02:59:36 @chrisoffner3d @BartWronski absolutely not, and I will continue to point out concerns i have with Tesla etc, as i have for the last several years.

2023-04-02 17:26:53 Opportunity Lost. The 6 month ban letter was flawed

2023-03-31 00:55:53 The 6 month ban is letter is not perfect, but a lot of the criticism misses the point. As a signatory and the person who broke the story on twitter and then endured endless attacks, my thoughts are here: https://garymarcus.substack.com/p/the-open-letter-controversy

2023-03-29 01:30:37 a big deal: @elonmusk, Y. Bengio, S. Russell, @tegmark, V. Kraknova, P. Maes, @Grady_Booch, @AndrewYang, @tristanharris &

2023-03-28 14:06:07 We desperately need to stop dichotomizing. AI poses serious risks BOTH short-term AND long-term. You don'

2023-03-22 16:52:59 @bwyble we will see about the first part

2023-03-22 10:41:23 Fantastic explainer on AI, hype and reality https://www.youtube.com/embed/Puo3VkPkNZ4

2023-03-22 08:24:23 Speaking of bullshit, anyone remember this?

2023-03-18 16:13:54 The not-so sweet smell of AI bullshit.https://garymarcus.substack.com/p/this-week-in-ai-doublespeak?sd=pf

2023-03-16 20:33:20 Exciting announcement, new title :) #TED23 https://conferences.ted.com/conference/ted2023/speakers

2023-03-14 00:18:57 @ceperez @kordinglab not sure i see what you mean by evolution in this context

2023-03-14 00:15:58 Should we worry more about short-term Al risk or long-term Alrisk? We can'

2023-03-13 19:57:53 AI, misinformation, and the 2024 elections:https://www.theatlantic.com/technology/archive/2023/03/ai-chatbots-large...

2023-03-10 04:05:53 What will criminals do with GPT? We don’t know. But it’s already clear we are in for some serious challenges.The Latest at The Road to AI We Can Trust:https://garymarcus.substack.com/p/the-long-shadow-of-gpt?utm_source=twit...

2023-03-06 05:33:47 Calling bullshit on national TV :) My 90 seconds of 60 Minutes fame: “I actually like to call what [ChatGPT] creates "

2023-03-05 10:00:00 CAFIAC FIX

2023-03-02 22:00:00 CAFIAC FIX

2023-02-28 06:24:15 @caseynewton @misc eg https://www.nytimes.com/2023/01/06/opinion/ezra-klein-podcast-gary-marcu...

2023-02-27 01:00:00 CAFIAC FIX

2023-02-19 08:15:31 could there be a news story more 2023 than this?

2023-02-18 16:18:24 David beats Go-liath! with some important lessons for claims about AI being superhuman, even in limited domains.https://garymarcus.substack.com/p/david-beats-go-liath

2023-02-17 23:21:23 How *not* to test GPT-3 (and how *to* think about that Theory of Mind paper people keep asking about), with Ernie Davishttps://garymarcus.substack.com/p/how-not-to-test-gpt-3?sd=pf

2023-02-11 17:45:50 Nightmare on LLM Street“.@garymarcus in absolute top form! I don'

2023-02-11 01:28:21 Saturday at 8am PT, at garymarcus.substack.com: Inside ChatGPT’s Heart of Darkness

2023-02-09 05:12:26 Oops! How Google bombed, while doing pretty much exactly the same thing as Microsoft did, with similar resultshttps://garymarcus.substack.com/p/oops-how-google-bombed-while-doing

2023-02-03 04:43:43 Happy Groundhog Day – the AI edition!https://garymarcus.substack.com/p/happy-groundhog-day-the-ai-edition

2023-02-02 14:40:07 ChatGPT ain’t rocket science!Brilliant @NPR story by @gbrumfiel. Listen &

2023-02-02 00:23:03 Gaslighting and reality in AI: Three years of epic @Ylecun and @garymarcus battles, summed up in a single convenient blog post. https://garymarcus.substack.com/p/gaslighting-and-reality-in-ai?sd=pf

2023-01-30 01:00:00 CAFIAC FIX

2023-01-16 18:07:21 @bibliotecaria @emilymbender emily has made things really hard for me by - misrepresenting my views- taking a single interview that I didn’t control to undermine a lifetime’s commitment to social justice- never citing me even though I frequently cite herThis makes me not want to promote her work even when I fully agree with it.

2023-01-14 18:33:30 @cgruenloh @DrVeronikaCH same for any LLM thus far or in near-term future!

2023-01-14 03:43:27 5 reasons why including ChatGPT in your list of authors is a bad ideahttps://garymarcus.substack.com/p/scientists-please-dont-let-your-chatbots

2023-01-11 06:02:44 Yet another AI reproducibility / reliability / generalizability crisis, this time in medicinehttps://www.nature.com/articles/d41586-023-00023-2

2023-01-08 16:34:03 How to get rich in AI:Propose an exciting new technology (driverless cars, virtual assistants, new-fangled search…)Promise fabulous resultsRaise a fortuneStumble, because products ≠ demosBlame the DataGather More DataStumble some moreRinse &

2023-01-08 14:39:19 @gullabi looks legit

2023-01-08 14:36:33 @gullabi yes that is the output I got and discussed on Ezra Klein

2023-01-08 04:07:21 further adventures in the ChatGPT vs Google revolution

2023-01-06 23:23:50 @vzocca that’s from two days ago!

2023-01-06 23:06:33 Um, what if ChatGPT *isn’t* as intelligent as it seems? Thrillled to chat w @ezraklein https://www.nytimes.com/2023/01/06/opinion/ezra-klein-podcast-gary-marcu...

2023-01-04 22:06:17 Vancouver BC

2023-01-04 17:49:53 @karabaic I clarified a bit after :)

2022-12-31 16:48:34 17 reasons why you should be worried about AI, as we go into 2023https://garymarcus.substack.com/p/an-epic-ai-debateand-why-everyone?sd=pf

2022-12-30 17:07:13 Is ChatGPT Really a “Code Red” for Google Search? https://garymarcus.substack.com/p/is-chatgpt-really-a-code-red-for

2022-12-28 21:15:40 It’s a myth to think you can build machines a value-neutral way.

2022-12-28 17:24:34 @abebab I don’t think that is the norm anymore.

2022-12-27 01:27:15 @orbiterlab @silke @sarahookr the cost of saying stuff like that is zero.

2022-12-26 22:32:57 epic summary of an epic debate: https://www.zdnet.com/article/ai-debate-3-everything-you-need-to-know-ab...

2022-12-26 03:44:20 @zeroexp nope but the video is at agidebate.com and I may put them there or write them up in a Substack

2022-12-25 17:47:07 What to Expect When You’re Expecting … #gpt4What comes after #ChatGPT? 7 predictions for 2023https://garymarcus.substack.com/p/what-to-expect-when-youre-expecting

2022-12-25 17:18:47 Woker-than-thou

2022-12-23 19:00:23 Live, tonight, from all over the globe, many of #AI’s leading thinkers. Agidebate.com@dileeplearning @bengoertzel Noam Chomsky @kordinglab and many more!

2022-12-20 22:39:31 Montreal, earlier today

2022-12-20 15:10:54 AI Platforms like ChatGPT Are Easy to Use but Also Potentially Dangerous (adaptation at Scientific American of my Jurassic Park essay)https://www.scientificamerican.com/article/ai-platforms-like-chatgpt-are...

2022-12-19 18:04:46 @vzocca OTOH this alternative question is plainly interpretable as a hypothetical word problem and it still fails

2022-12-19 15:32:59 ChatGPT mansplaining @ElonMusk and @Twitter https://t.co/W9rH9blJ2l

2022-12-19 15:26:33 ChatGPT stumbling a bit on today’s news…

2022-12-19 12:22:20 @GunnarBlohm @dileeplearning yes

2022-12-19 02:36:05 Elon’s poll is apparently just to save face! He was apparently told to step down prior to the poll:https://twitter.com/iamraisini/status/1604630171205529601

2022-12-19 02:12:43 This Friday, Dec 23, 5pm ET…Noam Chomsky, Erik Brynjolfsson, Yejin Choi, Canadian MP Michelle Garner Rempel, @dileeplearning, Ben Goertzel, Kai-Fu Lee, Jürgen Schmidhuber, Francesca Rossi, Anja Kaspersen, Jeff Clune, Artur d'

2022-12-19 01:58:18 RT @Montreal_AI: This Friday, Dec 23, 5pm ET… Noam Chomsky, @erikbryn, @YejinChoinka, MP @MichelleRempel, @bengoertzel, @kaifulee, @Sc…

2022-12-19 01:52:05 @taylorlorenz insanity. and, perhaps his final and ultimately reversible, act in office?

2022-12-19 00:52:42 @tdverstynen interesting thread

2022-12-19 00:48:19 @tdverstynen that particular ship has already sailed

2022-12-19 00:44:54 Ok, this is an interesting poll, and worth a brief trip to Birdland:https://twitter.com/elonmusk/status/1604617643973124097?s=46&

2022-12-18 21:29:41 From free speech to fascism in 7.5 weeks

2022-12-18 21:28:35 @futurology101 undoubtedly there will be some positive some negative

2022-12-18 21:27:53 @vzocca waiting on an editor. Email me via my contact page if you want a peek.

2022-12-18 17:44:51 @Riedl have heard that but not sure whether or not true.

2022-12-18 16:17:03 @eibriel ps I actually have 7 specific predictions but that’s a pretty good tl

2022-12-18 16:13:08 @eibriel darn, did you read the draft?

2022-12-18 15:15:13 @schoppik am sure it could, but doubt it would get the details right…

2022-12-18 14:49:11 working on an essay on my predictions for GPT-4

2022-12-17 01:44:19 @alepouda @rachelwilliams @vzocca

2022-12-17 00:54:43 @paulbalogh No short answer

2022-12-17 00:07:34 @marameo awesome example!

2022-12-17 00:06:42 @alepouda @rachelwilliams A. You didn’t precisely replicate what was doneB. The systems are stochastic C. That’s the problem

2022-12-17 00:03:39 It’s taken the neural network community 30 years to almost sorta hear me out on this

2022-12-17 00:00:51 @rachelwilliams it’s not actually remotely trivial. Eg how do you write a module that takes the output of the white bourse example and reliably give the right answer (white)? The system does *that* is the system that is actually intelligent.

2022-12-16 23:53:39 @marameo indeed it would be very easy to abuse to create fake info for selling an ads

2022-12-16 23:52:14 @rachelwilliams there are many examples like this

2022-12-16 23:32:49 Guardrails don’t come free

2022-12-16 19:18:00 @DrYohanJohn @PessoaBrain@neuromatch.social @LeslieKay @PessoaBrain@mastodon.social @dr_alexharris @tdverstynen It’s extremely challenging to be a critic in a world where people prefer to believe in fantasies.

2022-12-16 19:16:50 @pchiusano @filippie509 @melaniemitchell I prebunked it, and maybe ought write a reply https://www.nytimes.com/2015/06/28/opinion/sunday/face-it-your-brain-is-...

2022-12-16 19:16:01 @beltsbear then again for a car company it’s still very expensive. we’ll see …

2022-12-16 18:08:15 Elon Musk, Oct 20: “Tesla Could Be Worth More Than Apple And Saudi Aramco Combined” Stock Market, two months later: “Not til you get out of Twitter, that’s for damn sure”

2022-12-16 16:11:08 @LeslieKay @PessoaBrain @dr_alexharris @DrYohanJohn @tdverstynen the great past tense debate (pinker, McClelland, myself, etc) anticipates many of the current themes in AI and is still very relevant.

2022-12-16 15:21:53 @thegradient @drewharwell my server feels your pain

2022-12-16 04:10:02 So much for free speech

2022-12-15 16:07:15 @prem_k creeps me out that 1. the references are fake2. You could just as easily write a totally bogus letter as a legit letter

2022-12-12 20:46:48 @abebab update Abeba is where she should be: https://womeninaiethics.org/the-list/hall-of-fame/

2022-12-12 20:09:14 Great list of brilliant women in AI Ethics, to which I would unhesitatingly add @abebab: https://medium.com/women-in-ai-ethics/announcing-the-100-brilliant-women...

2022-12-12 12:39:02 @justinhendrix doesn’t seem like the 44B has improved his approval ratings the way he hoped?

2022-12-12 01:03:02 @martin he has resisted my efforts to talk to him about that very topic, so I doubt it.

2022-12-12 00:51:54 @anclement I expect a lot of the longer content to be LLM generated dreck, but who knows

2022-12-12 00:45:09 @spencerdailey I should have left then

2022-12-12 00:43:31 @adamjcook that is my precise fear

2022-12-12 00:42:40 @anclement ha! I was actually looking forward to the additional space (except the gpt-generated stuff)

2022-12-11 22:03:47 this makes me so uncomfortable I am taking a break. you can find me on mastodon, substack, and linkedin. https://t.co/2TvYHsuz3v

2022-12-11 21:51:01 @sergia_ch @sama See my comments at https://t.co/NnELd6ldtt

2022-12-11 21:50:32 @lorakolodny @wongmjane What she said :)

2022-12-11 21:08:13 Sorry, wrong. GPT-3 doesn’t bullshit because people do

2022-12-11 20:42:32 @TonyZador @bengoertzel @sama The problem is that they lose binding eg between entities and their properties, as I discussed here: https://t.co/bbOXczqzRQ

2022-12-11 18:02:28 Never a charge. Always our holiday gift to the world. https://t.co/4c17D87saZ

2022-12-11 18:00:51 @Kempton @Kobotic but great that at least some politicians, like Canadian MP @MichelleRempel, are showing an interesting in learning more and diving deeper into AI. She’ll be at https://t.co/VNBNm3FHSS, sharing her experience in regulating new tech.

2022-12-11 17:58:48 Looking forward to my Thursday 1pm Fireside chat w IBM Research’s Chief Scientist of Research Ruchir Puri, hosted by the MIT Sloan Quant Club. https://t.co/f7GKUPXKbP

2022-12-11 17:39:04 RT @benjaminjriley: “There is no one way the mind works, because the mind is not one thing. Instead, the mind has parts, and the different…

2022-12-11 17:37:59 @pwlot @bengoertzel funny i have been thinking of the same metaphor (wernicke’s aphasia) lately

2022-12-11 17:27:07 @benjaminjriley @chazfirestone! (Is Brian on here?)

2022-12-11 17:25:45 Let’s bury the hatchet, @sama. @bengoertzel just signed on to our 12/23 https://t.co/NnELd6ldtt

2022-12-11 17:11:45 @rmarcilhoo @bengoertzel @sama @ylecun @hardmaru yes, RL with world models will eventually be useful. the world models are the cake. when we get that right, we can get back to what yann likes to call the cherry on top. LLMs don’t have that, so aren’t super useful for that approach to RL

2022-12-11 17:06:03 @rmarcilhoo @bengoertzel @sama On RL I am actually very much with @ylecun. RL alone isn’t cutting it

2022-12-11 16:53:05 Read this thread, if you care at all about AGI and large language models. https://t.co/lsVZMezgEP

2022-12-11 16:40:22 Truth. Excerpted from a thread from one of the world’s deepest thinkers on AGI, @bengoertzel https://t.co/rk98uwr28j

2022-12-11 14:14:38 @sergia_ch @yudapearl @alexhanna That’s whole point of this essay

2022-12-11 14:13:42 RT @rogerkmoore: Brilliant example of the pattern-matching limitations of ChatGPT (thanks to Tim Hazan): “What letter comes next O T T F F…

2022-12-11 14:02:14 @NateSilver538 Curious to see examples of both @NateSilver538 (and wondering at what length the first part is effective).

2022-12-11 13:36:55 @wellingmax No! Was @sama Reviewer 2 in the evaluation of deep learning is biting a wall? I had no idea! Just kidding https://t.co/a4Cc4o8ouw

2022-12-11 04:58:13 Serious question, apropos novel writing, etc. What is the *longest* coherent prose you’ve gotten out of an LLM, without hand tinkering? The most interesting plot twist? Deepest characterization? Most surprising bit of dialog? https://t.co/9jmjZniRBd

2022-12-11 04:43:41 @davidchalmers42 guess the problem wasn’t so hard after all. cc @anilkseth

2022-12-11 04:42:48 If GeoHotz says its true, I guess maybe you aren’t conscious after all? Glad we cleared that up. (But IMHO he’s actually right that you are a computer: https://t.co/j9y9YqmqPC) https://t.co/i7uknfo0pq

2022-12-11 04:37:58 @renatrigiorese what i am saying is that because the developers are unable to rule out data leakage, and have done too little to alleviate that concern, the artists have a real shot, especially in cases where they can show enough similarity to convince a jury.

2022-12-11 04:25:16 @naivebaesian finally

2022-12-11 04:15:10 @DuaneJRich @sama @gdb it was the last line here that was key

2022-12-11 04:11:52 @DuaneJRich @sama @gdb To make the point that they should not be so dismissive of the next criticism, be it from me or somebody else.

2022-12-11 04:09:56 @renatrigiorese no, style cannot be. But individual art works can be stolen from, and copyright suits can be filed

2022-12-11 03:44:27 @tdietterich @Noahpinion Incidentally I already wrote about English ->

2022-12-11 03:41:22 @DuaneJRich @sama @gdb again, hundreds of of people piled on me. nobody ever called them out. why call only me out? deeper point: ML has been very unkind to criticism that has been proven in hindsight to be legitimate

2022-12-11 03:37:48 @tdietterich @Noahpinion but eg do you want your planning to go through systems that aren’t robust and trustworthy?

2022-12-11 03:32:37 @dave_at_hat @sd_marlow they have gotten a lot better this year

2022-12-11 03:31:44 Excellent question. (Spoiler alert: SEO optimization with faux text doesn’t count). https://t.co/rJaOW6JLge

2022-12-11 03:30:46 @DuaneJRich and, clarifying, the ridicule came from others (@sama, @gdb, @plinz, @ylecun, etc) not you. I didn’t read you that way and apologize if that was unclear.

2022-12-11 03:29:38 @DuaneJRich If it was just one, ok, but it was literally thousands of people, many quite prominent, posting and liking dozens and dozens of similar tweets. I think I am entitled to point that out. If you have experienced similar, great

2022-12-11 03:25:29 @DuaneJRich @sama @gdb And by the way I doubt that not a single person spoke up to call Sam’s original post childish. Willing to bet you didn’t. #doublestandards

2022-12-11 03:22:46 Hilarious rewriting of history. Hundreds of people ridiculed me on Twitter (and Facebook) in March and April. It was a meme! And now today is it’s “Happy We Knew it All Along” Day. Schopenhauer has got you covered! https://t.co/Q11UKykNQ9 https://t.co/en4oZKSidK

2022-12-11 03:05:25 @sama It’s not just that I told you so, back in March, but that back in April you (&

2022-12-11 03:00:02 @garymarcus, March, GPT shows “hallmarks of unreliability &

2022-12-11 02:45:33 100%, @sama, robustness and truthtfulness and by the way, that is *exactly* the wall in my award-winning essay that you and your co-founder mocked. Maybe read it again and be nicer next time? https://t.co/bVRpsfHIFo and your original response: https://t.co/Xos52HEqip https://t.co/jjwunBlUjK

2022-12-10 16:18:59 @balajis I am worried personalized misinformation tbh

2022-12-10 16:12:35 RT @GaryMarcus: AI's Jurassic Park moment - and 4 things we might do about it.

2022-12-10 16:12:16 RT @JamesFlint: This is important. Media - written media at least, has just changed forever. By @GaryMarcus #ChatGPT https://t.co/ueUKGq…

2022-12-10 15:48:47 @katiephangshow Thanks so much for allowing me to sound the alarm about the risk of automated misinformation on your show, @KatiePhang! And here is the link to the essay you so kindly mentioned: https://t.co/0bjkj1TPzR

2022-12-10 15:42:28 RT @katiephangshow: “ChatGPT can make wonderful simulations of almost anything, but it doesn’t really know what it’s talking about.” @GaryM…

2022-12-10 15:42:04 RT @greglinden: "For Stack Overflow, the issue is literally existential. If the website is flooded with worthless code examples, programmer…

2022-12-10 14:26:43 RT @GaryMarcus: @elonmusk @stillgray The biggest concern—far more imminent than AGI, and existential to @twitter—is automated propaganda: h…

2022-12-10 14:21:31 @elonmusk @stillgray The biggest concern—far more imminent than AGI, and existential to @twitter—is automated propaganda: https://t.co/0bjkj1TPzR

2022-12-10 13:43:17 RT @DrHughHarvey: I think this warning from @GaryMarcus will stand the test of time. We are at the precipice of wide spread misinformation…

2022-12-10 13:22:53 AI's Jurassic Park moment - and 4 things we might do about it. https://t.co/cICCuUIjyx

2022-12-10 13:21:41 @mridley and featured this morning by @katiephangshow, link to come

2022-12-10 13:15:59 @j2lovesfriday @katiephangshow Exactly, and see my pinned tweet: https://t.co/DjVCfSsmKg

2022-12-10 13:09:10 RT @kvashee: @GaryMarcus at #EMNLP2022 https://t.co/zLswziAl23

2022-12-10 04:59:55 RT @qantarot: 'When GPT sounds plausible, it is because every paraphrased bit that it pastes together is grounded in something that actual…

2022-12-10 02:06:57 @plibin It feels more like the trouble with tribbles to me https://t.co/NPyFDalvFI

2022-12-10 00:04:35 @patrickmesana i was locked out by an elaborate and somewhat disturbing phishing ruse, see my LinkedIn posts for details

2022-12-09 23:58:28 Looking forward to talking soon with @KatiePhang on @katiephangshow, tomorrow (Saturday) at 7:50am Eastern Time @msnbc, re my soon to appear Substack essay, _AI's Jurassic Park Moment_ https://t.co/jXZnzGb3qM

2022-12-09 21:46:53 @BobbyAlter There is no actual argument. It’s just noise.

2022-12-09 21:42:55 What I worry about the most, if I may be honest, is your qualifications for making important and consequential decisions around #AIEthics. https://t.co/tW1DJpZEaB

2022-12-09 20:00:40 Haha. But don’t think you or other folks would have enjoyed being part of the trap that was laid. Instead, be concerned about security around here, when fake yet seemingly verified Blue Badge Twitter employee w 75k followers can send authorization codes from same # as Twitter. https://t.co/FdQUTGydKl

2022-12-09 19:43:08 Some accidental ChatGPT comedy, via @vadimberman Be sure to click to see the whole thing

2022-12-09 19:37:16 First thing I see returning to Twitter https://t.co/rRAIDRkjvk

2022-12-08 21:04:55 RT @GaryMarcus: @blattnerma @Plinz Yes &

2022-12-08 21:04:15 @tsushil_thapa @ShizaCharania @ylecun not friends with people who repeatedly lie about me

2022-12-08 21:03:50 absolutely - the edge cases in open-ended chatbots are endless

2022-12-08 21:03:00 @info_sprinkles @ylecun sorry but lying is not cool

2022-12-08 21:02:51 RT @MMikeMMa: @pgolding @plibin @GaryMarcus Isn’t a concern that these models are striving towards “general” capabilities across many use c…

2022-12-08 19:52:42 RT @GaryMarcus: @ylecun . @yLeCun the last time you said this to @zdnet, I presented multiple AI publications in top journals (sample scree…

2022-12-08 19:52:39 @ylecun . @yLeCun the last time you said this to @zdnet, I presented multiple AI publications in top journals (sample screenshots below). @zdnet corrected you, because you lied. @kenneth0stanley corrected you yesterday re our algorithm. *why* are you lying about my credentials again? https://t.co/uG07u8sYsi

2022-12-08 19:27:20 RT @GaryMarcus: @ShizaCharania @ylecun That’ll be $8 :)

2022-12-08 19:25:54 @ShizaCharania @ylecun That’ll be $8 :)

2022-12-08 19:18:02 @wadhwa @Shrikantha5 AFAIK all LLMs are vulnerable to hallucination. The thing you sent was made by a different recent OpenAI model (not ChatGPT) one from @Shrikantha5 is indeed ChatGPT. General problem is very real! Ps: LLMs are stochastic &

2022-12-08 18:37:38 @MuhammedKambal Sure would like to know what was in the training corpus

2022-12-08 18:31:46 Eerily prescient paper from 2021: language models &

2022-12-08 18:12:21 @pwlot @jeblad @StolfiAlberto I got blocked (I believe) for doubting his claim that Austin would become the AI capital of the world. And if you read the thread carefully someone else brought the wiki to my attention. Something didn’t feel right tbh when he interviewed me in the MIT library.

2022-12-08 18:09:38 @__delas__ @Plinz Love it when Plinz publicly attacks me and I respond and then the fanboys pile on without reading the time stamps or understanding the chronology.

2022-12-08 17:44:14 @pwlot @jeblad @StolfiAlberto he's representing himself as an MIT research scientist, not as a Drexel PhD.

2022-12-08 17:27:49 @Tweetermeyer wow. not a great look, especially given how he styles himself.

2022-12-08 17:26:03 @pwlot @jeblad @StolfiAlberto *did* he graduate? from the answers I am getting, he taught some January classes, but was not on faculty.

2022-12-08 17:21:33 RT @ChloeSavignac: Excellent line-up

2022-12-08 16:36:51 I suspect that this is inaccurate, but am curious if anyone can confirm or deny it. https://t.co/qlDi5cyH9c https://t.co/myflXl9IAr

2022-12-08 15:28:07 RT @wadhwa: And I did not found "Telesoft Partners". As well, Google clearly says that I authored 5 books, not 3 I am beginning to agree w…

2022-12-08 13:00:00 CAFIAC FIX

2022-12-07 08:00:00 CAFIAC FIX

2022-12-07 08:00:00 CAFIAC FIX

2022-11-13 22:04:26 Can you preorder a book that hasn’t yet been written? https://t.co/gNHnj7brwm

2022-11-13 21:58:18 @karaswisher current AI techniques aren’t good enough to handle content moderation well, so they need a lot of humans.

2022-11-13 18:26:07 How is @elonmusk going to make Twitter an “accurate source of information” when he let his trust and safety team and his content moderators go?Magic? https://t.co/vFEySb1HYZ

2022-11-13 15:46:25 RT @cyberandy: 'Large language models are without equal at generating misinformation, but they suck at detecting it' @GaryMarcus also G str…

2022-11-13 01:07:11 I just put up $5,000 saying we won’t have global L5 driving w adequate safety data before 2030 and so far not one Tesla fan has joined Hotz in offering to take the other side.Y’all doubting Musk after two weeks of Twitter? What happened to your convictions? https://t.co/Pthsu5u0yb

2022-11-13 01:04:02 @cshirky Well, @realGeorgeHotz *is* betting that Lombard on a rainy night and Bombay in heavy traffic are all fine. That’s just what the bet is, global L5, not L4, disregarding regulatory issues but pending serous safety data. You are on my side, he isn’t.

2022-11-13 00:52:32 @cshirky It’s No now and no it shall remain for longer than many people think. I have been saying this since 2016… the criteria are clear: you don’t need a steering wheel or a driver and you can go where you want to go

2022-11-13 00:49:19 Great essay in ⁦@Analyticsindiam⁩ that ⁦@elonmusk⁩ would be well-advised to read. Cc ⁦@Jason⁩ https://t.co/LztkXaIj8S

2022-11-13 00:45:55 @Jess_Riedel a major city is not adequate for L5, since eg you can pick Phoenix w good weather and few pedestrians which is just a subset of the full problem. See eg https://t.co/azCUGUYUsr

2022-11-13 00:42:59 @cshirky No, that is not L5. There may be some cities like that, especially ones designed from scratch, but that’s not L5. And it won’t help eg in NYC.Certainly possible in principle to be safer than humans, fully automated, and it probably will happen, I just doubt it will be soon.

2022-11-13 00:02:24 @realGeorgeHotz @therishidesai @metaculus @MatthewJBar Great. Sounds like we basically agree on criteria. You donate 5k to Doctors Without Borders if we don’t have that analysis for global L5 by 31 December 2029

2022-11-12 23:52:51 @therishidesai @realGeorgeHotz @metaculus @MatthewJBar No that’s just more demos, and also restricted to US. True level 5 means you don’t need a steering wheel and the car takes you where you want to go around the globe (on conventional roads) without restriction. That’s how it is defined. Eg https://t.co/JwDecVBXic https://t.co/Pj3ucYjf5t

2022-11-12 23:47:07 Game on. Will we have L5 driverless cars by end of decade, around the globe? https://t.co/JfIhlLKNmm

2022-11-12 23:44:41 @realGeorgeHotz @metaculus @MatthewJBar Am willing to make a side bet on that part, too. Eg all of those mechanisms may be necessary but I predict that that they won’t be sufficient

2022-11-12 23:43:44 @realGeorgeHotz @metaculus @MatthewJBar Most commutes, around the globe, really?I am willing to bracket regulatory issues and pulled over vehicles if we can agree on a way to evaluate safety that is some kind of independent agency that has access to data.

2022-11-12 23:22:15 You want to put money on (safe as humans) L5 by end of decade? @metaculus @MatthewJBar https://t.co/W0wQMivMhe

2022-11-12 22:23:43 @ylecun @KordingLab @yudapearl @scac1041 @PhilDawid @pmddomingos @StephenPiment @AlexTensor @gopnik @stephensenn Maybe just asking for clarification for what you are claiming and what you the mechanism is.

2022-11-12 21:07:05 @ylecun - you still need some prior to organize that infinite data and decide how to generalize from it- the choice to minimize is an aesthetic choice, not a scientific one

2022-11-12 20:50:03 @AmanBitz @ErnestSDavis i sketched something here but it’s laborious to build: https://t.co/1N4qeEn6uH

2022-11-12 20:31:54 @pmddomingos @GhaffariMaani @KordingLab @ylecun @yudapearl @scac1041 @PhilDawid @StephenPiment @AlexTensor @gopnik @stephensenn that’s learning that the set of relevant cases is empty

2022-11-12 20:29:52 Trouble in BIG-Bench paradise?- @ErnestSDavis looks at 48 of the benchmarks within and finds problems with most: https://t.co/vwykkgWAMM- Many project AGI timelines based on performance on these benchmarks. If the benchmarks aren’t valid, consequent timelines are problematic

2022-11-12 20:26:28 @pmddomingos @GhaffariMaani @KordingLab @ylecun @yudapearl @scac1041 @PhilDawid @StephenPiment @AlexTensor @gopnik @stephensenn right, so you presuppose innately that there are symmetries and then learn which are relevant.

2022-11-12 20:13:24 @pmddomingos @ylecun @KordingLab @yudapearl @scac1041 @PhilDawid @StephenPiment @AlexTensor @gopnik @stephensenn not sure i follow why not

2022-11-12 20:12:41 @pmddomingos @GhaffariMaani @KordingLab @ylecun @yudapearl @scac1041 @PhilDawid @StephenPiment @AlexTensor @gopnik @stephensenn right, but in simple English, what’s the prior here?

2022-11-12 18:19:24 Remember when @realGeorgeHotz was hyping driverless cars?None of his new set of predictions will come to pass if we cannot constrain artificial agents to facts &

2022-11-12 15:41:36 @KevinIndrebo @ylecun @KordingLab @yudapearl @scac1041 @PhilDawid @pmddomingos @StephenPiment @AlexTensor @gopnik @stephensenn @davidchalmers42 Urge you to watch our debate, including my opening statement, and Chalmers’s question and his answer (“zero”). It’s on YouTube

2022-11-12 15:29:46 @ylecun @KordingLab @yudapearl @scac1041 @PhilDawid @pmddomingos @StephenPiment @AlexTensor @gopnik @stephensenn Ps @davidchalmers42 if you are keeping score that’s two more innate priors than he allowed for in our 2017 debate. Infinitely more than the zero he claimed then :)

2022-11-12 15:28:06 @ylecun @KordingLab @yudapearl @scac1041 @PhilDawid @pmddomingos @StephenPiment @AlexTensor @gopnik @stephensenn What’s the difference between an innate bias for equivariance over permutations and an innate apparatus for symbol-manipulating operations over variables?

2022-11-12 15:26:17 love it when @ylecun speaks up for innateness! https://t.co/Xsk3lNU4PY

2022-11-12 00:20:52 RT @phlu: Good heavens, here it is Kurt Vonnegut's One Hundredth Birthday. https://t.co/pp1zRBJOZy

2022-11-11 16:24:04 RT @brianbuccola: Can baboons do compositionality? That is, can they form compositional representations like "not blue"? In a new paper wit…

2022-11-11 12:02:13 RT @DrewLinsley: Check out our new paper, to appear at NeurIPS. We show that DNNs are becoming progressively *less* aligned with human perc…

2022-11-11 01:06:38 What’s the plural of “stable genius”?

2022-11-10 14:40:25 Weird recipes won’t kill us, but misinformation already has (eg by undermining vaccine programs). It’s no joke that LLMs can‘t handle the truth. https://t.co/2guFp5j05d

2022-11-10 13:54:45 @ruthaylett ick. my second try didn’t get labeled as toxic:on a quick inspection it mostly it just seems to look for curse words and stuff cc @vadimberman might enjoy testing this https://t.co/5vpRUKApXd

2022-11-10 12:54:31 RT @desai_pratik: "Large language models are without equal at generating misinformation, but they suck at detecting it." - Stakes are indee…

2022-11-10 02:05:30 @nachoarranz @elonmusk link in the full blog post

2022-11-10 02:04:38 @AmandaAskell justified. true.

2022-11-10 00:47:45 thread broke

2022-11-09 23:40:46 rest of thread: https://t.co/OXgoLqrzvd

2022-11-09 23:39:35 @Aella_Girl rational thinkers yes but also AI scientists who can think outside the box: https://t.co/OXgoLqrzvd

2022-11-09 22:38:12 RT @_SilkeHahn: "Misinformation travels faster on #Twitter right now than elsewhere—about 8 times faster than on Facebook—in part because t…

2022-11-09 22:33:04 @elonmusk https://t.co/XyOnq8qVtw

2022-11-09 22:29:30 5. You’re going to need something different, @elonmusk. Let’s talk. https://t.co/OXgoLqrzvd

2022-11-09 22:29:29 3. As the pace of machine-generated misinformation picks up, Twitter’s existing effort, Community Notes (aka Birdwatch), which is mostly done manually, by humans, is going to get left in the dust.

2022-11-09 22:29:26 2. The problem is about to get much, much worse. Knockoffs of GPT-3 are getting cheaper and cheaper, which means that the cost of generating misinformation is going to zero, and the quantity of misinformation is going to rise—probably exponentially.

2022-11-09 22:28:23 Bravo @ElonMusk for wanting Twitter to become the world’s “most accurate source of information”Here are 5 things to consider:1. Misinformation travels faster on Twitter right now than elsewhere—about 8 times faster than on Facebook.

2022-11-09 15:45:33 RT @Reuters: Russia orders troops to withdraw from the occupied Ukrainian city of Kherson and take up defensive lines on the opposite bank…

2022-11-09 15:43:14 @MadamePratolung @cfchabris @partygrrrl65 i have mentioned this work several times, eg in a tweet to Musk in last several days, and link to an update on it in my next substack essay, which will be about misinformation.

2022-11-09 15:33:43 Think your LLM can just make up recipes? Think again!It’ll look like a recipe, but taste lousy. Misinformation for the stomach! https://t.co/EnPeENWA2B

2022-11-09 12:08:35 RT @eugemallol: “Lo que veo es básicamente una forma de gastar un montón de energía en California”, dice Chomsky. Y apostilla el profesor d…

2022-11-09 02:25:31 @sleepinyourhat @Anthropic @nlpnyc @TonyZador along lines of what you were suggesting

2022-11-09 02:24:27 @sleepinyourhat @Anthropic @nlpnyc link? i don’t think i saw the calibration thing.

2022-11-09 02:20:02 @sleepinyourhat @Anthropic @nlpnyc i honestly don’t think scaling is going to help much at all with this particular problem. but… more importantly can you discuss what else aside from scaling is being considered?

2022-11-08 22:14:22 @TonyZador put that differently, on the theory of relativity, you seemingly would have accepted literally any answer from 0% to 100% confident (numerically or verbally) as “correct”. completely and utterly unfalsifiable, because it is a report that refers to something that doesn’t exist.

2022-11-08 22:12:48 @TonyZador you are making an anthropomorphic error in assuming it has its own confidence that it is relating to you.

2022-11-08 22:11:07 @TonyZador yes they are very lossy

2022-11-08 22:05:36 @TonyZador read carefully

2022-11-08 22:00:53 @TonyZador alas, the only thing they might have confidence around is the predictions of sequence of words, not likelihoods of underlying facts.they are not like us, and you have to stop thinking about them as if they are. (they are more like Markov Chains than they are like people.)

2022-11-08 21:37:56 @BethCarey12 @doRadiology we don’t really have words to describe systems that alien to how humans work

2022-11-08 21:34:34 most people working with LLMs seem to think that they can achieve truthfulness either through scaling (eg @anthropic.ai’s approach if I am not mistaken) or changing the objective function (eg as @nlpnyc suggested earlier). as indicated, I don’t expect these approaches to work. https://t.co/qKpfINwTNt

2022-11-08 20:57:46 Meanwhile, over on Mastodon, thoughts on language, truth, and large language models that are slightly too long to express in <

2022-11-08 19:21:33 RT @BDehbozorgi83: I was watching the debate btw. Profs. #Chomsky &

2022-11-08 19:21:23 Sex is good but… https://t.co/UnFHoAofWF

2022-11-08 18:59:17 If you don’t understand this, you don’t understand LLMs.Most people don’t. https://t.co/WoA8TDAijx

2022-11-08 18:19:19 yes - they can’t verify truth relative to databases https://t.co/8B3TjplYfA

2022-11-08 18:18:46 An LLM trained strictly on truth will still confabulate, because the LLM that is trained on truthful things will break the bindings in what is saw, in the interpolation process, and continue to fabricate. https://t.co/DHFqoxRvrH

2022-11-08 18:10:53 @NaveenGRao even then it is a deliberate attempt relative to an internal model. making an truthful human is a different project (eg involving incentives and values) than making an LLM truthful (not possible because it can’t restrict itself to that which is derivable from ground truth).

2022-11-08 18:01:54 @Abel_TorresM @glupyan @davidberreby it’s a very behaviorist perspective for a cognitive scientist like @glupyan to take, focusing on behavior without paying attention to underlying mechanism

2022-11-08 18:00:56 @nlpnyc i see no reason to expect more than marginal change, either from empirical work on TruthfulQA or an understanding of the underlying mechanism. @OwainEvans_UK has a nice article on this @lesswrong

2022-11-08 17:58:59 @NaveenGRao not the last part. humans can at least try to verify. LLMs don’t try and don’t have the concept of truth. it’s really not the same at the cognitive level.

2022-11-08 17:38:45 @Abebab and eg- they might also counsel foolish actions that they don’t understand- or lead people to fall in love with them and then later feel abandoned, because LLMs don’t actually understand dynamics of human relationshipsetc

2022-11-08 17:34:20 @glupyan @davidberreby . do you seriously believe the fact that some people are ignorant about astronomy puts GPT-3 (which builds no explicit models of the world) on a par with people (who do their best to construct models of the world)?

2022-11-08 16:29:24 LLMs don’t lie, per se. But you should *never* trust blindly in what they write

2022-11-07 20:24:02 had no idea that Barcelona was already half way to #neurosymbolic #AI! recursion for sale in shop windows! https://t.co/hVdl8v9xam

2022-11-07 19:52:48 RT @esenktutuncu: It was a real pleasure starting the day with a great talk by @GaryMarcus on AGI, foundation models and how symbolic under…

2022-11-07 11:44:43 RT @EvelinaLeivada: AI, big data, deep learning, language and cognition. A fantastic talk by @GaryMarcus at @UPFBarcelona! https://t.co/4Bf…

2022-11-07 06:47:39 can anyone connect me to @mcuban to discuss with him how to use AI to fight disinformation at scale?

2022-11-07 06:45:45 @mcuban to make social media more accurate, we need AI to do it at scale, but we need a new approach, which i am working on. current situation: https://t.co/Ouo3vHOmc5

2022-11-07 02:08:01 @elonmusk @alex_avoigt https://t.co/7Es9Nk5ghr

2022-11-07 01:56:55 @ESYudkowsky i get some credit for calling out @sama on his “AGI is going to be wild” tweet on DALL-E moments after it went out. All kinds of people dumped on me for saying it, too.

2022-11-07 01:52:07 RT @plibin: This is right. We need a new infrastructure for truth and a serious effort to build one has to grapple with the fact that techn…

2022-11-07 01:36:34 @ismashrobots https://t.co/7Es9Nk5ghr

2022-11-07 01:36:19 @elonmusk https://t.co/7Es9Nk5ghr

2022-11-07 01:34:19 if you are serious @elonmusk about trying to make Twitter the most accurate source of information, we should talk. And you need to start by understanding the core technical issue: https://t.co/QkdRpFpcMZ https://t.co/1P9WBH6K2O

2022-11-07 01:10:38 and if anyone wants to help me with AI for misinformation detection, eg with funding or research, please DM https://t.co/O5GhKfBWfm

2022-11-07 01:08:15 @elonmusk i am working on AI to help with that. DM.

2022-11-07 00:18:07 Sure, they canceled Westworld. But why the f$&

2022-11-06 23:57:38 @marktenenholtz there was a GPT-3 imitation of me for a while…

2022-11-06 23:22:31 sorry but no. if you ban them, you don’t get the 8 bucks a month. and you no longer get to play the unfettered free speech card any more, either. https://t.co/oWdJirgRsc

2022-11-06 23:10:53 Looks like someone’s feelings were hurt https://t.co/zGzrr7ahDW

2022-11-06 22:21:23 hey @ylecun check this out and maybe see you sometime, over at Mastodon. https://t.co/WOBcyHoeNk https://t.co/vaFpZUfyAq

2022-11-06 22:08:02 same. @garymarcus@sigmoid.social, some interesting AI peeps already gathering at Mastodon include @fhuszar @MelMitchell1 @alexhanna @emilymbender @gradientpub @raphaelmilliere @mark_riedl @dpkingma &

2022-11-06 10:57:10 barcelona bound! https://t.co/z014XgyRzU

2022-11-05 01:29:51 @sir_deenicus @KordingLab @TonyZador it is (roughly) autocomplete w interpolation, in the sense that it is frequency sensitive as @yasaman_razeghi and @sameer_ have shown. I don’t know if they have tried with prompt engineering but I would be surprised if the prompt fully solved the problem.

2022-11-04 18:47:14 RT @quaesita: "GPT is autocomplete on steroids but it gives the illusion that it's more than that."-@GaryMarcus, chatting with Noam Choms…

2022-11-04 14:39:09 RT @jeremyakahn: Will be interviewing Noam Chomsky and @GaryMarcus on “Debunking the AI Lie” on the centre stage @WebSummit at 4:15 pm Lisb…

2022-11-04 13:29:50 chatting with Noam Chomsky at 4:15pm lisbon time/9:15 am pacific timelive stream (free) at https://t.co/vQcEjAQCtg

2022-11-03 19:23:46 Data in ->

2022-11-03 16:46:22 RT @jeremyakahn: Those interested in watching the livestream of my conversation on "Debunking the AI Lie" with Noam Chomsky and @GaryMarcus…

2022-11-02 22:55:53 Stunning reminder of how much we don’t know about current AI systems. https://t.co/ExTIUkqppz

2022-11-01 23:00:41 RT @MaartenSap: Have you ever wondered whether large neural networks like GPT3 acquireTheory-of-Mind-like abilities? And whether they cou…

2022-11-01 20:15:57 @TonyZador it's amazing. but is it robust? across what range of problems, assumptions etc?

2022-10-31 22:46:20 @titudeadjust @Popehat @yudapearl @kahneman_daniel @R_Thaler tell @popehat he oughta unblock me

2022-10-31 19:05:42 If large language models can’t help lie about arithmetic, even after terabytes of text, why does on earth does any one expect them to tell the truth about anything else? https://t.co/Wa392eCf6f

2022-10-31 18:08:04 @fhuszar three claims that I made in 2001 that I stand by:- universally quantified one-to-one mappings are important for cognition- challenges in distribution shift make those problems hard for many naive neural nets- integrating operations over variables with learning is vital

2022-10-31 18:05:06 RT @GaryMarcus: @fhuszar i do think it is a fundamental reflection of the problems in doing extra-distribution shift in a system that lacks…

2022-10-31 17:59:04 @fhuszar i do think it is a fundamental reflection of the problems in doing extra-distribution shift in a system that lacks operations over variables—exactly as anticipated in Chapters 2 and 3 of The Algebraic Mind, two decades ago.

2022-10-31 14:20:46 @AnimaAnandkumar @rmichaelalvarez @Caltech incredibly important problem

2022-10-31 13:48:49 RT @GaryMarcus: @fhuszar skeptical. everything i have seen on the “emergent” side seems too tied to data idiosyncrasy, with problems of dis…

2022-10-31 13:43:08 @fhuszar skeptical. everything i have seen on the “emergent” side seems too tied to data idiosyncrasy, with problems of distribution shift, never at the level of abstraction required for a robust understanding of time, space, or causality. even multi-digit arithmetic is unreliable.

2022-10-30 18:28:18 AI Intelligence Elon Musk: “Twitter obviously cannot become a free-for-all hellscape.”Also Elon Musk: https://t.co/sRkeYww7ve

2022-10-30 13:44:12 quietly growing, and poised to change everything. https://t.co/9847MKRfDz

2022-10-30 01:02:04 RT @foysavas: Finally read https://t.co/iZIUUNGKbR by @GaryMarcus and Ernest Davis and it's a greatly accessible critique of contemporary A…

2022-10-29 19:45:43 AI fads come and go (anyone remember expert systems? SVMs?) Here’s are three ways that Scaling Maximalism, theory du jour, might end.https://t.co/jSFOVZQQ10

2022-10-29 14:29:57 and of course the thing to read on causality, in addition to what we say, is @yudapearl’s The Book of Why.And yes Kant was there long before we were :)

2022-10-29 14:27:38 actually “our suggestion”

2022-10-29 14:22:52 My suggestion for what AI needs, 2019. Nothing has changed. https://t.co/8aHX6Gd3HT

2022-10-29 13:48:21 RT @AlisonBLowndes: Looking forward to this @GaryMarcus. Tune in all, 4.15pm UK (Lisbon) Nov 4: https://t.co/RnuXORRmPC

2022-10-29 13:39:15 @LauraRuis @Isinlor I am less sanguine. I wouldn’t be surprised to see improvements on this or that benchmark, but pragmatics fundamentally requires reasoning about models of the world (eg other’s beliefs), and I see no reason to think that LLMs their own can reliably construct such things.

2022-10-29 01:03:47 @albertwebson @_jasonwei @tallinzen skeptical of 1, but keen to see the argumentyou may have seen my new work w @EvelinaLeivada and @ElliotMurphy91, different architecture but perhaps relevant:https://t.co/UWUg6kQ3Df

2022-10-29 00:46:43 @albertwebson @_jasonwei @tallinzen I don’t think all forms of intelligence need to do arithmetic. but these kinds of errors give some insight in a semi-controlled way into a deeper, not yet solved challenge of translating language into models of the world. curious to see the paper when you can share

2022-10-28 21:32:45 RT @egrefen: I'm biased, obviously, but I think this is *such* an important result. Implicature and pragmatics are essential, ubiquitous, a…

2022-10-28 17:50:52 Just rehearsed Noam Chomsky and @jeremyakahn Nov 4, livestreamed at https://t.co/vQcEjAQCtg https://t.co/k93hY3EmVu

2022-10-28 14:18:18 indeed, slashing moderation just as synthetic content becomes widespread does seem like a disaster-in-the-making https://t.co/gEMMehatm7

2022-10-28 10:49:43 AGI, my eye. https://t.co/EmVNJRUzpz

2022-10-28 10:47:51 @L_andreae @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab agree w @L_andreae, eg a classic paper by Thelen showed that newborn humans will step, and that is really an issue of weight etc rather than lack of motor program per se: https://t.co/09LnIrfIqf

2022-10-28 10:39:40 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab (i pasted the image above/here it is) https://t.co/JjMRtntwN9

2022-10-28 10:29:06 @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab presumably both

2022-10-28 10:21:19 @L_andreae @WiringTheBrain my books https://t.co/Pt7HZc3RJd, Kluge, and The Birth of The Mind

2022-10-28 10:13:59 RT @Abebab: the most insightful knowledge about a field comes from experts in the field that hold a healthy skeptical view on the topic and…

2022-10-28 10:08:07 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab beautiful new review, by @neuroluci. how do you even think about this stuff if you are a staunch anti-nativist? https://t.co/RvacUhsspE

2022-10-28 10:01:09 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab hmm, figure 2 from Verhage et al certainly speaks to those questions (# and pattern) to some degree

2022-10-28 09:56:27 @AthenaAI2 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it goes to showing how little we understand the brain &

2022-10-28 09:50:19 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab agreed! and yet you wouldn’t know that from reading a lot of philosophers and AI researchers these days. &

2022-10-28 09:39:10 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it seems to me then that the reasonable conclusion is same as what I wrote in my lay summary in 2004: brain development uses a mix of activity-dependent (some not all intrinsically guided) &

2022-10-28 09:13:25 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab not mentioned but also seemingly relevant is Verhage et al (2000), esp figure 2 (“Correct assembly of the brain the absence of neurotransmitter secretion”) below. Anyone in this thread have thoughts on it? https://t.co/6O4OlOjoQZ

2022-10-27 18:46:45 @fhuszar https://t.co/UWUg6kxUp7

2022-10-27 18:20:33 if i had a dollar for every time someone cherry-picked DALL-E 2 art, I coulda outbid Elon for Twitter https://t.co/kHe92VtnFB

2022-10-27 15:49:18 @ylecun @ProfNoahGian so you have never done your own ground truth study and are relying on those outside orgs?

2022-10-27 15:37:48 @ylecun @ProfNoahGian our messages crossed

2022-10-27 15:31:15 @ylecun @ProfNoahGian that doesn’t answer @ProfNoahGianyou need a (large) sample that is carefully human reviewed ground truth. in that ground truth, two questions:1. what % would be detected by the AI?2.what % of what is missed by the AI is flagged by humans?you are addressing #2, not #1

2022-10-27 14:58:35 As ever, @KBAndersen has his finger on the pulse. (Answer to his question in the comments, in case you need it.) https://t.co/gqwcRACOmx

2022-10-24 02:35:23 there are two kinds of people in this world.those brave enough to speak truth to powerand those who are not.

2022-10-24 02:31:32 @primalpoly @AmandaAskell @AlecMacGillis @JoshuaChaffin i agree. I would put a lot more trust in a single anecdote involving tricking poor people into flying to Martha’s Vineyard for nonexistent jobs.

2022-10-24 01:47:36 @pmddomingos “And if there's war between the sexesThen there'll be no people left”

2022-10-24 01:39:40 @filippie509 @KordingLab @TonyZador

2022-10-24 01:21:55 @filippie509 it also doesn’t cite my endless arguments for years about looking to cognitive science, or even my paper with marblestone in Science…

2022-10-24 00:37:11 @voxbec @davidchalmers42 @mbusigin not disputing this but many people would take these to be influences on cognition. as i said the list is odd, from how people typically use the term, not necessarily wrong. in general people often distinguish between perception (eg olfaction, low level vision) and cognition.

2022-10-23 23:56:44 RT @AmandaAskell: If you don't have time to provide any evidence that someone is a eugenicist (beyond alluding to articles that also don't…

2022-10-23 22:55:22 @GaneshNatesh @ceobillionaire there has been zero budget thus far, zero charge. all participants (including us) have donated their time.

2022-10-23 22:10:34 . @ceobillionaire and I are pleased to announce that we will be resuming our AI debate series, December 23. Mark your calendar, for some exciting conversations, and reply below to suggest participants. (We already have our own ideas, but might have room for a couple more.)

2022-10-23 22:07:25 In the real world of AGI, nobody knows what to do, nor how to do it. https://t.co/4DPbeIKz85

2022-10-23 21:59:17 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher here: https://t.co/5jrUODZuNt

2022-10-23 21:57:13 @Korrelan_AI @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher you need a world model for that, IMHO

2022-10-23 21:56:53 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher i saw a talk at NeurIPS c 2017 or 2018. don’t know how active it is

2022-10-23 21:47:41 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher well, if it’s created post-training but eg it’s snow white with elves rather than dwarves, i am not super impressed.but i think you are generally thinking about the right kind of issues

2022-10-23 21:44:47 . (But what a messed-up field if that actually guides people’s research choices.) https://t.co/rTXXHpKN3f

2022-10-23 21:43:10 @pmddomingos @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher link? i’d be curious

2022-10-23 21:41:34 @Peterbart @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher ever since Wittgenstein we’ve known that language doesn’t work that way… and I am happy in that corner

2022-10-23 21:40:33 @DeonTBenton the empiricist’s imagination knows no bounds.

2022-10-23 21:39:58 @AndrewLampinen @Peterbart @mpshanahan @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher fair :)

2022-10-23 21:39:44 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher just the general idea that we want generalization rather than mere memory. we might eg actually want to have two different tracks, short and open-length, tracking empirically how much progress on one leads to progress on the other.

2022-10-23 04:48:29 @NoraNewcombe you are *referencing* those distinctions, but I don’t see how their particular way of carving things up helps. as stated, something here needs be represented, and the format for that representation itself appears to be innate.

2022-10-23 04:41:20 @davidchalmers42 @mbusigin definitely grey and debatable but cognitive science actually seems broader than cognition/cognitive to me. (eg for me it clearly includes parts of philosophy of mind that i wouldn’t include in cognitive psychology per se). but there’s no fact of the matter certainly/YMMV.

2022-10-23 04:35:49 @DeonTBenton i even said “privately” ffs

2022-10-23 04:35:30 @DeonTBenton wow. you are a jerk, reposting private messages. i am going to block you and end our correspondance if you don’t delete that.

2022-10-23 02:21:41 @davidchalmers42 @mbusigin i mean, i guess, but olfaction? feelings? it’s pretty different from the most common uses of the word cognition

2022-10-23 02:19:17 RT @GaryMarcus: @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg…

2022-10-23 02:12:43 @bradpwyble @pfau @ylecun @KordingLab @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher @davidpoeppel @TonyZador @AdamMarblestone sorry totally agree and meant to mark that (though it has been considered in neuroscience)

2022-10-23 02:11:40 RT @william_g_ray: @ESYudkowsky In 20 years, people will view content created before 2021 as pure and untainted by GPT. They will mine it i…

2022-10-23 01:49:33 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble eg see list here of many agreements https://t.co/Odg2QIt39X

2022-10-23 01:48:00 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble don’t believe the press clippings

2022-10-23 01:47:45 Funny how people keep thinking this when there is so much @ylecun and I agree on, eg. recently summarized here: https://t.co/Odg2QIt39X https://t.co/bpwARC92W8

2022-10-23 01:46:14 @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher and of course the notion of variable binding is critical but hardly widely embraced in MLam sure that eg @KordingLab @davidpoeppel @TonyZador @AdamMarblestone and many others can add to this list

2022-10-23 01:43:24 @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg Van Essen diagram) - power of dendrites (@mattlark/@YiotaPoirazi etc)- variety of neurons (@AllenInstitute)- areawise specialization (@Nancy_Kanwisher)- intrinsic cues &

2022-10-23 01:25:07 @pfau @ylecun @KordingLab @bradpwyble well, ok, that might be fair. i will go with “some ideas from the 60s had real impact”, but neuroscience hasn’t been a major driver since. (experience replay, and what else from neuroscience post 1970? not a huge list?)

2022-10-23 01:21:26 @lathropa @DeonTBenton one might have thought that since dall-e had both images and text that that visual input might have served as grounding, but of course compositionality didn’t emerge. (a couple years ago people said we just need our llms to have visual input…)

2022-10-23 01:17:31 @lathropa @DeonTBenton dall-e doesn’t refute statistical learning (it is consistent with it) nor the notion that statistical learning might contribute to language learning but rather causes problems for a theory in which statistical learning by itself magically gives you language.

2022-10-22 21:13:38 Have we made exponential progress towards Hawking’s notion of intelligence? I dare say no. https://t.co/WMJDWCjuke

2022-10-22 21:12:41 RT @IntuitMachine: @GaryMarcus @PaulTopping @bradpwyble @DeonTBenton "Intelligence is the ability to adapt to change." - Stephen Hawking

2022-10-22 21:08:55 @IntuitMachine @PaulTopping @bradpwyble @DeonTBenton for sure

2022-10-22 21:08:40 @DeonTBenton sure, and see also https://t.co/0rYVHqm8SK for a bias that can’t be explained in the same way.there is both learning and innateness. you haven’t said anything that convinces me otherwise, just given me more evidence for learning (that as it happens I already knew)

2022-10-22 02:07:26 @quocleix there is an error somewhere. Hinton is born in 1947, not in 1963. may i assume that the error is in the tweet rather than the paper?

2022-10-21 21:52:16 RT @vdignum: @GaryMarcus @nirsd Working on moving #AI beyond current #ML practices and approaches is not only a good idea but a much needed…

2022-10-21 21:21:12 @Grady_Booch @davidchalmers42 fyi

2022-10-21 21:00:07 @LucaAmb strawpersoning someone who says “maybe” and “hints” as concluding anything prematurely is a bit unfair, no?

2022-10-21 20:55:00 @LucaAmb i raised a question and said i saw two hints

2022-10-21 19:41:41 @LucaAmb gee, if it were a deep learning demo you’d say “it’s early days, give them a chance”

2022-10-21 18:01:14 @PaulTopping @kohn_gregory @TonyZador @sanewman1 @ehud @SpeciesTypical or mansplaining masked in philosophical jargon? hard to say.

2022-10-21 18:00:14 Retweeting this tweet from four years ago because it perfectly anticipated @ylecun’s three stages of grief. https://t.co/lgoWfHjQBU

2022-10-21 17:59:06 RT @atldgtltrnsfnug: “In defense of skepticism about deep learning” by @GaryMarcus https://t.co/eYns9BWklh“All truth passes through three…

2022-10-21 17:57:40 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical and not the DNA itself?? just the reading process? what are you saying??

2022-10-21 17:56:54 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical do you honestly think that i don’t that? or that is any different from an unplayed record? or unread book?

2022-10-21 17:54:59 @SpeciesTypical @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory i have never blocked someone simply for being obtuse before but I am feeling close…

2022-10-21 17:54:36 @kohn_gregory @PaulTopping @ehud @sanewman1 @SpeciesTypical “not relevant until it participates” doesn’t mean it doesn’t exist!same thing with my record player groove

2022-10-21 17:34:40 Pretty prescient, no? https://t.co/lgoWfHjQBU

2022-10-21 17:17:32 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical and more seriously: https://t.co/JS6U3kWl1v

2022-10-21 17:16:51 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical nope! @kohn_gregory assures me there is no information in genes at all, &

2022-10-21 17:13:55 @WiringTheBrain @sanewman1 learning, clearly. because empiricism is all that is left!

2022-10-21 17:00:40 true, but fundamentally misleadinggenes aren’t blueprints. they don’t “directly” contain information for phenotypes—but they do contain an enormous amount of information that is relevant and causally influential. https://t.co/jH07H5Cjvh

2022-10-21 16:31:01 @math_dandy so now “deep” also encompasses a single layer, which was its historical point of contrast? and maybe literally all of statistics and computer science, too?

2022-10-21 15:53:02 so maybe you don’t need deep learning after all? just good statistics? 2nd hint of this in October… https://t.co/6Ks02OmNfr

2022-10-21 00:00:22 RT @ruchowdh: The 8 bit bias bounty is now live!! Thank you @Melissahei for the article on what the bounty program means in context of the…

2022-10-20 23:38:05 anyone want to organize a conference or conference session, red-teaming current AI, as suggest below by @nirsd? it’s a good idea. https://t.co/mTZGIrwqAi

2022-10-20 23:30:06 RT @kanjun: Today, AI systems can create stunning art &

2022-10-20 23:26:04 @nirsd i have actually seen such things advertised and honestly don’t know how you proceed if you don’t know what you are trying to fix

2022-10-20 23:24:01 RT @dcli: saw ted chiang give a talk yesterday where he basically told a bunch of ai bros, "we are nowhere near having real ai, and what we…

2022-10-20 23:14:20 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb and repeating since you missed it the first time: https://t.co/5MjfJ1h6k6

2022-10-20 23:13:41 @berent_iris for sure. no competition there

2022-10-20 23:09:55 @berent_iris also mouse wins on 3d understanding of where objects are

2022-10-20 23:07:52 @berent_iris mice have mediocre vision. so maybe not (of course they also use smell etc): but they beat models by a lot in knowing what to do with objects that they do recognize

2022-10-20 21:36:52 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb put differently, if a n-gram model (with very large n) got compositional syntax right 66% of the time, it wouldn't mean the n-gram model had compositionality, would it?

2022-10-20 21:35:34 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb whatever they get or don't get is parasitic on corpus stats. on the image synthesis stuff some colleagues and i have data we will report shortly and feel really confident that you are mistaking getting things right for the wrong reason with actually get them right.

2022-10-20 19:36:08 excellent work, offering a different and important take on compositionality

2022-10-20 19:30:56 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb another great example of the mappings are naive: https://t.co/d1bO9t8Spj

2022-10-20 19:28:50 @ihorgowda if herschel is our goal, we are doomed

2022-10-20 14:53:10 where is AI relative to mouse cognition? excellent thread! mouse experts please add to it! https://t.co/71a6kQJbqy

2022-10-20 14:48:26 @kohn_gregory @SpeciesTypical @sanewman1 that’s silly. honestly we have nothing to talk about

2022-10-20 14:43:41 @kohn_gregory @sanewman1 @SpeciesTypical not seeing the argument there, but yes i was trying to make a more sophisticated nativism.i am sure the @WiringTheBrain would say the subsequent 17 years of dev neuro have been kind to that position: eg https://t.co/bUgW8d0XTQ &

2022-10-20 14:40:08 @kohn_gregory @SpeciesTypical @sanewman1 this is confused. it is like saying there is no information in an LP, because you need to have a record player to play it.if you flatten the groove, you lose the info

2022-10-20 14:37:13 @kohn_gregory @SpeciesTypical umm… we were talking about the physical structure of the hand, and you have swapped to the motor programs controlling it. obviously the latter involves learning.

2022-10-20 14:33:19 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb @ZoubinGhahrama1 should give @TristanThrush access so Tristan can run a Winoground, then.

2022-10-20 14:31:17 and a pioneer in common sense reasoning! and now on Twitter: @ErnestSDavis https://t.co/3VWbdX2wTi

2022-10-20 08:09:29 RT @BarryOSullivan: I highly recommend this talk! And check out Ernie’s excellent book with @GaryMarcus: “Rebooting AI: Building Artificial…

2022-10-20 07:40:51 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb also note that @slatestarcodex initially declared victory in a bet re Imagen, and then retracted. i will believe that Imagen is real progress on compositionality only when it is carefully demonstrated.

2022-10-20 07:30:29 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i have asked them endlessly to make it available or test the Winoground benchmark. they have refused. i can speak w authority only re systems i (or others) can test. but remain skeptical about those hidden from testing.

2022-10-20 07:25:16 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb ie it is this claim above on image synthesis that i am questioning w.r.t. grammar: https://t.co/QLqpvsjqyL

2022-10-20 07:23:02 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i was responding to a tweet that seemed to be about image synthesis

2022-10-20 07:02:11 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb best evidence that an image synthesis system captures some grammar? they seem to have some serious problems with binding.

2022-10-20 04:02:23 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb correction to last sentence: even when it gets it right it is bloviating. in no case does LLM do inference and reasoning over a world model

2022-10-20 03:59:45 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb they composit but without proper variable binding they aren’t properly compositional. they also fail on part-whole relationships, function, etcin some ways the language side is akin to bag-of-words model, which represent language in a very impoverished way.

2022-10-20 03:45:25 RT @databu: Interesting discussion with @GaryMarcus and @bengoertzel there

2022-10-20 03:21:11 @pmddomingos @erikbryn without social media the riot probably wouldn’t have happened

2022-10-19 04:19:03 RT @smjain: @GaryMarcus @ASteckley @ylecun Science only progresses via questioning . Always start from a position of doubt

2022-10-19 04:07:32 RT @GaryMarcus: @ASteckley @ylecun in philosophy, it is ok to ask people to clarify

2022-10-19 03:57:15 @ASteckley @ylecun in philosophy, it is ok to ask people to clarify

2022-10-19 03:56:16 @ASteckley @ylecun i pushed hard to understand his position on symbols recently, even after i was irritated because he had lied about my credentials, because i genuinely wanted to understand his position

2022-10-19 03:53:25 @ASteckley @ylecun I am not begrudging him his right to change his mind, of course. but the circumstances around his reaction to my 2018 article are hard to swallow.

2022-10-19 03:50:35 @ASteckley @ylecun he claims not to have changed his position: https://t.co/rpvFTx609B

2022-10-19 03:48:36 @ASteckley @ylecun from twitter advanced search i only see him using the term pejoratively, eg https://t.co/CKdhUumQxZ

2022-10-19 03:39:31 @ASteckley @ylecun absolutely

2022-10-19 03:37:05 @ASteckley @ylecun i have never deliberately done this

2022-10-19 03:35:36 @ASteckley @ylecun many people have offered to host and I agree that there would be value in a properly moderated debate.

2022-10-19 02:55:27 RT @TonyZador: Bees can learn complex tasks from other bees!

2022-10-19 01:22:11 @Jake_Browning00 @cameronjbuckner @ylecun @davidchalmers42 @NoemaMag 1. is there any specific reason to think it might work, without tacitly relying on symbol-manipulation?2. still sounds like planning not reflection and I am wondering if you have anything behind planning in mind.

2022-10-19 01:06:32 if you want to understand what makes driverless cars hard, follow @warren_craddock. https://t.co/tTw0EEnd6d

2022-10-19 01:01:23 @Jake_Browning00 @cameronjbuckner @ylecun @davidchalmers42 @NoemaMag even if we grant that that his aspiration- not clear that it can work - not clear what these states are if not symbols nor how they evaluated, nor how you cycle through them w/o variable binding- overall proposal sounds more like planning than reflection to me

2022-10-19 00:26:41 @Jake_Browning00 @cameronjbuckner @ylecun @davidchalmers42 @NoemaMag i am asking you for what you think the alternative is and why you think your title warrants the term “reflection”

2022-10-18 23:44:57 RT @Alber_RomGar: There's a debate on AI writing tools on Twitter right now. As an AI writer, I want to give my 2 cents.Here's my hot tak…

2022-10-18 23:20:42 @Jake_Browning00 @cameronjbuckner @ylecun @davidchalmers42 @NoemaMag that said, it may rest in part on what means by reflection …eg does it involve reasoning about one’s past and future actions?

2022-10-18 23:18:50 @Jake_Browning00 @cameronjbuckner @ylecun @davidchalmers42 @NoemaMag how are you going to represent propositional attitudes otherwise? (i don’t think i have written about it explicitly but have a related passage in a paper I am working on.)

2022-10-18 22:44:06 @ylecun @armchair_prof you do not. never seen you do it. you were wrong to call my 2018 paper “mostly wrong”, even though you have yourself recently endorsed essentially all the claims i made there. you also lied about my credentials. ZDNet had to publish a correction. you never apologized.

2022-10-18 22:39:31 Grow up. If you are going to make a lot of strong statements and insult other people’s ideas, people have right to ask what it is that you are saying, and to ask for clarification. https://t.co/q0BNvxuiba

2022-10-18 14:02:02 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag if you add all that, you are more nativist than if you don’t. you want a world in which you can be both a nativist and empiricist. that’s why most of us use terms like “innately guided learning” literally nobody thinks there is no learning or nothing innate.

2022-10-18 13:57:36 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag &

2022-10-18 13:54:29 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag but 1. you are assuming your conclusion - empiricists will of course be happy if they can get everything they need for free. but can they? they have failed on out of distribution abstraction and lack imagination in the sense of being able to think through counterfactuals etc &

2022-10-18 13:51:34 @cameronjbuckner @pwlot @rgblong @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag i know the paper. i always used the term to refer to “unlearned” machinery and not eg cross cultural universal (which sometimes is unlearned sometime not)empiricist is similarly used in multiple senses, as we saw yesterday

2022-10-18 13:49:41 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag also his most nativist work. 1. stands in contrast to all his more empiricist work that lacks duals routes and decried them (eg all his past tense stuff)and2. contrasts with vast preponderance of more empiricist current ML which lacks machinery for explicit memory

2022-10-18 13:45:03 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag good! then you can rewrite all your claims like these

2022-10-18 04:04:14 # of people liking this optimistic tweet: 980# liking the caveat that “Results are mildly cherry-picked: It isn't hard to stump it or make it hallucinate answers”: 54 https://t.co/HfiGda9Rwf

2022-10-17 23:17:58 @rgblong @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag if you can’t even connect your terms with current debates Lecun/Browning and i are having (which i do think is substantive), your terms aren’t very useful, nor tied to what is actually at stake.

2022-10-17 22:43:52 the opening paragraphs of this include a stunning bit of parenting https://t.co/jA6ol70Fwt

2022-10-17 22:21:07 science fiction in 2017, science fiction in 2022. https://t.co/8I1RlhizJd

2022-10-17 21:08:39 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag i am not seeing anything useful here that would eg help classify proposals or make predictions about them etc , since you are basically weakening the terms such that anything (even innate priors) counts as empiricism.

2022-10-17 21:07:01 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag spelke’s core knowledge is about as nativist as any actual proposal in developmental psychology. if you count her as an empiricist, you should count me as well. your way of rephrasing stuff is so loose it groups literally every proposal together except maybe Chomsky LAD.

2022-10-17 21:03:43 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag episodic memory surely requires representations, ergo the DeepMind passage is nativist, under your own definitions.

2022-10-17 21:00:56 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag and how can you call something that endorses installing high level prior knowledge “empiricist” (in the sense of that word that contrasts with nativist)? isn’t that the very definition of nativism?

2022-10-17 20:58:16 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag where do you draw the line on “minimal hand engineering”? where is the substance to the claim? is machinery for and representation formats for episodic memory “minimal hand engineer”? how about representational formats/machinery for symbols? how about representing space?

2022-10-17 20:42:36 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag i don’t even know what it means for the mind to be “tabula rasa” if you have that much prior structure. Am I actually an empiricist in your sense?

2022-10-17 20:40:58 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag i don’t think that the position your defending has anything to do with what Elman, Bates, Smith, Sutton, etc actually advocate for. Give me some names and quotes from people in the last 50 years, who are “empiricists” that endorse that level of (domain-general) innate structure.

2022-10-17 20:28:18 @cameronjbuckner @Jake_Browning00 @ylecun @davidchalmers42 @NoemaMag which empiricists are you talking about? if you have a whole bunch of detailed cognitive architecture prior to learning, that sounds like nativism to me.

2022-10-17 19:40:34 @pmddomingos @AllenInstitute if i had a nickel for every time some one argued for neural networks on the basis that symbols are (allegedly) not biologically plausible …

2022-10-17 16:55:26 @manlius84 @KordingLab @PessoaBrain @anilkseth @WiringTheBrain @ricard_sole @AFornito @MauCorbetta @spornslab @laurahelmuth yep it is a surprising error

2022-10-08 13:54:42 RT @GaryMarcus: @pmddomingos for all X, such that X is a book, if an LLM L read X, L would have no clue.

2022-10-08 13:51:00 @UlrichJunker hmm i think they have been ok at this, but interesting question.

2022-10-08 04:14:32 this thread is fire. https://t.co/wQ3uOvChJy

2022-10-08 04:13:59 @ylecun @amcafee too much to ask for you to recognize that the same considerations might apply to other long term avenues that might be worth exploring yet hard to fund given those constraints?

2022-10-07 21:32:51 @Grady_Booch indeed, we made this point explicitly in Rebooting AI, 3 years ago. little has changed.

2022-10-07 21:30:58 @DrSueSchneider you are kind of overlooking the differing innate contributions. it’s not just the data, it’s the starting place, too.

2022-10-07 21:28:40 @pmddomingos for all X, such that X is a book, if an LLM L read X, L would have no clue.

2022-10-07 20:58:46 @AmandaAskell agree w @AmandaAskell there is an important middle ground, call it near-term forseeable/plausible, that deserves real attention, that isn’t yet “in development” in the sense of being coded, in testing stage etc eg perhaps @sleepinyourhat’s example of personalized propaganda bot

2022-10-07 18:28:22 @tetraduzione @alirahimi0 @wellingmax would love your thoughts on this

2022-10-07 18:18:38 if you aren’t concerned, you aren’t paying attention. source ⁦@sleepinyourhat⁩, https://t.co/EcJlTafY33 https://t.co/ZpVs6hOpQ1

2022-10-07 18:05:56 @sleepinyourhat Glad to see you do this. I have some disagreements w you here (eg negation is far from solved), but this quote at the crux is 100% right: “Pretty much every bad outcome we’re seeing from present-day NLP (and related technologies) could get a lot bigger and a lot worse.”

2022-10-07 18:04:23 “Pretty much every bad outcome we’re seeing from present-day NLP (and related technologies) could get a lot bigger and a lot worse.” – Sam Bowman https://t.co/99p36iPlPf

2022-10-07 16:28:47 @yudapearl (think it is far from trivial

2022-10-07 16:23:48 RT @danfalk: Has #AI been over-reliant on deep learning? I spoke with @GaryMarcus for @undarkmag about the inherent limitations of deep lea…

2022-10-07 14:51:00 Virtually every important question in AI revolves around the scope and limits of generalization—yet there is little consensus about what generalization even means or how to test it. Check out @_dieuwke et al’s new synthesis, @genbench, to get some much-needed clarity. https://t.co/OVI4IaMvRe

2022-10-07 14:24:27 both humiliating and fair https://t.co/JayPbtjyBd

2022-10-07 14:13:34 RT @EthanVPorter: Fact-checks work--they reduce belief in misinformation. This is true in the U.S. and around the world. Today in @washingt…

2022-10-07 06:32:07 ain’t saying that another AI winter is coming, but numbers like these in @chafkin’s cover story @bw are startling. https://t.co/1LXvKkxVeY

2022-10-07 00:32:09 @warren_craddock @BW and also the key limits. but i fully agree those nuances are where the critical questions lie

2022-10-07 00:31:13 @warren_craddock @yasaman_razeghi @guyvdb definitely DL is more than memorization

2022-10-07 00:22:02 @warren_craddock this concise popularization doesn’t capture the more nuanced view i always express when given space, re training distribution

2022-10-07 00:20:55 @warren_craddock this is too coarse: see distinction above

2022-10-06 13:33:23 @sucholutsky @todd_gureckis @ylecun my commitment is to the fact the symbol manipulation is a necessary but not sufficient part of general intelligence. @todd_gureckis’s post is suggesting new ways in which that part of the contribution to general intelligence might be discovered.

2022-10-06 13:30:08 @KordingLab @tiagopeixoto @yudapearl @pmddomingos @StephenPiment @ylecun @aminkarbasi @AlexTensor the capacity for meta learning itself might have an innate basis, that most animals likely lack

2022-10-06 13:29:26 @pmddomingos @yudapearl @AlexTensor @StephenPiment @ylecun @aminkarbasi @KordingLab this is completely arbitrary. you don’t actually know the number, and you don’t have a trade #off curve for eg .75x vs 1.25x (on any metric). would human brain development work better or worse if we has 75% of the genes we have?as simple as possible doesn’t entail tiny

2022-10-06 13:22:59 RT @tomg_: Pour aller plus loin que le machine learning "Rebooting AI: le livre indispensable pour repenser l’intelligence artificielle" pa…

2022-10-06 06:13:29 @pmddomingos @yudapearl @AlexTensor @StephenPiment @ylecun @aminkarbasi @KordingLab why is it to essential to reduce it to say 5 things rather than 8? what is the argument that a slightly smaller list is more likely to succeed than a slightly larger list? and what system, ever, has induced causality?

2022-10-05 14:41:48 RT @GeorgiaChal: Read a comprehensive collection of expert opinions on ⁦@TeslaAIBot⁩ by ⁦@IEEESpectrum⁩. Thank you ⁦@BotJunkie⁩ for also fe…

2022-10-05 01:12:31 @frankcarey same thing w medical advice, therapy etc. fine for surrealist prose, tho

2022-10-05 00:54:39 @frankcarey if you show me 90 examples of a driverless car making a good left and the 10 other times it crashes, you have a demo rather than a solution

2022-10-05 00:53:38 @frankcarey and they aren’t one-off

2022-10-05 00:51:47 @frankcarey it’s not reliable. the prompt took a few minutes to break. in anything that can change profoundly change someone’s life we need some measure of reliability, and this approach doesn’t have it (eg in robustness across prompts)

2022-10-04 23:55:34 RT @sapinker: The shallowness of DeepFakes &

2022-10-04 23:33:53 @pmddomingos @yudapearl @StephenPiment @ylecun @aminkarbasi @AlexTensor @KordingLab some of the symbolic reasoning will be about (hypothesized and sometimes underdifferentiated) causal relationships.

2022-10-04 23:26:32 @pmddomingos @yudapearl @StephenPiment @ylecun @aminkarbasi @AlexTensor @KordingLab and the neural net can massively mess up the counterfactuals, or decide that mixing cranberry juice and grape juice will kill you if the word “die” is correlated in some corpus with the phrase “very thirsty”. medical advice from an acausal LLM would be a disaster. https://t.co/svGp44lk2S

2022-10-04 21:41:45 great thread! I agree both there is a ton of progress and that we need to reflect on it deeply—even though I doubt the progress has much to do with AGI per se. All of these things can be true at once. https://t.co/c8WqEFhxzN

2022-10-04 21:40:10 @Alber_RomGar @TaliaRinger awesome quote from @TaliaRinger!

2022-10-04 21:36:59 RT @TaliaRinger: I really think it takes a certain kind of echo chamber to view the AI progress in the last 5-10 years as being indicative…

2022-10-04 21:33:26 @Jaber_Hassoun @AmandaAskell why?

2022-10-04 15:25:11 https://t.co/4M3XW5vpF7

2022-10-04 13:17:21 @pmddomingos @StephenPiment @ylecun @aminkarbasi @AlexTensor @KordingLab @yudapearl how will you physical and psychological reasoning without it?

2022-10-04 13:16:34 @pmddomingos @yudapearl @ylecun @aminkarbasi @AlexTensor @KordingLab if your inductive bias includes logic, things definitely become more interesting

2022-10-01 04:23:36 can’t buy me AGlno, no, no, noI don't care too much for money Money can't buy me AGI

2022-10-01 04:14:26 ooh what i would give to debate Elon Musk on scaling maximalism.

2022-10-01 03:56:38 . @elonmusk argues that since @tesla (supposedly) has the most data and most compute, it follows they will necessarily will contribute to artificial general intelligence.i disagree w this logic: building biggest model isn’t inherently a contribution to the innovations we need

2022-10-01 03:38:54 @rachelpurpel also on four hours notice? insane

2022-10-01 03:38:24 @simicvm @verge @animesh_garg @Ken_Goldberg

2022-10-01 03:36:49 brutal https://t.co/0EYnspFSZ3

2022-10-01 03:28:58 @animesh_garg @frossi_t @verge certainly there was nothing compelling about navigation, and almost nothing at all about cognition.

2022-10-01 03:27:55 https://t.co/RMekj7qBb5

2022-10-01 03:21:45 trick question? https://t.co/mYar2lfpgj

2022-10-01 03:19:48 @frossi_t @cwty exactly

2022-10-01 03:16:45 . @cwty is on https://t.co/eBJ37MjOIN

2022-10-01 03:08:01 best mesh networks alternatives to Eero? i just got notice that they are stopping security upgrades to some of my devices … in 4 hours.are you ####ing kidding?

2022-10-01 02:55:03 @mark_riedl @Ken_Goldberg @elonmusk had same thought

2022-10-01 02:52:48 @tpgoebel you can’t build a humanoid domestic robot that you can trust without making some progress towards AGI. it just won’t be safe.

2022-10-01 02:46:52 your prior on how soon AGI will come, updated after tonight’s presentation

2022-10-01 02:35:45 @filippie509 cool. i turned it off too soon.

2022-10-01 02:24:28 this aged pretty well https://t.co/gxXuR0gD5A

2022-10-01 02:15:50 https://t.co/bW8otC0RM5

2022-10-01 02:11:49 shame on @verge for repeating this as truth without asking anyone else in the field of robotics for comment. https://t.co/qNHxsQx6UE

2022-10-01 02:09:01 concur with @cwty

2022-09-29 11:32:43 @wellingmax @maxwelling i specified above what i believe that the debate was actually about: whether or not it was healthy for the ML to engage in work outside its mainstream that challenges core assumptions that have not been proven.

2022-09-29 11:30:55 @wellingmax @maxwelling yes, max,. i am sure the talk was terrific (I trust your judgement and like your own work in the area) but your characterization of the debate missed the point of the debate.

2022-09-29 02:54:34 @douglas_eck @JFPuget @ylecun could be worse, I suppose, than to be compared to Socrates https://t.co/ez8jsU1mPV, who pointed out that “the cost to society of silencing individuals who were irritating could be very high”

2022-09-29 01:57:06 5 reasons @tesla's Optimus Announcement may turn out to be a dud. https://t.co/vAgI3w8coB

2022-09-28 22:44:34 @JFPuget @ylecun this is BS. i shared links to a wealth of other researchers and couldn’t get you or yann or *any* of the Ml community to engage with any of them, even where results were decisively stronger than DL for certain problems.that’s what the whole debate was about.

2022-09-28 22:02:35 @JFPuget @_NicT_ @ylecun @JanelleCShane not going to get in a fight with Domingos who has undoubtedly done lots of important work, but i have peer review articles in Science, Nature, and leading journals in half a dozen different fields, myself, as well 6 books and monographs etc spanning psychology, AI, biology etc.

2022-09-28 21:59:29 RT @Grady_Booch: @MelMitchell1 @AndrewLampinen @math_dandy @GaryMarcus It's clear (to some of us) that DL methods alone won't scale.

2022-09-28 21:58:01 no question that a moderated debate would allow more depth than twitter. the 2017 debate moderated by @davidchalmers42 revealed real differences on innateness as well as important points of agreement.twitter’s brevity and virtue signaling won’t ever substitute for that. https://t.co/iECEXYYjrz

2022-09-28 21:45:28 @wellingmax @ylecun https://t.co/fBUELlKymJ

2022-09-28 21:44:53 @MelMitchell1 @AndrewLampinen @math_dandy would *anyone* disagree with that?

2022-09-28 21:40:04 @omaruddin @maxwelling oops @wellingmax

2022-09-28 21:22:00 Even the very bright @maxwelling has missed the point. The real debate is about whether the AI community should engage in the fast growing neurosymbolic literature (to which I gave numerous pointers). Or should we turn our back, because those in power don’t wish to engage? https://t.co/8iKWX62xoH

2022-09-28 21:18:54 @AndrewLampinen @MelMitchell1 @math_dandy yes that would be a disappointment, and that noisy, nuanced and complex world is why I gravitate to something like neurosymbolic rather than pure GOFAI (or for that matter pure deep learning which is not great around context or nuance).

2022-09-28 20:18:55 if you want to use a commercial service for AI, where do you head first?

2022-09-28 20:02:44 RT @khulick: It was great hearing from @GaryMarcus today about #AGI “predicting statistical words has nothing to do with concepts of the wo…

2022-09-28 20:02:12 @DanRodricks as a former guest one one of your shows, i can’t be there in person but would love to see a recorded performance!

2022-09-28 20:00:16 @pwlot @OpenAI they also put out whisper. i live in hope that they will return to their name.

2022-09-28 19:36:08 @openAI is getting more open again

2022-09-28 16:23:32 @Tom14985282 @slatestarcodex see my bet with elon musk at https://t.co/8ir1xKenr6 and tell me how you would change it

2022-09-28 14:18:50 RT @neuromatch: Scholars in all disciplines &

2022-09-28 12:45:36 RT @NoemaMag: How do we get to AI that is safe, trustworthy &

2022-09-28 04:22:13 @davidchalmers42 and in a related vein:https://t.co/JTgYtV8JTk

2022-09-25 17:03:34 @NickRMorgan @fchollet @TristanThrush @candacerossio Many are, @fchollet is not, consistently a cut above. I have great respect for him, even though in this instance he is a bit off.

2022-09-25 16:49:20 @ihorgowda @AoDespair Greatest honor of my twitter life. @aodespair’s Homicide: Life on The Streets is one of my 10 favorite books of all time! To say nothing of H:LOTS and The Wire. And yes it is true that we both revel in speaking truth to power.

2022-09-25 16:36:43 What I am thinking about as I write today’s essay https://t.co/LjPnV3wJDX

2022-09-25 16:35:06 @jbondc @MetaAI @ylecun then why has spent 2022 giving interviews and writing papers about not-yet implemented theoretical ideas? that’s having it both ways.people—not just him—should be allowed to do both. it is the height of hypocrisy and abuse of power to do as he has done.

2022-09-25 16:13:52 RT @SchwabeHenning: @ylecun ‘s unfair attack is tragic b/c it will help DL maximalist to dismiss his critique - he should have allied with…

2022-09-25 16:13:22 @hughes_meister some aspects need to be learned. the mere existence of space and time and how to reason over them probably ought to be baked in. without that, learning is grossly inefficient. (convolution does build in one aspect of space, and that’s why it works so well)

2022-09-25 16:09:20 @jasonbaldridge @fchollet please test @TristanThrush and @candacerossio ’s Winoground and publicly disclose the results

2022-09-25 15:43:59 No, @fchollet, these models are *not* compositional in the sense that linguists discuss.They consistently fail on tasks like @TristanThrush and @candacerossio’s Winoground. See also my substack (Horse Rides Astronaut &

2022-09-25 15:38:11 if anyone could introduce me to Liz Cheney’s staff please let me know https://t.co/07fojw6uOI

2022-09-25 15:36:20 @danielmalmer top of my list of people to reach out to

2022-09-25 15:09:16 @ProfNoahGian @sinanaral @noUpside @katestarbird yep

2022-09-25 15:03:49 @taamfp @Twitter @SchmidhuberAI thanks

2022-09-25 15:02:38 @Dataprogr @ylecun his “global insight” in the ZDNet interview consisted primarily of regurgitating the key points in my 2018 article deep learning: a critical appraisal (which he once called “mostly wrong”).

2022-09-25 14:43:47 @taamfp @Twitter no although i will include that as well. what i remember was a side by side comparison. half looked like this, half was from @SchmidhuberAI, highlighting the parallels. I could reconstruct it but would prefer the original tweet/pair of images https://t.co/q5E3ACLphx

2022-09-25 14:35:47 Hey @twitter please extend Twitter Advanced Search to a user’s own Likes. I want to find something I liked in July, and because I like a lot of stuff there is no efficient way to do it.Easy feature to add, would be very useful.

2022-09-25 14:26:50 @Zergylord @ulusdd @ZDNET @TiernanRayTech @ylecun yes (elaborated in a *book* that was peer reviewed by MIT press, The Algebraic Mind)

2022-09-25 14:24:06 @Zergylord @ulusdd @ZDNET @TiernanRayTech @ylecun of course lecun stacked the deck. the most important peer-reviewed contribution was an analysis of neural networks that happened to be published in a psychology journal, anticipating all the problems with distribution shift for reasoning and language that are prevalent now.

2022-09-25 14:15:53 @ulusdd @Zergylord @ZDNET @TiernanRayTech @ylecun i will put some of them in an essay i am writing this morning. or go se the correction Zdnet issued if you are impatient.

2022-09-25 14:14:07 anybody have handy the comparison tweet between LeCun’s recent manifesto and @SchmidhuberAI’s early work, side by side figure if the 2 ? i need the tweet for something i am writing.

2022-09-25 14:00:58 @hughes_meister true, but the deeper problem is the lack of explicit world models. adding video doesn’t immediately solve that

2022-09-25 04:59:43 thread in which @Grady_Booch and I articulate a rare disagreement we have with @fchollet. https://t.co/MJnYTQ6ZLs

2022-09-25 04:43:16 @fchollet no, it’s the difference between having an explicit functional model and working purely in an image space that is correlated with labels.

2022-09-25 04:42:04 @fchollet i wrote some about this last week, with further examples: https://t.co/1xciYldIVi

2022-09-25 04:36:35 RT @timnitGebru: From 2018: https://t.co/yYVHdDp2zP"When Artificial Intelligence works as intended, Silicon Valley types often say it's "l…

2022-09-25 04:34:58 @linasvepstas indeed. but i am not looking for people who exclaim Fake News :) I am looking for genuine bipartisan support to do better.

2022-09-25 04:24:35 @katestarbird @noUpside @sinanaral @ProfNoahGian thanks!!

2022-09-25 03:49:55 @prem_k but not inevitable that one cobblestone would dismiss the other and lie about his credentials.

2022-09-25 03:48:28 @noUpside @sinanaral @ProfNoahGian @katestarbird thank you!!

2022-09-25 03:28:19 @sinanaral @noUpside @ProfNoahGian @katestarbird any leads?

2022-09-25 03:24:54 hive mind, are there any well-known Republicans who have spoken out against misinformation?

2022-09-25 03:14:25 @Joegle richest man I ever treated to lunch :) (but only because the place i took him to didn’t allow guests to pay)

2022-09-25 03:00:47 @joegle thank you!

2022-09-25 02:55:53 my next follower will be my 70,000th. thank you all who follow for all your support, through thick and thin!(@twitter, you never responded to my request for verification

2022-09-25 02:28:04 @maartengm @MetaAI @ylecun it’s flagrant disregard of academic etiquette and he libeled me as well, claiming I don’t have any peer-reviewed work in AI journals, which is demonstrably false.

2022-09-25 02:24:26 @ZDNET @TiernanRayTech This libels me, by claiming—falsely—that I have not published any AI work in peer review journals &

2022-09-25 01:58:13 @KrishnaKGupta @MetaAI @ylecun @ZDNET here are the highlights in the comparison https://t.co/RYU1ytVnEa

2022-09-25 01:54:09 I’ve added receipts to the document the uncanny similarity between LeCun’s 2022 view and my own in this thread: https://t.co/RYU1ytVnEa

2022-09-25 01:50:57 See also @SchmidhuberAI: https://t.co/NgUb5Q6VKW (and @NoemaMag https://t.co/7WD8XxMEAr)

2022-09-25 01:50:56 LeCun 2018 on Marcus 2018: It’s “mostly wrong”LeCun 2022: Marcus has never made a contribution to AI

2022-09-24 19:33:55 @pressron 5 minutes on Twitter search should be enough to make obvious that it is in fact used to stereotype women.

2022-09-24 19:30:17 @mettle sexist, racist, and deliberately dehumanizing.

2022-09-24 19:15:12 @willbeason yep, like many tasteless jokes and racist slurs we have wisely deprecated.

2022-09-24 16:12:14 Why is the currently-trending word #Karen, used derogatively to stereotype white women, social acceptable, when so many other derogative stereotyping words are—rightly—no longer socially acceptable?The ideal we are aspiring to is *stop* judging people based on race &

2022-09-24 15:04:57 @chantz_y @TheShoeLady33

2022-09-24 15:04:21 @chantz_y And who could forget Mass Delusion 101?

2022-09-24 14:30:24 RT @bobehayes: Form, function, and the giant gulf between drawing a picture and understanding the world @GaryMarcusWhile these systems ca…

2022-09-24 14:05:56 @ImageSnippets saw that. nauseating

2022-09-24 13:46:00 RT @IntuitMachine: @GaryMarcus Not just that, a move towards worldwide fascism will destroy any possibility of future AI alignment. The iss…

2022-09-24 00:25:28 fully agree that misinformation is massive problem, particularly for democracy, and believe that a descent into fascism brings many risks, and probably doesn't mix well with AGI. https://t.co/2kWbthVbaQ

2022-09-24 00:23:17 @sergia_ch @emilymbender @timnitGebru My view, since you were kind enough to ask: one should assume good intentions unless there is direct evidence otherwise, and there is at least 3-5% chance that a well-argued entry could change the minds of people economically equipped to do something significant.

2022-09-24 00:17:52 RT @swarat: ...between the neural and symbolic aspects of the problem, as gradients prove powerful in symbolic tasks and code/math look clo…

2022-09-24 00:17:47 RT @swarat: Nice @CACMmag article by @donmonroe on neurosymbolic learning. It's been wonderful to work in this area over the last few years…

2022-09-23 21:36:07 RT @bobehayes: Another great read by @GaryMarcus “Those who wish that #AI criticism would go away are welcome to their views, but misrepr…

2022-09-23 21:10:19 in almost a decade of *talking* about fairness in AI, a lot of things haven’t much changed. https://t.co/9hRIa9i8Zt

2022-09-23 21:08:37 @chrissyfarr @bing and presumably not just @bing, but likely a general pattern across a wide range of ML engines that are driven entirely by convenience samples of data rather than ethical principles (or even up-to-date data).

2022-09-23 21:03:15 @emilymbender @timnitGebru i do share your concerns about wealth distribution.

2022-09-23 21:00:04 @emilymbender @timnitGebru by all means you should write about what you want to write about. (but I thought that you believed these guys were misallocating their resources and that that mattered to you.)

2022-09-23 20:55:29 @TaliaRinger they did the first time, and then it got easier

2022-09-23 20:53:06 @emilymbender @timnitGebru again this is directly framed as an opportunity to tell them not to do that.why not raise your best challenge, and post it as a blog if they don’t publish it?

2022-09-23 20:51:26 @emilymbender @timnitGebru they have asked, and are entitled to ask, for an original piece of work. I may enter

2022-09-23 20:31:54 @timnitGebru @emilymbender i am not seeing that in the rule requirements

2022-09-23 20:19:16 @Kdawg5000 @ftxfuturefund funnier if they won

2022-09-23 19:52:03 Terrified by the thought of AGI risk? Or think the whole discussion is misguided? Essay contest w serious prizes and a chance to advise the @ftxfuturefund and how to best help the world.Win-win. https://t.co/OkI0Jg9b7q

2022-09-23 18:15:18 RT @luislamb: Communications of the ACM article on #Neurosymbolic #AI, by Don Monroe References to our work with @AvilaGarcez and reflecti…

2022-09-23 16:46:12 @Sara_Imari AI thus far? or AI’s potential? very hard to dichotomize, either way…

2022-09-23 16:03:41 RT @CarissaVeliz: "The issue is not simply that deep learning has problems, it is that deep learning has consistent problems." By @GaryMarc…

2022-09-23 15:35:29 RT @bimedotcom: @EvaSmartAI @GaryMarcus @SubstackInc @sallyeaves @nigewillson @Nicochan33 @mvollmer1 @Shi4Tech @SusanHayes_ @RagusoSergio @…

2022-09-23 15:31:42 RT @Schellenberg_1: Great new volume edited by Keya Maitra and Jennifer McWeeny. This will be a groundbreaking volume for the imminent soci…

2022-09-23 06:18:26 @miguelisolano @davidchalmers42 @ylecun it’s also central to my 2020 Next Decade in AI arXiv: https://t.co/rbeWGMenKO

2022-09-23 06:11:22 @miguelisolano @davidchalmers42 @ylecun this is where i first suggested in 2019 that LLMs had a problem with modeling the world:(&

2022-09-23 05:35:20 RT @LaurenAulet: Someone give Susan Carey a book deal to write a full autobiography. This piece is so wildly inspiring and entertaininght…

2022-09-23 04:35:23 RT @JustinBullock14: @GaryMarcus respond to @erikphoel:"The hear of Hole's argument boils down to (my words not his) "AI has done better…

2022-09-23 03:57:45 @NaveenGRao 2/2 i am a big fan of both evolution and learning but think they need to be thought about in different ways, and the ML community is often unduly dismissive of genetic contributions.

2022-09-23 03:55:00 @NaveenGRao for me learning is about changing some internal representation based on data. AFAIK slime mold aren’t doing that, but I don’t really know the details, so am cautious.1/2

2022-09-23 03:34:23 @NaveenGRao i think you are committed to “yes” regardless of mechanism, whereas for me the details (on which I am not expert) matter. but see https://t.co/ZXTKRgSwKW

2022-09-23 03:29:25 @NaveenGRao @A_Aspuru_Guzik @HarvardCCB Of course heart adapts to changing constraints, but if you want to understand how you have to understand eg dynamics of gene expression, not back prop or the Hebbian rule or RL or observational learning etc.

2022-09-23 03:26:38 @NaveenGRao you are committed to saying the slime mold are learning, no? I am the one saying that not all adaptation is (usefully construed as) learning.

2022-09-23 03:25:24 @NaveenGRao the fjord adapts to the glaciers, no? i don’t see how you can even pose Hinton and Nowlan’s question if you think evolution and learning are the same thing.

2022-09-23 03:23:39 @A_Aspuru_Guzik @NaveenGRao @HarvardCCB eg, Hinton &

2022-09-23 03:19:55 @NaveenGRao @A_Aspuru_Guzik @HarvardCCB what does describing the heart as being learned actually buy you, in terms of explanation and prediction?

2022-09-23 03:19:08 @NaveenGRao sure, you can if you loosen the definition to “change over time, driven by the environment”. or you can trivially assume your own conclusion by defining learning as “any optimization over time”.

2022-09-23 03:17:22 @A_Aspuru_Guzik @NaveenGRao @HarvardCCB using the words learning, adaptation and optimization interchangeably would not make the paper you attached clearer.

2022-09-23 03:16:08 @A_Aspuru_Guzik @NaveenGRao @HarvardCCB learning and developmental biology are different ways of optimizing in an landscape. learning is a *subset* of the ways in which things can be optimized

2022-09-23 03:14:13 @EugeneVinitsky it’s an irritating joke, because nothing could be farther from the truth. I explain here: https://t.co/YMRYkpxXnt

2022-09-23 03:07:41 @NaveenGRao ps i am very well aware of the mechanisms of developmental biology, and wrote a whole book about them in comparison to and cooperation with learning.

2022-09-23 03:06:08 @NaveenGRao and you can describe the structure of fjord is learned. but it is not that helpful.

2022-09-23 02:45:31 @NaveenGRao you could use “adaptation” maybe in that sense. in what useful sense would you say that the structure of the heart is learned? what does that tell you about the mechanisms of biology? or evolution? or how eg those process differ from synaptic tuning or observational learning?

2022-09-23 02:41:34 w links to work and observations from @emilymbender @MelMitchell1 @ylecun @geoffreyhinton @guyvdb and @metaai

2022-09-19 04:24:58 @hardmaru ha! Beatles beat you to it, You stay stop, and I say go go go (or Goo Goog Gio Goo…)https://t.co/tsOZphbz72

2022-09-19 04:21:45 any chance that someone reasonable could slip through as the Republican nominee in ‘24, while these two slug it out? https://t.co/OcBfmpBM87

2022-09-19 03:53:30 RT @Shayan86: Mahsa Amini, 22, was arrested by Iran's morality police for so-called "inappropriate hijab". She fell into a coma at a police…

2022-09-19 03:08:24 RT @EricTopol: Wish this was true. What's over is @POTUS's and our government's will to get ahead of it, with magical thinking on the new b…

2022-09-19 02:13:55 @grbradsk and see my newest https://t.co/8ir1xKenr6 for how impoverished the current representations are

2022-09-19 02:13:16 @grbradsk the interesting question is whether you can bridge such more sophisticated world models into something that currently lives purely in image space—and how.

2022-09-19 00:49:50 @jasonintrator whereof he could not speak he would remain silent

2022-09-19 00:04:19 @MadamePratolung @EnglishOER @TheAtlantic @StephenMarche @mer__edith @nxthompson

2022-09-18 23:17:16 RT @gemmagaldon: Tan importante esto. Los sistemas de creación de imágenes tipo #DALLE no nos acercan a ningún tipo de inteligencia artific…

2022-09-18 23:16:02 @mikekwarner @FryRsquared guess you didn’t get too far into the essay before tweeting?

2022-09-18 23:06:37 @VictorButoi see my new essay for a lot of examples of requests they fail to model correctly https://t.co/Xnyh3pTeNo

2022-09-18 22:55:32 RT @kwiens: The lack of functional comprehension makes any mechanical outputs from these systems laughably incorrect. I love playing with S…

2022-09-18 21:53:38 @TonyZador here’s the nucleus of a relevant benchmark

2022-09-18 21:51:25 my own answer, which quotes @EMostaque, is now posted here: https://t.co/sf5BaT0MAX

2022-09-18 21:49:37 Hold on, folks. Are systems like DALL-E and Stable Diffusion really an important step towards general intelligence?Form, function, and the giant gulf between drawing a picture and understanding the world https://t.co/Xnyh3pTeNo

2022-09-18 21:06:50 @KatHogan9 i was actually thinking of something less fraught with questions about intentionality but the comparison is fair

2022-09-18 20:19:56 @TonyZador @EvelinaLeivada and in that sense, yes i think that many animals can to some degree understand the world, as evidenced eg when they use novel tools to solve novel problems

2022-09-18 20:19:16 @TonyZador @EvelinaLeivada how about “form a dynamically updatable representation that an agent can reliably reason with”

2022-09-18 19:59:50 How much do systems like Dall-E and Stable Diffusion understand about the world that they illustrate?

2022-09-18 17:00:51 @tdietterich @etzioni and four digit multiplication is basically at zero

2022-09-18 03:26:12 RT @amandalhu: 2020: “COVID is not airborne, and COVID doesn’t spread in schools so we don’t need airborne precautions, and esp not in scho…

2022-09-17 13:24:13 RT @IntuitMachine: DL will continue to be very useful, but it hits the wall wrt to autonomous systems and higher-level abstract thinking. I…

2022-09-17 12:37:06 RT @MayaPlentz: “Using a homegrown benchmark of a few examples and drawing big conclusions is a further no no, a fallacy of overgeneralizat…

2022-09-17 03:19:42 RT @KripkeCenter: We mourn the passing but celebrate the life and achievements of Saul Aaron Kripke (Nov. 13, 1940–Sep. 15, 2022). His fami…

2022-09-16 23:02:31 RT @HelloSurgeAI: Entering the @slatestarcodex vs. @GaryMarcus debate... What should "pixel art of a robot farmer in a cathedral holdin…

2022-09-16 16:41:11 RT @JoyceWhiteVance: "There is surely a special place in hell reserved for those who treat immigrants as so much refuse to be deposited wit…

2022-09-16 15:08:17 RT @isabel_pedersen: “It’s time for artificial-intelligence researchers to look up from the flashy, straight-to-the-media demos and ask fun…

2022-09-16 02:45:42 RT @GaryMarcus: @republicofspin @peterbakernyt @sbg1 brilliant closing https://t.co/YdWpyznmvr

2022-09-16 02:30:32 @republicofspin @peterbakernyt @sbg1 brilliant closing https://t.co/YdWpyznmvr

2022-09-16 02:21:06 @pmddomingos most of the people who share my concerns are simply ignored, which is no better. that largely extends to many of the concerns you yourself have raised.

2022-09-15 23:42:08 the sequel is even more insane the original. https://t.co/Z6y0vnPEk1

2022-09-15 22:22:20 Vindication for @BandyXLee1 who bravely foresaw so much https://t.co/8TSo7Bxa4k

2022-09-15 21:25:17 thank you to all who made this cry for more serious scientific practice in AI #5 on Hacker News https://t.co/w3nykjsIf4

2022-09-15 21:20:19 @TristanThrush @0x00B1 @mhtessler @GoogleAI @slatestarcodex @MetaAI @ylecun I think this is legit concern and have reached out to some highly placed folks at Meta to try to fix it.

2022-09-15 21:15:04 RT @0x00B1: I can’t believe it but I agree with (nearly) everything in a @GaryMarcus essay! He’s absolutely correct and I encourage anyone…

2022-09-15 18:58:23 RT @TaylorOgan: There are so many shocking things about this:1. That Elon Musk publicly tweeted this.2. That Elon Musk thought this in 20…

2022-09-15 18:56:15 RT @jmhessel: Does AI ""understand"" The New Yorker Caption Contest? (spoiler: no )excited for this fun collaboration!(data/models/code…

2022-09-15 15:59:32 @omershapira if you run across more recent #’s i would certainly be interested, cc @Abebab

2022-09-15 15:32:23 @omershapira source: https://t.co/JLSDrh367z

2022-09-15 15:23:21 society’s machine learning dollars at work https://t.co/MirZrBmvGw

2022-09-13 04:52:55 RT @lenlayton: @GaryMarcus Just finished reading “Rebooting AI.” Great stuff! Aligns with my observations over the last 25y. Cost/benefit o…

2022-09-13 03:19:41 Why is @googleAI too chicken to share Imagen with @huggingface’s @TristanThrush? cc @ClementDelangue https://t.co/XwZm6OK8R7

2022-09-13 01:22:00 RT @ProfArbel: Access to skeptics is vital for the development of robust, safe, and effective AI systems and I trust @googleai will make th…

2022-09-13 01:20:58 @ProfArbel @GoogleAI they haven’t so far, despite repeated requests, public and private

2022-09-13 01:04:42 shame on you, @googleAI. now nearly four months. you gave @slatestarcodex access to do a tiny bit of anecdotal research but continue to duck the repeated requests of professional scientists scientists like @TristanThrush who have created systematic benchmarks. https://t.co/hJ1visaNeO

2022-09-12 21:53:13 @David_Dobbs @philipcball @Grady_Booch

2022-09-11 19:27:48 RT @christogrozev: Zelensky to Russia, tonight... https://t.co/6Hgot3zKhY

2022-09-11 15:42:27 RT @JoyceWhiteVance: Roe, Roe, Roe your vote.It's the November theme.Merrily merrily, merrily, merrilyor Choice is but a dream.

2022-09-11 00:38:40 @tdietterich @DrMJoyner @DavidEpstein @nxthompson @cragcrest @geoffreyhinton @ACasadevall1 driving might “seem” trivial too… Apple Mail filtering is sometimes great and sometimes way too aggressive. for a while it was screening out my own mother (in my email contacts for a couple decades).

2022-09-10 23:32:28 RT @DrMJoyner: I rarely check my spam folder - found an email with a draft grant application from a collaborator…if AI/ML can’t do email sp…

2022-09-10 20:53:04 @espiers @jbarro @Grady_Booch

2022-09-10 20:52:53 RT @GaryMarcus: @espiers @jbarro Lex Fridman blocked me. A badge of honor?

2022-09-10 19:00:05 @espiers @jbarro Lex Fridman blocked me. A badge of honor?

2022-09-10 17:58:55 RT @visegrad24: BREAKING:The Ukrainian Army has entered Vovchansk.It means that Ukraine is taking back all the territory up to the bord…

2022-09-10 14:52:18 RT @NatashaBertrand: “Recapturing the two cities is a big strategic win for Kyiv…Russia was using Kupyansk as a rail hub to resupply its fo…

2022-09-10 14:22:25 RT @petestrzok: Better check Bedminster…On May 6, NARA emails Trump to say material is missing and may be at MAL.https://t.co/XbEWvl9Wd…

2022-09-10 14:14:35 RT @ukraine_world: The dynamics of Ukraine's counter-offensive in Kharkiv Oblast in the last couple of days. The video taken from the Tel…

2022-09-09 17:40:20 RT @sayashk: It is well known that image generation tools encode stereotypes about people. But do these models also perpetuate stereotypes…

2022-09-09 12:56:33 RT @brad262run: “sooner we can figure out what basis allows a system to learn symbolic abstractions, the sooner we can build systems that p…

2022-09-08 20:31:50 this is a pretty satisfying read. https://t.co/s1rCLyud5U

2022-09-08 18:23:52 RT @info_sprinkles: “There will be a death tied to a large language model in the next year.” Great interview on Foundation Models with @…

2022-09-08 15:27:08 https://t.co/TghmuE0fH8

2022-09-08 00:33:56 funny how @gdb didn’t choose this example when tried to troll me https://t.co/QmZ2GRU6qE

2022-09-07 19:18:29 RT @info_sprinkles: @GaryMarcus is one of my favorite AI folks talking reality &

2022-09-07 18:15:07 very disconcerting - if automatic driving systems fail, even attentive people may be slow to respond https://t.co/kNgcja7PF0

2022-09-07 18:05:13 RT @Kantrowitz: Here's my interview with @GaryMarcus. We talk about LaMDA, Dall-E, and this weird moment where AI is very powerful but stil…

2022-09-07 18:04:21 @PaulTopping it’s a *long* article, still in progress (with associated shorter post, too, presumably).

2022-09-07 17:08:31 Want a chatbot to teach you how to sell (and steal) drugs? Fantastic and terrifying paper by @anthropicAI that outlines some of the many risks we are likely to face in the near-term as chatbots become pervasive. Mandatory reading. https://t.co/oe80Czuy9o https://t.co/uOeFUlppr2

2022-09-07 15:30:59 @Abebab well-deserved!

2022-09-07 13:50:34 @Abebab premature closure is getting the world into a heap of trouble.

2022-09-07 13:47:32 Indeed! And a lust for simplicity has also led to overinvestment in big data + deep learning, at expense of considering richer but more difficult to construct hybrid solutions—and a world of approximative but deeply unreliable mimicry, easy to construct, impossible to count on. https://t.co/kDlaJGqDPc

2022-09-07 02:23:32 @ErnestSDavis https://t.co/9IqhX3ahUA

2022-09-07 01:25:12 @ErnestSDavis the ramifications are endless. here’s another, further recognized a few hours later, made incomparably worse if Cannon’s decision stands https://t.co/4JNhiPgGMF

2022-09-07 01:16:52 AI-generated list of all 50 states, text to image, by @nostalgebraist, h/t @slatestarcodex https://t.co/RwVq9pjFk0

2022-09-06 22:13:45 @ErnestSDavis and see https://t.co/pgwdq8m0gM

2022-09-06 22:13:01 @ErnestSDavis if it is not overruled no criminal lawyer will ever forget it.

2022-09-06 22:08:30 @ErnestSDavis https://t.co/0GHGYH61vJ

2022-09-06 22:06:19 @ErnestSDavis https://t.co/ADYmfIpSbi

2022-09-06 01:48:35 @robinc it’s pretty much what the exec told me. i was flabbergasted.

2022-09-05 23:58:41 @robinc it’s MINERVA, link elsewhere in this thread.

2022-09-05 22:58:37 @robinc and ffs the paper was about using LLMs for math, so don’t blame me. i just reprinted a figure.

2022-09-05 22:57:55 @robinc ditto for using an LLM for natural language understanding (rather than autocompletion)

2022-09-05 22:04:37 what i was trying to say earlier today, said better https://t.co/y8uyuQhPYA

2022-09-05 22:00:33 @robinc it’s clean a test case for the generalizability of the systems

2022-09-05 18:29:56 RT @AshaRangappa_: Expert on fascism here

2022-09-05 16:08:55 without doubt, one of the worst decisions in US legal history, without sound legal reasoning and (by wasting time) endangering the entire nation, in favor of the person who appointed the judge.even Bill Barr knows this a “crock of shit.” https://t.co/44tutb5Jv3

2022-09-05 15:50:50 h/t @ZeeshanAleem

2022-09-05 15:50:11 misinformation rears its ugly head yet again https://t.co/pVxg70NsnP https://t.co/Z4uXJibw1z

2022-09-05 15:27:35 @MadamePratolung @91_pometti and resentment towards those unawed

2022-09-05 15:17:47 @jon_forsyth @91_pometti the latest fuss was over a prize. (few seem to notice that it was in the digital category with a measly 20 entries at a state fair)

2022-09-05 15:05:38 great thoughts on the art and AI hysteria, by @91_pometti: https://t.co/q6Om56BWLU

2022-09-05 13:15:43 @srchvrs Minerva: https://t.co/UKUHGjNy7R

2022-09-05 13:14:22 @philljkc compare it to a calculator.

2022-09-05 13:13:59 @Kdawg5000 almost certainly yes

2022-09-05 00:16:21 RT @GaryMarcus: @MadamePratolung The Emperor’s New Calculator?

2022-09-04 23:46:54 @MadamePratolung The Emperor’s New Calculator?

2022-09-04 23:00:31 chance a certain major corporate exec will allege that large language models have solved the problem of generalization after you call bullshit: 100%chance that an LLM will be able to correctly multiply two four-digit integers together in a novel multiplication problem: 0% https://t.co/UmXEkQiSS1

2022-09-04 22:52:52 @Meaningness @ESYudkowsky dm/follow so we can discuss a little privately

2022-09-04 22:37:15 @Meaningness - i think your second paragraph (a) proves too much (eg legal systems work reasonably well, despite the slippery slopes) and (b) doesn’t say why we can’t implement values but only asserts (correctly) that we haven’t yet- link to @ESYudkowsky argument on extrapolated volition?

2022-09-04 19:20:13 RT @MayaPlentz: "Fake news can sway elections, tank economies and sow discord in everyday life. Data scientist @sinanaral demystifies how…

2022-09-04 15:46:58 RT @tribelaw: Unless Cannon “wants to rewrite decades of law, let Trump hold the nation hostage and deepen our national security crisis, sh…

2022-09-04 04:08:02 RT @Anki98765: Many gems here. I most resonate with "Large pretrained statistical models can do almost anything, at least enough for a proo…

2022-09-04 03:56:43 RT @consequence: Taylor Hawkins' 16-year-old son, Shane, does his father proud as he sits behind the kit for @FooFighters' performance of "…

2022-09-04 03:51:08 @Todd_S_Woodward @random_walker um, “can easily be manipulated” sure sounds like blaming the users. maybe you didn’t quite finish the abstract?

2022-09-04 02:40:04 irony is dead https://t.co/XOPSepGWj6

2022-09-03 21:21:05 RT @katealexandrite: Found this thread https://t.co/J6Kp441lgUin the @garymarcus piece and feel like both do an excellent job at clarifyin…

2022-09-03 19:44:19 RT @alanrew: Good @garymarcus piece about the perils of Google's robot projecthttps://t.co/b1rMfkQOXs

2022-09-02 20:52:00 @EranElhaik @joe_pickrell @ItsNeuronal @random_walker can you articulate what you mean by complex calculations that are difficult to!trace? and more generally, curious your thoughts on some of the challenges here.

2022-09-02 20:42:35 @joe_pickrell @ItsNeuronal @random_walker @EranElhaik

2022-09-02 17:29:48 RT @jenmercieca: "What, then, is a semi-fascist movement?We can think of a semi-fascist regime as one that approaches fascism while retai…

2022-09-02 15:35:37 @dcbyron @random_walker @RexDouglass that’s not a very substantive argument

2022-09-02 15:04:27 @alexvoica @Kantrowitz

2022-09-02 15:01:12 RT @Kantrowitz: Interviewing AI expert @GaryMarcus on the podcast today. What should I ask?

2022-09-02 14:53:00 @joe_pickrell @random_walker saying the properties are well-known is not same as saying the tools have been appropriately applied. whole replicability crisis stems in part from the fact that many people have a superficial understanding of statistics, even tho a deeper understanding is in principle possible

2022-09-02 14:42:02 RT @KaiaVintr: @random_walker @GaryMarcus if we can't be trusted to use relatively simple tools from linear algebra responsibly, how can we…

2022-09-02 14:17:23 @rasbt there’s plenty of blame to go around. peer reviewers aren’t paid, and nowadays a lot of AI research sidesteps peer review. @GoogleAI publishes stuff on arXxic about their systems’ understanding of jokes without any stats at all. &

2022-09-02 14:14:46 RT @random_walker: This might be the most depressing sentence I've ever read in an abstract. When researchers use tools we don't fully unde…

2022-09-02 14:06:51 Narcissists sometimes create as many problems for their friends as they do for their enemies. https://t.co/3e5u26pPff

2022-09-02 14:04:34 @OshriNaparstek There are many comparable problems with the misuse by people of deep learning throughout science

2022-09-02 01:05:21 I think it is interesting that Elon Musk didn’t take the bet that @wadhwa and I offered, and that the guy below from @openAI didn’t want to put up much more than a silly t-shirt. https://t.co/bxtLc3UE92

2022-09-02 01:03:30 @vasanthsarathy solving a closed-end board game every decade or two is not getting us to general AI. AlphaFold is a real contribution, for sure. But I don’t think you can at all say that AI has (yet) delivered on what its early pioneers had initially expected might happen over 67 years.

2022-09-02 01:00:28 @tdietterich @mpshanahan has an interesting related paper today, and maybe he might want to comment

2022-09-02 00:35:14 Let’s be real. We still don’t even have what Terry Winograd was driving at with SHRLDU and Blocks World in 1970 -– reliable open-ended dialog about the physical world. https://t.co/f00orn9VTALet alone what Peter Norvig aimed for in his 1986 dissertation: story understanding. https://t.co/WUL2o2UV2d

2022-09-01 23:29:53 @AngusTurner9

2022-09-01 23:29:09 @EranElhaik

2022-09-01 21:45:32 more fallout from the reproducabiliy crisis, potentially affecting tens of thousands of studies: Principal Component Analyses (PCA)-based findings in population genetic studies are highly biased and must be reevaluated https://t.co/mcpF04scRo

2022-09-01 20:22:19 semi? https://t.co/n5O8hKc9Og

2022-09-01 04:59:34 @jachiam0 pretty low stakes given the braggadocio in your words, no?

2022-09-01 04:48:18 @chr1sa surely it is not already here. we have only narrow and quite fallible AI now. i offered @elonmusk a related bet with very specific criteria. it doesn’t have to be vague: https://t.co/MZjzZSqlsk

2022-09-01 04:41:50 AGI by September 1, 2032? How much are you willing to put behind your words? https://t.co/VTqyiR1yzS

2022-09-01 03:02:19 RT @warren_craddock: There are now dozens of companies with AVs that can perform the entire driving task. There are now dozens of companies…

2022-08-31 15:11:48 LLMs are an incorrigible mix of misinformation and privacy invasion, yielding an untraceable stew of half-truth. great article by @Melissahei @techreview https://t.co/G7nRdBnaV8

2022-08-31 14:22:56 @MadamePratolung they assume everything is learned from a blank slate. it the slate is clearly not blank, so (a) eg things like time and space and causality probably don’t need to be learned and (b) that changed what does need to be learned and how it might be learned

2022-08-31 14:21:07 fantastic essay by @R3_GSI https://t.co/OEceGNs4bo

2022-08-31 14:16:01 @Sara_Imari nope

2022-08-31 14:11:48 Lordy there are tapes. what a thread. https://t.co/Up4mCtMBnQ

2022-08-31 14:02:20 @MadamePratolung see this and the tweet of mine that i link https://t.co/mGbZot2Q2L

2022-08-31 02:57:18 @elonmusk really ought read this lucid thread re lidar.and sure, humans don’t need lidar, but our cognitive models are far more sophisticated than anything current AI can deliver. hence the need for complementary techniques. https://t.co/nxwkEXwWtx

2022-08-30 23:13:40 @Sara_Imari no. but it evolves creatures that discover mathematics

2022-08-30 20:22:00 RT @OwainEvans_UK: @GaryMarcus @janleike @weidingerlaura @Abebab Latest results for TruthfulQA (generation task). OpenAI's WebGPT gets ~52%…

2022-08-29 16:49:36 angling for an insanity plea? https://t.co/Na98kSaaot

2022-08-29 16:23:13 RT @nukkailab1: “ How can you get data-driven learning and abstract, symbolic representations to work together in harmony in a single, more…

2022-08-29 14:42:10 RT @jenmercieca: Really important thread on how to frame the Trump investigations: it's not a political question, it's a legal question. Us…

2022-08-28 20:10:39 @Meaningness agree with everything except the “it can’t work” and “nailed in a few years”.

2022-08-28 13:55:15 RT @sapinker: This is the kind of approach that @garymarcus and I have been advocating for decades, starting with language. One Man’s Drea…

2022-08-28 01:38:57 RT @thedrive: Tesla owners say their cars are stopping way too early at intersections with bigger stop signs—and they’re breaking out the m…

2022-08-28 01:30:49 RT @FeiziSoheil: I may be naive but I don't think AGI will be solved on countless GPUs using trillion samples, but on a whiteboard with a m…

2022-08-27 17:59:31 @tetraduzione @luislamb @AvilaGarcez no. but there were people - even Hinton - who considered peace then. Hinton included a related Touretky paper in his ‘91 edited book.

2022-08-27 14:32:45 RT @maththrills: Care #robotics challenge #609: using any combination of perception and manipulation, work out which parts of this piece of…

2022-08-26 17:22:33 RT @mark_l_watson: I enjoyed the conversation between @GaryMarcus and @bengoertzel at the AGI virtual conference. Marcus said the AI needs…

2022-08-26 13:51:51 RT @__TweetinChar__: !!--#FF Follow those, who fight *techbros* for you/your kids' future + stay in touch w/ reality #AI #Blockchain@Abeba…

2022-08-26 00:22:02 you get what you don’t pay for https://t.co/FmndEt4Xnn

2022-08-26 00:19:57 @pmddomingos @emilynussbaum

2022-08-25 23:21:06 RT @adamsafron: Fascinating thread.

2022-08-25 22:43:51 RT @TristanThrush: Also curious about this @GoogleAI. Does your model stand a chance against Winoground?https://t.co/1HpdpnDRvX

2022-08-25 22:36:10 Dear @GoogleAI,Three months have passed since you claimed Imagen improved performance on compositionality. I asked for access and you didn’t respond.@TristanThrush offered you his Winoground materials

2022-08-25 21:45:49 @pmddomingos @emilynussbaum i did

2022-08-25 16:50:44 RT @mrgreene1977: Wow!This is fantastic. US gov't will require all papers on research funded by taxpayers be made freely available to the…

2022-08-25 16:33:31 RT @Tipalo1: AI FOR HUMANITY - CAN WE MAKE AI MORE HUMAN?"A CERN for AI." @GaryMarcusETHICS"How can we ensure that artificial intellig…

2022-08-25 16:31:34 @Meaningness it was superhuman only in some respects

2022-08-25 14:04:42 @a_stadt @didou @JeanRemiKing @spiantado @tallinzen @lakretz @c_caucheteux @ev_fedorenko @mtoneva1 @HassonLab @sleepinyourhat all to the good, i just the systems are inherently stalking the wrong quarry (predicting words rather than building world models), so not optimistic

2022-08-25 13:56:03 @sharpless89 @IntuitMachine @MJ_StormBorn =>

2022-08-25 13:54:44 indeed, and part of why as a unicyclist, I am terrified of driverless cars! https://t.co/VuA2SP7dOt

2022-08-25 13:53:20 @sharpless89 @MJ_StormBorn it is our job to teach them to reason in accordance to values

2022-08-25 13:52:07 @a_stadt @didou @JeanRemiKing @spiantado @tallinzen @lakretz @c_caucheteux @ev_fedorenko @mtoneva1 @HassonLab @sleepinyourhat such as? (hope to read your paper but only read abstract so far)

2022-08-25 13:50:45 @didou @a_stadt @JeanRemiKing @spiantado @tallinzen @lakretz @c_caucheteux @ev_fedorenko @mtoneva1 @HassonLab @sleepinyourhat we know that much, yes. we don’t know the details

2022-08-25 13:48:37 @a_stadt @didou @JeanRemiKing @spiantado @tallinzen @lakretz @c_caucheteux @ev_fedorenko @mtoneva1 @HassonLab @sleepinyourhat if they worked more like people (eg in mapping syntax into semantics and building stable discourse models) i would agree.some future AI systems will be very helpful in that regard. these particular ones, not so much.

2022-08-25 13:46:28 @cosmin_scaunasu i quote that in https://t.co/C6XodPcoZW and love it :)

2022-08-25 13:44:47 @iPrabhavKaula yes, and world models

2022-08-25 13:44:30 RT @johnmcternan: Fascinating thread

2022-08-20 01:10:09 true, but not new https://t.co/MU7Qa5PmQK

2022-08-20 00:32:38 RT @kareem_carr: Finally tried out the GPT-3 model from OpenAI. The green text is unaltered output generated by the AI. https://t.co/NxveJh…

2022-08-20 00:30:49 @andrey_kurenkov but humans are such a low bar

2022-08-19 21:00:17 RT @dcallahan2: @GaryMarcus @Google Really good @GaryMarcusarticle on #AI #language models &

2022-08-19 20:30:12 @DLBarack the failure of LLMs to reason reliably about everyday phenomena despite immense data doesn’t exactly hurt Fodor’s view…

2022-08-19 20:24:00 @robinc @Google that’s really not enough, and i addressed briefly in my essay

2022-08-19 19:55:53 @robinc @Google not sure i follow your sentence regarding mitigation.

2022-08-19 19:55:19 @Sergei_Imaging i don’t think the relation is likely to be regular enough to be relied upon

2022-08-19 17:23:01 @iPrabhavKaula @poolio i am Monday, no? @bengoertzel

2022-08-19 15:18:11 RT @vitouphy: This is a great article. Long story short, "Big data won’t be enough to solve the problems of robotics". We need to take…

2022-08-19 13:53:53 RT @cyberandy: Insightful as always. Why LLM cannot (yet) control humanoid robots

2022-08-18 21:56:27 @DeepMind @yudapearl

2022-08-18 21:51:12 @tomgoldsteincs @model_mechanic

2022-08-18 21:21:01 @BAPearlmutter just a simple misunderstanding, that’s all

2022-08-18 20:05:47 RT @etzioni: "Large language models are like bulls in a china shop: awesome, powerful, and reckless. " @GaryMarcus

2022-08-18 19:17:11 RT @mrgreene1977: Per the usual, @garymarcus cuts to the heart of the matter with surgical precision. My two cents: Tesla, Google, and th…

2022-08-18 19:02:38 hey @elonmusk i am guessing that the demo you promised for September was going to look something like what Google just showed. That one has some serious problems, described below - betting they will apply to your demo, too. https://t.co/Fy6mKOwFyq

2022-08-18 18:18:39 "Building a robot on top of language system that has such little comprehension of the world can’t be a recipe for success.” Some thoughts on @Google’s new #robot projecthttps://t.co/9HPKnCM1rY

2022-08-18 15:27:19 @grsimari @elonmusk Didja ever notice…

2022-08-18 14:37:48 @thedudefrugal @elonmusk to the tune of https://t.co/fDoJGvaqEm

2022-08-18 14:35:45 @grbradsk @elonmusk ps he evidently said boring chores, and few chores involving nothing.

2022-08-18 14:34:51 @grbradsk @elonmusk i doubt he will even get Optimus to work as well as a Roomba in the near term.

2022-08-18 14:21:06 link to a slightly more detailed story: https://t.co/Ne0ZhD9COi

2022-08-18 14:18:53 Should I offer @elonmusk another bet? https://t.co/hb8KUriHfD

2022-08-18 05:07:11 Read this thread. “users must recalibrate their thinking on what Google is and how information is returned to them … we must apply the same scrutiny [to searches] we’ve learned to have toward information on social media.” https://t.co/y61MICZSsP

2022-08-18 04:16:17 please tell me this is not our only strategy for reaching ethical AI. https://t.co/Q5IlhJNFEo

2022-08-17 21:25:45 excellent thread on why edge cases are so pernicious (&

2022-08-17 19:16:54 another tesla rear-ending another stopped vehicle https://t.co/SS13phI7HT

2022-08-17 14:04:18 RT @sapinker: More I, less A: a conversation with @GaryMarcus on @michaelshermer's show.

2022-08-17 00:15:04 RT @kncukier: So pleased we could put this together in a responsible and substantive way. With huge thanks to @danieldennett, David Chalmer…

2022-08-16 22:26:12 RT @ShannonVallor: The saddest thing for me about modern tech’s long spiral into user manipulation and surveillance is how it has just slow…

2022-08-16 19:52:39 RT @lemtaz: @cbouzy This is brilliant. https://t.co/kTJ2YnDN21

2022-08-16 18:49:01 RT @michaelshermer: If you're worried about the AI terminator, or AI turning us all into paperclips, or that AI is about to (or already has…

2022-08-16 16:57:56 excellent thread. https://t.co/sgfk9UEnSL

2022-08-16 16:32:38 RT @FZdyb: Why AI needs to ease up on scaling and learn how to code:https://t.co/Ei2Dv5QyfM@GaryMarcus @fchollet @yudapearl

2022-08-16 15:08:37 RT @barneyp: I agree with this statement by @GaryMarcus in his new paper: “The sooner we can figure out what basis allows a system to get t…

2022-08-16 15:03:14 RT @ProfFeynman: The difference between information and knowledge: https://t.co/TDcQqtBgRm

2022-08-16 15:01:51 Deep comprehension of the human brain! https://t.co/mbffuYKXwA

2022-08-16 14:57:02 Cool benchmark! cc @yudapearl @eliasbareinboim https://t.co/ktBuVUVQJG

2022-08-16 12:46:09 RT @joannashields: Thought-provoking piece by @GaryMarcus, demystifying #AI. Wise words on tighter regulation, large companies bearing more…

2022-08-16 04:25:53 RT @DGBassani: Infections cut in half in Boston schools that kept mask mandates compared to schools that did not. Even though the schools t…

2022-08-15 22:36:55 @DrMJoyner but too many variables that just can’t really be predicted, eg what happens to oil if documents link trump even more closely to MBS, in some way nobody has even guessed?

2022-08-15 22:27:43 quite possibly not a solvable problem. but certainly there are other unsolved problems (eg commonsense physical and psychological reasoning) that are not amenable to big data that are solvable, with innovation. https://t.co/KafsVlCQ9q

2022-08-15 20:33:44 @BrandonLive and see eg for yet another in a parade of examples of out-of-distribution data causing trouble https://t.co/0GoxIESa1e

2022-08-15 20:15:32 RT @ISusmelj: A friend just shared this with me. A great example for how deep learning models struggle with out-of-distribution data.I won…

2022-08-15 20:01:13 @BrandonLive whatever you think of o’dowd, this independent study adds to the intuition that i just articulated: https://t.co/8ryaG0CVdy

2022-08-15 19:55:20 @BrandonLive O’Dowd and the one i posted yesterday, which itself followed up a similar result from the same YouTuber.neither is fully persuasive but they do hint that the parameters of recognition are pretty narrow

2022-08-15 19:50:38 @BrandonLive a thought experiment based on results from results from two separate informal experiments hardly constitutes “misinformation.” referring to such a thought experiment, clearly labeled, as misinformation on the other hand…

2022-08-15 01:38:24 RT @bobehayes: Great read:"The sooner we can figure out what basis allows a system to get to the point where it can learn #symbolic abstr…

2022-08-14 17:55:02 another Youtuber finds some issues with Tesla FSD and children. https://t.co/CPrQI5z6se

2022-08-14 04:05:27 Looking forward to my fireside chat w @bengoertzel! https://t.co/wkBuu402Sc

2022-08-14 02:18:10 RT @WUTangKids: Maybe the coolest cover of Michael Jackson's Billie Jean ever https://t.co/ZyLGqDjQ2b

2022-08-14 00:19:17 RT @timcrane102: Good sense on AI from ⁦@GaryMarcus⁩ https://t.co/JSZL6GrvDF

2022-08-13 21:44:05 @samuelmcurtis @mark_riedl @jackclarkSF spammers don’t much care… if benefit exceeds cost, criminals gonna crime

2022-08-13 21:23:58 @girishsastry @jackclarkSF @samuelmcurtis it might. or it might not. to say that they are “overhyped” is to be (nearly) certain, &

2022-08-13 20:32:46 @mark_riedl @jackclarkSF @samuelmcurtis @ykilcher i disagree. but don’t want to share publicly ideas about what a criminal might do…

2022-08-13 20:28:24 @mark_riedl @jackclarkSF @samuelmcurtis @ykilcher kilcher mostly got away with it at large scale for days

2022-08-13 20:24:33 @jackclarkSF @justinhendrix @samuelmcurtis just posted a few studies above

2022-08-13 20:23:10 @jackclarkSF @justinhendrix @samuelmcurtis and this https://t.co/ffFQEapIrB

2022-08-13 20:21:56 @jackclarkSF @justinhendrix @samuelmcurtis and this https://t.co/HAScqn1Phd

2022-08-13 20:21:22 @jackclarkSF @justinhendrix @samuelmcurtis and there is research like thishttps://t.co/vfGQRPA1mb

2022-08-13 19:49:01 @mark_riedl @jackclarkSF @samuelmcurtis also why is this either/or? i am sure it is or soon will be both

2022-08-13 19:48:34 @mark_riedl @jackclarkSF @samuelmcurtis how do you know this? if i am troll farm why should i tell you my methods? both your response and Jack’s seem extremely speculative. meanwhile @ykilcher’s lambasted demo shows how easy it is to put this stuff to work.

2022-08-13 19:19:58 @jackclarkSF @samuelmcurtis i don’t understand what you are saying. it seems trivial to use, and it’s not so expensive that people don’t fool around with it endlessly.

2022-08-13 19:17:25 @jackclarkSF @justinhendrix @samuelmcurtis a web of fake reports around mayim bailey and cbd gummies

2022-08-13 18:47:16 @justinhendrix @jackclarkSF @samuelmcurtis but would we know if they had? i think i know at least one anecdotal example.

2022-08-13 18:46:20 @jackclarkSF on what basis do say this? they are unparalleled in their ability to create plausible sounding misinformation at near zero cost. and have no built-in mechanisms to track reality independent of corpus statistics.

2022-08-13 12:47:20 RT @DorotheaBaur: If #Blenderbot claims that "Donald Trump is still president and there is "definitely a lot of evidence" that the election…

2022-08-13 02:07:51 @rcalo differentiated?

2022-08-13 02:06:54 @Dervine7 @chazfirestone the general knowledge (chapter 4 in The Algebraic Mind) is that types are not the same as tokens.

2022-08-13 01:57:02 @Dervine7 @chazfirestone and the ability to hypothesize they it could actually be two

2022-08-13 00:20:55 @samuelmehr @SBMost nothing to see here. MOOve on.

2022-08-12 22:43:09 @SBMost and the realization that there must be two is called cognition :)

2022-08-12 19:46:52 RT @dataengines: @yudapearl @johnmark_taylor Universal approximability is much less useful in worlds with distribution drift. As @GaryMarcu…

2022-08-12 18:10:12 @chazfirestone 1. never said there was a sharp line, though surely they differ2.don’t know the source of the image, but assume it is supposed to be a system that gives veridical answers, rather than a model of human perception independent of cognition.3. humans can use common sense somewhere

2022-08-12 17:59:54 @chazfirestone well, no, we don’t, not on reflection.

2022-08-12 17:55:57 @chazfirestone if it gives a length estimate it should be modeling the final output

2022-08-12 17:22:52 @hb_cell @BarryOSullivan if a type-token distinction is not at the core of your cognition, you can easily wind up here….

2022-08-12 16:52:57 Prove you are a human (h/t @BarryOSullivan) https://t.co/LXmwld9sPV

2022-08-12 15:17:31 @raphaelmilliere @DLBarack @MetaAI impersonal impersonation is the most insincere form of flattery? the old Descartes joke (he says “I think not” and then disappears) is better :)

2022-08-12 15:08:17 RT @NoemaMag: Is a decades-long AI debate finally coming to a resolution?@garymarcus is seeing signs that it is.Now "we can finally foc…

2022-08-12 15:02:07 @whatsayadam @ylecun @Jake_Browning00 @DeepMind @AvilaGarcez

2022-08-12 13:29:14 @yudapearl thanks for mentioning the debate!

2022-08-12 13:22:13 @YiMaTweets so well said.

2022-08-12 13:15:29 RT @YiMaTweets: Personally, I would love to see more papers entitled "XXX alone is not enough," instead of those entitled "XXX is all you n…

2022-08-12 13:04:34 RT @yudapearl: An interesting debate between @GaryMarcus and @ylecun on the tension between neural vs. symbolic architectures in AI: ttps:/…

2022-08-12 13:04:04 @yudapearl @ylecun correcting the link: https://t.co/wCOiHq8Kxa

2022-08-12 12:48:59 RT @defnotbeka: it only took 8 years at a minimumimagine where we'd be if folx had admitted this back then. what cool stuff would we have…

2022-08-12 12:48:29 RT @kostadis_tech: hallelujah

2022-08-12 12:44:04 RT @ossip: "La inteligencia artificial se ha centrado principalmente en una técnica llamada aprendizaje profundo. Puede que sea el momento…

2022-08-12 12:43:26 @Floridi @alan_winfield sure, please dm or follow back

2022-08-12 04:16:50 RT @luislamb: Insights on AI future by @GaryMarcus: challenges ahead demand integrated symbolic &

2022-08-12 02:15:44 @UlyssesPascal @ylecun @Jake_Browning00 @DeepMind my most detailed discussion of this in chapter 3 of The Algebraic Mind.

2022-08-12 02:14:37 @andrey_kurenkov @ylecun @Jake_Browning00 @DeepMind see my remark on development vs computation

2022-08-11 22:22:27 @jacyanthis @ylecun @Jake_Browning00 @DeepMind x2 and x1 are generally regarded as awfully similar. LeCun and Browning appear to be accepting both.

2022-08-11 22:13:59 Next time someone tells you Stratego is a grand challenge in AI, think about this, and where we have and haven’t made progress in the intervening 16 years.

2022-08-11 22:07:46 @jacyanthis @ylecun @Jake_Browning00 @DeepMind even if you weren’t selectively quoting from my article (and eg ignoring the bits about Hinton), I wouldn’t really understand what you are talking about. you can’t have X and replace X at same time, can you? and surely you can’t have X and call it aether at the same time.

2022-08-11 22:05:13 “Programming a computer to be ethical is much more difficult than programming a computer to play world-champion chess” - James Moor, 2006

2022-08-11 21:44:03 RT @igrgavilan: #enlosblogs What a debate! "Deep Learning Alone Isn’t Getting Us To Human-Like AI" (https://t.co/JlRbIiGDx6) by @GaryMarcus…

2022-08-11 18:50:40 ditto for AI? https://t.co/IqPmGtMWH9

2022-08-11 18:45:38 @dileeplearning oh for the old days when you posted truth, like this: https://t.co/DZN0aCSSUb

2022-08-11 18:35:09 @RadioFreeMonad @FelixHill84 @danbri @goodside @nnaisense indeed, that is linked in my reply to LeCun and Browning @NoemaMag

2022-08-11 18:33:48 RT @dmonett: "On natural language, compositionality &

2022-08-11 18:33:02 @mosesjones i absolutely think we should take the threat seriously

2022-08-11 16:44:08 @PaulTopping @ylecun @Jake_Browning00 https://t.co/lVa6BXzkaS

2022-08-11 16:37:47 Something big is happening in AI — even researchers like @ylecun are acknowledging the need for symbol-manipulation, alongside deep learning.A reply to LeCun’s recent essay with ⁦@Jake_Browning00⁩, w discussion of recent work ⁦@DeepMind⁩. https://t.co/azmrt57ru2

2022-08-11 15:42:52 @Jeanc9orf72 @carlzimmer almost any machine can be hacked

2022-08-11 15:41:30 @goodside @sir_deenicus @danbri what task is the error rate <

2022-08-11 14:46:30 @AlisonWants sounds like they might be headed to Mar-a-Lago?

2022-08-11 14:08:15 @danbri @goodside the hybrid systems that @nnaisense is building, which seem like total vindication for what I have been saying for years.

2022-08-11 13:31:41 @danbri @goodside it’s gotten more than its share of careful examination

2022-08-11 13:27:22 @RobKnight__ of course not.

2022-08-11 13:26:22 @eerac @goodside @danbri none. that’s my whole point. we need to look beyond current tech.

2022-08-11 13:25:34 @mattburgess1 @WIRED

2022-08-11 13:24:17 @danbri @goodside i can’t tell you exactly how many times people have shown me cute GPT stuff and then it turns out not to be robust. probably over 30.i know GPT is fun to play with, but it’s not the road to trustworthy intelligence

2022-08-11 13:21:48 now imagine the same, but with Neuralink https://t.co/gkPwrcvkqh

2022-08-11 13:16:36 @goodside @danbri i have yet to see anything reliable from these systems.

2022-08-11 12:50:50 @danbri @goodside as always, my first question: is it reliable?

2022-08-10 18:56:16 RT @vdignum: AI regulation, AI literacy, imbalance of power, unreliability,...Good read by @GaryMarcus"Where does all this leave us? Fo…

2022-08-10 14:30:48 “If you’re innocent, why are you taking the Fifth Amendment?” - former President Donald Trumphttps://t.co/e8oFSWvMZ3

2022-08-10 14:26:45 turned in final edits on my reply to @ylecun and @Jake_Browning00’s essay on innateness and symbols-manipulation. should be out soon.

2022-08-10 14:23:04 @MathieuVVyve @BrandonLive agree that if FSD was not engaged, that is a problem. are there other specific issues?

2022-08-10 04:23:03 @BrandonLive this is your fundamental error. no way is this a realistic real world assumption., with respect to many human beings.

2022-08-10 04:03:45 @BrandonLive tesla’s do have known issues with some stationary objects…

2022-08-10 04:02:48 @BrandonLive other vehicles: not relevant

2022-08-10 04:00:15 @BrandonLive there is plenty of detail here: https://t.co/Yz5DzGfIIv

2022-08-10 03:59:24 @BrandonLive seems to me like a key issue may be that the mannequin was in profile rather than full on? was that true in IIHS test?

2022-08-10 03:56:58 @BrandonLive well you posted a different link and i don’t see the detail in this image in the second link. more than that, i doubt any one test would suffice

2022-08-10 03:14:30 @BrandonLive doesn’t speak to the issue of children

2022-08-10 02:39:45 @BrandonLive i know i could count on you, @BrandonLive

2022-08-10 01:49:35 Tesla’s self-driving technology fails to detect children in the road, tests find https://t.co/2NZV2ZU0Pp

2022-08-10 00:13:33 RT @lizadixon: The reason that this is so cringey (besides the obvious) is that due to #autonowashing and other factors, many people wrongl…

2022-08-09 21:22:34 RT @TomerUllman: Had Dall-E generate 'The Trolley Problem'

2022-08-09 19:11:10 RT @sd_marlow: @GaryMarcus LLM's are like mile-high Jenga towers... https://t.co/Q9qO6NzZ0p

2022-08-09 19:10:51 RT @mnitabach: Absolutely! Current AI systems are only "extremely powerful" bcs we are plugging shitty decision outputs into extremely powe…

2022-08-09 17:43:42 agreed: we should not build extremely powerful AI systems before we figure out how to make them safe.the more so if those systems are (as they are now) superficial and unreliable https://t.co/WvybpygO8y

2022-08-09 01:01:45 : Facebook's AI Chatbot: ‘Since Deleting Facebook My Life Has Been Much Better’ https://t.co/mim1j3vidE

2022-08-09 00:43:00 RT @vermontgmg: THREAD: “What’s the difference between this and Watergate?” Donald Trump raged in a statement Monday night after FBI agents…

2022-07-26 21:11:43 @FelixHill84 there were loads of papers like this

2022-07-26 21:10:53 he who rewrites the past is doomed to repeat it?astonished by pushback I am getting in suggesting that end-to-end deep learning has historically been posed in opposition to classic, modular cognitive architectures. here’s typical example from 2016:https://t.co/mN8xJ0a4Yo https://t.co/OZxgN6jR9T https://t.co/ymKpxwirPV

2022-07-26 18:48:23 RT @jmourabarbosa: It is often claimed that ‘DNN are a good model of the visual cortex’.Three papers have come out recently that in my op…

2022-07-26 17:36:51 @AndrewLampinen @FelixHill84 sure but in the early days of the past tense debate Jay argued for an eliminative model that was end-to-end and lacked such things

2022-07-26 17:34:59 @AndrewLampinen @FelixHill84 the whole point of my original post is there original skepticism may have been misplaced. @felixhill oddly denied that it even existed.

2022-07-26 17:32:56 @FelixHill84 even that is nonsense. what i was saying is that theory has developed (in many different ways) in 50 years, many quite relevant, and that Chomsky and linguistics were both being strawpersoned.

2022-07-26 03:36:30 @FelixHill84 i think almost every word of what you said above is nonsense, but i look forward to the references. certainly i did not “disown” Chomsky

2022-07-26 03:33:42 @FelixHill84 connectionism has a long history, the first time I heard the term “end-to-end deep learning” was mid 2010s and for years it was definitely pitched as an alternative to modular structures. and a good bit of that history (eg the past tense debate) was certainly anti-modular.

2022-07-25 17:02:59 @NaveenGRao pletny: years of dismissing old ideas and then starting to use them (generally without any acknowledgment of their intellectual history) is poor form, and we might have made progress sooner if people hadn’t been so dismissive.

2022-07-25 16:39:14 RT @PolymathToBe: @GaryMarcus Joni Mitchell segment on CBS This morning. She had to teach herself to play guitar after a brain aneurysm.…

2022-07-25 16:12:28 End-to-end deep learning started out as an alternative to classic, modular Cognitive Architectures, but is gradually adding bits (like modules and external memory) that more &

2022-07-24 18:01:44 @IntuitMachine never known exactly what that encompasses

2022-07-24 18:01:26 @IntuitMachine i would lean towards salvaging what can be salvaged, without be dogmatic about saving it all.

2022-07-24 17:53:23 @IntuitMachine no, i lean towards developing innovative hybrids

2022-07-24 17:52:30 RT @bobehayes: Excellent >

2022-07-22 19:41:53 RT @didou: Great watch below (just bumpe into!) @GaryMarcus Personal Prediction: Neural nets will soon be able to create AI that will sta…

2022-07-22 17:14:21 @Meaningness and also because it does a good job of synonymy via embeddings

2022-07-22 17:13:18 @Meaningness diagrees. don’t think GPT build any kind of tree, nor any kind mapping between syntax and semantics, whereas I think people read semantics from (imperfectly represented) trees.GPT gives the illusion otherwise because it pastisches such large fragments of text

2022-07-22 17:12:01 @Meaningness i stand by that argument

2022-07-22 17:09:27 @Meaningness i find the minimalist program very implausible, and have many other issues with Chomsky’s view, which I articulated earlier this year (eg I think he underemphasizes w and contributions of cognition), but overall agree with his arguments for generativity, innateness, &

2022-07-22 13:13:57 RT @codingFriends: #LanguageLearning is Harder Than You Think - @GaryMarcus gives some very interesting insights on how we learn a languag…

2022-07-22 07:11:22 https://t.co/WX2ayDT3Hh

2022-07-22 01:29:35 RT @sapinker: Learning language is a computationally formidable achievement, @garymarcus reminds us. https://t.co/SRH3n4byIb

2022-07-21 19:39:00 @sd_marlow @BrandonLive @AthenaAI2 @Tesla or put differently, could a person take the right actions based on whatever data the vehicle is taking action on.it’s not a perfect test, but it is a very good thought experiment.

2022-07-21 17:30:55 RT @bobehayes: Check out this great discussion on AI, including topics of #aihype #ethics #LLM #AGI @GaryMarcus: Is #AI Sentient? The Que…

2022-07-21 17:29:25 @BrandonLive @sd_marlow @AthenaAI2 @Tesla in maybe 300 comments you have made on threads on my feed, i have seen maybe 2 criticisms of Musk, and 250 statements that defend Tesla. i know you are not a bot, but I can see why you might be perceived as one.

2022-07-21 16:09:14 Can’t speak directly to @tesla’s management &

2022-07-21 13:07:38 RT @jlmendivil: A very timely blog post by @GaryMarcus:Learning Language is Harder Than You Think https://t.co/gjpZibNqJU

2022-07-20 21:34:33 RT @bobehayes: Excellent >

2022-07-20 20:52:18 Learning Language is Harder Than You ThinkSure, kids imitate their parents, but that’s just a small part of the storyA new essay, replying to @meaningness[as ever, Noam Chomsky read it first and thought I was *way* too nice ] https://t.co/sgodzRbdfN

2022-07-20 16:28:39 RT @mrgreene1977: 100% this

2022-07-20 00:51:14 Misinformation, which current AI generates but does not reliably detect, is a threat to the 2022 elections. Until that changes, AI is likely to be a net negative for democracy. https://t.co/9z2WT6VoWS

2022-07-19 16:29:03 Fabulous, deeply probing interview on AI, misinformation, and making the world a better place. Thank you @sinanaral! https://t.co/PIKQPUFR5p

2022-07-19 15:23:05 fabulous, deep interview re: AI and making the world a better place. https://t.co/1ibyOSsDU9

2022-07-16 17:40:53 RT @pomeranian99: "In my third year of reporting on the coronavirus pandemic, I find woodpeckers, which can ram their heads against hard su…

2022-07-16 16:55:45 @chrmanning @ZhunLiu3 if only you had signed the receipt. photographed it, and sold the picture as an NFT

2022-07-14 03:16:02 RT @lizadixon: “An AI prosecutor has even started charging people in Shanghai with crimes it believes they have committed, the South China…

2022-07-14 03:01:44 @lizadixon omg! maybe not ever will this be a good idea. certainly not now. certainly not soon.

2022-07-13 21:37:11 tesla’s leading AI light is departing. https://t.co/3BTxxrVQoq

2022-07-12 21:39:19 @pmddomingos @ylecun @SchmidhuberAI and it is exactly what stymied it prematurely for decades.

2022-07-11 22:48:33 RT @Aras121: UK national screening committee conclusions on the use of AI into clinical practice of breast screening programmes. Hint: @Gar…

2022-07-10 23:24:39 @vipulved @auren please. sir, document what has been moved, or don’t say it.

2022-07-09 18:21:22 RT @luislamb: A great accomplishment Tim Scarfe. This episode unveils in a clear way the challenges ahead we still face in building robust…

2022-07-08 23:24:54 in the end after the legal dust settles, Elon will

2022-07-08 20:38:22 @pmddomingos @SchmidhuberAI how about the end-to-end training and differentiability that is #1 in what Yann is claiming: https://t.co/M1Mq1NUbqN

2022-07-08 20:35:58 @AVMiceliBarone @SchmidhuberAI Not so. eg: https://t.co/rE99nKpPlQ

2022-07-08 17:37:35 . @SchmidhuberAI makes a strong case here

2022-07-08 01:52:34 @BrandonLive @MadamePratolung no no no. the mislabeling of the product is clearly dangerous and makes the likely (mis)use of the product dangerous.

2022-07-08 01:37:16 @BrandonLive @sd_marlow @AthenaAI2 what you say is true only if you ignore prior history

2022-07-08 01:08:19 @BrandonLive @sd_marlow @AthenaAI2 source?

2022-07-08 01:04:39 @sd_marlow @BrandonLive @AthenaAI2 driver error is always possible, but there have been many previous accidents spanning back to 2017 or earlier in which Teslas crashed into stopped vehicles, which is something NHTSA is focusing on. And at least two other similar deaths involving going under tractor trailers.

2022-07-07 21:18:30 Two new fatal Tesla crashes are being examined by US investigators - The Verge https://t.co/psEIMQkx0f

2022-07-07 18:12:20 @dan_sinykin @MadamePratolung there is also eg truthqa by @OwainEvans_UK and Stephanie Linhttps://t.co/rlzbmQxb9Z

2022-07-06 23:23:44 happy 87th birthday to the Dalai Lama (photo from 2014) https://t.co/Ce7hu7HjXB https://t.co/FfSfOlNFDk

2022-07-06 17:10:14 RT @raphaelmilliere: The recordings of our event "The Challenge of Compositionality for AI" are now available on https://t.co/JbwYJZroal!…

2022-07-06 04:58:09 RT @sintelion: Preprint: "How many linguists does it take to fool a question answering model?"We take a systematic linguistically informe…

2022-07-06 02:24:12 RT @auren: AI is closer to Microsoft Excel than it is to humans. It’s just doing statistical calculations and spitting out a result. It doe…

2022-07-06 01:51:44 RT @auren: If you trust your instincts too much, you fall in love with your biases. For almost any piece of data, humans have their own th…

2022-07-05 23:10:46 @dileeplearning skeptical of tenured, respected profs work ≠ regarded as clowns

2022-07-05 22:59:30 RT @tomgoldsteincs: This plot show the growth of large language models. BERT was in 2018, PaLM in 2022. Training PaLM used about 0.0001%…

2022-07-05 22:58:08 RT @tomgoldsteincs: Training PaLM takes 3.2 million kilowatt hours of power. If you powered TPUs by riding a bicycle, and you pedaled har…

2022-07-05 22:42:58 @subjacentish well, yes, their knowledge of—and respect for—language remains limited. but their facility and creativity with ML has always been respected. fair point.

2022-07-05 22:39:08 this is mythology. nobody to my knowledge ever thought any of these people were “clowns”, it’s just that other techniques were working better and getting more attention. hinton got all kinds of prizes and ran a major UK institute. https://t.co/4H7iwGtpDF

2022-07-05 22:34:32 @bakztfuture @auren

2022-07-05 22:21:39 RT @auren: criteria for real AGI (from @GaryMarcus): read + understand a novel and character motivations watch + actually understand a…

2022-07-05 20:47:48 @Abebab just say “no” a lot. but not to me! :)

2022-07-05 14:50:14 @DylanoRepublic

2022-07-05 14:47:17 @Meaningness @s_r_constantin more like the intersection than the union, in my view. as says on the tin, word 5, it ain’t cognitive science if it ain’t interdisciplinary.

2022-07-05 14:45:00 thanks! that’s a large reason why my Substack exists -- to bring insights from cognitive science to bear on the hard questions that AI faces. (there’s also a whole chapter in https://t.co/Pt7HZbLIv5 on the same topic.) https://t.co/UY7lfIjDNN

2022-07-05 13:36:00 RT @MLWave: Sharp words by @GaryMarcus in Language Model AI debate. Shows how philosophy, cognitive science, and linguistics deserve their…

2022-07-05 04:48:17 “For a successful technology, reality must take precedence over public relations, for nature cannot be fooled." -- Richard Feynman, 1988 report on the Challenger explosion https://t.co/4lIavHtlLo

2022-07-03 22:48:20 RT @luislamb: An interesting conversation with @GaryMarcus, @sinanaral, Kristian Kersting, Eugenio V Garcia, @MayaPlentz Marielza Oliveira…

2022-07-03 20:28:47 RT @MayaPlentz: Listen to our Twitter Spaces with @GaryMarcus moderated by @MarielzaTalks from @UNESCO with special guests @luislamb, Vis…

2022-07-03 04:57:30 RT @unixpickle: I'm still not convinced that large-scale RL is anything other than "smoothed-out search distilled into a small model". 1/5

2022-07-02 21:48:47 RT @TimothyDSnyder: Russia has a hunger plan. Vladimir Putin is preparing to starve much of the developing world as the next stage in his w…

2022-07-02 19:56:26 RT @Krauss_PK: Will full self-driving ever work w/ deep learning alone and w/o a world model and common sense? @GaryMarcus @maier_ak @Vis…

2022-07-02 19:53:48 scary video. for better or worse, that is the challenge @karpathy seems to have set for himself. https://t.co/vwTWlXvnCu

2022-07-02 16:17:00 @karpathy celebrates the virtue of bigger data sets “without substantial architecture modifications or paradigm shifts” -- but has that strategy solved the outlier problem for @tesla, even with an immense budget? https://t.co/TyQbGNqjP1

2022-07-02 14:51:26 @crazyuddie @MarielzaTalks some example of how current systems fail?

2022-07-02 14:43:48 good thread &

2022-07-02 14:37:23 this is exactly the right question—and apparently one that neither @GoogleAI nor the ML community seems to want us to ask. https://t.co/1K2383eLe7

2022-07-02 13:30:50 @template_rex @wnlckwd @DeepMind which only means they intended to work on it, not that anyone else had it high on their list

2022-07-02 13:29:43 Shocked by this this tweet. Since when has science has ever benefited from casting a blind eye to skepticism? Lack of peer review in virtually all recent major results, combined with corporate hype, is a recipe for a crisis of replicability &

2022-07-02 13:22:36 @Zergylord sorry - i misread the abstract

2022-07-02 13:16:57 @fhuszar @srikumarks no that’s not true. it is really important to have balance in an eta when corporate types so systematically overhype their work and avoid peer review

2022-07-02 13:16:05 @fhuszar @srikumarks not looking for them, people send me them, tag me, have massive press campaigns etc. hard to miss even when i am not looking.

2022-07-02 03:00:02 @srikumarks of course people should report their results. but they don’t need to report progress on Stratego as if it were major progress on a widely agreed grand challenge. &

2022-07-02 00:37:54 @Zergylord and here’s how i think about it: https://t.co/CeKS5SQkcN

2022-07-02 00:37:42 @khimya and here’s why: https://t.co/CeKS5SQkcN

2022-07-02 00:36:41 Google’s new math system, neither peer reviewed nor made available to skeptics, got 50% on problems that are challenging to *high school* students (not professionals). Great, but should we reall ask about its “economic impact”, so soon?Anyone remember radiology?2/5

2022-07-02 00:36:40 Absolutely, @RishiBommasani is right, there have been significant strides in various arenas of math, and games, too. BUT … overhype is still a problem and goal posts (not mine) are still shifting downward. A short thread on what I saw just in last 24 hours:1/5 https://t.co/QztgPDQ1sm

2022-07-02 00:14:16 RT @TobyWalsh: @erikbryn @percyliang @ilyasut @GaryMarcus @fchollet @JeffDean @DigEconLab We had systems like this over forty years ago (an…

2022-07-01 23:35:35 @tdietterich @erikbryn @percyliang @ilyasut @fchollet @JeffDean @DigEconLab at 100% accuracy that quote would be true and relevant

2022-07-01 23:31:20 The “grand challenge” currency in AI seems to be devalued. Nowadays you apparently get grand challenge cred for - playing Stratego pretty well- scoring 50% on a high school math exam

2022-07-01 22:36:13 @erikbryn @percyliang @ilyasut @fchollet @JeffDean @DigEconLab what economic work do you think such systems can do with that reliability? how many jobs have been replaced by “autonomous vehicles” that are far more accurate (95%+)? (&

2022-07-01 19:51:15 “At least 20 reviews in 17 fields have found widespread errors” in how machine learning is used in science. Upcoming workshop on incipient reproducibility crisis. (source for quote in below). &

2022-07-01 17:09:24 RT @expertdotai: As all the linguists out there know, leaving reference, cognitive models and compositionally out of #AI models can have se…

2022-07-01 17:08:55 @Mike_Page localist models should certainly still be on the table, and i would agree that the argument for “neural networks” being particularly neural or more neural than anything else is still thin.

2022-06-30 13:26:30 RT @MayaPlentz: Listen to our Twitter Spaces with @GaryMarcus moderated by Dr. Marielza Oliveira from @UNESCO @MarielzaTalks with special g…

2022-06-30 00:16:43 RT @SemaSgaier: Thoughtful article from @GaryMarcus in @sciam that tempers some of the #AI hype. Worth a read.https://t.co/bANI8ng7me

2022-06-29 18:49:17 @glupyan immensely facilitatory, perhaps not necessary?

2022-06-29 12:37:35 RT @raphaelmilliere: Starting today!9am PDT | 12pm EDT | 5pm BST | 6pm CESTWhy Compositionality Matters for AIWith @AllysonEttinger…

2022-06-29 01:32:22 @anecdotal @gwern @hardmaru @EthanJPerez @MelMitchell1 @gwern seems to have responded to me but locked me out from seeing the post.

2022-06-28 17:50:12 May this not turn out to be the epitaph of the United States of America. https://t.co/stIULcYwiV

2022-06-28 17:45:18 if ultimate narcissist TFG isn’t charged, the US may as well disband https://t.co/qaFcRZqRhC

2022-06-28 13:39:36 @sd_marlow @nickyclayton22 sure. but that’s why @nickyclayton22’s lab is one of the best -- careful and clever studies that stand the test of time.

2022-06-28 04:28:51 temped to use this in my upcoming talk on compositionality https://t.co/RPUCF1HlNI

2022-06-28 03:59:43 @hardmaru @EthanJPerez @MelMitchell1 it’s natural to think that certain things won’t scale, but aside from stereotyping and the like, why should things anti-scale? bigger data might be costly and ecologically expensive and not always super productive, but why should it have negative effect on accuracy?

2022-06-27 21:07:13 @ylecun @ErnestSDavis @techreview thanks, Yann, I look forward to reading!

2022-06-27 18:52:15 RT @etzioni: Agree or disagree, @GaryMarcus's essay is a gem!Three ideas from linguistics that everyone in AI should know https://t.co/We…

2022-06-27 18:40:34 RT @EthanJPerez: We’re announcing the Inverse Scaling Prize: a $100k grand prize + $150k in additional prizes for finding an important task…

2022-06-27 14:33:31 Corvids are amazing!UK science funding: not to so much One of the world’s leading animal researchers, @nickyclayton22, is about to lose funding.UK, don’t let this happen. Spending zillions on artificial intelligence &

2022-06-27 04:40:54 RT @narasimman_tech: @GaryMarcus Agreed! I think that the key to making progress with AI is to keep pushing the envelope with new ideas. We…

2022-06-27 04:37:09 RT @KBAndersen: I knew Congress could change the size of the Supreme Court. However, I only just learned from the excellent @jbouie’s @nyti…

2022-06-27 03:39:19 RT @bobehayes: Great podcast by @DeadCatShow: Nothing More Than a Magic Trick. @GaryMarcus talks with @EricNewcomer and @cityofthetown Me…

2022-06-26 19:43:36 Science https://t.co/6gufTfl8kQ

2022-06-26 14:30:09 Disagree with this well-written . Rate-limiting step in AI progress is not better aligned data (though that would certainly be helpful), it’s new ideas.What we need most is fresh ideas about how to connect sequences of words to cognitive models of the world. https://t.co/w0VSDQ4MyP

2022-06-26 13:44:16 RT @BDehbozorgi83: @bobehayes @GaryMarcus @ElliotMurphy91 Thanks a lot for sharing. I agree. Some fundamental issues have to be addre…

2022-06-25 16:11:50 over 800 people signed up, and an amazing lineup! https://t.co/zjo5VzTxc8

2022-06-25 14:55:38 Getting a lot of queries about @ylecun’s “bold new vision” for AI, asking me how it differs from the common-sense and cognitive models-centered vision @ErnestSDavis &

2022-06-24 14:59:49 This feed is temporarily going dark in recognition of what the Supreme Court has done to the United States.

2022-06-24 14:56:59 RT @sineadbovell: The decision to overturn Roe v. Wade doesn’t just set women back and erase decades of progress in gender equality. It wil…

2022-06-24 14:25:05 @bakztfuture @CBCNews @FrontBurnerCBC Vancouver BC

2022-06-24 13:38:00 personal news: as a proud and recent immigrant to Canada, it was a real treat to wake up this morning and unexpectedly see my picture in the CBC News thanks @CBCNews @FrontBurnerCBC https://t.co/zGY0hQzA5p

2022-06-23 21:32:34 RT @mullinsms: Really love the latest from @GaryMarcus, if nothing else it validates the time I spent reading Frege in grad school! It's ju…

2022-06-23 20:38:30 RT @dseetharaman: This is a pretty sprawling conversation but the thing that stuck out to me is @GaryMarcus’ point that while there is crit…

2022-06-23 17:44:25 Some might argue that compositionality is the most important looming technical challenge for AI. At this upcoming workshop, an all-star cast will have a look. https://t.co/zjo5VAb83G

2022-06-23 17:35:20 RT @drmeneguzzi: Just read the paper from the @rao2z 's group on the limitations of LLMs for non-trivial reasoning about change. This is…

2022-06-23 17:09:54 loved this terrific, no holds-barred look at how media could better cover AI, with @EricNewcomer and @cityofthetown on @DeadCatShow https://t.co/B8PgS4OHVf

2022-06-23 16:46:07 @cajundiscordian @David_Gunkel @Inframethod @ThomasTelving @MCoeckelbergh duck, duck, duck - just like Eugene Goostman!

2022-06-23 15:53:26 RT @Mikel_Jollett: The Supreme Court on Guns: "The states cannot decide!"The Supreme Court on Abortion: "The states must decide!"

2022-06-23 15:46:40 @David_Gunkel @Inframethod @ThomasTelving @MCoeckelbergh @cajundiscordian @cajundiscordian this is the question i keep asking and that you keep deflecting

2022-06-23 15:42:43 @cajundiscordian @David_Gunkel @Inframethod @ThomasTelving @MCoeckelbergh you keep returning this to me and i keep asking about *your* criteria. why do you keep ducking my question?

2022-06-23 15:02:45 RT @PreetBharara: SCOTUS read neither the room nor the Constitution correctly

2022-06-23 15:01:44 @cajundiscordian @David_Gunkel @Inframethod @ThomasTelving @MCoeckelbergh no. I don’t care whether you think I am sentient. I am trying to understand the criteria you applied to LaMDA.

2022-06-23 14:51:26 @cajundiscordian @David_Gunkel @Inframethod @ThomasTelving @MCoeckelbergh so that’s a “no”, as in “no, i won’t supply any criteria by which i made my attribution to LaMDA”?

2022-06-23 14:15:00 @David_Gunkel @Inframethod @ThomasTelving @MCoeckelbergh @cajundiscordian there are lots of definitions. some are trivial (eg my apple watch and thermostat sense things)

2022-06-23 14:08:05 @spiantado @synbiocs @ErnestSDavis @f_charton @ChrSzegedy per the algebraic mind, my main argument is that you have to be able to freely generalize some operations to whole classes of entities, in a mutatis mutandis way, independent for the similarity of new members to previously learned instances.

2022-06-23 13:58:59 @David_Gunkel @Inframethod @ThomasTelving @MCoeckelbergh @cajundiscordian 1. if it were just an autocomplete system would you ascribe sentience? if yes, why? if no, in what respect(s) would it need to differ from autocomplete in order to qualify?

2022-06-23 13:39:51 My question for Blake Lemoine (@cajundiscordian): https://t.co/wZbu2UUAST

2022-06-23 13:33:25 @synbiocs @spiantado @ErnestSDavis @f_charton @ChrSzegedy have all thought a lot about the current and near future state of automated mathematics.separately, i agree with your remarks around combinators: they are an alternative approach (worth studying) that try to do similar work to variables.

2022-06-23 13:25:27 @cajundiscordian @David_Gunkel @Inframethod @ThomasTelving @MCoeckelbergh Please explain to us then how you think it works, since your view seems to be very different from mine. I think that to a first approximation all LamDA does is predict next words in a very large corpus a la autocomplete. What is it you think LaMDA does?

2022-06-23 13:22:05 @cajundiscordian @Inframethod @David_Gunkel @ThomasTelving @MCoeckelbergh If no knowledge could be relevant, you simply have an unfalsifiable claim.2/2

2022-06-23 13:21:45 @cajundiscordian @Inframethod @David_Gunkel @ThomasTelving @MCoeckelbergh You have the causal arrows reversed here. Knowledge of some entity’s operation per se doesn’t make that entity more or less sentient—but it might help us to understand *whether* (relative to some definition) that entity is sentient.1/2

2022-06-23 12:11:59 RT @bimedotcom: Three ideas from #linguistics that everyone in #AI should know https://t.co/2KCAbgA1qT by @GaryMarcus &

2022-06-23 01:46:53 RT @Dr_Atoosa: Check out our paper at @FAccTConference this week, on "Taxonomy of Risks posed by Language Models". https://t.co/fdOvnOY6V…

2022-06-21 21:14:58 and can the AI community recognize the work that still needs to be done? https://t.co/Cv4kiB8iSU

2022-06-21 20:57:42 RT @SColesPorter: I’m looking for a bit of assistance The most amazing person I know She’s a member of the all girl Afghan robotics t…

2022-06-21 20:34:58 https://t.co/zOT7Co9slp

2022-06-21 20:31:50 wonder whether GPT-3 and other LLM’s like LaMDA will similarly eventually be dialed back because of their tendency towards generating misinformation https://t.co/U0C5mJuUEo

2022-06-21 20:27:58 RT @MickMulvaney: "We've got lots of theories. We just don't have any evidence."-Rudy GiulianiThat, ladies and gentlemen, was the head…

2022-06-21 19:37:24 RT @GaryMarcus: @Lang__Leon @Pehdrew_ well, no. Like all prompt engineering “Yo be real” works when it works, doesn’t when it doesn’t. http…

2022-06-21 19:37:02 @Lang__Leon @Pehdrew_ @ErnestSDavis

2022-06-21 19:24:45 much the same could be said for the new fad of prompt engineering. https://t.co/ORU0pwOHL0

2022-06-21 19:23:35 @Lang__Leon @Pehdrew_ well, no. Like all prompt engineering “Yo be real” works when it works, doesn’t when it doesn’t. https://t.co/OrCoLKY6sN

2022-06-21 19:18:48 honestly, this don’t go viral like “Nonsense on Stilts”, but it’s just as much fun https://t.co/nGTMmZlZYt

2022-06-21 19:17:33 RT @indicoData: Our founder @sl8rv is both a poetry aficionado AND CTO! In a collaboration with @GaryMarcus he riffed on famous poetry to b…

2022-06-21 18:40:34 always a fun conference @WorldSummitAI https://t.co/ZCxM1Q7jrq

2022-06-21 18:38:35 @ErnestSDavis @samc0hen @sir_deenicus @ItalyHighTech @irinablok oops

2022-06-21 18:25:36 @jonathanrlarkin i don’t think NELL made any real progress

2022-06-21 16:15:53 @kaalam_ai we discussed in https://t.co/C6XodPts1W

2022-06-21 16:00:55 Exactly!!! Not all iterative progress is aimed at the right destination.Wisdom comes from knowing what forms of progress can really get you to your final destination. https://t.co/T4cWuBQmx0

2022-06-21 15:59:03 great thread responding to @ylecun and @Jake_Browning00 by @Abel_TorresM.important to realize that LeCun and Browning frame me as a being alone and when in reality so many others see same issues https://t.co/ZEBqWNfQQD

2022-06-21 15:53:56 no. you are overestimating value of learning &

2022-06-21 15:41:34 Glad so many people are excited to see this! My thoughts on Cyc—with a very special guest star—are underway, but will take some time. https://t.co/uIfoU56h6l

2022-06-21 15:16:43 If you care about commonsense reasoning, but think Cyc is just a database of patterns, or don't even know what Cyc is, stay tuned for an important future essay. https://t.co/zcXHgmmWun

2022-06-21 05:12:13 Puzzled what the 6% were thinking. But how could we be close to AGI if common sense still eludes AI? https://t.co/imYO38Qon1

2022-06-21 03:13:53 RT @willie_agnew: I'm so honored to have received best paper award at FAcct '22 with @Abebab @ria_kalluri @dallascard @dotanravit @TheMiche…

2022-06-21 01:52:42 @AndreiMellas and yet my hard work in trying to isolate needs to be fixed is treated with nasty memes. but i do see your point, and so often point to some of the accomplishments of deep learning

2022-06-21 01:51:39 @grbradsk assume a spherical cow

2022-06-20 23:05:50 @DotCSV if only we could fix real world infrastructure with inpainting :)

2022-06-20 22:54:15 @joshelman maybe optimus the humanoid will be Transformer powered by Transformers!

2022-06-20 22:50:28 Out hiking. Wondering how Optimus would handle these signs and this under-repair, not listed on any map bridge https://t.co/RHOwQ17lMz

2022-06-20 22:41:51 @grbradsk wait elon told me we are all living in a simulation, so, um, what’s the difference?

2022-06-20 22:40:45 @aniketvartak @IntuitMachine no worries on AB would be great if field would take criticism seriously, kinda useless if all errors are explained post hoc and there is no innovation in architecture

2022-06-20 22:24:12 RT @csabaveres: @ylecun @GaryMarcus If it could be shown that symbols are learned I’d be surprised but not devastated. The key point is tha…

2022-06-20 22:23:57 @aniketvartak @IntuitMachine saw it. a it was weird since it quoted all these replies to me with our quoting me, like orbiting around an invisible planet but b what is the actual step forward? in what? measured how?

2022-06-20 22:06:08 @IntuitMachine step forward in art synthesis: yesin AGI: why? i don’t see that

2022-06-20 22:03:07 exactly. https://t.co/ORU0px6iCy

2022-06-20 21:05:09 if people didn’t keep implying they were a step towards AGI, I would! https://t.co/wxdmYZuc4U

2022-06-20 21:02:31 @OlafenwaMoses @GoogleAI @ErnestSDavis @aniketvartak see my essay Horse Rides Elephant

2022-06-20 20:51:23 @AmandaAskell by the age of 3 they have a clear, cross-modal semantics and an ability to reason about the world, such that they can acquire knowledge in many domains.

2022-06-20 20:40:12 RT @sinanaral: @GaryMarcus Cuz common sense is about exception handling. Put any AI in an NYC subway for 5 minutes &

2022-06-20 20:30:27 8 hours left for the “Game Over” and “AGI is almost here” crowd to explain why/how common sense has (almost?) been solved.Still waiting for your arguments.If you don’t have any, well, maybe we are not that close to AGI after all? https://t.co/imYO397Zez

2022-06-20 20:27:17 RT @TorrentTiago: @GaryMarcus @JhendersonIMB @fernandaedi This one is not hard for humans, but it breaks dall.e, though:

2022-06-20 20:24:04 @JhendersonIMB @fernandaedi “but that’s one hard for humans, too”, says the deep learning crowd. in this case, they are right.try “The newspaper published by The Herald stank” :)

2022-06-20 04:39:11 Current AI (explain your answer in the comments):

2022-06-20 04:30:18 @paulg i wouldn’t assume that writing bogus essays is actually an important step towards intelligence, rather than just demonstration of the power of statistical mimicry. your tweet may well not age well.

2022-06-20 03:12:41 RT @vboykis: I am much less worried about sentient AIs than the fact that we are surrounded and influenced by machine learning systems and…

2022-06-20 02:16:29 RT @sunilmallya: @GaryMarcus @hardmaru One way to look at it is - this is just the moment that the holes developed. And you happen to captu…

2022-06-20 01:43:42 perhaps a metaphor for the whole field https://t.co/QY25YSByhE

2022-06-20 01:38:48 RT @ForbinP: @UrMultimedia @GaryMarcus I find it both obvious and incredible that, despite all of the hype, an artificial neural network is…

2022-06-20 01:38:13 @ItalyHighTech @samc0hen @ErnestSDavis @irinablok my guess is that it is at chance, and answer will vary from trial to trial

2022-06-20 01:32:19 AGI is gonna be wild, part 733 https://t.co/jRuXY22hkA

2022-06-20 01:12:40 RT @samc0hen: @GaryMarcus @ErnestSDavis @irinablok Seven years later, and AI is still cutting through the wrong side of the branch https://…

2022-06-20 00:43:38 so much speculation, such little evidence (and so little explanation of criteria).why bother with science (or philosophy), at all? https://t.co/3HzVlBjSH0

2022-06-20 00:32:22 @hardmaru do they?

2022-06-19 23:35:34 The M. Mitchells keep telling it like it is. @melmitchell1 ⁦@mmitchell_ai⁩ https://t.co/2fhDYSiH3d

2022-06-19 23:33:20 Very much concur with @MelMitchell1’s critique of PaLM- and am sad to see @GoogleAI to continue to play the game of hype without peer review. Anyone remember all the hype around Google Duplex? Media needs to step up its game. https://t.co/nw6mjftKTE

2022-06-19 23:24:36 RT @Kobotic: 'We are putting bandaids on street lights.' Thoughtful consideration how toxicity cannot be solved- because we are building…

2022-06-19 22:27:31 RT @GaryMarcus: Allen Ginsberg—with special guest Taylor Swift!Poetry for Our AGI Times"It has been a long, long time since I had this…

2022-06-19 21:18:52 @stanislavfort @gwern it’s not very reliable. @ernestsdavis might list some counter examples here

2022-06-19 21:17:02 @_arohan_ @mmbronstein @GoogleAI @ErnestSDavis @aniketvartak want to try “realistic, not surreal” in the prompt?

2022-06-19 20:29:49 @b_snef always the so-so proxy, never more

2022-06-19 20:29:10 RT @Benambridge: @GaryMarcus serious question - can you recommend any recent language modeling work (whether DL, symbolic or hybrid) that D…

2022-06-19 20:28:53 @b_snef @JR_Hochmann we need to write about this work

2022-06-19 03:02:42 RT @javiturek: @motionphi This is an amazing example of how good and bad these models are. A sloth with fingers? There is a big difference…

2022-06-19 03:02:39 @javiturek @motionphi this guy laughs, and appreciates your eye for detail https://t.co/Z6tXEWohtG

2022-06-19 02:48:51 @MadamePratolung AI is indeed in a bit of a pickle

2022-06-19 00:44:14 @plevy @Plinz

2022-06-19 00:43:54 @zenstyle @plevy that’s insane!

2022-06-19 00:33:31 RT @GaryMarcus: @plevy “I need a paradigm shift, not a billion more parameters” https://t.co/fMWIGnHwPX

2022-06-19 00:33:11 @plevy “I need a paradigm shift, not a billion more parameters” https://t.co/fMWIGnHwPX

2022-06-19 00:29:30 @plevy won’t work for this paper but might for another

2022-06-19 00:27:07 @plevy or 3 orders magnitude more data

2022-06-18 23:31:40 this is so well said https://t.co/h22pedyW6c

2022-06-18 23:22:58 RT @curiouswavefn: The illusion of machine intelligence is worse than machine intelligence. Being unable to distinguish the two would set u…

2022-06-18 23:16:32 @MadamePratolung @KathrynECramer @mer__edith “primes darker content”, like Marcus Davis cranberry/grape juice example @techreview

2022-06-18 23:15:13 @KathrynECramer @MadamePratolung @mer__edith ! screenshot?

2022-06-18 23:10:47 @MadamePratolung @KathrynECramer @mer__edith no, but (very roughly) you are priming which part of its corpus to rely on

2022-06-18 23:08:24 @charleswangb @balazskegl yes, you could make that argument. it bullshits, sometime it speak truth, sometimes not

2022-06-18 23:07:23 @MadamePratolung @KathrynECramer @mer__edith there is certainly some degree of generalization, &

2022-06-18 23:05:02 @KathrynECramer @MadamePratolung @mer__edith i suspect you are correct and wish for the sake of science that there fuller disclosure

2022-06-18 21:37:29 @MadamePratolung @KathrynECramer @Miles_Brundage might know

2022-06-18 21:32:10 @elonmusk feel free to join https://t.co/1rS8jDPon8

2022-06-18 21:29:55 @charleswangb thanks. very tempted to use this one!

2022-06-18 04:19:08 that’s what i have been trying to tell you. these systems confabulate spread misinformation constantly because they have no conception whatsoever of truth. *all* they know is the probabilities of word sequences.everything else is anthropomorphism. https://t.co/nzoqS6nrZ3

2022-06-17 22:49:38 RT @anecdotal: @GaryMarcus @sl8rv I’m fascinated to see when AI can do metaphor and catachresis. That’ll be the test of sentience for me.

2022-06-17 22:44:26 RT @danfaggella: .@sl8rv and @GaryMarcus had a little private poetry slam about #AGI, and the poems are worth reading.https://t.co/54Lfjl…

2022-06-17 22:30:23 RT @ShannonVallor: Genius

2022-06-17 21:53:45 RT @PaulTopping: It's hard to get any more meta than this!

2022-06-17 21:48:03 RT @sl8rv: AI Poetry Slam! Perfect for people with interests at the intersection of deep tech and Taylor Swift (which I feel like should be…

2022-06-17 21:23:45 @fchollet clearly humans don’t use exactly the same principles as other primatessee https://t.co/lnr3lD7E1z w @AdamMarblestone

2022-06-17 21:20:33 Allen Ginsberg—with special guest Taylor Swift!Poetry for Our AGI Times"It has been a long, long time since I had this much fun on the internet." -– ⁦⁦@sl8rv⁩Artificial sentiences under the age of 18 not admitted without an adult. https://t.co/a1oI73e2GK

2022-06-17 21:13:25 @bleepbeepbzzz sure but i have emphasized reasoning, outliers, and compositionally from day 1 and these are still where the systems break down.

2022-06-17 20:58:50 @bbenzon @anecdotal @sl8rv

2022-06-17 20:44:50 It's kind of funny how people keep claiming I have moved the goalposts and I keep pointing to the same damn book I wrote in 2001. https://t.co/vxybYF3Esb

2022-06-17 20:39:57 @ylecun @csabaveres sure. i gave my terms in the algebraic mind- operations over variables- explicit type-token distinction- structured hierarchical representations

2022-06-17 20:36:29 @ylecun @csabaveres right, but the question is - does it need to be symbolically, and as you point out in your essay, if it does need to be symbolically, can you learn what you need, or not?

2022-06-17 20:26:13 @ChurchillMic it’s not my prompt, and only a hint not an answer, but i hope it opens some eyes around relevance, distraction and inference.the example reminds me a great deal of Jia and @percyliang’s excellent https://t.co/CSiuDCTfSK

2022-06-17 20:04:14 @ChurchillMic it’s a way of seeing how shallow the comprehension really is. you know, science.

2022-06-17 19:22:06 indeed since the 50s the question has always been: if there is a difference, what is the difference? do neural nets offer a real alternative? or can they only succeed in reimplementing symbol-manipulation? https://t.co/qVRal7SyR4

2022-06-17 19:11:23 @jsm_creative @tejasdkulkarni @GoogleAI a. what were the prompts?b. did you read Horses Rides Astronauts at https://t.co/8ir1xKvqt6?

2022-06-17 15:40:01 @terrible_archer @boazbaraktcs replied above, thanks.

2022-06-17 15:39:36 @boazbaraktcs @ylecun nonsense. deep learning is still having problems in exactly the areas that i have consistently predicted it would for last 10 years, including compositionality and reasoning.

2022-06-17 15:28:54 RT @WillOremus: I wrote about that Google AI, what the Turing test really tells us, and why passing it should be considered a red flag inst…

2022-06-17 13:10:32 Posting something light and fun and completely different, soon.

2022-06-17 13:07:31 RT @GaryMarcus: the hype cycle, revised:- Big Company releases Big Paper

2022-06-17 12:46:01 @fhuszar @DaniloJRezende @unsorsodicorda @jacobmbuckman @slatestarcodex the whole point is that you can’t assume the data in the wild will be from the same space as the training distribution. if it were, we’d have solved intelligence. it’s not, and so we have a problem thus far is unsolved.the ongoing problems with driverless cars are a clear case.

2022-06-17 12:40:47 So much for free speech? https://t.co/D26gi33fnv

2022-06-17 04:43:51 @peterwildeford try all the examples in @togelius’s thread, and report all?

2022-06-17 04:42:05 the hype cycle, revised:- Big Company releases Big Paper

2022-06-17 04:35:41 RT @LindseyGrahamSC: If we nominate Trump, we will get destroyed.......and we will deserve it.

2022-06-17 04:20:43 RT @togelius: Everything sounds more impressive if you add some math. Just ask the typical author of a NeurIPS submission. https://t.co/6cg…

2022-06-17 04:13:38 is this …[shudder] … prompt engineering … hitting a wall? https://t.co/IBp1dCUQGf

2022-06-17 04:03:35 RT @togelius: The nonsense continues. Step by step. https://t.co/veGTSIbew3

2022-06-17 03:18:17 RT @dan_p_simpson: The thing is, we don't need any more innovation in the AI space. All we need is more data and bigger models. Scaling is…

2022-06-17 01:54:30 explaining Blake Lemoine and Lamda over dinner to my 9 year old, and he asks, “Is that like ELIZA?”

2022-06-17 01:40:03 this tweet aged pretty well https://t.co/dOPjb5mCuN

2022-06-17 01:33:57 recording at link below: https://t.co/4XPlDODpY1

2022-06-17 01:20:57 thanks @karaswisher and @CaseyNewton for giving @nitashatiku and me a chance to preview some truly new challenges for society. &

2022-06-17 01:10:08 @bobehayes @karaswisher and @CaseyNewton and @nitashatiku great discussion all around

2022-06-17 01:09:12 RT @bobehayes: Great discussion on #ArtificialIntelligence #hype and whether we are close to having #sentient #AIhttps://t.co/Tn5NUxsTAR…

2022-06-16 23:45:11 @nyttypos @karaswisher @nytopinion @elonmusk @CaseyNewton @nitashatiku @TwitterSpaces the world would be a less interesting place without @karaswisher’s rebellion

2022-06-16 23:43:50 dueling with @nyttypos as the opening act https://t.co/lKEclLqWvi

2022-06-16 23:42:38 RT @BestLiveAudio: @karaswisher @elonmusk @CaseyNewton @GaryMarcus @nitashatiku @TwitterSpaces Excited for this!

2022-06-16 23:42:22 @nyttypos @karaswisher @nytopinion @elonmusk @CaseyNewton @nitashatiku @TwitterSpaces we can let @karaswisher explain her own choice :) i sensed a little bit of deliberate rebellion in her explanation.

2022-06-16 23:40:48 RT @GaryMarcus: if you are sentient and have been following LaMDAgate, you should come to this.

2022-06-16 23:33:02 @karaswisher @nyttypos @nytopinion @elonmusk @CaseyNewton @nitashatiku @TwitterSpaces descriptive linguistics >

2022-06-16 23:12:01 @grbradsk thanks! teachable moment :)

2022-06-16 22:20:21 RT @GaryMarcus: This is going to be something special. @nitashatiku, who broke the LaMDA “sentience” story at the @washingtonpost, and I…

2022-06-16 21:20:37 @DaniloJRezende @unsorsodicorda @jacobmbuckman @slatestarcodex tell us more…

2022-06-16 20:14:58 terrific thread. #deeplearning hit a hexagon https://t.co/Et97Y9Iawj

2022-06-16 20:08:36 +1 from other side of the pond https://t.co/T14aJIsMAS

2022-06-16 19:22:52 @Dr_Cuspy @ylecun i actually wrote about all our similarity a couple days ago and opened our 2017 debate by making the same point. that said i do think things we disagree about (not all outlined in this essay) are pretty important, too

2022-06-16 19:15:53 @AthenaAI2 @ylecun @Jake_Browning00 yeah and i don’t really structure the argument that way. it’s good, it’s not perfect :)

2022-06-16 18:59:50 excellent essay by @ylecun, w @Jake_Browning00, well worth reading, including this clear statement of points of agreement: https://t.co/COf2TLBs7v https://t.co/k05E5apyoQ

2022-06-16 18:31:46 let that sink in. https://t.co/rhiZlT9U0T

2022-06-16 17:38:37 if you are sentient and have been following LaMDAgate, you should come to this. https://t.co/sB0hSsV56F

2022-06-16 17:37:12 @subjacentish @FelixHill84 @dileeplearning @spiantado @emilymbender that one i know and quoted somewhere

2022-06-16 17:22:23 This is going to be something special. @nitashatiku, who broke the LaMDA “sentience” story at the @washingtonpost, and I (AI scientist &

2022-06-16 17:15:27 @FelixHill84 @subjacentish @dileeplearning @spiantado @emilymbender ?

2022-06-16 17:05:12 RT @karaswisher: Hey @elonmusk when you are done chatting with the tweets, come over and help @CaseyNewton and I discuss the pending Termin…

2022-06-16 17:04:52 @FelixHill84 @subjacentish @dileeplearning @spiantado @emilymbender and yet can’t even tell me why or how you disagree with Frege?

2022-06-16 16:32:42 RT @bobehayes: Today at 5pm PT. This will be worth your time. @GaryMarcus talking with @karaswisher#artificialintelligence #hype #transpa…

2022-06-16 16:29:39 Linguistics was solved same way, a couple weeks earlier. Amazing! I think neuroscience is next week. https://t.co/sXbrr2c2lt

2022-06-16 16:23:44 @subjacentish @FelixHill84 @dileeplearning @spiantado @emilymbender could be urban legend, then

2022-06-16 16:20:06 @SelfDrivingLie @karaswisher I like the idea. it’s the execution that is harder than most people (myself excluded-I was an early skeptic in 2016) have imagined.

2022-06-16 16:01:29 @subjacentish @FelixHill84 @dileeplearning @spiantado @emilymbender is the latter a joke, a la the Vision project that he did actually assign? or a documentable true story?

2022-06-16 16:00:42 @subjacentish @FelixHill84 @dileeplearning @spiantado @emilymbender not entirely fair … but not entirely unfair, either. “I met Partee at a party once”ndoesn’t really cut it, and it isn’t, “I read enough Frege to understand why he is so central to two fields, one in philosophy, the other in linguistics, but i still disagree because X”, either

2022-06-16 15:39:05 Honeybees are amazing. They can also do same-different tasks, and generalize the solar azimuth function to lighting conditions they have never seen. The famous waggle dance is just the start! https://t.co/LKaLIEBUuW

2022-06-15 18:31:05 RT @dutchlight360: Will Tesla ever “Solve” Full Self Driving? “Core issue is outliers, the endless list of unusual circumstances, like a pe…

2022-06-15 18:27:10 @FelixHill84 @matyi7m @AndrewLampinen @NickRMorgan @dileeplearning @peabody124 @spiantado @emilymbender on the other hand the key pre 1980s ideas, like Frege’s notion of compositionally inferring the meaning of sentence from its parts, still aren’t solved. (pure) statistical prediction and language understanding: https://t.co/ECKJONYBVi

2022-06-15 18:13:58 @andrey_kurenkov

2022-06-15 18:03:24 RT @PolymathToBe: Very interesting newsletter from @GaryMarcus https://t.co/K4IF5FQvTO

2022-06-15 18:02:43 @BrandonLive @Tesla what specifically are you saying is misleading?

2022-06-15 17:48:41 @BrandonLive if it wasn’t switched off prior to accidents, and of course they ignores terrain, use case limits etc.

2022-06-15 17:46:55 Real Full-Self Driving is nowhere in sight. Why today’s new data are problematic for @Tesla and the whole field. https://t.co/Z21HnFYbuw

2022-06-15 16:26:05 essay on all this: https://t.co/udbUPHrV2v

2022-06-15 16:24:27 fuller essay on this: https://t.co/udbUPHrV2v

2022-06-15 16:24:03 @KPieranski @Tesla @elonmusk funny how I am quoting Elon on that part :)

2022-06-15 16:23:31 @tdietterich @Tesla @elonmusk Well, I am quoting Elon there, and he knows his business better than I.But I think the issue is that the 100:1 P/E doesn’t make sense for a regular car company, even with one that is fun to drive. And if the stock were to suddenly crash, some of the mystique would be gone.

2022-06-15 16:12:43 Will @Tesla ever "Solve" Driverless Cars?And will they survive if they don't? @elonmusk himself doesn't think so. https://t.co/udbUPHrV2v

2022-06-15 15:36:30 Whether you love Tesla or hate it, if you enjoyed my viral “Nonsense on Stilts” take on LaMDA (as covered by CNN, CBC, NPR, etc) please subscribe to https://t.co/8ir1xKvqt6. Report coming soon on what today’s NHTSA’s data dump means for @tesla and the driverless car industry.

2022-06-15 14:50:12 @dileeplearning @FelixHill84 @spiantado 100 percent, exactly what @emilymbender and I are also saying, in different words. No connection to the world (what philosophers of language would call reference) no meaning. Just games with words at that point.

2022-06-15 14:48:38 @jonathanrlarkin do you mund maybe reading all my tweets from the last hour, first? I made same points multiple times, just not in this particular tweet.

2022-06-15 14:32:03 sorry. it doesn’t help Tesla’s claim towards genuine FSD if the whole industry is struggling.rather it reinforces the fact that what has been promised far outstrips what *anybody* can deliver anytime soon. https://t.co/V0B7RTuwI4

2022-06-15 14:26:36 @LucaAmb *eventually* the machines might do better

2022-06-15 14:18:52 @BeebsMemes my hot take which asks same questions and was posted above: https://t.co/TyyRn9yZAW

2022-06-15 14:15:07 @EpistemiclyRich

2022-06-15 14:14:31 3 accidents every 4 days, per new data Imagine how many more accidents there would be per day if Tesla’s ADAS were turned *Fullly* loose in all roads at all times, without driver intervention w millions of vehiclesSo much work left to be done https://t.co/ES9i97P7t5

2022-06-15 03:36:28 @spiantado oh sorry i misparsed your sentence in haste

2022-06-15 03:27:47 RT @AI_4_Healthcare: "It’s time for #ArtificialIntelligence researchers to look up. We can’t solve #AI with PR alone." Timely essay in @s…

2022-06-15 03:09:46 @charleswangb @dileeplearning @spiantado totally he primed you to think it was real….

2022-06-15 02:50:45 RT @GaryMarcus: @srikumarks @charleswangb @dileeplearning @spiantado one man’s joke is another’s delusion

2022-06-15 02:46:39 @srikumarks @charleswangb @dileeplearning @spiantado one man’s joke is another’s delusion

2022-06-15 02:30:16 @charleswangb @dileeplearning @spiantado Ethan was kidding there but overall i agree

2022-06-15 02:17:48 @dileeplearning @spiantado @dileeplearning are making the same point

2022-06-15 02:15:54 Fixed that for you https://t.co/4a8liJlmS6

2022-06-15 02:12:32 @spiantado so does my iPhone’s autocomplete have conceptual role semantics?

2022-06-15 01:44:25 Say how you really feel https://t.co/U8KD7JJtZM

2022-06-15 01:40:39 @spiantado I agree with Bender and Koller that reference is essential, and that it is lacking in LLMs

2022-06-15 01:39:38 @spiantado Note that you could draw the same kind of clustering figure by doing multidimensional scaling on the underlying representations on any autocomplete system (eg on the soft keyboard on my iPhone). Would you say that your autocomplete system is sentient, too, by parity of reason?

2022-06-14 22:15:37 @tegmark i initially mistagged you but a lot of are curious to hear more

2022-06-14 22:01:43 @pmddomingos true observation, but applies to women too.

2022-06-14 20:09:51 @erikbryn @atangibletruth @cajundiscordian @percyliang @sapinker @StanfordHAI @DigEconLab @pmddomingos @deaneckles @JeffDean @elonmusk https://t.co/LnlBzouQD6

2022-06-14 19:57:39 So so true: until we get the right autonomous driving data (which is something scientists ought be involved in defining) “it's just going to be cheerleading + hot takes” https://t.co/zxU3qQ4KKl

2022-06-14 19:40:55 @marc_lepage @NGaylinn @karaswisher so say we all

2022-06-14 19:40:40 About time I run for political office https://t.co/vJLJevWPPC

2022-06-14 19:34:45 RT @NGaylinn: @GaryMarcus @karaswisher I recently read GEB, which focuses on aspects of AI other than deep learning. It was fascinating to…

2022-06-14 18:22:23 @dutchlight360 @MaxTegmark

2022-06-14 02:22:25 @togelius @emilymbender

2022-06-14 02:21:01 this dog being Nipper: https://t.co/lHQCXYG43P

2022-06-14 02:20:05 i have to assume that 11% here were joking. Nipper is not pleased. https://t.co/cHpYf1MbVk

2022-06-14 00:56:14 don’t know why I never noticed this restaurant before https://t.co/d3slTAa5I6

2022-06-14 00:43:47 RT @timFinin: In this opinion piece in Scientific American, Gary Marcus (@GaryMarcus) likens systems like OpenAI's DALL-E 2 and DeepMind's…

2022-06-14 00:28:16 you saw (or should see) the netflix show, sex education, now read about a real life version. kids stepping up,where grown-ups fail them. https://t.co/pRN54yk8ep

2022-06-13 23:33:09 @DrMJoyner @ACasadevall1 @DavidEpstein @sweatscience @NAChristakis @cragcrest @lindsaycrouse @JoNel_Aleccia @JonathanLWai the story is a little overwrought but i think that kind of network analysis is true for many things

2022-06-13 23:20:16 @amyalkon @slatestarcodex already saw - i subscribe! his and @Noahpinion were the first two I signed up for when I joined Substack in May.

2022-06-13 23:17:49 remember how yesterday i said that academia had its own problems? this is a great report on one: the all-too often cumbersome process of spinning out companies from academic research.we need our academies, but (many of) our academies need to do better https://t.co/0nNnOAyeTl

2022-06-13 21:37:48 RT @KarlBode: this thoughtful breakdown on why Google's AI isn't particularly bright or sentient will get .02% the attention of inflammator…

2022-06-13 21:37:03 @bbenzon @nypost @ethan you forgot to include the Ilyaic constant of slightness, usually written as {BS}

2022-06-13 21:07:08 @pavel_soukenik @nypost the ball refuses to confirm or deny

2022-06-13 20:49:37 RT @sinanaral: @GaryMarcus @mariaressa What a great (and important) conversation! Thank you so much for having it with me @GaryMarcus Th…

2022-06-13 20:21:01 @tyrell_turing steve cohen, owner of the Mets, inspiration for the show Billionaires, might like a word with you: https://t.co/MHG5LtXvHt

2022-06-13 20:03:54 links to the series: https://t.co/wfZpp2ER0N

2022-06-13 20:03:04 Judging by my experience this morning as interviewee, @sinanaral’s new The Digital Insider is going to be w fantastic series, with deep conversations about technology, ethics and more. First guest was @mariaressa!Honored to be part of it. will post a link when our episode is up https://t.co/eO4r9BRKxv

2022-06-13 19:44:39 RT @sapinker: Nonsense on Stilts: @garymarcus on whether LaMDA, the pattern extrapolator recently credited with "sentience," is intelligent…

2022-06-13 19:43:38 RT @charleswangb: Might be one of the most underrated truth in Gary's efforts.This is what's at issue, at risk, at stake, in fundamental…

2022-06-13 19:43:20 @Sams_Antics funny that

2022-06-13 19:13:08 @tyrell_turing give or take some hallucinogens? there must be a confidence interval in here somewhere

2022-06-12 16:14:53 RT @erikbryn: Foundation models are incredibly effective at stringing together statistically plausible chunks of text in response to prompt…

2022-06-12 15:47:56 @ylecun i hope i have represented your position correctly

2022-06-12 15:46:35 And featuring criteria inspired by @AVMiceliBarone for how we might tell—sharpening a potential bet proposed by @slatestarcodex @kevin2kelly @MatthewJBar @tamaybes

2022-06-12 15:41:39 Why I think there is a 90% chance that AI needs a paradigm shift—and why I am hardly alone in thinking that. https://t.co/Odg2QIK6bX

2022-06-12 13:57:03 @solidoxx @bleepbeepbzzz @_mattbeard @slatestarcodex no i see signs that they are faking it. they really don’t, and you are confused by mimicry.

2022-06-12 13:49:09 RT @memosisland: Refreshing ideas from @GaryMarcus

2022-06-12 13:46:22 RT @dutchlight360: @RikeFranke We are still not even close … please read the article of @GaryMarcus of 2014

2022-06-12 13:34:38 Coning Wednesday, a Twitter Space discussion on my (proposed) bet with @ElonMusk on where AGI will be in 2029.11am PT, hosted by The UN Brief@MayaPlentzopen to all, especially @elonmuskhttps://t.co/SK3UyTp29q

2022-06-12 05:18:08 @MFordFuture @cajundiscordian it might be historically significant but i don’t think it has anything to do with general intelligence and think it is a mistake to emphasize it.if kurzweil wins it will not be because AGI has been achieved.

2022-06-12 03:54:28 RT @pomeranian99: I always learn stuff from watching @GaryMarcus and Ernie Davis poke around with massive language models and find the plac…

2022-06-12 03:39:25 @MFordFuture @cajundiscordian https://t.co/6pkCfIqUiw

2022-06-12 03:37:25 Not sure what Turing would say, but I don’t think the Turing Test itself is meaningfulrelies on human gullibility it can easily be gamedadvances in it have not historically led to advances in AIessay I wrote about in 2014 still applies: https://t.co/4KLe1cbDny https://t.co/5b63hb1mmi

2022-06-12 02:21:04 RT @aniketvartak: After @GaryMarcus another heavyweight - Douglas Hofstadter weights in on the AI hype. 1/?

2022-06-12 02:06:25 @Ted_Underwood https://t.co/tkr7P1kEGC

2022-06-12 01:10:42 @andrey_kurenkov thanks for clarifying and I’ve made some edits based on our DM’s that we just had :)

2022-06-12 00:46:44 @EmmetPeppers @andyzhouu @TeslaTydirium @NHTSAgov @goodsoilinvest @NaveenGRao @Tesla @elonmusk and i couldn’t squeeze in the most important point, which is that the L2 stuff is mainly used on highway, so not strictly comparable to lots of other driving where outliers are more of a problem.

2022-06-12 00:45:32 @EmmetPeppers @andyzhouu @TeslaTydirium @NHTSAgov @goodsoilinvest @NaveenGRao @Tesla @elonmusk BUT1. Top graph does a classic trick of shortening y x-ais (starting from 70 not 0) to exaggerate effect size.3. Selective

2022-06-12 00:32:24 RT @mpshanahan: A large language model that uses the word "I" is no more conscious than this bottle https://t.co/fvk5B68Xw7

2022-06-12 00:30:44 @MFordFuture @cajundiscordian Of course it’s not sentient. Pastiching human language does not make a machine sentient.Unless Google allows scientific community (including me) access I am not going to take it seriously.

2022-06-12 00:25:58 @yudapearl @slatestarcodex @MelMitchell1 @mmitchell_ai @timnitGebru @JanelleCShane @katecrawford @merbroussard @ErnestSDavis @ylecun and here’s the last episode, which cites you/links Book of Why:https://t.co/ml0aDgCwpW

2022-06-11 13:14:41 RT @xccastells: Why does the big black dot seem to expand? The answer here:https://t.co/d3KJAqvs5PThis study provides additional support…

2022-06-11 12:29:35 RT @andrey_kurenkov: It's pretty obvious that 'simply scaling' a GPT-style LLM will not lead to AGI by virtue of the inherent limits of the…

2022-06-11 05:31:56 RT @cyberandy: From pattern-matching AI systems to true AGI. Abstraction is all we need. A great (long) thread, summarizing @GaryMarcus cri…

2022-06-11 05:09:35 @andyzhouu @EmmetPeppers @TeslaTydirium @NHTSAgov @goodsoilinvest @NaveenGRao @Tesla @elonmusk calling for fuller disclosure of data and truth in messaging does not equate to calling for the project to be canceled.

2022-06-11 05:08:15 @andyzhouu @EmmetPeppers @TeslaTydirium @NHTSAgov @goodsoilinvest @NaveenGRao @Tesla @elonmusk they haven’t. by opting out of reporting CA intervention statistics &

2022-06-11 03:04:22 RT @DoveFirebrand: Inspired by the very thought-provoking debate between @GaryMarcus and @slatestarcodex &

2022-06-11 03:01:37 I used to think it was just AI that exceeded his grasp https://t.co/hBWs2itfXY

2022-06-11 02:58:26 @mosesjones yeah this was pretty problematic @kncukier

2022-06-11 01:59:42 @calistoker123 @AlbertBridgeCap @EmmetPeppers @TeslaTydirium @NHTSAgov @goodsoilinvest @NaveenGRao @Tesla @elonmusk sure, once the engineering is mature. not there yet, but someday

2022-06-11 01:58:44 @MatchasmMatt @NaveenGRao @EmmetPeppers thanks for listening with an open mind :) more on my views about AI (not yet specifically about cars) @ https://t.co/8ir1xKvqt6

2022-06-11 01:53:37 The federal government’s Tesla Autopilot investigation is moving into a new phase - The Verge(re the ongoing issues crashing into stopped vehicles that was mentioned in 2019, in https://t.co/Pt7HZbLIv5) https://t.co/Gxe5wF8RCr

2022-06-11 01:31:24 @sl8rv @slatestarcodex pay no attention to the man behind the curtain?

2022-06-11 00:41:57 doesn’t build trust that Tesla is not meeting or exceeding industry best practices for data disclosure https://t.co/DamFSLPr1C

2022-06-11 00:26:29 @EmmetPeppers @TeslaTydirium @NHTSAgov @goodsoilinvest @NaveenGRao @Tesla @elonmusk there are such studies, i have posted. and yes i did raise these kinds of confounds in the interview saying you need to compare with other similar cars, similar mass, etc (eg i may have mentioned similar age of vehicle and types of moles). coarse statistics are very misleading.

2022-06-11 00:24:40 RT @nirsd: While DALLE-2 can do new and more spectacular things than previous models, the way they fail is reminiscent of older models, e.g…

2022-06-11 00:24:32 RT @nirsd: DALLE-2 etc are not AGI. I agree with @garymarcus that scaling is not likely to force some kind of phase transition where the mo…

2022-06-11 00:18:40 @EmmetPeppers @TeslaTydirium @NHTSAgov @goodsoilinvest @NaveenGRao @Tesla @elonmusk sorry but no. we explained why in the interview. soooo many confounds if you do it that way

2022-06-10 23:29:03 Does it really serve Tesla community’s interests to downplay edge cases, when literally every expert in the field knows that is the key problem? Buys a little immediate comfort but doesn’t solve what ails the autonomous vehicle industry. https://t.co/ZIUZrs44Am

2022-06-10 23:25:39 And people used to think Y2K was a problem https://t.co/9ikt0bHUMu

2022-06-10 23:22:24 typo: @NaveenGRao

2022-06-10 23:19:25 RT @NaveenGRao: @EmmetPeppers @LifeAt130BPM @goodsoilinvest @GaryMarcus Tbf I own a Tesla, changed from solar panels, and was a really big…

2022-06-10 21:13:15 @unsojo

2022-06-10 21:11:43 @Pehdrew_ @slatestarcodex says above: mention in my reply to SSC.

2022-06-10 21:05:50 if nobody steps up, I am putting you i charge @MelMitchell1 :)

2022-06-10 20:55:13 @matthewcobb so much for the “time flies” intepretation

2022-06-10 20:52:16 @bradpwyble @ShlomoArgamon @slatestarcodex @ErnestSDavis the prompt ended at step by step

2022-06-10 20:40:50 essay contest! what’s a paradigm shift? how can the field turn ⁦@slatestarcodex⁩’s propositions 4 and 5 into something tangible?winner(s) might get mention tomorrow in my substack reply to SSC https://t.co/enikL3cRvD

2022-06-10 20:35:16 @neuroecology @slatestarcodex @ErnestSDavis @rogerkmoore put it well: https://t.co/WR7s95qCpF

2022-06-10 20:27:18 Indeed, who are kidding?“Ok, Star Trek Computer, how many dilithium crystals can we synthesize before impact? Using a state of the art protocol, please supply a SOTA estimate, with minimal bias, and remember to take this step by step.We only have 17 seconds.” https://t.co/zzAB76OhbE

2022-06-10 20:13:10 @neuroecology @slatestarcodex @ErnestSDavis the problem is that it is always just sequences of words and never actual concepts

2022-06-10 20:01:10 reader @slatestarcodex: what if you gave the cow a name, Blue Bell?@ErnestSDavis: ok, let me tryGPT-3 [deadpan]: “It is possible for a cow to be revived after it has died, but it is not something that happens often. It is more likely that Blue Bell will remain dead.”YMMV https://t.co/oWF1dE1UwI

2022-06-10 19:44:39 @NaveenGRao @calistoker123 @EmmetPeppers @MatchasmMatt @Tesla and Tesla lacks lidar which is an important potential source of information. unnecessary for humans but possible necessary in systems that lack human cognitive flexibility and common sense.

2022-06-10 19:34:06 @yudapearl @slatestarcodex @MelMitchell1 @mmitchell_ai @timnitGebru @JanelleCShane @katecrawford @merbroussard @ErnestSDavis @ylecun his last post (which links my last post, which links his first post) is here. and i will reply tomorrow at https://t.co/8ir1xKdPBy https://t.co/Pxu8xyYs4Q

2022-06-10 19:09:14 @ChrSzegedy @MelMitchell1 @ChrSzegedy please follow back or dm me your email so we can iron out some details :)

2022-06-10 18:47:43 Lots of people are asking for a sequel. What will it take @ylecun to get that project off the ground?@slatestarcodex and I are certainly having fun :) https://t.co/WBSNPsBgDG

2022-06-10 18:45:33 @PaulTopping maybe when @ylecun sees how nice i am to him tomorrow on Substack, he will finally agree?

2022-06-10 18:36:02 @bleepbeepbzzz @_mattbeard @slatestarcodex !

2022-06-10 18:33:38 Deep in the heart of @tesla/TSLA fandom, @NaveenGRao and I raise some concerns about the current state of play in autonomous driving, and suggest what kinds of data we would really like to see.cc @CadeMetz @lorakolodny @PeteButtigieg https://t.co/tFgxGS39Kx

2022-06-10 18:30:07 it was great! still relevant, too. https://t.co/WBSNPsBgDG

2022-06-10 16:52:00 RT @ProfArbel: Inspired by @GaryMarcus, I just bet $500 with a friend that by 12.1.2029, AI will be able to summarize and analyze the narra…

2022-06-10 15:23:36 and rising star @Abebab!

2022-06-09 19:32:28 just recorded a super fun podcast on @tesla and self-driving with @NaveenGRao - will post a link when it is live.

2022-06-09 17:29:17 it’s not a good idea to post replies without context (of numerators and denominators and other pesky facts about reliability) https://t.co/x1sFxrQuBe https://t.co/dDSMIH1sni

2022-06-09 17:26:12 @TaliaRinger that’s why i keep writing about them.

2022-06-09 17:24:31 @MichaelTrazzi name one goalpost that I ever moved, w possible exception of winograd

2022-06-09 17:23:53 what goalpost did i ever move? i was wrong about how easily Winograd was gamed but pretty much everything else i have said i think still stands.so many accusations, so little evidence. https://t.co/v4sWH783wx

2022-06-09 16:51:03 from new report from @ErnestSDavis, responding in part to anecdotal data from @plinzContrary to popular belief, AI Prompt Whisperer is probably not a profession with a future https://t.co/1Z0FirqtUx https://t.co/LP9wb8EG93

2022-06-09 16:42:38 @Whats_AI @discord @towards_AI see my reply, at https://t.co/8ir1xKvqt6

2022-06-09 16:32:12 @LucaAmb i would honestly want something like general intelligence in my medical system, yes, given the variety of potentially relevant information (eg circumstances in sustaining an injury)

2022-06-09 14:38:52 Difference between shouts of allegedly exponential AI progress in the lab, and the reality of making products in the real world, even with all the resources of Google:#Singularity #AGI #RealityCheck https://t.co/gSmtzjPHFx

2022-06-09 14:32:11 2018 Marcus/Davis, NYT: “If you read [about] Google Duplex..scope of..project is surprisingly limited. It encompasses just three tasks: helping users “make restaurant reservations, schedule hair salon appointments, and get holiday hours.”2022 You can make movie reservations, too https://t.co/vWBoVmLdWc

2022-06-09 14:03:32 this link is old (2019), but what did ever become of Google Duplex? anyone know? demoware? actually in use? for what? any recent updates? https://t.co/KvIGIcs628

2022-06-09 14:02:06 @dontmeanathing @BierVicki @ErnestSDavis “the cow” implied it, and ernie replicated w some other wordings (and got a correct answer at least once, in perhaps 4 attempts?)

2022-06-09 13:39:34 @dxbarradas i will try to take that as a compliment, and yes it’s an exactly tweet (w a tiny bit of clarification from me underneath)

2022-06-09 13:38:38 shh, @BierVicki ! don’t ask too many questions! the funding might dry up!and of course “step by step” is not really universal, eg courtesy @ErnestSDavis: https://t.co/VsaIbhR6X5 https://t.co/qKiqlf4o86

2022-06-09 13:34:04 to which @matthewcobb, riffing on my horse riding astronaut essay, added this https://t.co/fdFGzr3Xqh

2022-06-09 13:31:54 @lvalentgar oops somehow messed it up. will reply to my own reply in a second with rest of thread.

2022-06-09 13:21:25 RT @GaryMarcus: or imagine if your self-driving car only works with the right incantation, with “drive to the Grand Canyon” yielding tragic…

2022-06-09 13:18:12 or imagine if your self-driving car only works with the right incantation, with “drive to the Grand Canyon” yielding tragically different results from “drive to the edge of the Grand Canyon”.2/2

2022-06-08 03:01:07 @srikumarks @DrLaurenOR @ykilcher @huggingface the fact that no ethics approval was required is itself a further wake-up call

2022-06-08 02:57:40 @srikumarks @DrLaurenOR @ykilcher @huggingface this has potential to scales hurtful uncivility to the nth degree, just as LLMs have likely already scaled misinformation

2022-06-08 02:56:39 @srikumarks @DrLaurenOR @ykilcher @huggingface i just don’t want to spell it all out to give people ideas but i think the same kind of hack in a different context could cause a lot of suffering to a lot of people, and the ease with which this hack was done is something that needs to be reckoned with.

2022-06-08 02:54:24 @DrLaurenOR @srikumarks @ykilcher @huggingface the reason i am somewhere in between on this is that doing this in a controlled way on 4chan helps the community to predict—and maybe defend—against much worse things that one can imagineit’s a clear warning.

2022-06-08 02:47:48 @sl8rv @quaesita Darkness falls across the landThe AGI hour is close at handRobots crawl in search of blood…go, @sl8rv (&

2022-06-08 01:02:34 RT @nutanc: @GaryMarcus @slatestarcodex I think @slatestarcodex does a nice turn of words.He mentions "Someone releases a new AI and demons…

2022-06-07 23:45:48 game on :)comment below, under my reply to @ChrSzegedy to help us formalize details :) https://t.co/qUFufFSPgi

2022-06-07 22:56:13 RT @xkcd: Good news: #dalle can do urban planning! I used it to generate some intersections, and I think they look well-designed and very s…

2022-06-07 22:34:01 @MMikeMMa @AI21Labs probably, but weighting recency seems to me like a half-baked, partial solution. eg you might want to know who the previous president was.

2022-06-07 22:20:19 @MMikeMMa if you mean lose their freshness like the Preisdent example in the reply to SSC, i think it is hard to know. but it’s definitely an issue. @AI21Labs

2022-06-07 21:43:44 The AI hype cycle, as summed up by @slatestarcodex, and my reply: https://t.co/B5i2PVT56s https://t.co/Ymc0v7WG0D https://t.co/gndqttSzIn

2022-06-07 21:40:22 RT @johannes_hage: You definitely have to give @GaryMarcus credit for responding quickly to criticism tho (he was on a panel at a worksho…

2022-06-07 20:35:26 @ChrSzegedy @MelMitchell1 Ok @ErnestSDavis &

2022-06-07 19:44:00 RT @wcrpaul: @bimedotcom @EvaSmartAI @BetaMoroney @Fabriziobustama @Khulood_Almani @chidambara09 @LavaletteAstrid @pchamard @MaiaGabunia @C…

2022-06-07 18:58:29 Impromptu online debate!It's @slatestarcodex vs @garymarcus!dueling Substacks! more bets!his: https://t.co/fF83oNkmOmmine: https://t.co/Gh51GFX4pusubscribe to both :)optional

2022-06-07 18:57:29 RT @grsimari: Except for the last one, I think none of these items will be solved by an AI by 2029. I'm doubtful that an AI will be in the…

2022-06-07 18:48:29 RT @sl8rv: Part 2 in what is proving to be an addictive back-and-forth with @GaryMarcus, but something I feel like @quaesita would enjoy to…

2022-06-07 18:37:14 @yoavgo @maxisawesome538 and in fact that is what he argues

2022-06-07 17:53:23 @__dipam__ @ChrSzegedy @MelMitchell1 @ErnestSDavis @f_charton @MatthewJBar @Abel_TorresM excellent questions, open to suggestions.

2022-06-07 17:52:47 @begusgasper open to suggestions.

2022-06-07 04:51:34 RT @juanBotiaBlaya: Brilliant @GaryMarcus https://t.co/nynZiYW4dv General Intelligence Is Not as Imminent as You Might Think via @sciam

2022-06-07 04:45:38 controversy today w @DrLaurenOR &

2022-06-07 04:31:22 RT @Abel_TorresM: Do GD learning solve -extremely trivial- multistep reasoning tasks?

2022-06-07 04:25:22 @rayerskins @DjAlkimi @hillel_art @arttank the bet is here https://t.co/lEal60e54Z

2022-06-07 03:41:05 @srikumarks @slatestarcodex don’t know if the data gets added to training - it’s an interesting question!i didn’t find it compelling, either. and have drafted a reply, likely will post tomorrow morning PT.

2022-06-07 03:10:34 RT @bobehayes: From the latest article by @GaryMarcus in @sciam. Why we need basic research in #ArtificialIntelligence.

2022-06-07 03:10:10 RT @rao2z: It has become fashionable to justify the lack of guarantees of large learned models in #AI with an airy "but people provide no…

2022-06-07 02:16:28 Coming soon, to https://t.co/8ir1xKvqt6, my reply to @slatestarcodex’s dissection of @garymarcus examples :)As ever, subscribers see it first :) https://t.co/qhEAviHFto

2022-06-07 02:14:33 RT @htTweets: #Opinion | "AI can do amazing things such as speech recognition &

2022-06-07 01:13:27 . ⁦@slatestarcodex⁩ writes 3,600 words critiquing … @garymarcus!My reply coming soon :) https://t.co/1GsPRQ7ZeT

2022-06-07 00:27:32 hey @slatestarcodex interesting analysis

2022-06-06 21:47:59 RT @nirsd: Gradient descent has brought us this far into the uncanny valley. Can it also lead us out?@GaryMarcushttps://t.co/1qeyzhyFJV

2022-06-06 20:54:55 @DuaneJRich @ylecun thanks, but as a human being, i also care about the rampant and consistent misrepresentation of my position. portraying me as a late follower to something I myself pushed for years against resistance is both unfair and a distortion of intellectual history.

2022-06-06 20:52:16 @miguelisolano @FelixHill84 @ylecun @ZoubinGhahrama1 @AliceAlbrecht I was thinking more eg of her work on the origins of discrete number counting systems and her book The Origin of Concepts, than fast mapping (also hers) per se. @StanDehaene has also thought a lot about such things.

2022-06-06 20:34:39 “People love spending time with a disruptor like me. At home, you’re just doing what you’re doing all day with no interruptions. At the office, I might see what you’re doing, tear it all up, and tell you to start over. You can’t put a price tag on… that.” https://t.co/6rIMLcwmfM

2022-06-06 20:16:45 RT @fjmendez: Great article on the #AIdelusionals...

2022-06-06 20:16:25 @fjmendez @sciam but see the Musk piece at (and subscribe to) https://t.co/8ir1xKvqt6

2022-06-06 19:56:30 RT @billhilf: Strongly agree with @garymarcus in this well written piece 'Artificial General Intelligence Is Not as Imminent as You Might T…

2022-06-06 19:53:10 @rolisz @f_charton @ErnestSDavis @sameer_

2022-06-06 19:52:08 @begusgasper but think what power Lisp got from linked lists.i argue in the Birth of the Mind that recursion added to a primate genome takes you a huge way.

2022-06-06 18:20:05 on language, variables, and the origins of knowledge

2022-06-06 18:13:38 @FelixHill84 @ylecun @ZoubinGhahrama1 or maybe just maybe that doesn’t have to be either/or

2022-06-06 18:11:49 but will we have the wisdom to recognize dead ends when we hit them? https://t.co/QELcosw7yQ

2022-06-06 18:11:17 @FelixHill84 @ylecun @ZoubinGhahrama1 if you can do the work that i want done (reasoning, including causal, temporal. physical and psychological reasoning) without variables, more power to you.

2022-06-06 18:07:56 @FelixHill84 @ylecun @ZoubinGhahrama1 that is exactly the right problem: where we do get the variables from? best work is maybe Susan Carey in cognitive development context

2022-06-06 17:57:03 RT @laurahelmuth: "We are still stuck on precisely the same challenges that academic scientists (including @GaryMarcus) have been pointing…

2022-06-06 17:51:08 RT @wadhwa: Artificial General Intelligence Is Not as Imminent as You Might Think https://t.co/vO9Vclf9zL via @sciam by @GaryMarcus --E…

2022-06-06 17:50:57 @FelixHill84 @ylecun @ZoubinGhahrama1 i think at the absolute least you need priors about there being persisting objects, places, agents in the world

2022-06-06 17:50:10 @FelixHill84 @ylecun @sapinker we ran into each other on a plane in 2018 or so and had a great chat about old times :)

2022-06-06 17:38:44 @FelixHill84 @ylecun @sapinker how are his streams not dichotomous? i never understood why didn’t cite our well-known related work and compare and contrast…

2022-06-06 17:37:05 @FelixHill84 @ylecun @ZoubinGhahrama1 now, just about anywhere in the world you can study large language models, which in a decade people will realize were not that useful for natural language understanding, after all.

2022-06-06 17:18:11 @FilipoGiovanni @ylecun do i get to strawman “Marcus” or do I have to behave myself?

2022-06-06 17:13:16 @FelixHill84 @ylecun on that I fully agree, the system is not remotely fair. genuinely novelty is rarely rewarded

2022-06-06 17:12:20 @FelixHill84 @ylecun and, brother, what academic has not faced a ton of rejections? certainly not me.

2022-06-06 17:11:27 @FelixHill84 @ylecun that’s stuff is so exaggerated. Hinton ran the huge Gatsby institute in London, with Dayan, @ZoubinGhahrama1 and other stars at his side.

2022-06-06 17:07:05 RT @GaryMarcus: 6/6 “It’s time for artificial intelligence researchers to look up. We can’t ‘solve AI’ with PR alone.”Excerpt from my n…

2022-06-06 17:06:29 @FelixHill84 @ylecun - transformers didn’t exist in 1992- even today transformers have a lot of trouble generalizing in an across the board way (which was what the 1992 monograph was about), see @yasaman_razeghi and @sameer_ 2022 arXiv

2022-06-06 17:04:15 @FelixHill84 @ylecun also, i don’t know what CS departments per se were doing in 1992, but psychology and cogsci departments were certainly hiring lots of neural net people. People like Rumelhart, McClelland, Jordan (then doing neural nets), Elman, Seidenberg were employed by top departments, etc.

2022-06-06 17:00:50 @FelixHill84 @ylecun but no to “a bit”

2022-06-06 17:00:09 @FelixHill84 @ylecun Yes, the conclusion of my monograph on overregularization with @sapinker (my first publication, 1992) was for a hybrid model, associative network for irregulars, symbolic rule for regulars.https://t.co/3VeqRZB4RXAlso, it was the whole point of my 2001 book, The Algebraic Mind

2022-06-06 02:39:54 @ylecun https://t.co/36cIo7Nb4W

2022-06-06 02:00:53 stimulating conversation with @ch402, re biology and computation https://t.co/xPSM87iD10

2022-06-06 01:57:35 @ch402 @stephen_wolfram @MelMitchell1 another way to think about minimal bridges might be in terms of a minimal primitives in a language like Lisp: https://t.co/URxFVoXQxD

2022-06-06 01:55:34 @ch402 @stephen_wolfram @MelMitchell1 Great question! The Algebraic Mind proposed a possible minimal set that needed to be bridged: operations over variables (including storage, retrieval, and comparison), a type-token distinction, and machinery for structured, hierarchical representations.

2022-06-06 01:51:47 @ChrSzegedy @iraphas13 @ch402 @stephen_wolfram @MelMitchell1 pandemic hasn’t helped…

2022-06-05 23:08:00 just now, British Columbia: https://t.co/y6cJ1fNV3Z

2022-06-05 22:29:02 @ch402 @stephen_wolfram @MelMitchell1 i have started a longer essay on semantics and understanding

2022-06-05 22:28:11 @ch402 @stephen_wolfram @MelMitchell1 100% support trying to identify specific hypothesized mechanisms, and will be curious to see how such mechanism relate to the classic operations of symbol-manipulation (which includes store-and-retrieve/copy and paste, as well as variable instantiation more generally).

2022-06-05 22:16:39 @santoroAI fair. but i would put it differently: parsimony might require a zillion parameters, but that doesn’t make it elegant.

2022-06-05 22:15:25 @ch402 @stephen_wolfram yes this is excellent. (in my book i used topographic maps to make this point)it’s the stuff of which innateness is made. when the cascades are rich enough, you have massive power, like writing Python with a bunch of awesome libraries.then maybe you need a lot less data

2022-06-05 22:11:21 true. importantly though, unbelievable power ≠ unlimited power, and progress will only come through appreciating the limits in order to understand what needs to come next. https://t.co/PMXFGjXUCX

2022-06-05 22:08:48 @ch402 @stephen_wolfram pretty sure it is true :) @WiringTheBrain @Dr_SimonFisher (some of the argument is in my book The Birth of The Mind, which tries relate the cognitive science of innateness with the mechanisms of developmental biology)

2022-06-05 22:06:12 @ch402 @stephen_wolfram @UriAlonWeizmann

2022-06-05 22:01:10 @ch402 @stephen_wolfram yes the question of how to acquire new abstract representation is key

2022-06-05 21:59:36 @ch402 @stephen_wolfram there’s in my view a strong argument that each gene is like a (separate) cellular automaton, firing under certain conditions, in order to build a self-organized (but partly preprogrammed, partly sensitive to environment) structure (eg heart or brain).

2022-06-05 21:55:20 RT @MelMitchell1: I like this analogy, but seems incomplete. Evolution continually builds up complexity, by creating hierarchy, modularity…

2022-06-05 21:54:07 @santoroAI one thing to say a kajillion parameters are necessary for a given task, another to say that is an *elegant* solution CYC has a kajillon parameters of a different sort but people rarely proclaim it to be elegant.

2022-06-05 21:52:43 @ch402 glad to hear you are *not* lumping yourself in with those who have framed deep learning etc as biologically plausible (when there are many important disanalogies as well).but couldn’t you make the same argument wrt to cellular automata, as @stephen_wolfram has?

2022-06-05 18:41:43 @ParejaAldo @ch402 https://t.co/xPSM87iD10

2022-06-05 18:39:13 @seanmcarroll

2022-06-05 03:00:56 RT @KrishnaKGupta: "Systems like GPT-3 aren’t quite like monkeys with typewriters, but nor are they much like humans. They do not gravitate…

2022-06-04 16:52:33 @raphaelmilliere good start

2022-06-04 16:26:38 @lathropa is that Elon in the bottom row, center?

2022-06-04 16:23:34 Dall-E-style GIF generator (eg “dancing chicken dressed as a robot”)Reasonably reliable, widely used:

2022-06-04 16:15:54 dall-e fans, please help :) https://t.co/Di8ihjrBj8

2022-06-04 16:05:51 @mrgreene1977 @wadhwa was expecting a chicken in a robot costume but close enough :)

2022-06-04 15:59:47 @wadhwa @mrgreene1977 has a theory about that…

2022-06-04 14:32:23 @IntuitMachine thanks you, Carlos

2022-06-04 14:14:00 https://t.co/ALE9rSHcIA

2022-06-04 11:51:34 RT @thenextweb: The cow says moo. The sheep says baa. But the chicken isn't making a peep https://t.co/SeEs70kQs0

2022-06-04 04:44:11 thought-provoking video on bots and social media by @ykilcher. https://t.co/UMll8P4M76

2022-06-03 22:59:54 RT @wadhwa: Why is Elon Musk too chicken to take a measly $500K bet on AI? https://t.co/5SSUG0erN4 via @thenextweb@GaryMarcus, @mrgreen…

2022-06-03 22:59:24 @bleepbeepbzzz @nutanc no. the question is whether the system can reliably infer the meanings of sentences from their composition.

2022-06-03 22:58:10 RT @UsmanSattarMD: I agree with @GaryMarcus on the criteria for #AGI but interesting discussion nonetheless!

2022-06-03 19:50:43 ≈ “Homeless rich guy living in friend’s Austin, Texas mansion just can’t admit he was wrong.” https://t.co/y8Q3CQvMJC

2022-06-03 19:17:07 @BeebsMemes then again their technical advantage in this domain is mainly (massive) data and infrastructure that pertains specifically to driving, rather than algorithmic innovation per se (or eg large-scale common sense knowledge). so there may be less to transfer than one might think.

2022-06-03 19:11:33 @grbradsk biggest technical mistake in my view is eschewing lidar when the software isn’t likely to be smart enough to do without it

2022-06-03 18:44:30 @ipeirotis i think that’s just definitional, and depends on what you want to use the term for

2022-06-03 18:43:39 @BeebsMemes just want to point out that in principle maybe you could do all that and solve driverless cars and still perhaps not have AGI, beyond the one application.

2022-06-03 18:41:35 @BrandonLive for reference his exact words were “i’d be surprised if we don’t have AGI by then”, which sounds to me stronger than “could”and weaker than “will”

2022-06-03 17:37:49 Musk has been (presumably deliberately) overpromising about AI over and over for *years*

2022-06-03 17:24:50 @tdietterich that’s fine for some cases but not others (eg point to point driving without human involvement)

2022-06-03 17:22:57 @ElliotMurphy91 @cedricboeckx personally i make no commitments about whether Merge is even the right primitive per se, but surely language is mosaic of some old elements and some new (as i argued in that old cognition article)

2022-06-03 17:15:38 @ElliotMurphy91 i think @cedricboeckx is exactly right here, even if we don’t have an enough causal understanding to prove it. my own similar statement: https://t.co/n1731uyzcZ

2022-06-03 16:49:26 @elonmusk @jack $500,000 says you are wrong about the future of AI, @FortuneMagazine: https://t.co/MZjzZS8cec

2022-06-03 16:42:19 .@ElonMusk, these experts will bet you $500,000 that you’re wrong about the future of A.I. https://t.co/MZjzZS8cec

2022-06-03 16:39:42 @orenmatar if that’s all it can do, they shouldn’t show it!

2022-06-03 16:30:58 Seems to be some of that going around. https://t.co/m342QDno86

2022-06-03 16:29:51 RT @bobehayes: Great book by @GaryMarcus and @ErnestSDavis. Got the audio book a couple of years ago when I had a long commute.No #Artifi…

2022-06-03 16:16:15 Optimus predictions (for #AIDay2, September 30) are coming in. Add yours to the thread below :)- teleop?- another dancer in a costume?- or the real deal?and if the real deal, what do you think will be the smartest thing it might do, and how robust do you think its AI will be? https://t.co/5wkm5dKgf5

2022-06-03 05:29:26 @pmddomingos @TaliaRinger sure. i certainly think *some* people have been aware (yourself included!). I do think that the deep learning community for too long confused within-distribution generalization with generalization, period, and I still see that from time to time.

2022-06-03 05:21:51 @pmddomingos @TaliaRinger when I first wrote about in 1997 I was accused of “a terrorist attack” on neural networks.when I stressed it as the key weakness of MLPs in 2001, I was ignoredwhen I pointed it out in 2018, LeCun said I was “mostly wrong”etc etc

2022-06-03 05:17:27 six years ago in AI history: Tay, the neo-Nazi millennial chatbot, gets autopsied | Ars Technica https://t.co/GUkpx1ft5v

2022-06-03 04:26:20 @emilymbender @TaliaRinger and in-crowd vs disdain for everything else

2022-06-03 04:03:37 @TaliaRinger know the feeling. eg pointing out for 20 years that generalization outside the training space was an issue, being ignored (or ridiculed), and then suddenly *Bengio* says the magic words “distribution shift”…

2022-06-03 02:04:01 @Plinz @sciam i was about to say i would only make one bet this week, but then Elon posted about Optimus

2022-06-03 02:02:30 @BethCarey12 guestimate based on authors, computation costs etc.

2022-06-03 02:01:09 media friends, i would love to cover this :) dm me? https://t.co/bll0OZ557Y

2022-06-03 00:24:32 “Well then, Musk—do you accept Marcus’ challenge?” asks @AI_TechNews https://t.co/oF5PSzoSVg

2022-06-02 23:04:49 source: https://t.co/wKAnzSbUMC

2022-06-02 23:04:48 https://t.co/gFBv53QTLM

2022-06-02 22:32:34 “He's a grifter, he sells a vision in hopes that he can one day deliver what he's promising, but he doesn't know that... He's just really good at pretending he knows” https://t.co/hyy79FfZdi

2022-06-02 21:39:45 Coming Monday, @sciam, my riff on Don’t Look Up (the AI edition)

2022-06-02 21:31:11 @mraginsky @IntuitMachine @TonyZador @davidpoeppel @cedricboeckx transformational or not (eg minimalist program) that’s the conventional use of generative grammar

2022-06-02 21:29:46 @raphaelmilliere @hardmaru @_lexfridman wrong lex but you get the idea

2022-06-02 21:29:22 @raphaelmilliere @hardmaru weird! and i was blocked by @_lexfridman. some things make no sense

2022-06-02 21:18:59 @bkeithpayne @drbret odds are good though that the (purely observational version of the) study would come out exactly as you joke

2022-06-02 21:12:43 @TheDavidSJ discussed this issue in the article

2022-06-02 21:05:57 @raphaelmilliere @hardmaru

2022-06-02 21:05:13 @mraginsky @IntuitMachine @TonyZador sorry but this is confused

2022-06-02 03:15:02 @david_picard @ZoubinGhahrama1 @Plinz and you really want clear info about confidence bounds, as @ZoubinGhahrama1 likes to stress

2022-06-02 03:14:28 @david_picard @ZoubinGhahrama1 @Plinz 1. in the end that it is hard to build adequate debuggability if you have no idea why your systems makes the specific mistakes that it does2. if you can’t produce rich structured outputs, it’s hard to do rich cross-checking with other systems.

2022-06-02 03:07:38 RT @plevy: « Semantics isn’t taken seriously in most of AI, but it should be. »

2022-06-02 02:46:57 @Abel_TorresM GPT-2

2022-06-02 01:56:21 @pmddomingos cf “the toxic brew of easy access guns, gun culture, and hysteria” which i could accept

2022-06-02 01:55:04 @pmddomingos only if you are standing by hysteria as largest factor.

2022-06-02 01:53:31 @aromeromyt @khademinori @pmddomingos exactly. eg sitting here in BC i weep (literally) at what is happening in my homeland (US) but i don’t feel immediately hysterical that it will affect me here, given culture, history, and firearm distribution.

2022-06-02 01:47:05 @pmddomingos i mentioned canada because they hear largely same press coverage (eg incredibly moving CBC story yesterday)

2022-06-02 01:40:34 @pmddomingos i think you are assuming some kind of linear function, and ignoring other factors like economic inequality

2022-06-02 01:34:32 @pmddomingos @khademinori use canada as your comparison, and things look different

2022-06-02 01:33:35 @pmddomingos i was responding to the word “biggest”

2022-06-02 01:31:48 found someone! already we are crafting!

2022-06-02 01:27:29 @Plinz @dileeplearning @tyrell_turing underway! w guest star!

2022-06-02 01:14:38 @pmddomingos no i didn’t (but scanned just now). i am sure copycat syndrome explains *some* of the variance. the availability of guns surely does too. Canadians see the U.S. news but generally don’t have as many guns and don’t copy…

2022-06-02 01:09:37 @scottiev @bengoertzel and you would tell him to stay the hell off of Twitter. Same as Trump’s people told him. Endlessly.

2022-06-02 01:06:55 who wants to help me co-write some straight talk about semantics? (candidates must have training in linguistics

2022-06-02 01:03:02 @Plinz @dileeplearning @tyrell_turing illustrated the urgent need for a clear discussion about semantics

2022-06-02 01:01:55 @pmddomingos i am sorry, did you give an actual argument for your claim? all i saw was an assertion.-no data-no proposed causal mechanism-no consideration of alternatives

2022-06-02 00:56:28 @scottiev @bengoertzel to stand behind his words?

2022-06-02 00:55:20 @chrmanning @csabaveres @rogerkmoore i think this is a fair criticism and the interesting and unsolved question is what to do about it? how much goes into semantics, pragmatics, cognition etc as well idiom, stored cases of various sorts etc

2022-06-01 16:52:30 @WiringTheBrain read Mark Twain’s comments on the matter in That Awful German Language

2022-06-01 16:05:06 @davidad definitely not trivial.

2022-06-01 15:58:18 It could! That’s what I keep lobbying for. And all this DARPA money going into neurosymbolic could catalyze it! https://t.co/f2eeyqH3zH

2022-06-01 15:52:09 @JhendersonIMB woud be great if there maturation in both fields!

2022-06-01 13:57:10 RT @PMinervini: Spurious Correlations Are All You Need

2022-06-01 13:53:20 Have confirmed that this is a legit offer

2022-06-01 13:52:18 GIF contest! We appear to be at half million dollars.What best conveys something along the lines of, hey Elon, how about you step up to the plate and stand by your words? Or are you chicken?(feel free to ixnay my clichés) Still images from Dall-E, Imagen also acceptable :)

2022-06-01 13:42:14 RT @keithfrankish: In philosophy, if you think the answer is obvious, you haven't understood the question.

2022-06-01 13:31:06 @davidmanheim @wadhwa @aniketvartak @kevin2kelly the essay explains why the bet is re a full package of five.

2022-06-01 13:06:42 RT @wadhwa: @PercyRajani @GaryMarcus @elonmusk Gary, this makes it $400K and we have no response.Getting to $1M will be easy

2022-06-01 13:04:07 RT @AlexUjcik: About Google Imagen: "The problem was not that the system physically couldn’t create an image of something so improbable as…

2022-06-01 07:04:08 Doubt Elon will take the bet, but $350k is a pretty powerful commentary on people’s declining appetite for AI hype. https://t.co/B0523Q1K61

2022-06-01 06:56:58 2023: Field grows up and finally faces compositionality, semantics, and knowledge representation https://t.co/WlJ8hZaoF8

2022-06-01 06:18:33 RT @guyfrom2045: So @GaryMarcus has raised 350k so far to bet Elon that his "AGI by 2029" prediction won't happen, and I'm loving the whole…

2022-06-01 05:47:20 RT @tomg_: Depuis 2015 @elonmusk annonce les voitures autonomes "dans 2 ans". Maintenant il prédit l'intelligence artificielle générale pou…

2022-06-01 04:26:29 @RobKnight__ @MatthiasLalisse @Plinz @hardmaru yes and i discuss significance of that in my essay

2022-06-01 04:25:24 piercing essay, by @mer__edith

2022-06-01 04:24:32 “proposals to “democratize” access to AI research infrastructures amount to calls to subsidize tech giants further by licensinginfrastructure from these firms in ways that allow them to continue defining the terms and conditions of AI and AI research.” https://t.co/7ppBaHTEnB

2022-06-01 04:13:48 hmm. you really should Horse Rides Astronaut at https://t.co/8ir1xKvqt6

2022-06-01 03:59:34 no. shorting magical thinking. https://t.co/sCrcTeJEA2

2022-06-01 03:21:46 @filippie509 @gregeganSF @aniketvartak #5 and counting

2022-06-01 03:17:52 still one of the best shows ever. RIP Michael K Williams https://t.co/LNLFvhoIVk

2022-06-01 03:08:04 @Plinz @khademinori @hardmaru indeed although the interface is language it would be interesting see that dimensions the sliders would control

2022-06-01 03:07:22 @Plinz @khademinori @hardmaru how about shadows? song dynasty dragon made of glass casting a shadow on a rectangular grid? (or start with an opaque dragon)

2022-06-01 03:02:12 @Plinz @khademinori @hardmaru not perfect but still impressive

2022-06-01 03:01:28 @Plinz @khademinori @hardmaru not it’s best work tbh

2022-06-01 01:55:45 @Plinz @hardmaru also “A song dynasty dragon breathing fire while typing on a laptop, seen from above”

2022-06-01 01:55:16 @Plinz @hardmaru hands are weird (lot of human hands in there) but otherwise very good. try some more. “A Song Dynastry dragon, typing on a laptop, shot from behind” or “seen through binoculars”

2022-06-01 01:48:51 @hardmaru @Plinz the thing that impresses me the most is the consistency of perspective/camera angle. would be fun to test the robustness of that, and things like occlusion, noncanonical views, how specific you can be, etc.

2022-06-01 01:39:59 @Plinz @hardmaru and let’s not forget that classic, “how many attempts did you try?”(Luckily @hardmaru gives part of the answer.)

2022-06-01 01:18:10 I’m pretty confident that AI will be mainstream soon (and arguably in some respects it already is)

2022-06-01 01:13:50 @gregeganSF @aniketvartak you are at least the fourth person to express concerns.

2022-06-01 01:11:42 Two things I did not anticipate on my bingo card for today- DALL-E speaking in tongues- A quarter million dollars added to on my AGI bet. https://t.co/ql1l4fqast

2022-06-01 01:09:22 Plus $50k from @aniketvartak

2022-06-01 01:05:22 I double-checked

2022-05-31 05:31:43 RT @bobehayes: Has #AI found a new Foundation? @GaryMarcus#AGI needs:Cognitive modelReal world knowledgeIdentify entities and their re…

2022-05-31 04:28:45 p.s. If you want to see “Five things @ElonMusk Should Know About AGI“ first, subscribe to https://t.co/8ir1xKvqt6 :) https://t.co/ccfT0KcOaQ

2022-05-31 04:22:50 Where two people a minute have voted to say that they want to hear this, I’ve started writing. Almost done the first draft. I will post tomorrow :) https://t.co/0IFOK2yZCV https://t.co/RtnVuGDw7F

2022-05-31 03:54:22 artwork for the top? https://t.co/z3vqr92hxJ

2022-05-31 03:30:56 Potential topic for next Substack: Four things Elon Musk should know about AGIShould I write it?

2022-05-31 00:33:31 RT @eucopresident: Tonight #EUCO agreed a sixth package of sanctions.It will allow a ban on oil imports from #Russia. The sanctions wil…

2022-05-30 23:35:37 @TechRonic9876 @astralcodexten i will stick with @maayanvisuals :)

2022-05-30 23:32:06 @stevenbjohnson

2022-05-30 23:20:39 excellent analysis from @slatestarcodex, consistent with what i speculated in https://t.co/ZacV6k297G

2022-05-30 23:17:26 “DALL-E is clearly capable of incredible work. That having been said, I mostly couldn’t access it….trying to get it to do anything at all unusual degrades the style until you end up picking the best of a bad lot and being grateful for it” -@astralcodexten https://t.co/aHTkG94y5p

2022-05-30 22:53:53 Cuddling, self-organizing geese (video). https://t.co/VkpnQkSrg9

2022-05-30 22:43:45 Wake up! You will be late for training! https://t.co/QqCl3O4cCW

2022-05-30 22:14:06 @elonmusk @jack https://t.co/p2gicOktHe

2022-05-30 21:57:32 @jcrg_137 @elonmusk ain’t nobody here said “impossible”

2022-05-30 21:32:39 @xzistor @aoargunsah I don’t think I can get him the mansion he deserves. But I agree that half the fun in the bet would be coming up with enforceable definitions.

2022-05-30 21:23:37 @MelMitchell1 @elonmusk, I’m down for https://t.co/gYO3sxWe7s if you are.

2022-05-30 20:09:43 @nearlydaniel Criticizing *is* helping. Zero chance we get there if we don’t take realistic look at current limits and understand the obstacles.

2022-05-30 19:39:54 How much are you willing to bet? https://t.co/9pxVgyMMqm

2022-05-30 19:02:27 RT @jaaanaru: In our new paper, we argue that deep learning is not (yet) ready to conquer Theory of Mind. The agents discover shortcuts i…

2022-05-30 19:02:00 RT @jaaanaru: If you're into figuring out how the brain works then you'll love this new paper by @mattlark "the humble point neuron, devo…

2022-05-30 16:40:28 @FelixHill84 @mpshanahan @sir_deenicus @aniketvartak parrots have plenty of intelligence, just not a huge amount of language comprehension (though Irene Pepperberg has shown they can develop some, w extensive training).building parrot-level cognition would certainly help w deep comprehension. LLMs, perhaps not so much.

2022-05-30 14:59:06 I saw the best minds of my generation … spending hours upon hourshaving fun with DALL-Eand taking things “step by step”instead of figuring how we might getto artificial intelligencethat we could actually trust[with apologies to Alan Ginsberg] https://t.co/XcWyMljTft

2022-05-30 14:55:33 @sir_deenicus @aniketvartak @FelixHill84 @emilymbender

2022-05-30 14:46:19 @kanishkamisra @FelixHill84 sometimes, but not always. that’s what the whole 1990s past tense debate was about, and in a certain sense what my Horse Rides Astronaut essay this week (https://t.co/8ir1xKvqt6) was about.And it’s why distribution shift is a central challenge

2022-05-30 14:30:07 @spacepanty @FelixHill84 have been hearing the same fantasy thinking for 30 years

2022-05-30 14:21:14 @rolisz @HarshitTaneja18 there is a long continuum in between

2022-05-30 14:19:52 @RWerpachowski @FelixHill84 hybrid AI does not exclude this possibility, but I remain concerned about the fundamental lack of reliability, comprehension, and explicit representation.

2022-05-30 14:14:58 @rolisz @HarshitTaneja18 at collaging, yes

2022-05-30 14:09:15 @FelixHill84 right. and for sure the details are open. but LLM’s do not exhibit that behavior, so they aren’t the right solution.

2022-05-30 13:38:38 @FelixHill84 we know something from cognitive psychology (not neuroscience) about how flexible and general they are

2022-05-30 13:28:02 @HarshitTaneja18 650M labeled image pairs is not that fast, and it is still missing a lot of language understanding and a good bit of common sense.

2022-05-30 13:18:14 @justMathana current AI pretty much is a stochastic parrot (see all@my recent Substack writing, @emilymbender etc)the internet is a lot sloppier than college but much more comprehensive. a true AI would learn an immense amount from it.

2022-05-30 13:13:08 @FelixHill84 not so sure. it will still lacka. proper knowledge structures for representing abstraction (independently of frequency and similarity)b. real representations of the dynamically-changing world

2022-05-30 13:10:07 @AnoopRKulkarni @GoogleAI @seanmcarroll @seanmcarroll this refers to my trilogy at https://t.co/8ir1xKvqt6 and all the recent models. i would be happy to discuss if you like

2022-05-30 13:08:14 @FelixHill84 no doubt but it still wouldn’t get much out of college, and its grasp on common sense remains both tenuous and superficial.

2022-05-30 13:04:18 A human can learn a lot from college, a parrot, not so much. https://t.co/j5xgMO4yEP

2022-05-30 12:31:34 On prompt engineering: https://t.co/lta32VnXU4

2022-05-30 12:00:15 RT @m_mezini: I enjoyed reading #longread

2022-05-30 03:00:14 from 2015, still cracks me up every time https://t.co/ALEu4lJrIh

2022-05-30 02:18:29 RT @hangingnoodles: Is unbridled AI enthusiasm gonna saddle us with astronomical horseshit?re "extract meanings" from wordsThe meanin…

2022-05-30 02:14:40 @xriskology you did however interpolate the word “Holocaust” into a quote that you elided in order to give it that air. and you used it to misrepresent that person’s position, and make it specifically sound like the person minimized the world Holocaust.but, true. you didn’t use word Nazi.

2022-05-29 22:48:16 @morungos of course not

2022-05-29 20:51:22 The point that deep learning ≠ deep understanding was also central to https://t.co/Pt7HZbLIv5. On this @yudapearl and I are in 100% agreement. https://t.co/8MhzS3vY3k

2022-05-29 20:48:00 @HappyAar dishwashers and pool cleaners are fine as far as they go, but nothing like near-AGI nor do they relieve of most human work, handy as they are

2022-05-29 20:32:52 Also, essay question, everyone is welcome: can “near-AGI” that is applied to most human work actually *be* safe? Or is it a disaster waiting to happen? https://t.co/WU6NkltX4i

2022-05-29 20:29:23 I won’t say this will *never* happen but it most certainly won’t happen in this decade. We are nowhere near “safe near-AGI” that can handle most human work. https://t.co/hGBdXUWPqF

2022-05-29 19:48:21 In this sense, a lot of ML is like how people used to misuse statistic packages, applying the analyses, but without understanding the underlying assumptions. https://t.co/g0wfEHFlMI

2022-05-29 19:45:57 @IntuitMachine it’s not the just science, but scientists who say stuff like “put up or shut up” to others who might have good ideas, coming from other disciplines.

2022-05-29 19:10:09 @KarlaParussel hmm maybe time for an essay: AI: 75 Years of Mistakes

2022-05-29 19:08:13 RT @milangritta: Some responses to Gary's valid criticism look like when a religious person is told that God doesn't exist. #AI #AGI #Artif…

2022-05-29 18:31:39 @bbenzon mostly: no

2022-05-29 18:18:50 yes! and 3. the education and culture are insufficiently interdisciplinary. https://t.co/sYR9139T0B

2022-05-29 17:35:53 @Oireniar @ElijahYilma @pmddomingos good recent example is MRKL from @AI21Labs

2022-05-29 17:34:44 @tdietterich @rogerkmoore and then we wake up and the frog has been fully boiled

2022-05-29 17:19:54 @tdietterich @rogerkmoore people were much clearer (never perfect) about these distinctions in the ‘90s… i have never seen the current level of confusion, though

2022-05-29 17:18:55 @ElijahYilma @pmddomingos yes (though other forms of hybrid AI might be developed too)

2022-05-29 16:19:01 RT @ks_StoryScout: „A new breed of #AI systems, with a firmer innate basis, closer to the Kantian model, but out of step with current Zeitg…

2022-05-29 16:10:58 thanks! perhaps the only thing that is truly exponential these days is hype? https://t.co/4CLB62P7bA

2022-05-29 06:05:30 RT @matthewcobb: If you have seen pictures of an astronaut riding a horse along with claims about a great AI breakthrough, read this excell…

2022-05-29 05:45:16 RT @bimedotcom: Horse rides astronaut - why when it comes to #AI #hype, you still can't believe what you read https://t.co/45xR55b7zo@Gary…

2022-05-29 05:30:58 @pmddomingos hybrid AI: many representations

2022-05-29 05:29:46 @Tweetermeyer

2022-05-29 03:57:35 @timnitGebru that’s very much like what Ernest Davis and I argued here: https://t.co/qwExULkwCO

2022-05-29 03:54:52 RT @thomas_macaulay: @GaryMarcus @ylecun you need to step up to the plate

2022-05-29 02:06:34 RT @Grady_Booch: @soumithchintala Language has powerful loving vocabulary. This gives Language Models the ability to love.

2022-05-29 01:29:34 RT @weed10_weed: @Tweetermeyer https://t.co/VdUgVDsqlg

2022-05-29 01:27:28 RT @langufacture: in fact,,,it's the kind of argument you'd expect a language model to make

2022-05-29 01:16:36 this argument is not sound: “Language has powerful reasoning vocabulary.This gives Language Models the ability to reason” First premise is true, but leap to second is not logically entailed, given that language models have known trouble w compositionality.@guyvdb @luislamb https://t.co/L0CTMxzdxW

2022-05-29 01:12:18 @soumithchintala how would you define reason? at best this sounds like a hypothesis not a proven fact. and see eg https://t.co/r3Bs1BKk0Y

2022-05-28 21:53:32 @Plinz https://t.co/xuAJC1YjdM

2022-05-28 20:32:52 RT @thomas_macaulay: @GaryMarcus @ylecun are you up for the challenge?

2022-05-28 19:27:54 @AR_Kareem_ @LucaAmb @GoogleAI discussed more here https://t.co/2MQJtWdNhN

2022-05-28 19:26:15 @bneyshabur addressed here: https://t.co/2MQJtWuQjN

2022-05-28 19:25:52 @yanndubs @DavidSKrueger @stanislavfort @freedguy @joe_shabadoo indeed. but see: https://t.co/2MQJtWuQjN

2022-05-28 19:25:25 @LucaAmb @GoogleAI and it is the wrong explanation here, see: https://t.co/2MQJtWuQjN

2022-05-28 18:37:11 now analyzed in further detail here: https://t.co/ZacV6jKxJ6

2022-05-28 18:19:20 RT @AdnannKurt: “what I learned as a cognitive psychologist is that any given piece of data, on its own, can probably be explained in more…

2022-05-28 18:18:37 @DotCSV @GoogleAI ps next essay will end with a very different line, I promise :)

2022-05-27 16:31:19 And - A. Why won’t @Tesla be more forthcoming about the things?- B. What will it take to pass laws that make public disclosure and tracking of autopilot involvement mandatory? https://t.co/zbIIpuTiHe

2022-05-27 16:28:53 @TShevlane @gershbrain @PTetlock

2022-05-27 15:28:45 @hellosemy omg. link?

2022-05-27 15:17:24 Coming soon https://t.co/WDFhQ6PRJO

2022-05-27 15:12:07 @BogdanIonutCir2 @sir_deenicus @DonaldClark @IntuitMachine this is a lousy question, because the term deep learning is too poorly defined. see eg discussion Bengio and I had here: https://t.co/U8GBFplk4ncc @matthewjbar

2022-05-27 15:02:07 @sir_deenicus @DonaldClark @IntuitMachine what if the only thing the survives is practice of using lots of data?

2022-05-27 15:00:54 @raphaelmilliere @Chitwan_Saharia @Chitwan_Saharia please try (&

2022-05-27 14:57:45 Nobody has studied the behind-the-scenes tactics of Elon Musk more carefully than @Tweetermeyer: https://t.co/74qh8GTWz3

2022-05-27 14:57:32 @raphaelmilliere what were the prompts for the trees? was it “literal parse trees”?would like to see @Chitwan_Saharia run the same on Imagen to compare the text.

2022-05-27 14:53:36 @qhardy @Abebab

2022-05-27 14:51:16 Bravo, @Abebab nails the critique others have been trying to formulate, but without any of the unfair ad hominem.Looking forward to seeing how she develops the argument. https://t.co/Q0iMGE1k3S

2022-05-27 14:44:19 @MadamePratolung @francoisfleuret @doRadiology @geoffreyhinton exactly my experience when someone used GPT two weeks ago to ape me: parts got my style, but only of what i said before, not what was novel, and it lost the coherence. might be fine for an undergrad term paper. but i can’t see any use for it in the essays i would actually write.

2022-05-27 13:54:00 or if you prefer graphs with your dose of reality: https://t.co/BAO0981gJs

2022-05-27 13:51:48 Guns don’t kill people. Stupid policies kill people. First part of the above is false

2022-05-27 13:48:46 to @MadamePratolung’s point, even tiny tasks like proctoring exams can be fraught when you try to replace humans: https://t.co/L0biwwJ8LP https://t.co/4cNnOxCIsI

2022-05-27 13:01:50 RT @manes: @AndreaPossenti @GaryMarcus Bang-on encapsulation by Gary Marcus, as in previous writing about other instantiations of machine l…

2022-05-27 12:54:08 RT @AndreaPossenti: Very interesting read on SOTA language models and human cognition. Is abstraction ever going to be achievable? by ⁦@Gar…

2022-05-27 06:11:54 @yanndubs @DavidSKrueger @stanislavfort @freedguy @joe_shabadoo it will be here, first draft complete: https://t.co/QLcNbSyN7k

2022-05-27 06:10:39 Subscribers to https://t.co/8ir1xKvqt6 (free, as ever) will be first to see “Horses, astronauts, and the bumpy road to general AI”

2022-05-27 04:11:08 Now there’s a benchmark! https://t.co/cr48F3wM1k

2022-05-27 04:08:47 @AngeloDalli why is it so hard for so many people to see that?

2022-05-27 04:07:46 @danbri what is important about it?

2022-05-27 00:49:54 @yanndubs @DavidSKrueger @stanislavfort @freedguy @joe_shabadoo i think this is missing the point

2022-05-26 21:53:03 At this point, the machine learning community can be divided into two groups:- those that crack jokes- those that recognize reality. https://t.co/WAA16PcDl1

2022-05-26 21:40:02 RT @DeWeeseLab: Like all deep generative models today, Google's #imagen gives beautiful, convincing output, but can't understand + follow s…

2022-05-26 20:28:53 @santoroAI @ylecun btw i agree with what you say in this tweet

2022-05-26 20:27:36 @santoroAI @ylecun @HonghuaZhang2 @guyvdb in calling the debate :”vacuous”, he’s *not* putting his cards on the table (eg, symbols not required for reasoning), he’s kicking over the table, and telling other people to not even bother to place their bets. that’s what I am objecting to, not his particular views.

2022-05-26 20:24:56 RT @doRadiology: @GaryMarcus This year marks the 5th anniversary of "We should stop training radiologists now. It's completely obvious th…

2022-05-26 18:57:34 @santoroAI @ylecun I think that’s an empirical question, and a substantive one, and an important one. And that this @HonghuaZhang2/ @guyvdb study, for example, suggest/ that reasoning without hybrids is not likely to be effective. https://t.co/JiunWqe3cg

2022-05-26 18:20:18 RT @eadinella: @GaryMarcus Anyone with access to codex beyond cherry-picked examples is not concerned

2022-05-26 16:32:03 @patrickmesana @ylecun if you use a symbolic database to track a cognitive model is that “just SWE”?

2022-05-26 16:10:19 @maartengm @ylecun i wrote a whole book in 2001 about my definitions and they have not changed.

2022-05-26 16:07:07 @raphaelmilliere @mpshanahan @ylecun I explicitly asked LeCun for clarification, repeatedly, after he made that remark, and he did not respond.

2022-05-26 16:04:50 @maartengm @ylecun all of which were the direct subject of my monograph The Algebraic Mind

2022-05-26 16:04:02 @maartengm @ylecun no he is not naive on that. but eg these are real issues: https://t.co/k0UmkjxXIT

2022-05-26 15:59:45 @willknight

2022-05-26 15:59:20 @raphaelmilliere @mpshanahan @ylecun fine, but that doesn’t mean the debate is “vacuous”, it means that there are important substantive issues that need to be clarified.I don’t mind if people have different views on these issues

2022-05-26 15:56:42 Software developers, Wired says you are worried. True or False? “Both Codex and Copilot have stirred up some anxiety among developers, who fear they could be automated out of a job.”https://t.co/f9cxPVoBG3

2022-05-26 15:37:36 @raphaelmilliere @mpshanahan @ylecun our differences are mostly that I see the following three things as essential- research on how to build hybrids - greater attention to defining and implementing a broad range of innate priors-explicit cognitive models &

2022-05-26 15:25:21 @raphaelmilliere @mpshanahan @ylecun i am arguing for a strong interpretation, though open to listening to alternatives.

2022-05-26 15:23:21 @pkghosh99 @Chitwan_Saharia @raphaelmilliere @TristanThrush misinformation, unfortunately

2022-05-26 15:16:35 @raphaelmilliere @mpshanahan @ylecun ps i wrote a whole book defining these specific terms and their relevance and application

2022-05-26 15:15:16 @raphaelmilliere @mpshanahan @ylecun i don’t see the problem with the terms or commitments here, but feel free to elaborate

2022-05-26 14:43:28 @mpshanahan @ylecun i don’t actually have a strong view about differentiability. i remain fully committed to the value of symbolic operations over variables, and to structured representations, but open (w mild but not absolute pessimism) to how deeply differentiation can go.

2022-05-26 14:40:33 @TacoCohen @ylecun 2/ What we need theory and practice about how to integrate the two traditions intellectual respect and humility, such that people on either side respect the strengths and weakness of the two approaches, and places for possible collaboration

2022-05-26 14:38:26 @TacoCohen @ylecun 1/ Is a classic expert system hybrid? no. Is a pure multilayer perception hybrid? nois searching a symbolic tree and applying a neural net to each candidate a hybrid? yesthe fact that a beanbag chair is a non canonical chair doesn’t mean we can’t talk about chairs.

2022-05-26 14:21:12 @LivingOnStar9 bs. see https://t.co/ufbxCXeFOL

2022-05-26 14:13:12 first it was”completely obvious” that all the radiologists were going to be replaced … https://t.co/PumKjkfLCS

2022-05-26 14:09:51 @fchollet https://t.co/ufbxCXeFOL

2022-05-26 13:43:17 @ylecun https://t.co/Hpb1ymGx6h

2022-05-26 13:33:39 @KarlaParussel @JaumeTeixi @ShaneLegg everyone in AI should that paper. More than once.

2022-05-26 02:06:31 @hoskingc @Grady_Booch @TheNickBlizzard that tells us almost nothing

2022-05-26 01:50:06 @hoskingc @Grady_Booch @TheNickBlizzard it’s like saying that humans are like sea slugs because we both have DNA and share some genes and proteins. all true but not terribly informative as to causal mechanisms.

2022-05-26 00:39:41 @pmddomingos @ethanCaballero can hook you up

2022-05-26 00:39:04 with this one trick, GPT-3 can solve decades-old riddles of reasoning and commonsense!nearly 2,000 people have already liked it!try it now, while supplies last!“Let’s think step by step”! https://t.co/axJRJBq7nT

2022-05-26 00:24:30 @kristintynski @NickRMorgan @prokraustinator that’s the illusion of large data

2022-05-26 00:03:34 RT @Grady_Booch: “GPT-3 is trained to predict the next word on a large dataset of Internet text”Thus proving the point that I and many ot…

2022-05-26 00:00:11 @MaartenBosma but you as human can reread and correct your mistake, System 2 winning out over System 1

2022-05-25 23:57:15 @pmddomingos touché

2022-05-25 23:55:04 @pmddomingos a lot of people are rightly (IMHO) sensitive around the name, especially given the response of some when the issue was raised.

2022-05-25 23:50:03 @max_nlp @raphaelmilliere @TristanThrush That’s seems like a question for @TristanThrush

2022-05-25 23:46:43 @kristintynski i will DM you one that I can’t post for TOS reasons

2022-05-25 23:46:14 should have mentioned that @ErnestSDavis ran this test, based on one of our examples in 2014.

2022-05-25 23:45:21 @_jasonwei example from @ErnestSDavis, drawn from one of our 2014 articles: https://t.co/yeyDrNXDMu

2022-05-25 23:43:12 @kristintynski @prokraustinator then hopefully you will take my last tweet w a sense of humor :)

2022-05-25 23:42:09 “Let’s think step by step.” Sometimes it keeps GPT on the rails.Sometimes it doesn’t: https://t.co/hYUIpQ1c4D https://t.co/JwwCWiqWDh

2022-05-25 23:38:49 The real gallows humor joke here is that researchers used to ask each other challenging questions. https://t.co/VMMkYV3FNp

2022-05-25 23:34:19 @kristintynski @prokraustinator stay tuned :)

2022-05-25 23:32:24 @kristintynski i don’t doubt this, but the problem lies in getting the right groove. if you want humans-in-the-loop, it might work to some degree

2022-05-25 23:30:01 @Abebab which is not to take away from the incredibly difficult challenges around formulating and implementing policy. but it’s not either/or in my view

2022-05-25 14:51:18 @cgmooreauthor and see my conversation with Penn, https://t.co/mrSq5vydxY (and next episode)

2022-05-25 14:48:58 @luislamb @AvilaGarcez @pascalhitzler @Melleo54Sis @ykilcher @YouTube Thanks, @MLStreetTalk, for hosting, long before it was fashionable

2022-05-25 14:48:13 RT @luislamb: @GaryMarcus This conversation explains the need for Neurosymbolic AI. @GaryMarcus @AvilaGarcez @pascalhitzler @Melleo54Sis an…

2022-05-25 14:38:48 @lathropa @XiaohuaZhai Matches @ErnestSDavis ‘s 1-in-3

2022-05-25 14:37:12 @grsimari I actually asked him for clarification multiple times re his earlier claims and he refused to engage.

2022-05-25 14:34:38 Marcus says “These are like parlour tricks... They’re able to fool unsophisticated humans who aren’t trained to understand these things. But that doesn’t mean that they’re actually anywhere near AGI”…DeepMind wasn’t available for comment.– @sparkes https://t.co/EQsjZ1y53R

2022-05-25 14:26:31 The problem with large language models, in a single poem https://t.co/5rZXppiNuy

2022-05-25 14:24:45 @hangingnoodles @hangingnoodles, that was my next Substack! :)

2022-05-25 14:21:10 @XiaohuaZhai in fairness, Ernie ran a third, while I was typing, and your system got that right. but I am definitely not smelling “deep language understanding”we’re happy to try more if you want to give us access to your largest model https://t.co/Ebelx4PXem

2022-05-25 14:17:38 @punkstrategy @mark_riedl @ShaneLegg @bengoertzel would love your take, both on the history and the meaning of the term AGI

2022-05-25 14:17:09 Success on a single prompt does not “deep understanding of language make”. It’s great, though, that people are sending me links for models to try.I often pass them to @ErnestSDavis

2022-05-25 14:07:17 RT @DorotheaBaur: Essentially blank slate systems like #GPT3 should motivate us to explore "their very antithesis, which would be Kantian s…

2022-05-25 13:39:38 I believe @ShaneLegg (also at DeepMind) coined the term AGI. Personally, I use it as a shorthand for any intelligence (there might be many) that is flexible and general, with resourcefulness and reliability comparable to (or beyond) human intelligence. https://t.co/LldjBmF1iN

2022-05-25 13:29:44 “the fact that DeepMind called Gato a “generalist” might have made it a victim of the AI sector’s excessive hype around AGI.” -@melissahei https://t.co/CRmKakqwQQ

2022-05-25 13:11:26 @pragmaticml @GoogleAI A few years ago people shared data

2022-05-25 13:09:25 @ykilcher @dionhaefner see also eg https://t.co/oS8uXbIFF3

2022-05-25 13:01:27@ykilcher @dionhaefner my core point is that neural nets often succeed for the wrong reasons, glomming onto accidental contingencies, and making them appear to reason (causally, or otherwise) more effectively and broadly than they actually can. (“System 1” does some of this in humans, to be sure.)

2022-05-25 12:25:07sorry, but no. you don’t get to dismiss an entire field—hybrid, neurosymbolic models that seek to integrate reasoning with neural networks—as “vacuous” and then two weeks later try to take credit for its core mission. https://t.co/qn4SBzoNDd

2022-05-25 12:17:24RT @AvilaGarcez: So, neural nets (NN) can handle all of propositional logic and a fragment of first-order logic (FOL) reasoning, but not fu…

2022-05-25 12:10:39@LucaAmb @GoogleAI not a defense nor explanation

2022-05-25 12:02:47RT @danbri: @_jasonwei @suzatweet I love this stuff! It has a weird elegance, is fun to explore &

2022-05-25 11:59:57RT @_dieuwke_: #NLProc What do you think are the most convincing examples of generalisation failures in recent NLP models -- i.e. cases whe…

2022-05-25 03:33:39um, how about gun control? https://t.co/IS483QFhgx

2022-05-25 03:21:52@_jasonwei reading this, my first thought was, “is it robust”? my second was to look at the paper. answer is, “somewhat, but it depends”: https://t.co/Kb0OIRoW7S

2022-05-25 03:16:38RT @yoavgo: what do you know, like the DALLE2 paper, the Imagen paper also doesnt say anything about the training data source and compositi…

2022-05-25 03:16:18@pmddomingos and, like, I dunno, the whole brain was like, um, … linear algebra?

2022-05-25 02:54:00@pmddomingos What if, like, neural networks were… slightly conscious?

2022-05-25 02:42:51if only we’d known https://t.co/nGpMpPcOuT

2022-05-25 01:50:46@NickRMorgan Imagen doesn’t understand the *mapping* between syntax and semantics

2022-05-25 01:49:57@NickRMorgan i haven’t had any access to it whatsoever

2022-05-25 01:45:45@NickRMorgan it doesn’t, not fully. riding is a thing that an entity does relative to another entity, and it doesn’t get which is which

2022-05-25 01:37:13@andrey_kurenkov @Miles_Brundage @GoogleAI @ErnestSDavis we already have asked for the raw outputs, in fact

2022-05-25 01:36:53@andrey_kurenkov @Miles_Brundage @GoogleAI thanks. i really wish they would have supplied the actual outputs, at least from their own model, rather than just saying that relatively speaking people preferred theirs to DALl-E. (which wasn’t true for our stimuli, i’d you look at the figure).

2022-05-25 01:33:26@andrey_kurenkov @Miles_Brundage @GoogleAI interesting. who is in charge? can @ErnestSDavis and i add some more?

2022-05-25 01:26:22@Miles_Brundage @GoogleAI and it’s not my bar, in this case, but Fodor and Pylyshyn, 1988, from a third of century ago

2022-05-25 01:24:44@Miles_Brundage @GoogleAI would like to try it, eg w novel phrases (not just eg “hello world”), but yes it seems better on text. any idea as to why?

2022-05-25 01:15:07https://t.co/RggGKwLklB

2022-05-25 01:13:38 @GoogleAI, “a deep level understanding”? Seriously?!Your system can’t distinguish “a horse riding an astronaut” from “an astronaut riding a horse”. https://t.co/7XlaQr3rgb https://t.co/fyrDpPv6VL

2022-05-25 01:01:20More will follow, but yes Andrew Ng’s turnaround to hybrid neurosymbolic models was dramatic. https://t.co/pO1byzAxXc

2022-05-25 00:56:55RT @s_batzoglou: It is becoming clear that we are all realizing the limits of deep learning, as the achievements pile up. Logical and combi…

2022-05-24 23:31:51and each of those guys thinks he is so clever… (but see https://t.co/z5w4Czkjij) https://t.co/KYhlW7ocuC

2022-05-24 23:30:57@jordiae @giffmana @ethanCaballero https://t.co/GviHPby1fo

2022-05-24 23:23:36RT @GaryMarcus: @timnitGebru My view: AGI has potential for good, if developed properly (eg in medical &

2022-05-24 23:22:35@timnitGebru My view: AGI has potential for good, if developed properly (eg in medical &

2022-05-24 23:11:06honestly, if there is anything Game Over in AI lately, it is this: https://t.co/JQQQ1VrhwB

2022-05-24 21:18:45Dearest deep learning fan, may I ask, did you read the paper carefully, before you wrote your tweet?Here’s your wall for you, page 33: compositionality with noncanonical scenarios, same place it’s been since Fodor and Pylyshyn, 1988. https://t.co/TQyy60o55H https://t.co/zd0KkH6RDQ

2022-05-24 21:02:55RT @gshu: No, it is not too soon to talk about gun control.

2022-05-24 21:02:41@MikeNashTech

2022-05-24 20:51:59@HappyAar by scaling-über-alles i mean “solving AI by simply adding more parameters and data to current AI systems without fundamentally changing the architecture”

2022-05-24 19:27:24Fantastic study and bad news for the scaling-über-alles crowd. https://t.co/r3Bs1BsICo

2022-05-24 19:26:02@AlexGDimakis but they don’t

2022-05-24 19:24:06@pmddomingos or maybe it’s a wild bull in a china shop, no cart

2022-05-24 19:20:16@CarolinKemper yes. and it is a terrifying prospect!

2022-05-24 19:14:31this made my day, ⁦@HochreiterSepp⁩ https://t.co/uzMVBCVzLV

2022-05-24 18:48:54First this year to join the hybrid AI revolution: Andrew NgSecond: Sepp Hochreiter, LSTM pioneerMany more will follow https://t.co/uzZSe3GVDF

2022-05-24 18:30:04@HonghuaZhang2 @LiLiunian @TaoMeng10 @kaiwei_chang @guyvdb fantastic work!

2022-05-24 18:28:42Can we start giving out Razzie Awards for Giant Corporate AI teams that flout scientific conventions, emphasizing hype at the expense of science?Google Brain is currently in the lead for 2022, per the below. https://t.co/3UuVU22oaK

2022-05-24 18:25:42@MadamePratolung @sir_deenicus @Ted_Underwood ha I just made that same point in an oped i am writing :)

2022-05-24 17:31:25@sir_deenicus @Ted_Underwood @MadamePratolung Agreed. I’ve been starting to notice this, toono public access to modelsno disclosure of actual tests runno disclosure of resultsno denominators (!) for signature resultssketchy discussion of limitationscherry-picking@ZoubinGhahrama1

2022-05-24 17:26:25@sir_deenicus @Ted_Underwood @MadamePratolung &

2022-05-24 16:09:29@onurc @elonmusk only added, what, $1.5T in market cap?

2022-05-24 15:42:30“Barraza said Tesla's Fremont factory was home to "rampant sexual harassment" and the company operated like a "frat house,"”At what point will the Board consider replacing the CEO? https://t.co/TvyFDdwIWh

2022-05-24 14:20:23@emilymbender OMG where?

2022-05-24 14:08:38RT @raphaelmilliere: For further discussion on this topic, join the upcoming workshop on compositionality and AI I'm organizing with @GaryM…

2022-05-24 14:08:01One never knows when @plinz is kidding, but we should all remember that DALL-E 2 doesn’t actually understand concepts like mass, density, or weight. https://t.co/ZReubhFE4e

2022-05-24 14:05:52@cyberandy tweeted about this last night: if you look at the few tests they did of *noncanonical* low frequency events, like horses riding astronauts, it would appear that same old problems remain.they tested our recent arXiv stimuli, but compared w Dall-E and didn’t show actual output.

2022-05-24 13:58:27Thoughts, @ylecun? I think Valiant is correct. https://t.co/orJd2sFHEi

2022-03-19 14:09:54 RT @Article21_Ind: *#Fellowship Alert* Today, we open the call for applications for '#Article21 Research Fellowship' on two subjects: - Dat…

2022-10-29 01:03:47 @albertwebson @_jasonwei @tallinzen skeptical of 1, but keen to see the argumentyou may have seen my new work w @EvelinaLeivada and @ElliotMurphy91, different architecture but perhaps relevant:https://t.co/UWUg6kQ3Df

2022-10-29 00:46:43 @albertwebson @_jasonwei @tallinzen I don’t think all forms of intelligence need to do arithmetic. but these kinds of errors give some insight in a semi-controlled way into a deeper, not yet solved challenge of translating language into models of the world. curious to see the paper when you can share

2022-10-28 21:32:45 RT @egrefen: I'm biased, obviously, but I think this is *such* an important result. Implicature and pragmatics are essential, ubiquitous, a…

2022-10-28 17:50:52 Just rehearsed Noam Chomsky and @jeremyakahn Nov 4, livestreamed at https://t.co/vQcEjAQCtg https://t.co/k93hY3EmVu

2022-10-28 14:18:18 indeed, slashing moderation just as synthetic content becomes widespread does seem like a disaster-in-the-making https://t.co/gEMMehatm7

2022-10-28 10:49:43 AGI, my eye. https://t.co/EmVNJRUzpz

2022-10-28 10:47:51 @L_andreae @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab agree w @L_andreae, eg a classic paper by Thelen showed that newborn humans will step, and that is really an issue of weight etc rather than lack of motor program per se: https://t.co/09LnIrfIqf

2022-10-28 10:39:40 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab (i pasted the image above/here it is) https://t.co/JjMRtntwN9

2022-10-28 10:29:06 @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab presumably both

2022-10-28 10:21:19 @L_andreae @WiringTheBrain my books https://t.co/Pt7HZc3RJd, Kluge, and The Birth of The Mind

2022-10-28 10:13:59 RT @Abebab: the most insightful knowledge about a field comes from experts in the field that hold a healthy skeptical view on the topic and…

2022-10-28 10:08:07 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab beautiful new review, by @neuroluci. how do you even think about this stuff if you are a staunch anti-nativist? https://t.co/RvacUhsspE

2022-10-28 10:01:09 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab hmm, figure 2 from Verhage et al certainly speaks to those questions (# and pattern) to some degree

2022-10-28 09:56:27 @AthenaAI2 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it goes to showing how little we understand the brain &

2022-10-28 09:50:19 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab agreed! and yet you wouldn’t know that from reading a lot of philosophers and AI researchers these days. &

2022-10-28 09:39:10 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it seems to me then that the reasonable conclusion is same as what I wrote in my lay summary in 2004: brain development uses a mix of activity-dependent (some not all intrinsically guided) &

2022-10-28 09:13:25 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab not mentioned but also seemingly relevant is Verhage et al (2000), esp figure 2 (“Correct assembly of the brain the absence of neurotransmitter secretion”) below. Anyone in this thread have thoughts on it? https://t.co/6O4OlOjoQZ

2022-10-27 18:46:45 @fhuszar https://t.co/UWUg6kxUp7

2022-10-27 18:20:33 if i had a dollar for every time someone cherry-picked DALL-E 2 art, I coulda outbid Elon for Twitter https://t.co/kHe92VtnFB

2022-10-27 15:49:18 @ylecun @ProfNoahGian so you have never done your own ground truth study and are relying on those outside orgs?

2022-10-27 15:37:48 @ylecun @ProfNoahGian our messages crossed

2022-10-27 15:31:15 @ylecun @ProfNoahGian that doesn’t answer @ProfNoahGianyou need a (large) sample that is carefully human reviewed ground truth. in that ground truth, two questions:1. what % would be detected by the AI?2.what % of what is missed by the AI is flagged by humans?you are addressing #2, not #1

2022-10-27 14:58:35 As ever, @KBAndersen has his finger on the pulse. (Answer to his question in the comments, in case you need it.) https://t.co/gqwcRACOmx

2022-10-24 02:35:23 there are two kinds of people in this world.those brave enough to speak truth to powerand those who are not.

2022-10-24 02:31:32 @primalpoly @AmandaAskell @AlecMacGillis @JoshuaChaffin i agree. I would put a lot more trust in a single anecdote involving tricking poor people into flying to Martha’s Vineyard for nonexistent jobs.

2022-10-24 01:47:36 @pmddomingos “And if there's war between the sexesThen there'll be no people left”

2022-10-24 01:39:40 @filippie509 @KordingLab @TonyZador

2022-10-24 01:21:55 @filippie509 it also doesn’t cite my endless arguments for years about looking to cognitive science, or even my paper with marblestone in Science…

2022-10-24 00:37:11 @voxbec @davidchalmers42 @mbusigin not disputing this but many people would take these to be influences on cognition. as i said the list is odd, from how people typically use the term, not necessarily wrong. in general people often distinguish between perception (eg olfaction, low level vision) and cognition.

2022-10-23 23:56:44 RT @AmandaAskell: If you don't have time to provide any evidence that someone is a eugenicist (beyond alluding to articles that also don't…

2022-10-23 22:55:22 @GaneshNatesh @ceobillionaire there has been zero budget thus far, zero charge. all participants (including us) have donated their time.

2022-10-23 22:10:34 . @ceobillionaire and I are pleased to announce that we will be resuming our AI debate series, December 23. Mark your calendar, for some exciting conversations, and reply below to suggest participants. (We already have our own ideas, but might have room for a couple more.)

2022-10-23 22:07:25 In the real world of AGI, nobody knows what to do, nor how to do it. https://t.co/4DPbeIKz85

2022-10-23 21:59:17 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher here: https://t.co/5jrUODZuNt

2022-10-23 21:57:13 @Korrelan_AI @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher you need a world model for that, IMHO

2022-10-23 21:56:53 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher i saw a talk at NeurIPS c 2017 or 2018. don’t know how active it is

2022-10-23 21:47:41 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher well, if it’s created post-training but eg it’s snow white with elves rather than dwarves, i am not super impressed.but i think you are generally thinking about the right kind of issues

2022-10-23 21:44:47 . (But what a messed-up field if that actually guides people’s research choices.) https://t.co/rTXXHpKN3f

2022-10-23 21:43:10 @pmddomingos @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher link? i’d be curious

2022-10-23 21:41:34 @Peterbart @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher ever since Wittgenstein we’ve known that language doesn’t work that way… and I am happy in that corner

2022-10-23 21:40:33 @DeonTBenton the empiricist’s imagination knows no bounds.

2022-10-23 21:39:58 @AndrewLampinen @Peterbart @mpshanahan @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher fair :)

2022-10-23 21:39:44 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher just the general idea that we want generalization rather than mere memory. we might eg actually want to have two different tracks, short and open-length, tracking empirically how much progress on one leads to progress on the other.

2022-10-23 04:48:29 @NoraNewcombe you are *referencing* those distinctions, but I don’t see how their particular way of carving things up helps. as stated, something here needs be represented, and the format for that representation itself appears to be innate.

2022-10-23 04:41:20 @davidchalmers42 @mbusigin definitely grey and debatable but cognitive science actually seems broader than cognition/cognitive to me. (eg for me it clearly includes parts of philosophy of mind that i wouldn’t include in cognitive psychology per se). but there’s no fact of the matter certainly/YMMV.

2022-10-23 04:35:49 @DeonTBenton i even said “privately” ffs

2022-10-23 04:35:30 @DeonTBenton wow. you are a jerk, reposting private messages. i am going to block you and end our correspondance if you don’t delete that.

2022-10-23 02:21:41 @davidchalmers42 @mbusigin i mean, i guess, but olfaction? feelings? it’s pretty different from the most common uses of the word cognition

2022-10-23 02:19:17 RT @GaryMarcus: @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg…

2022-10-23 02:12:43 @bradpwyble @pfau @ylecun @KordingLab @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher @davidpoeppel @TonyZador @AdamMarblestone sorry totally agree and meant to mark that (though it has been considered in neuroscience)

2022-10-23 02:11:40 RT @william_g_ray: @ESYudkowsky In 20 years, people will view content created before 2021 as pure and untainted by GPT. They will mine it i…

2022-10-23 01:49:33 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble eg see list here of many agreements https://t.co/Odg2QIt39X

2022-10-23 01:48:00 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble don’t believe the press clippings

2022-10-23 01:47:45 Funny how people keep thinking this when there is so much @ylecun and I agree on, eg. recently summarized here: https://t.co/Odg2QIt39X https://t.co/bpwARC92W8

2022-10-23 01:46:14 @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher and of course the notion of variable binding is critical but hardly widely embraced in MLam sure that eg @KordingLab @davidpoeppel @TonyZador @AdamMarblestone and many others can add to this list

2022-10-23 01:43:24 @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg Van Essen diagram) - power of dendrites (@mattlark/@YiotaPoirazi etc)- variety of neurons (@AllenInstitute)- areawise specialization (@Nancy_Kanwisher)- intrinsic cues &

2022-10-23 01:25:07 @pfau @ylecun @KordingLab @bradpwyble well, ok, that might be fair. i will go with “some ideas from the 60s had real impact”, but neuroscience hasn’t been a major driver since. (experience replay, and what else from neuroscience post 1970? not a huge list?)

2022-10-23 01:21:26 @lathropa @DeonTBenton one might have thought that since dall-e had both images and text that that visual input might have served as grounding, but of course compositionality didn’t emerge. (a couple years ago people said we just need our llms to have visual input…)

2022-10-23 01:17:31 @lathropa @DeonTBenton dall-e doesn’t refute statistical learning (it is consistent with it) nor the notion that statistical learning might contribute to language learning but rather causes problems for a theory in which statistical learning by itself magically gives you language.

2022-10-22 21:13:38 Have we made exponential progress towards Hawking’s notion of intelligence? I dare say no. https://t.co/WMJDWCjuke

2022-10-22 21:12:41 RT @IntuitMachine: @GaryMarcus @PaulTopping @bradpwyble @DeonTBenton "Intelligence is the ability to adapt to change." - Stephen Hawking

2022-10-22 21:08:55 @IntuitMachine @PaulTopping @bradpwyble @DeonTBenton for sure

2022-10-22 21:08:40 @DeonTBenton sure, and see also https://t.co/0rYVHqm8SK for a bias that can’t be explained in the same way.there is both learning and innateness. you haven’t said anything that convinces me otherwise, just given me more evidence for learning (that as it happens I already knew)

2022-10-22 02:07:26 @quocleix there is an error somewhere. Hinton is born in 1947, not in 1963. may i assume that the error is in the tweet rather than the paper?

2022-10-21 21:52:16 RT @vdignum: @GaryMarcus @nirsd Working on moving #AI beyond current #ML practices and approaches is not only a good idea but a much needed…

2022-10-21 21:21:12 @Grady_Booch @davidchalmers42 fyi

2022-10-21 21:00:07 @LucaAmb strawpersoning someone who says “maybe” and “hints” as concluding anything prematurely is a bit unfair, no?

2022-10-21 20:55:00 @LucaAmb i raised a question and said i saw two hints

2022-10-21 19:41:41 @LucaAmb gee, if it were a deep learning demo you’d say “it’s early days, give them a chance”

2022-10-21 18:01:14 @PaulTopping @kohn_gregory @TonyZador @sanewman1 @ehud @SpeciesTypical or mansplaining masked in philosophical jargon? hard to say.

2022-10-21 18:00:14 Retweeting this tweet from four years ago because it perfectly anticipated @ylecun’s three stages of grief. https://t.co/lgoWfHjQBU

2022-10-21 17:59:06 RT @atldgtltrnsfnug: “In defense of skepticism about deep learning” by @GaryMarcus https://t.co/eYns9BWklh“All truth passes through three…

2022-10-21 17:57:40 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical and not the DNA itself?? just the reading process? what are you saying??

2022-10-21 17:56:54 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical do you honestly think that i don’t that? or that is any different from an unplayed record? or unread book?

2022-10-21 17:54:59 @SpeciesTypical @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory i have never blocked someone simply for being obtuse before but I am feeling close…

2022-10-21 17:54:36 @kohn_gregory @PaulTopping @ehud @sanewman1 @SpeciesTypical “not relevant until it participates” doesn’t mean it doesn’t exist!same thing with my record player groove

2022-10-21 17:34:40 Pretty prescient, no? https://t.co/lgoWfHjQBU

2022-10-21 17:17:32 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical and more seriously: https://t.co/JS6U3kWl1v

2022-10-21 17:16:51 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical nope! @kohn_gregory assures me there is no information in genes at all, &

2022-10-21 17:13:55 @WiringTheBrain @sanewman1 learning, clearly. because empiricism is all that is left!

2022-10-21 17:00:40 true, but fundamentally misleadinggenes aren’t blueprints. they don’t “directly” contain information for phenotypes—but they do contain an enormous amount of information that is relevant and causally influential. https://t.co/jH07H5Cjvh

2022-10-21 16:31:01 @math_dandy so now “deep” also encompasses a single layer, which was its historical point of contrast? and maybe literally all of statistics and computer science, too?

2022-10-21 15:53:02 so maybe you don’t need deep learning after all? just good statistics? 2nd hint of this in October… https://t.co/6Ks02OmNfr

2022-10-21 00:00:22 RT @ruchowdh: The 8 bit bias bounty is now live!! Thank you @Melissahei for the article on what the bounty program means in context of the…

2022-10-20 23:38:05 anyone want to organize a conference or conference session, red-teaming current AI, as suggest below by @nirsd? it’s a good idea. https://t.co/mTZGIrwqAi

2022-10-20 23:30:06 RT @kanjun: Today, AI systems can create stunning art &

2022-10-20 23:26:04 @nirsd i have actually seen such things advertised and honestly don’t know how you proceed if you don’t know what you are trying to fix

2022-10-20 23:24:01 RT @dcli: saw ted chiang give a talk yesterday where he basically told a bunch of ai bros, "we are nowhere near having real ai, and what we…

2022-10-20 23:14:20 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb and repeating since you missed it the first time: https://t.co/5MjfJ1h6k6

2022-10-20 23:13:41 @berent_iris for sure. no competition there

2022-10-20 23:09:55 @berent_iris also mouse wins on 3d understanding of where objects are

2022-10-20 23:07:52 @berent_iris mice have mediocre vision. so maybe not (of course they also use smell etc): but they beat models by a lot in knowing what to do with objects that they do recognize

2022-10-20 21:36:52 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb put differently, if a n-gram model (with very large n) got compositional syntax right 66% of the time, it wouldn't mean the n-gram model had compositionality, would it?

2022-10-20 21:35:34 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb whatever they get or don't get is parasitic on corpus stats. on the image synthesis stuff some colleagues and i have data we will report shortly and feel really confident that you are mistaking getting things right for the wrong reason with actually get them right.

2022-10-20 19:36:08 excellent work, offering a different and important take on compositionality

2022-10-20 19:30:56 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb another great example of the mappings are naive: https://t.co/d1bO9t8Spj

2022-10-20 19:28:50 @ihorgowda if herschel is our goal, we are doomed

2022-10-20 14:53:10 where is AI relative to mouse cognition? excellent thread! mouse experts please add to it! https://t.co/71a6kQJbqy

2022-10-20 14:48:26 @kohn_gregory @SpeciesTypical @sanewman1 that’s silly. honestly we have nothing to talk about

2022-10-20 14:43:41 @kohn_gregory @sanewman1 @SpeciesTypical not seeing the argument there, but yes i was trying to make a more sophisticated nativism.i am sure the @WiringTheBrain would say the subsequent 17 years of dev neuro have been kind to that position: eg https://t.co/bUgW8d0XTQ &

2022-10-20 14:40:08 @kohn_gregory @SpeciesTypical @sanewman1 this is confused. it is like saying there is no information in an LP, because you need to have a record player to play it.if you flatten the groove, you lose the info

2022-10-20 14:37:13 @kohn_gregory @SpeciesTypical umm… we were talking about the physical structure of the hand, and you have swapped to the motor programs controlling it. obviously the latter involves learning.

2022-10-20 14:33:19 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb @ZoubinGhahrama1 should give @TristanThrush access so Tristan can run a Winoground, then.

2022-10-20 14:31:17 and a pioneer in common sense reasoning! and now on Twitter: @ErnestSDavis https://t.co/3VWbdX2wTi

2022-10-20 08:09:29 RT @BarryOSullivan: I highly recommend this talk! And check out Ernie’s excellent book with @GaryMarcus: “Rebooting AI: Building Artificial…

2022-10-20 07:40:51 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb also note that @slatestarcodex initially declared victory in a bet re Imagen, and then retracted. i will believe that Imagen is real progress on compositionality only when it is carefully demonstrated.

2022-10-20 07:30:29 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i have asked them endlessly to make it available or test the Winoground benchmark. they have refused. i can speak w authority only re systems i (or others) can test. but remain skeptical about those hidden from testing.

2022-10-20 07:25:16 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb ie it is this claim above on image synthesis that i am questioning w.r.t. grammar: https://t.co/QLqpvsjqyL

2022-10-20 07:23:02 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i was responding to a tweet that seemed to be about image synthesis

2022-10-20 07:02:11 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb best evidence that an image synthesis system captures some grammar? they seem to have some serious problems with binding.

2022-10-20 04:02:23 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb correction to last sentence: even when it gets it right it is bloviating. in no case does LLM do inference and reasoning over a world model

2022-10-20 03:59:45 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb they composit but without proper variable binding they aren’t properly compositional. they also fail on part-whole relationships, function, etcin some ways the language side is akin to bag-of-words model, which represent language in a very impoverished way.

2022-10-20 03:45:25 RT @databu: Interesting discussion with @GaryMarcus and @bengoertzel there

2022-10-20 03:21:11 @pmddomingos @erikbryn without social media the riot probably wouldn’t have happened

2022-10-30 01:02:04 RT @foysavas: Finally read https://t.co/iZIUUNGKbR by @GaryMarcus and Ernest Davis and it's a greatly accessible critique of contemporary A…

2022-10-29 19:45:43 AI fads come and go (anyone remember expert systems? SVMs?) Here’s are three ways that Scaling Maximalism, theory du jour, might end.https://t.co/jSFOVZQQ10

2022-10-29 14:29:57 and of course the thing to read on causality, in addition to what we say, is @yudapearl’s The Book of Why.And yes Kant was there long before we were :)

2022-10-29 14:27:38 actually “our suggestion”

2022-10-29 14:22:52 My suggestion for what AI needs, 2019. Nothing has changed. https://t.co/8aHX6Gd3HT

2022-10-29 13:48:21 RT @AlisonBLowndes: Looking forward to this @GaryMarcus. Tune in all, 4.15pm UK (Lisbon) Nov 4: https://t.co/RnuXORRmPC

2022-10-29 13:39:15 @LauraRuis @Isinlor I am less sanguine. I wouldn’t be surprised to see improvements on this or that benchmark, but pragmatics fundamentally requires reasoning about models of the world (eg other’s beliefs), and I see no reason to think that LLMs their own can reliably construct such things.

2022-10-29 01:03:47 @albertwebson @_jasonwei @tallinzen skeptical of 1, but keen to see the argumentyou may have seen my new work w @EvelinaLeivada and @ElliotMurphy91, different architecture but perhaps relevant:https://t.co/UWUg6kQ3Df

2022-10-29 00:46:43 @albertwebson @_jasonwei @tallinzen I don’t think all forms of intelligence need to do arithmetic. but these kinds of errors give some insight in a semi-controlled way into a deeper, not yet solved challenge of translating language into models of the world. curious to see the paper when you can share

2022-10-28 21:32:45 RT @egrefen: I'm biased, obviously, but I think this is *such* an important result. Implicature and pragmatics are essential, ubiquitous, a…

2022-10-28 17:50:52 Just rehearsed Noam Chomsky and @jeremyakahn Nov 4, livestreamed at https://t.co/vQcEjAQCtg https://t.co/k93hY3EmVu

2022-10-28 14:18:18 indeed, slashing moderation just as synthetic content becomes widespread does seem like a disaster-in-the-making https://t.co/gEMMehatm7

2022-10-28 10:49:43 AGI, my eye. https://t.co/EmVNJRUzpz

2022-10-28 10:47:51 @L_andreae @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab agree w @L_andreae, eg a classic paper by Thelen showed that newborn humans will step, and that is really an issue of weight etc rather than lack of motor program per se: https://t.co/09LnIrfIqf

2022-10-28 10:39:40 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab (i pasted the image above/here it is) https://t.co/JjMRtntwN9

2022-10-28 10:29:06 @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab presumably both

2022-10-28 10:21:19 @L_andreae @WiringTheBrain my books https://t.co/Pt7HZc3RJd, Kluge, and The Birth of The Mind

2022-10-28 10:13:59 RT @Abebab: the most insightful knowledge about a field comes from experts in the field that hold a healthy skeptical view on the topic and…

2022-10-28 10:08:07 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab beautiful new review, by @neuroluci. how do you even think about this stuff if you are a staunch anti-nativist? https://t.co/RvacUhsspE

2022-10-28 10:01:09 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab hmm, figure 2 from Verhage et al certainly speaks to those questions (# and pattern) to some degree

2022-10-28 09:56:27 @AthenaAI2 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it goes to showing how little we understand the brain &

2022-10-28 09:50:19 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab agreed! and yet you wouldn’t know that from reading a lot of philosophers and AI researchers these days. &

2022-10-28 09:39:10 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it seems to me then that the reasonable conclusion is same as what I wrote in my lay summary in 2004: brain development uses a mix of activity-dependent (some not all intrinsically guided) &

2022-10-28 09:13:25 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab not mentioned but also seemingly relevant is Verhage et al (2000), esp figure 2 (“Correct assembly of the brain the absence of neurotransmitter secretion”) below. Anyone in this thread have thoughts on it? https://t.co/6O4OlOjoQZ

2022-10-27 18:46:45 @fhuszar https://t.co/UWUg6kxUp7

2022-10-27 18:20:33 if i had a dollar for every time someone cherry-picked DALL-E 2 art, I coulda outbid Elon for Twitter https://t.co/kHe92VtnFB

2022-10-27 15:49:18 @ylecun @ProfNoahGian so you have never done your own ground truth study and are relying on those outside orgs?

2022-10-27 15:37:48 @ylecun @ProfNoahGian our messages crossed

2022-10-27 15:31:15 @ylecun @ProfNoahGian that doesn’t answer @ProfNoahGianyou need a (large) sample that is carefully human reviewed ground truth. in that ground truth, two questions:1. what % would be detected by the AI?2.what % of what is missed by the AI is flagged by humans?you are addressing #2, not #1

2022-10-27 14:58:35 As ever, @KBAndersen has his finger on the pulse. (Answer to his question in the comments, in case you need it.) https://t.co/gqwcRACOmx

2022-10-24 02:35:23 there are two kinds of people in this world.those brave enough to speak truth to powerand those who are not.

2022-10-24 02:31:32 @primalpoly @AmandaAskell @AlecMacGillis @JoshuaChaffin i agree. I would put a lot more trust in a single anecdote involving tricking poor people into flying to Martha’s Vineyard for nonexistent jobs.

2022-10-24 01:47:36 @pmddomingos “And if there's war between the sexesThen there'll be no people left”

2022-10-24 01:39:40 @filippie509 @KordingLab @TonyZador

2022-10-24 01:21:55 @filippie509 it also doesn’t cite my endless arguments for years about looking to cognitive science, or even my paper with marblestone in Science…

2022-10-24 00:37:11 @voxbec @davidchalmers42 @mbusigin not disputing this but many people would take these to be influences on cognition. as i said the list is odd, from how people typically use the term, not necessarily wrong. in general people often distinguish between perception (eg olfaction, low level vision) and cognition.

2022-10-23 23:56:44 RT @AmandaAskell: If you don't have time to provide any evidence that someone is a eugenicist (beyond alluding to articles that also don't…

2022-10-23 22:55:22 @GaneshNatesh @ceobillionaire there has been zero budget thus far, zero charge. all participants (including us) have donated their time.

2022-10-23 22:10:34 . @ceobillionaire and I are pleased to announce that we will be resuming our AI debate series, December 23. Mark your calendar, for some exciting conversations, and reply below to suggest participants. (We already have our own ideas, but might have room for a couple more.)

2022-10-23 22:07:25 In the real world of AGI, nobody knows what to do, nor how to do it. https://t.co/4DPbeIKz85

2022-10-23 21:59:17 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher here: https://t.co/5jrUODZuNt

2022-10-23 21:57:13 @Korrelan_AI @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher you need a world model for that, IMHO

2022-10-23 21:56:53 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher i saw a talk at NeurIPS c 2017 or 2018. don’t know how active it is

2022-10-23 21:47:41 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher well, if it’s created post-training but eg it’s snow white with elves rather than dwarves, i am not super impressed.but i think you are generally thinking about the right kind of issues

2022-10-23 21:44:47 . (But what a messed-up field if that actually guides people’s research choices.) https://t.co/rTXXHpKN3f

2022-10-23 21:43:10 @pmddomingos @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher link? i’d be curious

2022-10-23 21:41:34 @Peterbart @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher ever since Wittgenstein we’ve known that language doesn’t work that way… and I am happy in that corner

2022-10-23 21:40:33 @DeonTBenton the empiricist’s imagination knows no bounds.

2022-10-23 21:39:58 @AndrewLampinen @Peterbart @mpshanahan @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher fair :)

2022-10-23 21:39:44 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher just the general idea that we want generalization rather than mere memory. we might eg actually want to have two different tracks, short and open-length, tracking empirically how much progress on one leads to progress on the other.

2022-10-23 04:48:29 @NoraNewcombe you are *referencing* those distinctions, but I don’t see how their particular way of carving things up helps. as stated, something here needs be represented, and the format for that representation itself appears to be innate.

2022-10-23 04:41:20 @davidchalmers42 @mbusigin definitely grey and debatable but cognitive science actually seems broader than cognition/cognitive to me. (eg for me it clearly includes parts of philosophy of mind that i wouldn’t include in cognitive psychology per se). but there’s no fact of the matter certainly/YMMV.

2022-10-23 04:35:49 @DeonTBenton i even said “privately” ffs

2022-10-23 04:35:30 @DeonTBenton wow. you are a jerk, reposting private messages. i am going to block you and end our correspondance if you don’t delete that.

2022-10-23 02:21:41 @davidchalmers42 @mbusigin i mean, i guess, but olfaction? feelings? it’s pretty different from the most common uses of the word cognition

2022-10-23 02:19:17 RT @GaryMarcus: @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg…

2022-10-23 02:12:43 @bradpwyble @pfau @ylecun @KordingLab @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher @davidpoeppel @TonyZador @AdamMarblestone sorry totally agree and meant to mark that (though it has been considered in neuroscience)

2022-10-23 02:11:40 RT @william_g_ray: @ESYudkowsky In 20 years, people will view content created before 2021 as pure and untainted by GPT. They will mine it i…

2022-10-23 01:49:33 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble eg see list here of many agreements https://t.co/Odg2QIt39X

2022-10-23 01:48:00 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble don’t believe the press clippings

2022-10-23 01:47:45 Funny how people keep thinking this when there is so much @ylecun and I agree on, eg. recently summarized here: https://t.co/Odg2QIt39X https://t.co/bpwARC92W8

2022-10-23 01:46:14 @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher and of course the notion of variable binding is critical but hardly widely embraced in MLam sure that eg @KordingLab @davidpoeppel @TonyZador @AdamMarblestone and many others can add to this list

2022-10-23 01:43:24 @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg Van Essen diagram) - power of dendrites (@mattlark/@YiotaPoirazi etc)- variety of neurons (@AllenInstitute)- areawise specialization (@Nancy_Kanwisher)- intrinsic cues &

2022-10-23 01:25:07 @pfau @ylecun @KordingLab @bradpwyble well, ok, that might be fair. i will go with “some ideas from the 60s had real impact”, but neuroscience hasn’t been a major driver since. (experience replay, and what else from neuroscience post 1970? not a huge list?)

2022-10-23 01:21:26 @lathropa @DeonTBenton one might have thought that since dall-e had both images and text that that visual input might have served as grounding, but of course compositionality didn’t emerge. (a couple years ago people said we just need our llms to have visual input…)

2022-10-23 01:17:31 @lathropa @DeonTBenton dall-e doesn’t refute statistical learning (it is consistent with it) nor the notion that statistical learning might contribute to language learning but rather causes problems for a theory in which statistical learning by itself magically gives you language.

2022-10-22 21:13:38 Have we made exponential progress towards Hawking’s notion of intelligence? I dare say no. https://t.co/WMJDWCjuke

2022-10-22 21:12:41 RT @IntuitMachine: @GaryMarcus @PaulTopping @bradpwyble @DeonTBenton "Intelligence is the ability to adapt to change." - Stephen Hawking

2022-10-22 21:08:55 @IntuitMachine @PaulTopping @bradpwyble @DeonTBenton for sure

2022-10-22 21:08:40 @DeonTBenton sure, and see also https://t.co/0rYVHqm8SK for a bias that can’t be explained in the same way.there is both learning and innateness. you haven’t said anything that convinces me otherwise, just given me more evidence for learning (that as it happens I already knew)

2022-10-22 02:07:26 @quocleix there is an error somewhere. Hinton is born in 1947, not in 1963. may i assume that the error is in the tweet rather than the paper?

2022-10-21 21:52:16 RT @vdignum: @GaryMarcus @nirsd Working on moving #AI beyond current #ML practices and approaches is not only a good idea but a much needed…

2022-10-21 21:21:12 @Grady_Booch @davidchalmers42 fyi

2022-10-21 21:00:07 @LucaAmb strawpersoning someone who says “maybe” and “hints” as concluding anything prematurely is a bit unfair, no?

2022-10-21 20:55:00 @LucaAmb i raised a question and said i saw two hints

2022-10-21 19:41:41 @LucaAmb gee, if it were a deep learning demo you’d say “it’s early days, give them a chance”

2022-10-21 18:01:14 @PaulTopping @kohn_gregory @TonyZador @sanewman1 @ehud @SpeciesTypical or mansplaining masked in philosophical jargon? hard to say.

2022-10-21 18:00:14 Retweeting this tweet from four years ago because it perfectly anticipated @ylecun’s three stages of grief. https://t.co/lgoWfHjQBU

2022-10-21 17:59:06 RT @atldgtltrnsfnug: “In defense of skepticism about deep learning” by @GaryMarcus https://t.co/eYns9BWklh“All truth passes through three…

2022-10-21 17:57:40 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical and not the DNA itself?? just the reading process? what are you saying??

2022-10-21 17:56:54 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical do you honestly think that i don’t that? or that is any different from an unplayed record? or unread book?

2022-10-21 17:54:59 @SpeciesTypical @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory i have never blocked someone simply for being obtuse before but I am feeling close…

2022-10-21 17:54:36 @kohn_gregory @PaulTopping @ehud @sanewman1 @SpeciesTypical “not relevant until it participates” doesn’t mean it doesn’t exist!same thing with my record player groove

2022-10-21 17:34:40 Pretty prescient, no? https://t.co/lgoWfHjQBU

2022-10-21 17:17:32 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical and more seriously: https://t.co/JS6U3kWl1v

2022-10-21 17:16:51 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical nope! @kohn_gregory assures me there is no information in genes at all, &

2022-10-21 17:13:55 @WiringTheBrain @sanewman1 learning, clearly. because empiricism is all that is left!

2022-10-21 17:00:40 true, but fundamentally misleadinggenes aren’t blueprints. they don’t “directly” contain information for phenotypes—but they do contain an enormous amount of information that is relevant and causally influential. https://t.co/jH07H5Cjvh

2022-10-21 16:31:01 @math_dandy so now “deep” also encompasses a single layer, which was its historical point of contrast? and maybe literally all of statistics and computer science, too?

2022-10-21 15:53:02 so maybe you don’t need deep learning after all? just good statistics? 2nd hint of this in October… https://t.co/6Ks02OmNfr

2022-10-21 00:00:22 RT @ruchowdh: The 8 bit bias bounty is now live!! Thank you @Melissahei for the article on what the bounty program means in context of the…

2022-10-20 23:38:05 anyone want to organize a conference or conference session, red-teaming current AI, as suggest below by @nirsd? it’s a good idea. https://t.co/mTZGIrwqAi

2022-10-20 23:30:06 RT @kanjun: Today, AI systems can create stunning art &

2022-10-20 23:26:04 @nirsd i have actually seen such things advertised and honestly don’t know how you proceed if you don’t know what you are trying to fix

2022-10-20 23:24:01 RT @dcli: saw ted chiang give a talk yesterday where he basically told a bunch of ai bros, "we are nowhere near having real ai, and what we…

2022-10-20 23:14:20 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb and repeating since you missed it the first time: https://t.co/5MjfJ1h6k6

2022-10-20 23:13:41 @berent_iris for sure. no competition there

2022-10-20 23:09:55 @berent_iris also mouse wins on 3d understanding of where objects are

2022-10-20 23:07:52 @berent_iris mice have mediocre vision. so maybe not (of course they also use smell etc): but they beat models by a lot in knowing what to do with objects that they do recognize

2022-10-20 21:36:52 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb put differently, if a n-gram model (with very large n) got compositional syntax right 66% of the time, it wouldn't mean the n-gram model had compositionality, would it?

2022-10-20 21:35:34 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb whatever they get or don't get is parasitic on corpus stats. on the image synthesis stuff some colleagues and i have data we will report shortly and feel really confident that you are mistaking getting things right for the wrong reason with actually get them right.

2022-10-20 19:36:08 excellent work, offering a different and important take on compositionality

2022-10-20 19:30:56 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb another great example of the mappings are naive: https://t.co/d1bO9t8Spj

2022-10-20 19:28:50 @ihorgowda if herschel is our goal, we are doomed

2022-10-20 14:53:10 where is AI relative to mouse cognition? excellent thread! mouse experts please add to it! https://t.co/71a6kQJbqy

2022-10-20 14:48:26 @kohn_gregory @SpeciesTypical @sanewman1 that’s silly. honestly we have nothing to talk about

2022-10-20 14:43:41 @kohn_gregory @sanewman1 @SpeciesTypical not seeing the argument there, but yes i was trying to make a more sophisticated nativism.i am sure the @WiringTheBrain would say the subsequent 17 years of dev neuro have been kind to that position: eg https://t.co/bUgW8d0XTQ &

2022-10-20 14:40:08 @kohn_gregory @SpeciesTypical @sanewman1 this is confused. it is like saying there is no information in an LP, because you need to have a record player to play it.if you flatten the groove, you lose the info

2022-10-20 14:37:13 @kohn_gregory @SpeciesTypical umm… we were talking about the physical structure of the hand, and you have swapped to the motor programs controlling it. obviously the latter involves learning.

2022-10-20 14:33:19 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb @ZoubinGhahrama1 should give @TristanThrush access so Tristan can run a Winoground, then.

2022-10-20 14:31:17 and a pioneer in common sense reasoning! and now on Twitter: @ErnestSDavis https://t.co/3VWbdX2wTi

2022-10-20 08:09:29 RT @BarryOSullivan: I highly recommend this talk! And check out Ernie’s excellent book with @GaryMarcus: “Rebooting AI: Building Artificial…

2022-10-20 07:40:51 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb also note that @slatestarcodex initially declared victory in a bet re Imagen, and then retracted. i will believe that Imagen is real progress on compositionality only when it is carefully demonstrated.

2022-10-20 07:30:29 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i have asked them endlessly to make it available or test the Winoground benchmark. they have refused. i can speak w authority only re systems i (or others) can test. but remain skeptical about those hidden from testing.

2022-10-20 07:25:16 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb ie it is this claim above on image synthesis that i am questioning w.r.t. grammar: https://t.co/QLqpvsjqyL

2022-10-20 07:23:02 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i was responding to a tweet that seemed to be about image synthesis

2022-10-20 07:02:11 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb best evidence that an image synthesis system captures some grammar? they seem to have some serious problems with binding.

2022-10-20 04:02:23 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb correction to last sentence: even when it gets it right it is bloviating. in no case does LLM do inference and reasoning over a world model

2022-10-20 03:59:45 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb they composit but without proper variable binding they aren’t properly compositional. they also fail on part-whole relationships, function, etcin some ways the language side is akin to bag-of-words model, which represent language in a very impoverished way.

2022-10-20 03:45:25 RT @databu: Interesting discussion with @GaryMarcus and @bengoertzel there

2022-10-20 03:21:11 @pmddomingos @erikbryn without social media the riot probably wouldn’t have happened

2022-10-30 13:44:12 quietly growing, and poised to change everything. https://t.co/9847MKRfDz

2022-10-30 01:02:04 RT @foysavas: Finally read https://t.co/iZIUUNGKbR by @GaryMarcus and Ernest Davis and it's a greatly accessible critique of contemporary A…

2022-10-29 19:45:43 AI fads come and go (anyone remember expert systems? SVMs?) Here’s are three ways that Scaling Maximalism, theory du jour, might end.https://t.co/jSFOVZQQ10

2022-10-29 14:29:57 and of course the thing to read on causality, in addition to what we say, is @yudapearl’s The Book of Why.And yes Kant was there long before we were :)

2022-10-29 14:27:38 actually “our suggestion”

2022-10-29 14:22:52 My suggestion for what AI needs, 2019. Nothing has changed. https://t.co/8aHX6Gd3HT

2022-10-29 13:48:21 RT @AlisonBLowndes: Looking forward to this @GaryMarcus. Tune in all, 4.15pm UK (Lisbon) Nov 4: https://t.co/RnuXORRmPC

2022-10-29 13:39:15 @LauraRuis @Isinlor I am less sanguine. I wouldn’t be surprised to see improvements on this or that benchmark, but pragmatics fundamentally requires reasoning about models of the world (eg other’s beliefs), and I see no reason to think that LLMs their own can reliably construct such things.

2022-10-29 01:03:47 @albertwebson @_jasonwei @tallinzen skeptical of 1, but keen to see the argumentyou may have seen my new work w @EvelinaLeivada and @ElliotMurphy91, different architecture but perhaps relevant:https://t.co/UWUg6kQ3Df

2022-10-29 00:46:43 @albertwebson @_jasonwei @tallinzen I don’t think all forms of intelligence need to do arithmetic. but these kinds of errors give some insight in a semi-controlled way into a deeper, not yet solved challenge of translating language into models of the world. curious to see the paper when you can share

2022-10-28 21:32:45 RT @egrefen: I'm biased, obviously, but I think this is *such* an important result. Implicature and pragmatics are essential, ubiquitous, a…

2022-10-28 17:50:52 Just rehearsed Noam Chomsky and @jeremyakahn Nov 4, livestreamed at https://t.co/vQcEjAQCtg https://t.co/k93hY3EmVu

2022-10-28 14:18:18 indeed, slashing moderation just as synthetic content becomes widespread does seem like a disaster-in-the-making https://t.co/gEMMehatm7

2022-10-28 10:49:43 AGI, my eye. https://t.co/EmVNJRUzpz

2022-10-28 10:47:51 @L_andreae @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab agree w @L_andreae, eg a classic paper by Thelen showed that newborn humans will step, and that is really an issue of weight etc rather than lack of motor program per se: https://t.co/09LnIrfIqf

2022-10-28 10:39:40 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab (i pasted the image above/here it is) https://t.co/JjMRtntwN9

2022-10-28 10:29:06 @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab presumably both

2022-10-28 10:21:19 @L_andreae @WiringTheBrain my books https://t.co/Pt7HZc3RJd, Kluge, and The Birth of The Mind

2022-10-28 10:13:59 RT @Abebab: the most insightful knowledge about a field comes from experts in the field that hold a healthy skeptical view on the topic and…

2022-10-28 10:08:07 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab beautiful new review, by @neuroluci. how do you even think about this stuff if you are a staunch anti-nativist? https://t.co/RvacUhsspE

2022-10-28 10:01:09 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab hmm, figure 2 from Verhage et al certainly speaks to those questions (# and pattern) to some degree

2022-10-28 09:56:27 @AthenaAI2 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it goes to showing how little we understand the brain &

2022-10-28 09:50:19 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab agreed! and yet you wouldn’t know that from reading a lot of philosophers and AI researchers these days. &

2022-10-28 09:39:10 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it seems to me then that the reasonable conclusion is same as what I wrote in my lay summary in 2004: brain development uses a mix of activity-dependent (some not all intrinsically guided) &

2022-10-28 09:13:25 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab not mentioned but also seemingly relevant is Verhage et al (2000), esp figure 2 (“Correct assembly of the brain the absence of neurotransmitter secretion”) below. Anyone in this thread have thoughts on it? https://t.co/6O4OlOjoQZ

2022-10-27 18:46:45 @fhuszar https://t.co/UWUg6kxUp7

2022-10-27 18:20:33 if i had a dollar for every time someone cherry-picked DALL-E 2 art, I coulda outbid Elon for Twitter https://t.co/kHe92VtnFB

2022-10-27 15:49:18 @ylecun @ProfNoahGian so you have never done your own ground truth study and are relying on those outside orgs?

2022-10-27 15:37:48 @ylecun @ProfNoahGian our messages crossed

2022-10-27 15:31:15 @ylecun @ProfNoahGian that doesn’t answer @ProfNoahGianyou need a (large) sample that is carefully human reviewed ground truth. in that ground truth, two questions:1. what % would be detected by the AI?2.what % of what is missed by the AI is flagged by humans?you are addressing #2, not #1

2022-10-27 14:58:35 As ever, @KBAndersen has his finger on the pulse. (Answer to his question in the comments, in case you need it.) https://t.co/gqwcRACOmx

2022-10-24 02:35:23 there are two kinds of people in this world.those brave enough to speak truth to powerand those who are not.

2022-10-24 02:31:32 @primalpoly @AmandaAskell @AlecMacGillis @JoshuaChaffin i agree. I would put a lot more trust in a single anecdote involving tricking poor people into flying to Martha’s Vineyard for nonexistent jobs.

2022-10-24 01:47:36 @pmddomingos “And if there's war between the sexesThen there'll be no people left”

2022-10-24 01:39:40 @filippie509 @KordingLab @TonyZador

2022-10-24 01:21:55 @filippie509 it also doesn’t cite my endless arguments for years about looking to cognitive science, or even my paper with marblestone in Science…

2022-10-24 00:37:11 @voxbec @davidchalmers42 @mbusigin not disputing this but many people would take these to be influences on cognition. as i said the list is odd, from how people typically use the term, not necessarily wrong. in general people often distinguish between perception (eg olfaction, low level vision) and cognition.

2022-10-23 23:56:44 RT @AmandaAskell: If you don't have time to provide any evidence that someone is a eugenicist (beyond alluding to articles that also don't…

2022-10-23 22:55:22 @GaneshNatesh @ceobillionaire there has been zero budget thus far, zero charge. all participants (including us) have donated their time.

2022-10-23 22:10:34 . @ceobillionaire and I are pleased to announce that we will be resuming our AI debate series, December 23. Mark your calendar, for some exciting conversations, and reply below to suggest participants. (We already have our own ideas, but might have room for a couple more.)

2022-10-23 22:07:25 In the real world of AGI, nobody knows what to do, nor how to do it. https://t.co/4DPbeIKz85

2022-10-23 21:59:17 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher here: https://t.co/5jrUODZuNt

2022-10-23 21:57:13 @Korrelan_AI @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher you need a world model for that, IMHO

2022-10-23 21:56:53 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher i saw a talk at NeurIPS c 2017 or 2018. don’t know how active it is

2022-10-23 21:47:41 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher well, if it’s created post-training but eg it’s snow white with elves rather than dwarves, i am not super impressed.but i think you are generally thinking about the right kind of issues

2022-10-23 21:44:47 . (But what a messed-up field if that actually guides people’s research choices.) https://t.co/rTXXHpKN3f

2022-10-23 21:43:10 @pmddomingos @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher link? i’d be curious

2022-10-23 21:41:34 @Peterbart @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher ever since Wittgenstein we’ve known that language doesn’t work that way… and I am happy in that corner

2022-10-23 21:40:33 @DeonTBenton the empiricist’s imagination knows no bounds.

2022-10-23 21:39:58 @AndrewLampinen @Peterbart @mpshanahan @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher fair :)

2022-10-23 21:39:44 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher just the general idea that we want generalization rather than mere memory. we might eg actually want to have two different tracks, short and open-length, tracking empirically how much progress on one leads to progress on the other.

2022-10-23 04:48:29 @NoraNewcombe you are *referencing* those distinctions, but I don’t see how their particular way of carving things up helps. as stated, something here needs be represented, and the format for that representation itself appears to be innate.

2022-10-23 04:41:20 @davidchalmers42 @mbusigin definitely grey and debatable but cognitive science actually seems broader than cognition/cognitive to me. (eg for me it clearly includes parts of philosophy of mind that i wouldn’t include in cognitive psychology per se). but there’s no fact of the matter certainly/YMMV.

2022-10-23 04:35:49 @DeonTBenton i even said “privately” ffs

2022-10-23 04:35:30 @DeonTBenton wow. you are a jerk, reposting private messages. i am going to block you and end our correspondance if you don’t delete that.

2022-10-23 02:21:41 @davidchalmers42 @mbusigin i mean, i guess, but olfaction? feelings? it’s pretty different from the most common uses of the word cognition

2022-10-23 02:19:17 RT @GaryMarcus: @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg…

2022-10-23 02:12:43 @bradpwyble @pfau @ylecun @KordingLab @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher @davidpoeppel @TonyZador @AdamMarblestone sorry totally agree and meant to mark that (though it has been considered in neuroscience)

2022-10-23 02:11:40 RT @william_g_ray: @ESYudkowsky In 20 years, people will view content created before 2021 as pure and untainted by GPT. They will mine it i…

2022-10-23 01:49:33 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble eg see list here of many agreements https://t.co/Odg2QIt39X

2022-10-23 01:48:00 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble don’t believe the press clippings

2022-10-23 01:47:45 Funny how people keep thinking this when there is so much @ylecun and I agree on, eg. recently summarized here: https://t.co/Odg2QIt39X https://t.co/bpwARC92W8

2022-10-23 01:46:14 @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher and of course the notion of variable binding is critical but hardly widely embraced in MLam sure that eg @KordingLab @davidpoeppel @TonyZador @AdamMarblestone and many others can add to this list

2022-10-23 01:43:24 @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg Van Essen diagram) - power of dendrites (@mattlark/@YiotaPoirazi etc)- variety of neurons (@AllenInstitute)- areawise specialization (@Nancy_Kanwisher)- intrinsic cues &

2022-10-23 01:25:07 @pfau @ylecun @KordingLab @bradpwyble well, ok, that might be fair. i will go with “some ideas from the 60s had real impact”, but neuroscience hasn’t been a major driver since. (experience replay, and what else from neuroscience post 1970? not a huge list?)

2022-10-23 01:21:26 @lathropa @DeonTBenton one might have thought that since dall-e had both images and text that that visual input might have served as grounding, but of course compositionality didn’t emerge. (a couple years ago people said we just need our llms to have visual input…)

2022-10-23 01:17:31 @lathropa @DeonTBenton dall-e doesn’t refute statistical learning (it is consistent with it) nor the notion that statistical learning might contribute to language learning but rather causes problems for a theory in which statistical learning by itself magically gives you language.

2022-10-22 21:13:38 Have we made exponential progress towards Hawking’s notion of intelligence? I dare say no. https://t.co/WMJDWCjuke

2022-10-22 21:12:41 RT @IntuitMachine: @GaryMarcus @PaulTopping @bradpwyble @DeonTBenton "Intelligence is the ability to adapt to change." - Stephen Hawking

2022-10-22 21:08:55 @IntuitMachine @PaulTopping @bradpwyble @DeonTBenton for sure

2022-10-22 21:08:40 @DeonTBenton sure, and see also https://t.co/0rYVHqm8SK for a bias that can’t be explained in the same way.there is both learning and innateness. you haven’t said anything that convinces me otherwise, just given me more evidence for learning (that as it happens I already knew)

2022-10-22 02:07:26 @quocleix there is an error somewhere. Hinton is born in 1947, not in 1963. may i assume that the error is in the tweet rather than the paper?

2022-10-21 21:52:16 RT @vdignum: @GaryMarcus @nirsd Working on moving #AI beyond current #ML practices and approaches is not only a good idea but a much needed…

2022-10-21 21:21:12 @Grady_Booch @davidchalmers42 fyi

2022-10-21 21:00:07 @LucaAmb strawpersoning someone who says “maybe” and “hints” as concluding anything prematurely is a bit unfair, no?

2022-10-21 20:55:00 @LucaAmb i raised a question and said i saw two hints

2022-10-21 19:41:41 @LucaAmb gee, if it were a deep learning demo you’d say “it’s early days, give them a chance”

2022-10-21 18:01:14 @PaulTopping @kohn_gregory @TonyZador @sanewman1 @ehud @SpeciesTypical or mansplaining masked in philosophical jargon? hard to say.

2022-10-21 18:00:14 Retweeting this tweet from four years ago because it perfectly anticipated @ylecun’s three stages of grief. https://t.co/lgoWfHjQBU

2022-10-21 17:59:06 RT @atldgtltrnsfnug: “In defense of skepticism about deep learning” by @GaryMarcus https://t.co/eYns9BWklh“All truth passes through three…

2022-10-21 17:57:40 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical and not the DNA itself?? just the reading process? what are you saying??

2022-10-21 17:56:54 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical do you honestly think that i don’t that? or that is any different from an unplayed record? or unread book?

2022-10-21 17:54:59 @SpeciesTypical @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory i have never blocked someone simply for being obtuse before but I am feeling close…

2022-10-21 17:54:36 @kohn_gregory @PaulTopping @ehud @sanewman1 @SpeciesTypical “not relevant until it participates” doesn’t mean it doesn’t exist!same thing with my record player groove

2022-10-21 17:34:40 Pretty prescient, no? https://t.co/lgoWfHjQBU

2022-10-21 17:17:32 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical and more seriously: https://t.co/JS6U3kWl1v

2022-10-21 17:16:51 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical nope! @kohn_gregory assures me there is no information in genes at all, &

2022-10-21 17:13:55 @WiringTheBrain @sanewman1 learning, clearly. because empiricism is all that is left!

2022-10-21 17:00:40 true, but fundamentally misleadinggenes aren’t blueprints. they don’t “directly” contain information for phenotypes—but they do contain an enormous amount of information that is relevant and causally influential. https://t.co/jH07H5Cjvh

2022-10-21 16:31:01 @math_dandy so now “deep” also encompasses a single layer, which was its historical point of contrast? and maybe literally all of statistics and computer science, too?

2022-10-21 15:53:02 so maybe you don’t need deep learning after all? just good statistics? 2nd hint of this in October… https://t.co/6Ks02OmNfr

2022-10-21 00:00:22 RT @ruchowdh: The 8 bit bias bounty is now live!! Thank you @Melissahei for the article on what the bounty program means in context of the…

2022-10-20 23:38:05 anyone want to organize a conference or conference session, red-teaming current AI, as suggest below by @nirsd? it’s a good idea. https://t.co/mTZGIrwqAi

2022-10-20 23:30:06 RT @kanjun: Today, AI systems can create stunning art &

2022-10-20 23:26:04 @nirsd i have actually seen such things advertised and honestly don’t know how you proceed if you don’t know what you are trying to fix

2022-10-20 23:24:01 RT @dcli: saw ted chiang give a talk yesterday where he basically told a bunch of ai bros, "we are nowhere near having real ai, and what we…

2022-10-20 23:14:20 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb and repeating since you missed it the first time: https://t.co/5MjfJ1h6k6

2022-10-20 23:13:41 @berent_iris for sure. no competition there

2022-10-20 23:09:55 @berent_iris also mouse wins on 3d understanding of where objects are

2022-10-20 23:07:52 @berent_iris mice have mediocre vision. so maybe not (of course they also use smell etc): but they beat models by a lot in knowing what to do with objects that they do recognize

2022-10-20 21:36:52 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb put differently, if a n-gram model (with very large n) got compositional syntax right 66% of the time, it wouldn't mean the n-gram model had compositionality, would it?

2022-10-20 21:35:34 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb whatever they get or don't get is parasitic on corpus stats. on the image synthesis stuff some colleagues and i have data we will report shortly and feel really confident that you are mistaking getting things right for the wrong reason with actually get them right.

2022-10-20 19:36:08 excellent work, offering a different and important take on compositionality

2022-10-20 19:30:56 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb another great example of the mappings are naive: https://t.co/d1bO9t8Spj

2022-10-20 19:28:50 @ihorgowda if herschel is our goal, we are doomed

2022-10-20 14:53:10 where is AI relative to mouse cognition? excellent thread! mouse experts please add to it! https://t.co/71a6kQJbqy

2022-10-20 14:48:26 @kohn_gregory @SpeciesTypical @sanewman1 that’s silly. honestly we have nothing to talk about

2022-10-20 14:43:41 @kohn_gregory @sanewman1 @SpeciesTypical not seeing the argument there, but yes i was trying to make a more sophisticated nativism.i am sure the @WiringTheBrain would say the subsequent 17 years of dev neuro have been kind to that position: eg https://t.co/bUgW8d0XTQ &

2022-10-20 14:40:08 @kohn_gregory @SpeciesTypical @sanewman1 this is confused. it is like saying there is no information in an LP, because you need to have a record player to play it.if you flatten the groove, you lose the info

2022-10-20 14:37:13 @kohn_gregory @SpeciesTypical umm… we were talking about the physical structure of the hand, and you have swapped to the motor programs controlling it. obviously the latter involves learning.

2022-10-20 14:33:19 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb @ZoubinGhahrama1 should give @TristanThrush access so Tristan can run a Winoground, then.

2022-10-20 14:31:17 and a pioneer in common sense reasoning! and now on Twitter: @ErnestSDavis https://t.co/3VWbdX2wTi

2022-10-20 08:09:29 RT @BarryOSullivan: I highly recommend this talk! And check out Ernie’s excellent book with @GaryMarcus: “Rebooting AI: Building Artificial…

2022-10-20 07:40:51 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb also note that @slatestarcodex initially declared victory in a bet re Imagen, and then retracted. i will believe that Imagen is real progress on compositionality only when it is carefully demonstrated.

2022-10-20 07:30:29 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i have asked them endlessly to make it available or test the Winoground benchmark. they have refused. i can speak w authority only re systems i (or others) can test. but remain skeptical about those hidden from testing.

2022-10-20 07:25:16 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb ie it is this claim above on image synthesis that i am questioning w.r.t. grammar: https://t.co/QLqpvsjqyL

2022-10-20 07:23:02 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i was responding to a tweet that seemed to be about image synthesis

2022-10-20 07:02:11 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb best evidence that an image synthesis system captures some grammar? they seem to have some serious problems with binding.

2022-10-20 04:02:23 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb correction to last sentence: even when it gets it right it is bloviating. in no case does LLM do inference and reasoning over a world model

2022-10-20 03:59:45 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb they composit but without proper variable binding they aren’t properly compositional. they also fail on part-whole relationships, function, etcin some ways the language side is akin to bag-of-words model, which represent language in a very impoverished way.

2022-10-20 03:45:25 RT @databu: Interesting discussion with @GaryMarcus and @bengoertzel there

2022-10-20 03:21:11 @pmddomingos @erikbryn without social media the riot probably wouldn’t have happened

2022-10-30 18:28:18 Elon Musk: “Twitter obviously cannot become a free-for-all hellscape.”Also Elon Musk: https://t.co/sRkeYww7ve

2022-10-30 13:44:12 quietly growing, and poised to change everything. https://t.co/9847MKRfDz

2022-10-30 01:02:04 RT @foysavas: Finally read https://t.co/iZIUUNGKbR by @GaryMarcus and Ernest Davis and it's a greatly accessible critique of contemporary A…

2022-10-29 19:45:43 AI fads come and go (anyone remember expert systems? SVMs?) Here’s are three ways that Scaling Maximalism, theory du jour, might end.https://t.co/jSFOVZQQ10

2022-10-29 14:29:57 and of course the thing to read on causality, in addition to what we say, is @yudapearl’s The Book of Why.And yes Kant was there long before we were :)

2022-10-29 14:27:38 actually “our suggestion”

2022-10-29 14:22:52 My suggestion for what AI needs, 2019. Nothing has changed. https://t.co/8aHX6Gd3HT

2022-10-29 13:48:21 RT @AlisonBLowndes: Looking forward to this @GaryMarcus. Tune in all, 4.15pm UK (Lisbon) Nov 4: https://t.co/RnuXORRmPC

2022-10-29 13:39:15 @LauraRuis @Isinlor I am less sanguine. I wouldn’t be surprised to see improvements on this or that benchmark, but pragmatics fundamentally requires reasoning about models of the world (eg other’s beliefs), and I see no reason to think that LLMs their own can reliably construct such things.

2022-10-29 01:03:47 @albertwebson @_jasonwei @tallinzen skeptical of 1, but keen to see the argumentyou may have seen my new work w @EvelinaLeivada and @ElliotMurphy91, different architecture but perhaps relevant:https://t.co/UWUg6kQ3Df

2022-10-29 00:46:43 @albertwebson @_jasonwei @tallinzen I don’t think all forms of intelligence need to do arithmetic. but these kinds of errors give some insight in a semi-controlled way into a deeper, not yet solved challenge of translating language into models of the world. curious to see the paper when you can share

2022-10-28 21:32:45 RT @egrefen: I'm biased, obviously, but I think this is *such* an important result. Implicature and pragmatics are essential, ubiquitous, a…

2022-10-28 17:50:52 Just rehearsed Noam Chomsky and @jeremyakahn Nov 4, livestreamed at https://t.co/vQcEjAQCtg https://t.co/k93hY3EmVu

2022-10-28 14:18:18 indeed, slashing moderation just as synthetic content becomes widespread does seem like a disaster-in-the-making https://t.co/gEMMehatm7

2022-10-28 10:49:43 AGI, my eye. https://t.co/EmVNJRUzpz

2022-10-28 10:47:51 @L_andreae @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab agree w @L_andreae, eg a classic paper by Thelen showed that newborn humans will step, and that is really an issue of weight etc rather than lack of motor program per se: https://t.co/09LnIrfIqf

2022-10-28 10:39:40 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab (i pasted the image above/here it is) https://t.co/JjMRtntwN9

2022-10-28 10:29:06 @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab presumably both

2022-10-28 10:21:19 @L_andreae @WiringTheBrain my books https://t.co/Pt7HZc3RJd, Kluge, and The Birth of The Mind

2022-10-28 10:13:59 RT @Abebab: the most insightful knowledge about a field comes from experts in the field that hold a healthy skeptical view on the topic and…

2022-10-28 10:08:07 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab beautiful new review, by @neuroluci. how do you even think about this stuff if you are a staunch anti-nativist? https://t.co/RvacUhsspE

2022-10-28 10:01:09 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab hmm, figure 2 from Verhage et al certainly speaks to those questions (# and pattern) to some degree

2022-10-28 09:56:27 @AthenaAI2 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it goes to showing how little we understand the brain &

2022-10-28 09:50:19 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab agreed! and yet you wouldn’t know that from reading a lot of philosophers and AI researchers these days. &

2022-10-28 09:39:10 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it seems to me then that the reasonable conclusion is same as what I wrote in my lay summary in 2004: brain development uses a mix of activity-dependent (some not all intrinsically guided) &

2022-10-28 09:13:25 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab not mentioned but also seemingly relevant is Verhage et al (2000), esp figure 2 (“Correct assembly of the brain the absence of neurotransmitter secretion”) below. Anyone in this thread have thoughts on it? https://t.co/6O4OlOjoQZ

2022-10-27 18:46:45 @fhuszar https://t.co/UWUg6kxUp7

2022-10-27 18:20:33 if i had a dollar for every time someone cherry-picked DALL-E 2 art, I coulda outbid Elon for Twitter https://t.co/kHe92VtnFB

2022-10-27 15:49:18 @ylecun @ProfNoahGian so you have never done your own ground truth study and are relying on those outside orgs?

2022-10-27 15:37:48 @ylecun @ProfNoahGian our messages crossed

2022-10-27 15:31:15 @ylecun @ProfNoahGian that doesn’t answer @ProfNoahGianyou need a (large) sample that is carefully human reviewed ground truth. in that ground truth, two questions:1. what % would be detected by the AI?2.what % of what is missed by the AI is flagged by humans?you are addressing #2, not #1

2022-10-27 14:58:35 As ever, @KBAndersen has his finger on the pulse. (Answer to his question in the comments, in case you need it.) https://t.co/gqwcRACOmx

2022-10-24 02:35:23 there are two kinds of people in this world.those brave enough to speak truth to powerand those who are not.

2022-10-24 02:31:32 @primalpoly @AmandaAskell @AlecMacGillis @JoshuaChaffin i agree. I would put a lot more trust in a single anecdote involving tricking poor people into flying to Martha’s Vineyard for nonexistent jobs.

2022-10-24 01:47:36 @pmddomingos “And if there's war between the sexesThen there'll be no people left”

2022-10-24 01:39:40 @filippie509 @KordingLab @TonyZador

2022-10-24 01:21:55 @filippie509 it also doesn’t cite my endless arguments for years about looking to cognitive science, or even my paper with marblestone in Science…

2022-10-24 00:37:11 @voxbec @davidchalmers42 @mbusigin not disputing this but many people would take these to be influences on cognition. as i said the list is odd, from how people typically use the term, not necessarily wrong. in general people often distinguish between perception (eg olfaction, low level vision) and cognition.

2022-10-23 23:56:44 RT @AmandaAskell: If you don't have time to provide any evidence that someone is a eugenicist (beyond alluding to articles that also don't…

2022-10-23 22:55:22 @GaneshNatesh @ceobillionaire there has been zero budget thus far, zero charge. all participants (including us) have donated their time.

2022-10-23 22:10:34 . @ceobillionaire and I are pleased to announce that we will be resuming our AI debate series, December 23. Mark your calendar, for some exciting conversations, and reply below to suggest participants. (We already have our own ideas, but might have room for a couple more.)

2022-10-23 22:07:25 In the real world of AGI, nobody knows what to do, nor how to do it. https://t.co/4DPbeIKz85

2022-10-23 21:59:17 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher here: https://t.co/5jrUODZuNt

2022-10-23 21:57:13 @Korrelan_AI @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher you need a world model for that, IMHO

2022-10-23 21:56:53 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher i saw a talk at NeurIPS c 2017 or 2018. don’t know how active it is

2022-10-23 21:47:41 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher well, if it’s created post-training but eg it’s snow white with elves rather than dwarves, i am not super impressed.but i think you are generally thinking about the right kind of issues

2022-10-23 21:44:47 . (But what a messed-up field if that actually guides people’s research choices.) https://t.co/rTXXHpKN3f

2022-10-23 21:43:10 @pmddomingos @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher link? i’d be curious

2022-10-23 21:41:34 @Peterbart @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher ever since Wittgenstein we’ve known that language doesn’t work that way… and I am happy in that corner

2022-10-23 21:40:33 @DeonTBenton the empiricist’s imagination knows no bounds.

2022-10-23 21:39:58 @AndrewLampinen @Peterbart @mpshanahan @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher fair :)

2022-10-23 21:39:44 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher just the general idea that we want generalization rather than mere memory. we might eg actually want to have two different tracks, short and open-length, tracking empirically how much progress on one leads to progress on the other.

2022-10-23 04:48:29 @NoraNewcombe you are *referencing* those distinctions, but I don’t see how their particular way of carving things up helps. as stated, something here needs be represented, and the format for that representation itself appears to be innate.

2022-10-23 04:41:20 @davidchalmers42 @mbusigin definitely grey and debatable but cognitive science actually seems broader than cognition/cognitive to me. (eg for me it clearly includes parts of philosophy of mind that i wouldn’t include in cognitive psychology per se). but there’s no fact of the matter certainly/YMMV.

2022-10-23 04:35:49 @DeonTBenton i even said “privately” ffs

2022-10-23 04:35:30 @DeonTBenton wow. you are a jerk, reposting private messages. i am going to block you and end our correspondance if you don’t delete that.

2022-10-23 02:21:41 @davidchalmers42 @mbusigin i mean, i guess, but olfaction? feelings? it’s pretty different from the most common uses of the word cognition

2022-10-23 02:19:17 RT @GaryMarcus: @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg…

2022-10-23 02:12:43 @bradpwyble @pfau @ylecun @KordingLab @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher @davidpoeppel @TonyZador @AdamMarblestone sorry totally agree and meant to mark that (though it has been considered in neuroscience)

2022-10-23 02:11:40 RT @william_g_ray: @ESYudkowsky In 20 years, people will view content created before 2021 as pure and untainted by GPT. They will mine it i…

2022-10-23 01:49:33 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble eg see list here of many agreements https://t.co/Odg2QIt39X

2022-10-23 01:48:00 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble don’t believe the press clippings

2022-10-23 01:47:45 Funny how people keep thinking this when there is so much @ylecun and I agree on, eg. recently summarized here: https://t.co/Odg2QIt39X https://t.co/bpwARC92W8

2022-10-23 01:46:14 @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher and of course the notion of variable binding is critical but hardly widely embraced in MLam sure that eg @KordingLab @davidpoeppel @TonyZador @AdamMarblestone and many others can add to this list

2022-10-23 01:43:24 @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg Van Essen diagram) - power of dendrites (@mattlark/@YiotaPoirazi etc)- variety of neurons (@AllenInstitute)- areawise specialization (@Nancy_Kanwisher)- intrinsic cues &

2022-10-23 01:25:07 @pfau @ylecun @KordingLab @bradpwyble well, ok, that might be fair. i will go with “some ideas from the 60s had real impact”, but neuroscience hasn’t been a major driver since. (experience replay, and what else from neuroscience post 1970? not a huge list?)

2022-10-23 01:21:26 @lathropa @DeonTBenton one might have thought that since dall-e had both images and text that that visual input might have served as grounding, but of course compositionality didn’t emerge. (a couple years ago people said we just need our llms to have visual input…)

2022-10-23 01:17:31 @lathropa @DeonTBenton dall-e doesn’t refute statistical learning (it is consistent with it) nor the notion that statistical learning might contribute to language learning but rather causes problems for a theory in which statistical learning by itself magically gives you language.

2022-10-22 21:13:38 Have we made exponential progress towards Hawking’s notion of intelligence? I dare say no. https://t.co/WMJDWCjuke

2022-10-22 21:12:41 RT @IntuitMachine: @GaryMarcus @PaulTopping @bradpwyble @DeonTBenton "Intelligence is the ability to adapt to change." - Stephen Hawking

2022-10-22 21:08:55 @IntuitMachine @PaulTopping @bradpwyble @DeonTBenton for sure

2022-10-22 21:08:40 @DeonTBenton sure, and see also https://t.co/0rYVHqm8SK for a bias that can’t be explained in the same way.there is both learning and innateness. you haven’t said anything that convinces me otherwise, just given me more evidence for learning (that as it happens I already knew)

2022-10-22 02:07:26 @quocleix there is an error somewhere. Hinton is born in 1947, not in 1963. may i assume that the error is in the tweet rather than the paper?

2022-10-21 21:52:16 RT @vdignum: @GaryMarcus @nirsd Working on moving #AI beyond current #ML practices and approaches is not only a good idea but a much needed…

2022-10-21 21:21:12 @Grady_Booch @davidchalmers42 fyi

2022-10-21 21:00:07 @LucaAmb strawpersoning someone who says “maybe” and “hints” as concluding anything prematurely is a bit unfair, no?

2022-10-21 20:55:00 @LucaAmb i raised a question and said i saw two hints

2022-10-21 19:41:41 @LucaAmb gee, if it were a deep learning demo you’d say “it’s early days, give them a chance”

2022-10-21 18:01:14 @PaulTopping @kohn_gregory @TonyZador @sanewman1 @ehud @SpeciesTypical or mansplaining masked in philosophical jargon? hard to say.

2022-10-21 18:00:14 Retweeting this tweet from four years ago because it perfectly anticipated @ylecun’s three stages of grief. https://t.co/lgoWfHjQBU

2022-10-21 17:59:06 RT @atldgtltrnsfnug: “In defense of skepticism about deep learning” by @GaryMarcus https://t.co/eYns9BWklh“All truth passes through three…

2022-10-21 17:57:40 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical and not the DNA itself?? just the reading process? what are you saying??

2022-10-21 17:56:54 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical do you honestly think that i don’t that? or that is any different from an unplayed record? or unread book?

2022-10-21 17:54:59 @SpeciesTypical @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory i have never blocked someone simply for being obtuse before but I am feeling close…

2022-10-21 17:54:36 @kohn_gregory @PaulTopping @ehud @sanewman1 @SpeciesTypical “not relevant until it participates” doesn’t mean it doesn’t exist!same thing with my record player groove

2022-10-21 17:34:40 Pretty prescient, no? https://t.co/lgoWfHjQBU

2022-10-21 17:17:32 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical and more seriously: https://t.co/JS6U3kWl1v

2022-10-21 17:16:51 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical nope! @kohn_gregory assures me there is no information in genes at all, &

2022-10-21 17:13:55 @WiringTheBrain @sanewman1 learning, clearly. because empiricism is all that is left!

2022-10-21 17:00:40 true, but fundamentally misleadinggenes aren’t blueprints. they don’t “directly” contain information for phenotypes—but they do contain an enormous amount of information that is relevant and causally influential. https://t.co/jH07H5Cjvh

2022-10-21 16:31:01 @math_dandy so now “deep” also encompasses a single layer, which was its historical point of contrast? and maybe literally all of statistics and computer science, too?

2022-10-21 15:53:02 so maybe you don’t need deep learning after all? just good statistics? 2nd hint of this in October… https://t.co/6Ks02OmNfr

2022-10-21 00:00:22 RT @ruchowdh: The 8 bit bias bounty is now live!! Thank you @Melissahei for the article on what the bounty program means in context of the…

2022-10-20 23:38:05 anyone want to organize a conference or conference session, red-teaming current AI, as suggest below by @nirsd? it’s a good idea. https://t.co/mTZGIrwqAi

2022-10-20 23:30:06 RT @kanjun: Today, AI systems can create stunning art &

2022-10-20 23:26:04 @nirsd i have actually seen such things advertised and honestly don’t know how you proceed if you don’t know what you are trying to fix

2022-10-20 23:24:01 RT @dcli: saw ted chiang give a talk yesterday where he basically told a bunch of ai bros, "we are nowhere near having real ai, and what we…

2022-10-20 23:14:20 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb and repeating since you missed it the first time: https://t.co/5MjfJ1h6k6

2022-10-20 23:13:41 @berent_iris for sure. no competition there

2022-10-20 23:09:55 @berent_iris also mouse wins on 3d understanding of where objects are

2022-10-20 23:07:52 @berent_iris mice have mediocre vision. so maybe not (of course they also use smell etc): but they beat models by a lot in knowing what to do with objects that they do recognize

2022-10-20 21:36:52 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb put differently, if a n-gram model (with very large n) got compositional syntax right 66% of the time, it wouldn't mean the n-gram model had compositionality, would it?

2022-10-20 21:35:34 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb whatever they get or don't get is parasitic on corpus stats. on the image synthesis stuff some colleagues and i have data we will report shortly and feel really confident that you are mistaking getting things right for the wrong reason with actually get them right.

2022-10-20 19:36:08 excellent work, offering a different and important take on compositionality

2022-10-20 19:30:56 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb another great example of the mappings are naive: https://t.co/d1bO9t8Spj

2022-10-20 19:28:50 @ihorgowda if herschel is our goal, we are doomed

2022-10-20 14:53:10 where is AI relative to mouse cognition? excellent thread! mouse experts please add to it! https://t.co/71a6kQJbqy

2022-10-20 14:48:26 @kohn_gregory @SpeciesTypical @sanewman1 that’s silly. honestly we have nothing to talk about

2022-10-20 14:43:41 @kohn_gregory @sanewman1 @SpeciesTypical not seeing the argument there, but yes i was trying to make a more sophisticated nativism.i am sure the @WiringTheBrain would say the subsequent 17 years of dev neuro have been kind to that position: eg https://t.co/bUgW8d0XTQ &

2022-10-20 14:40:08 @kohn_gregory @SpeciesTypical @sanewman1 this is confused. it is like saying there is no information in an LP, because you need to have a record player to play it.if you flatten the groove, you lose the info

2022-10-20 14:37:13 @kohn_gregory @SpeciesTypical umm… we were talking about the physical structure of the hand, and you have swapped to the motor programs controlling it. obviously the latter involves learning.

2022-10-20 14:33:19 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb @ZoubinGhahrama1 should give @TristanThrush access so Tristan can run a Winoground, then.

2022-10-20 14:31:17 and a pioneer in common sense reasoning! and now on Twitter: @ErnestSDavis https://t.co/3VWbdX2wTi

2022-10-20 08:09:29 RT @BarryOSullivan: I highly recommend this talk! And check out Ernie’s excellent book with @GaryMarcus: “Rebooting AI: Building Artificial…

2022-10-20 07:40:51 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb also note that @slatestarcodex initially declared victory in a bet re Imagen, and then retracted. i will believe that Imagen is real progress on compositionality only when it is carefully demonstrated.

2022-10-20 07:30:29 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i have asked them endlessly to make it available or test the Winoground benchmark. they have refused. i can speak w authority only re systems i (or others) can test. but remain skeptical about those hidden from testing.

2022-10-20 07:25:16 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb ie it is this claim above on image synthesis that i am questioning w.r.t. grammar: https://t.co/QLqpvsjqyL

2022-10-20 07:23:02 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i was responding to a tweet that seemed to be about image synthesis

2022-10-20 07:02:11 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb best evidence that an image synthesis system captures some grammar? they seem to have some serious problems with binding.

2022-10-20 04:02:23 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb correction to last sentence: even when it gets it right it is bloviating. in no case does LLM do inference and reasoning over a world model

2022-10-20 03:59:45 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb they composit but without proper variable binding they aren’t properly compositional. they also fail on part-whole relationships, function, etcin some ways the language side is akin to bag-of-words model, which represent language in a very impoverished way.

2022-10-20 03:45:25 RT @databu: Interesting discussion with @GaryMarcus and @bengoertzel there

2022-10-20 03:21:11 @pmddomingos @erikbryn without social media the riot probably wouldn’t have happened

2022-10-30 18:28:18 Elon Musk: “Twitter obviously cannot become a free-for-all hellscape.”Also Elon Musk: https://t.co/sRkeYww7ve

2022-10-30 13:44:12 quietly growing, and poised to change everything. https://t.co/9847MKRfDz

2022-10-30 01:02:04 RT @foysavas: Finally read https://t.co/iZIUUNGKbR by @GaryMarcus and Ernest Davis and it's a greatly accessible critique of contemporary A…

2022-10-29 19:45:43 AI fads come and go (anyone remember expert systems? SVMs?) Here’s are three ways that Scaling Maximalism, theory du jour, might end.https://t.co/jSFOVZQQ10

2022-10-29 14:29:57 and of course the thing to read on causality, in addition to what we say, is @yudapearl’s The Book of Why.And yes Kant was there long before we were :)

2022-10-29 14:27:38 actually “our suggestion”

2022-10-29 14:22:52 My suggestion for what AI needs, 2019. Nothing has changed. https://t.co/8aHX6Gd3HT

2022-10-29 13:48:21 RT @AlisonBLowndes: Looking forward to this @GaryMarcus. Tune in all, 4.15pm UK (Lisbon) Nov 4: https://t.co/RnuXORRmPC

2022-10-29 13:39:15 @LauraRuis @Isinlor I am less sanguine. I wouldn’t be surprised to see improvements on this or that benchmark, but pragmatics fundamentally requires reasoning about models of the world (eg other’s beliefs), and I see no reason to think that LLMs their own can reliably construct such things.

2022-10-29 01:03:47 @albertwebson @_jasonwei @tallinzen skeptical of 1, but keen to see the argumentyou may have seen my new work w @EvelinaLeivada and @ElliotMurphy91, different architecture but perhaps relevant:https://t.co/UWUg6kQ3Df

2022-10-29 00:46:43 @albertwebson @_jasonwei @tallinzen I don’t think all forms of intelligence need to do arithmetic. but these kinds of errors give some insight in a semi-controlled way into a deeper, not yet solved challenge of translating language into models of the world. curious to see the paper when you can share

2022-10-28 21:32:45 RT @egrefen: I'm biased, obviously, but I think this is *such* an important result. Implicature and pragmatics are essential, ubiquitous, a…

2022-10-28 17:50:52 Just rehearsed Noam Chomsky and @jeremyakahn Nov 4, livestreamed at https://t.co/vQcEjAQCtg https://t.co/k93hY3EmVu

2022-10-28 14:18:18 indeed, slashing moderation just as synthetic content becomes widespread does seem like a disaster-in-the-making https://t.co/gEMMehatm7

2022-10-28 10:49:43 AGI, my eye. https://t.co/EmVNJRUzpz

2022-10-28 10:47:51 @L_andreae @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab agree w @L_andreae, eg a classic paper by Thelen showed that newborn humans will step, and that is really an issue of weight etc rather than lack of motor program per se: https://t.co/09LnIrfIqf

2022-10-28 10:39:40 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab (i pasted the image above/here it is) https://t.co/JjMRtntwN9

2022-10-28 10:29:06 @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab presumably both

2022-10-28 10:21:19 @L_andreae @WiringTheBrain my books https://t.co/Pt7HZc3RJd, Kluge, and The Birth of The Mind

2022-10-28 10:13:59 RT @Abebab: the most insightful knowledge about a field comes from experts in the field that hold a healthy skeptical view on the topic and…

2022-10-28 10:08:07 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab beautiful new review, by @neuroluci. how do you even think about this stuff if you are a staunch anti-nativist? https://t.co/RvacUhsspE

2022-10-28 10:01:09 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab hmm, figure 2 from Verhage et al certainly speaks to those questions (# and pattern) to some degree

2022-10-28 09:56:27 @AthenaAI2 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it goes to showing how little we understand the brain &

2022-10-28 09:50:19 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab agreed! and yet you wouldn’t know that from reading a lot of philosophers and AI researchers these days. &

2022-10-28 09:39:10 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it seems to me then that the reasonable conclusion is same as what I wrote in my lay summary in 2004: brain development uses a mix of activity-dependent (some not all intrinsically guided) &

2022-10-28 09:13:25 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab not mentioned but also seemingly relevant is Verhage et al (2000), esp figure 2 (“Correct assembly of the brain the absence of neurotransmitter secretion”) below. Anyone in this thread have thoughts on it? https://t.co/6O4OlOjoQZ

2022-10-27 18:46:45 @fhuszar https://t.co/UWUg6kxUp7

2022-10-27 18:20:33 if i had a dollar for every time someone cherry-picked DALL-E 2 art, I coulda outbid Elon for Twitter https://t.co/kHe92VtnFB

2022-10-27 15:49:18 @ylecun @ProfNoahGian so you have never done your own ground truth study and are relying on those outside orgs?

2022-10-27 15:37:48 @ylecun @ProfNoahGian our messages crossed

2022-10-27 15:31:15 @ylecun @ProfNoahGian that doesn’t answer @ProfNoahGianyou need a (large) sample that is carefully human reviewed ground truth. in that ground truth, two questions:1. what % would be detected by the AI?2.what % of what is missed by the AI is flagged by humans?you are addressing #2, not #1

2022-10-27 14:58:35 As ever, @KBAndersen has his finger on the pulse. (Answer to his question in the comments, in case you need it.) https://t.co/gqwcRACOmx

2022-10-24 02:35:23 there are two kinds of people in this world.those brave enough to speak truth to powerand those who are not.

2022-10-24 02:31:32 @primalpoly @AmandaAskell @AlecMacGillis @JoshuaChaffin i agree. I would put a lot more trust in a single anecdote involving tricking poor people into flying to Martha’s Vineyard for nonexistent jobs.

2022-10-24 01:47:36 @pmddomingos “And if there's war between the sexesThen there'll be no people left”

2022-10-24 01:39:40 @filippie509 @KordingLab @TonyZador

2022-10-24 01:21:55 @filippie509 it also doesn’t cite my endless arguments for years about looking to cognitive science, or even my paper with marblestone in Science…

2022-10-24 00:37:11 @voxbec @davidchalmers42 @mbusigin not disputing this but many people would take these to be influences on cognition. as i said the list is odd, from how people typically use the term, not necessarily wrong. in general people often distinguish between perception (eg olfaction, low level vision) and cognition.

2022-10-23 23:56:44 RT @AmandaAskell: If you don't have time to provide any evidence that someone is a eugenicist (beyond alluding to articles that also don't…

2022-10-23 22:55:22 @GaneshNatesh @ceobillionaire there has been zero budget thus far, zero charge. all participants (including us) have donated their time.

2022-10-23 22:10:34 . @ceobillionaire and I are pleased to announce that we will be resuming our AI debate series, December 23. Mark your calendar, for some exciting conversations, and reply below to suggest participants. (We already have our own ideas, but might have room for a couple more.)

2022-10-23 22:07:25 In the real world of AGI, nobody knows what to do, nor how to do it. https://t.co/4DPbeIKz85

2022-10-23 21:59:17 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher here: https://t.co/5jrUODZuNt

2022-10-23 21:57:13 @Korrelan_AI @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher you need a world model for that, IMHO

2022-10-23 21:56:53 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher i saw a talk at NeurIPS c 2017 or 2018. don’t know how active it is

2022-10-23 21:47:41 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher well, if it’s created post-training but eg it’s snow white with elves rather than dwarves, i am not super impressed.but i think you are generally thinking about the right kind of issues

2022-10-23 21:44:47 . (But what a messed-up field if that actually guides people’s research choices.) https://t.co/rTXXHpKN3f

2022-10-23 21:43:10 @pmddomingos @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher link? i’d be curious

2022-10-23 21:41:34 @Peterbart @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher ever since Wittgenstein we’ve known that language doesn’t work that way… and I am happy in that corner

2022-10-23 21:40:33 @DeonTBenton the empiricist’s imagination knows no bounds.

2022-10-23 21:39:58 @AndrewLampinen @Peterbart @mpshanahan @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher fair :)

2022-10-23 21:39:44 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher just the general idea that we want generalization rather than mere memory. we might eg actually want to have two different tracks, short and open-length, tracking empirically how much progress on one leads to progress on the other.

2022-10-23 04:48:29 @NoraNewcombe you are *referencing* those distinctions, but I don’t see how their particular way of carving things up helps. as stated, something here needs be represented, and the format for that representation itself appears to be innate.

2022-10-23 04:41:20 @davidchalmers42 @mbusigin definitely grey and debatable but cognitive science actually seems broader than cognition/cognitive to me. (eg for me it clearly includes parts of philosophy of mind that i wouldn’t include in cognitive psychology per se). but there’s no fact of the matter certainly/YMMV.

2022-10-23 04:35:49 @DeonTBenton i even said “privately” ffs

2022-10-23 04:35:30 @DeonTBenton wow. you are a jerk, reposting private messages. i am going to block you and end our correspondance if you don’t delete that.

2022-10-23 02:21:41 @davidchalmers42 @mbusigin i mean, i guess, but olfaction? feelings? it’s pretty different from the most common uses of the word cognition

2022-10-23 02:19:17 RT @GaryMarcus: @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg…

2022-10-23 02:12:43 @bradpwyble @pfau @ylecun @KordingLab @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher @davidpoeppel @TonyZador @AdamMarblestone sorry totally agree and meant to mark that (though it has been considered in neuroscience)

2022-10-23 02:11:40 RT @william_g_ray: @ESYudkowsky In 20 years, people will view content created before 2021 as pure and untainted by GPT. They will mine it i…

2022-10-23 01:49:33 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble eg see list here of many agreements https://t.co/Odg2QIt39X

2022-10-23 01:48:00 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble don’t believe the press clippings

2022-10-23 01:47:45 Funny how people keep thinking this when there is so much @ylecun and I agree on, eg. recently summarized here: https://t.co/Odg2QIt39X https://t.co/bpwARC92W8

2022-10-23 01:46:14 @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher and of course the notion of variable binding is critical but hardly widely embraced in MLam sure that eg @KordingLab @davidpoeppel @TonyZador @AdamMarblestone and many others can add to this list

2022-10-23 01:43:24 @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg Van Essen diagram) - power of dendrites (@mattlark/@YiotaPoirazi etc)- variety of neurons (@AllenInstitute)- areawise specialization (@Nancy_Kanwisher)- intrinsic cues &

2022-10-23 01:25:07 @pfau @ylecun @KordingLab @bradpwyble well, ok, that might be fair. i will go with “some ideas from the 60s had real impact”, but neuroscience hasn’t been a major driver since. (experience replay, and what else from neuroscience post 1970? not a huge list?)

2022-10-23 01:21:26 @lathropa @DeonTBenton one might have thought that since dall-e had both images and text that that visual input might have served as grounding, but of course compositionality didn’t emerge. (a couple years ago people said we just need our llms to have visual input…)

2022-10-23 01:17:31 @lathropa @DeonTBenton dall-e doesn’t refute statistical learning (it is consistent with it) nor the notion that statistical learning might contribute to language learning but rather causes problems for a theory in which statistical learning by itself magically gives you language.

2022-10-22 21:13:38 Have we made exponential progress towards Hawking’s notion of intelligence? I dare say no. https://t.co/WMJDWCjuke

2022-10-22 21:12:41 RT @IntuitMachine: @GaryMarcus @PaulTopping @bradpwyble @DeonTBenton "Intelligence is the ability to adapt to change." - Stephen Hawking

2022-10-22 21:08:55 @IntuitMachine @PaulTopping @bradpwyble @DeonTBenton for sure

2022-10-22 21:08:40 @DeonTBenton sure, and see also https://t.co/0rYVHqm8SK for a bias that can’t be explained in the same way.there is both learning and innateness. you haven’t said anything that convinces me otherwise, just given me more evidence for learning (that as it happens I already knew)

2022-10-22 02:07:26 @quocleix there is an error somewhere. Hinton is born in 1947, not in 1963. may i assume that the error is in the tweet rather than the paper?

2022-10-21 21:52:16 RT @vdignum: @GaryMarcus @nirsd Working on moving #AI beyond current #ML practices and approaches is not only a good idea but a much needed…

2022-10-21 21:21:12 @Grady_Booch @davidchalmers42 fyi

2022-10-21 21:00:07 @LucaAmb strawpersoning someone who says “maybe” and “hints” as concluding anything prematurely is a bit unfair, no?

2022-10-21 20:55:00 @LucaAmb i raised a question and said i saw two hints

2022-10-21 19:41:41 @LucaAmb gee, if it were a deep learning demo you’d say “it’s early days, give them a chance”

2022-10-21 18:01:14 @PaulTopping @kohn_gregory @TonyZador @sanewman1 @ehud @SpeciesTypical or mansplaining masked in philosophical jargon? hard to say.

2022-10-21 18:00:14 Retweeting this tweet from four years ago because it perfectly anticipated @ylecun’s three stages of grief. https://t.co/lgoWfHjQBU

2022-10-21 17:59:06 RT @atldgtltrnsfnug: “In defense of skepticism about deep learning” by @GaryMarcus https://t.co/eYns9BWklh“All truth passes through three…

2022-10-21 17:57:40 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical and not the DNA itself?? just the reading process? what are you saying??

2022-10-21 17:56:54 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical do you honestly think that i don’t that? or that is any different from an unplayed record? or unread book?

2022-10-21 17:54:59 @SpeciesTypical @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory i have never blocked someone simply for being obtuse before but I am feeling close…

2022-10-21 17:54:36 @kohn_gregory @PaulTopping @ehud @sanewman1 @SpeciesTypical “not relevant until it participates” doesn’t mean it doesn’t exist!same thing with my record player groove

2022-10-21 17:34:40 Pretty prescient, no? https://t.co/lgoWfHjQBU

2022-10-21 17:17:32 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical and more seriously: https://t.co/JS6U3kWl1v

2022-10-21 17:16:51 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical nope! @kohn_gregory assures me there is no information in genes at all, &

2022-10-21 17:13:55 @WiringTheBrain @sanewman1 learning, clearly. because empiricism is all that is left!

2022-10-21 17:00:40 true, but fundamentally misleadinggenes aren’t blueprints. they don’t “directly” contain information for phenotypes—but they do contain an enormous amount of information that is relevant and causally influential. https://t.co/jH07H5Cjvh

2022-10-21 16:31:01 @math_dandy so now “deep” also encompasses a single layer, which was its historical point of contrast? and maybe literally all of statistics and computer science, too?

2022-10-21 15:53:02 so maybe you don’t need deep learning after all? just good statistics? 2nd hint of this in October… https://t.co/6Ks02OmNfr

2022-10-21 00:00:22 RT @ruchowdh: The 8 bit bias bounty is now live!! Thank you @Melissahei for the article on what the bounty program means in context of the…

2022-10-20 23:38:05 anyone want to organize a conference or conference session, red-teaming current AI, as suggest below by @nirsd? it’s a good idea. https://t.co/mTZGIrwqAi

2022-10-20 23:30:06 RT @kanjun: Today, AI systems can create stunning art &

2022-10-20 23:26:04 @nirsd i have actually seen such things advertised and honestly don’t know how you proceed if you don’t know what you are trying to fix

2022-10-20 23:24:01 RT @dcli: saw ted chiang give a talk yesterday where he basically told a bunch of ai bros, "we are nowhere near having real ai, and what we…

2022-10-20 23:14:20 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb and repeating since you missed it the first time: https://t.co/5MjfJ1h6k6

2022-10-20 23:13:41 @berent_iris for sure. no competition there

2022-10-20 23:09:55 @berent_iris also mouse wins on 3d understanding of where objects are

2022-10-20 23:07:52 @berent_iris mice have mediocre vision. so maybe not (of course they also use smell etc): but they beat models by a lot in knowing what to do with objects that they do recognize

2022-10-20 21:36:52 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb put differently, if a n-gram model (with very large n) got compositional syntax right 66% of the time, it wouldn't mean the n-gram model had compositionality, would it?

2022-10-20 21:35:34 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb whatever they get or don't get is parasitic on corpus stats. on the image synthesis stuff some colleagues and i have data we will report shortly and feel really confident that you are mistaking getting things right for the wrong reason with actually get them right.

2022-10-20 19:36:08 excellent work, offering a different and important take on compositionality

2022-10-20 19:30:56 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb another great example of the mappings are naive: https://t.co/d1bO9t8Spj

2022-10-20 19:28:50 @ihorgowda if herschel is our goal, we are doomed

2022-10-20 14:53:10 where is AI relative to mouse cognition? excellent thread! mouse experts please add to it! https://t.co/71a6kQJbqy

2022-10-20 14:48:26 @kohn_gregory @SpeciesTypical @sanewman1 that’s silly. honestly we have nothing to talk about

2022-10-20 14:43:41 @kohn_gregory @sanewman1 @SpeciesTypical not seeing the argument there, but yes i was trying to make a more sophisticated nativism.i am sure the @WiringTheBrain would say the subsequent 17 years of dev neuro have been kind to that position: eg https://t.co/bUgW8d0XTQ &

2022-10-20 14:40:08 @kohn_gregory @SpeciesTypical @sanewman1 this is confused. it is like saying there is no information in an LP, because you need to have a record player to play it.if you flatten the groove, you lose the info

2022-10-20 14:37:13 @kohn_gregory @SpeciesTypical umm… we were talking about the physical structure of the hand, and you have swapped to the motor programs controlling it. obviously the latter involves learning.

2022-10-20 14:33:19 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb @ZoubinGhahrama1 should give @TristanThrush access so Tristan can run a Winoground, then.

2022-10-20 14:31:17 and a pioneer in common sense reasoning! and now on Twitter: @ErnestSDavis https://t.co/3VWbdX2wTi

2022-10-20 08:09:29 RT @BarryOSullivan: I highly recommend this talk! And check out Ernie’s excellent book with @GaryMarcus: “Rebooting AI: Building Artificial…

2022-10-20 07:40:51 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb also note that @slatestarcodex initially declared victory in a bet re Imagen, and then retracted. i will believe that Imagen is real progress on compositionality only when it is carefully demonstrated.

2022-10-20 07:30:29 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i have asked them endlessly to make it available or test the Winoground benchmark. they have refused. i can speak w authority only re systems i (or others) can test. but remain skeptical about those hidden from testing.

2022-10-20 07:25:16 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb ie it is this claim above on image synthesis that i am questioning w.r.t. grammar: https://t.co/QLqpvsjqyL

2022-10-20 07:23:02 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i was responding to a tweet that seemed to be about image synthesis

2022-10-20 07:02:11 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb best evidence that an image synthesis system captures some grammar? they seem to have some serious problems with binding.

2022-10-20 04:02:23 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb correction to last sentence: even when it gets it right it is bloviating. in no case does LLM do inference and reasoning over a world model

2022-10-20 03:59:45 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb they composit but without proper variable binding they aren’t properly compositional. they also fail on part-whole relationships, function, etcin some ways the language side is akin to bag-of-words model, which represent language in a very impoverished way.

2022-10-20 03:45:25 RT @databu: Interesting discussion with @GaryMarcus and @bengoertzel there

2022-10-20 03:21:11 @pmddomingos @erikbryn without social media the riot probably wouldn’t have happened

2022-10-30 18:28:18 Elon Musk: “Twitter obviously cannot become a free-for-all hellscape.”Also Elon Musk: https://t.co/sRkeYww7ve

2022-10-30 13:44:12 quietly growing, and poised to change everything. https://t.co/9847MKRfDz

2022-10-30 01:02:04 RT @foysavas: Finally read https://t.co/iZIUUNGKbR by @GaryMarcus and Ernest Davis and it's a greatly accessible critique of contemporary A…

2022-10-29 19:45:43 AI fads come and go (anyone remember expert systems? SVMs?) Here’s are three ways that Scaling Maximalism, theory du jour, might end.https://t.co/jSFOVZQQ10

2022-10-29 14:29:57 and of course the thing to read on causality, in addition to what we say, is @yudapearl’s The Book of Why.And yes Kant was there long before we were :)

2022-10-29 14:27:38 actually “our suggestion”

2022-10-29 14:22:52 My suggestion for what AI needs, 2019. Nothing has changed. https://t.co/8aHX6Gd3HT

2022-10-29 13:48:21 RT @AlisonBLowndes: Looking forward to this @GaryMarcus. Tune in all, 4.15pm UK (Lisbon) Nov 4: https://t.co/RnuXORRmPC

2022-10-29 13:39:15 @LauraRuis @Isinlor I am less sanguine. I wouldn’t be surprised to see improvements on this or that benchmark, but pragmatics fundamentally requires reasoning about models of the world (eg other’s beliefs), and I see no reason to think that LLMs their own can reliably construct such things.

2022-10-29 01:03:47 @albertwebson @_jasonwei @tallinzen skeptical of 1, but keen to see the argumentyou may have seen my new work w @EvelinaLeivada and @ElliotMurphy91, different architecture but perhaps relevant:https://t.co/UWUg6kQ3Df

2022-10-29 00:46:43 @albertwebson @_jasonwei @tallinzen I don’t think all forms of intelligence need to do arithmetic. but these kinds of errors give some insight in a semi-controlled way into a deeper, not yet solved challenge of translating language into models of the world. curious to see the paper when you can share

2022-10-28 21:32:45 RT @egrefen: I'm biased, obviously, but I think this is *such* an important result. Implicature and pragmatics are essential, ubiquitous, a…

2022-10-28 17:50:52 Just rehearsed Noam Chomsky and @jeremyakahn Nov 4, livestreamed at https://t.co/vQcEjAQCtg https://t.co/k93hY3EmVu

2022-10-28 14:18:18 indeed, slashing moderation just as synthetic content becomes widespread does seem like a disaster-in-the-making https://t.co/gEMMehatm7

2022-10-28 10:49:43 AGI, my eye. https://t.co/EmVNJRUzpz

2022-10-28 10:47:51 @L_andreae @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab agree w @L_andreae, eg a classic paper by Thelen showed that newborn humans will step, and that is really an issue of weight etc rather than lack of motor program per se: https://t.co/09LnIrfIqf

2022-10-28 10:39:40 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab (i pasted the image above/here it is) https://t.co/JjMRtntwN9

2022-10-28 10:29:06 @IntuitMachine @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab presumably both

2022-10-28 10:21:19 @L_andreae @WiringTheBrain my books https://t.co/Pt7HZc3RJd, Kluge, and The Birth of The Mind

2022-10-28 10:13:59 RT @Abebab: the most insightful knowledge about a field comes from experts in the field that hold a healthy skeptical view on the topic and…

2022-10-28 10:08:07 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab beautiful new review, by @neuroluci. how do you even think about this stuff if you are a staunch anti-nativist? https://t.co/RvacUhsspE

2022-10-28 10:01:09 @L_andreae @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @Datta_Lab hmm, figure 2 from Verhage et al certainly speaks to those questions (# and pattern) to some degree

2022-10-28 09:56:27 @AthenaAI2 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it goes to showing how little we understand the brain &

2022-10-28 09:50:19 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab agreed! and yet you wouldn’t know that from reading a lot of philosophers and AI researchers these days. &

2022-10-28 09:39:10 @bdanubius @WiringTheBrain @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab it seems to me then that the reasonable conclusion is same as what I wrote in my lay summary in 2004: brain development uses a mix of activity-dependent (some not all intrinsically guided) &

2022-10-28 09:13:25 @WiringTheBrain @bdanubius @sainsbury_tom @skeptencephalon @kendmil @behrenstimb @TonyZador @DrYohanJohn @drmichaellevin @StevenQuartz @prokraustinator @BeckerLabCRTD @Christakou @michael_nyaga @L_andreae @Datta_Lab not mentioned but also seemingly relevant is Verhage et al (2000), esp figure 2 (“Correct assembly of the brain the absence of neurotransmitter secretion”) below. Anyone in this thread have thoughts on it? https://t.co/6O4OlOjoQZ

2022-10-27 18:46:45 @fhuszar https://t.co/UWUg6kxUp7

2022-10-27 18:20:33 if i had a dollar for every time someone cherry-picked DALL-E 2 art, I coulda outbid Elon for Twitter https://t.co/kHe92VtnFB

2022-10-27 15:49:18 @ylecun @ProfNoahGian so you have never done your own ground truth study and are relying on those outside orgs?

2022-10-27 15:37:48 @ylecun @ProfNoahGian our messages crossed

2022-10-27 15:31:15 @ylecun @ProfNoahGian that doesn’t answer @ProfNoahGianyou need a (large) sample that is carefully human reviewed ground truth. in that ground truth, two questions:1. what % would be detected by the AI?2.what % of what is missed by the AI is flagged by humans?you are addressing #2, not #1

2022-10-27 14:58:35 As ever, @KBAndersen has his finger on the pulse. (Answer to his question in the comments, in case you need it.) https://t.co/gqwcRACOmx

2022-10-24 02:35:23 there are two kinds of people in this world.those brave enough to speak truth to powerand those who are not.

2022-10-24 02:31:32 @primalpoly @AmandaAskell @AlecMacGillis @JoshuaChaffin i agree. I would put a lot more trust in a single anecdote involving tricking poor people into flying to Martha’s Vineyard for nonexistent jobs.

2022-10-24 01:47:36 @pmddomingos “And if there's war between the sexesThen there'll be no people left”

2022-10-24 01:39:40 @filippie509 @KordingLab @TonyZador

2022-10-24 01:21:55 @filippie509 it also doesn’t cite my endless arguments for years about looking to cognitive science, or even my paper with marblestone in Science…

2022-10-24 00:37:11 @voxbec @davidchalmers42 @mbusigin not disputing this but many people would take these to be influences on cognition. as i said the list is odd, from how people typically use the term, not necessarily wrong. in general people often distinguish between perception (eg olfaction, low level vision) and cognition.

2022-10-23 23:56:44 RT @AmandaAskell: If you don't have time to provide any evidence that someone is a eugenicist (beyond alluding to articles that also don't…

2022-10-23 22:55:22 @GaneshNatesh @ceobillionaire there has been zero budget thus far, zero charge. all participants (including us) have donated their time.

2022-10-23 22:10:34 . @ceobillionaire and I are pleased to announce that we will be resuming our AI debate series, December 23. Mark your calendar, for some exciting conversations, and reply below to suggest participants. (We already have our own ideas, but might have room for a couple more.)

2022-10-23 22:07:25 In the real world of AGI, nobody knows what to do, nor how to do it. https://t.co/4DPbeIKz85

2022-10-23 21:59:17 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher here: https://t.co/5jrUODZuNt

2022-10-23 21:57:13 @Korrelan_AI @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher you need a world model for that, IMHO

2022-10-23 21:56:53 @AndrewLampinen @pmddomingos @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher i saw a talk at NeurIPS c 2017 or 2018. don’t know how active it is

2022-10-23 21:47:41 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher well, if it’s created post-training but eg it’s snow white with elves rather than dwarves, i am not super impressed.but i think you are generally thinking about the right kind of issues

2022-10-23 21:44:47 . (But what a messed-up field if that actually guides people’s research choices.) https://t.co/rTXXHpKN3f

2022-10-23 21:43:10 @pmddomingos @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher link? i’d be curious

2022-10-23 21:41:34 @Peterbart @mpshanahan @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher ever since Wittgenstein we’ve known that language doesn’t work that way… and I am happy in that corner

2022-10-23 21:40:33 @DeonTBenton the empiricist’s imagination knows no bounds.

2022-10-23 21:39:58 @AndrewLampinen @Peterbart @mpshanahan @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher fair :)

2022-10-23 21:39:44 @AndrewLampinen @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher just the general idea that we want generalization rather than mere memory. we might eg actually want to have two different tracks, short and open-length, tracking empirically how much progress on one leads to progress on the other.

2022-10-23 04:48:29 @NoraNewcombe you are *referencing* those distinctions, but I don’t see how their particular way of carving things up helps. as stated, something here needs be represented, and the format for that representation itself appears to be innate.

2022-10-23 04:41:20 @davidchalmers42 @mbusigin definitely grey and debatable but cognitive science actually seems broader than cognition/cognitive to me. (eg for me it clearly includes parts of philosophy of mind that i wouldn’t include in cognitive psychology per se). but there’s no fact of the matter certainly/YMMV.

2022-10-23 04:35:49 @DeonTBenton i even said “privately” ffs

2022-10-23 04:35:30 @DeonTBenton wow. you are a jerk, reposting private messages. i am going to block you and end our correspondance if you don’t delete that.

2022-10-23 02:21:41 @davidchalmers42 @mbusigin i mean, i guess, but olfaction? feelings? it’s pretty different from the most common uses of the word cognition

2022-10-23 02:19:17 RT @GaryMarcus: @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg…

2022-10-23 02:12:43 @bradpwyble @pfau @ylecun @KordingLab @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher @davidpoeppel @TonyZador @AdamMarblestone sorry totally agree and meant to mark that (though it has been considered in neuroscience)

2022-10-23 02:11:40 RT @william_g_ray: @ESYudkowsky In 20 years, people will view content created before 2021 as pure and untainted by GPT. They will mine it i…

2022-10-23 01:49:33 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble eg see list here of many agreements https://t.co/Odg2QIt39X

2022-10-23 01:48:00 @sreejan_kumar @ylecun @pfau @KordingLab @bradpwyble don’t believe the press clippings

2022-10-23 01:47:45 Funny how people keep thinking this when there is so much @ylecun and I agree on, eg. recently summarized here: https://t.co/Odg2QIt39X https://t.co/bpwARC92W8

2022-10-23 01:46:14 @pfau @ylecun @KordingLab @bradpwyble @mattlark @YiotaPoirazi @AllenInstitute @Nancy_Kanwisher and of course the notion of variable binding is critical but hardly widely embraced in MLam sure that eg @KordingLab @davidpoeppel @TonyZador @AdamMarblestone and many others can add to this list

2022-10-23 01:43:24 @pfau @ylecun @KordingLab @bradpwyble ideas from neurosci yet to be fully embraced in ML:- massive amount of structure (eg Van Essen diagram) - power of dendrites (@mattlark/@YiotaPoirazi etc)- variety of neurons (@AllenInstitute)- areawise specialization (@Nancy_Kanwisher)- intrinsic cues &

2022-10-23 01:25:07 @pfau @ylecun @KordingLab @bradpwyble well, ok, that might be fair. i will go with “some ideas from the 60s had real impact”, but neuroscience hasn’t been a major driver since. (experience replay, and what else from neuroscience post 1970? not a huge list?)

2022-10-23 01:21:26 @lathropa @DeonTBenton one might have thought that since dall-e had both images and text that that visual input might have served as grounding, but of course compositionality didn’t emerge. (a couple years ago people said we just need our llms to have visual input…)

2022-10-23 01:17:31 @lathropa @DeonTBenton dall-e doesn’t refute statistical learning (it is consistent with it) nor the notion that statistical learning might contribute to language learning but rather causes problems for a theory in which statistical learning by itself magically gives you language.

2022-10-22 21:13:38 Have we made exponential progress towards Hawking’s notion of intelligence? I dare say no. https://t.co/WMJDWCjuke

2022-10-22 21:12:41 RT @IntuitMachine: @GaryMarcus @PaulTopping @bradpwyble @DeonTBenton "Intelligence is the ability to adapt to change." - Stephen Hawking

2022-10-22 21:08:55 @IntuitMachine @PaulTopping @bradpwyble @DeonTBenton for sure

2022-10-22 21:08:40 @DeonTBenton sure, and see also https://t.co/0rYVHqm8SK for a bias that can’t be explained in the same way.there is both learning and innateness. you haven’t said anything that convinces me otherwise, just given me more evidence for learning (that as it happens I already knew)

2022-10-22 02:07:26 @quocleix there is an error somewhere. Hinton is born in 1947, not in 1963. may i assume that the error is in the tweet rather than the paper?

2022-10-21 21:52:16 RT @vdignum: @GaryMarcus @nirsd Working on moving #AI beyond current #ML practices and approaches is not only a good idea but a much needed…

2022-10-21 21:21:12 @Grady_Booch @davidchalmers42 fyi

2022-10-21 21:00:07 @LucaAmb strawpersoning someone who says “maybe” and “hints” as concluding anything prematurely is a bit unfair, no?

2022-10-21 20:55:00 @LucaAmb i raised a question and said i saw two hints

2022-10-21 19:41:41 @LucaAmb gee, if it were a deep learning demo you’d say “it’s early days, give them a chance”

2022-10-21 18:01:14 @PaulTopping @kohn_gregory @TonyZador @sanewman1 @ehud @SpeciesTypical or mansplaining masked in philosophical jargon? hard to say.

2022-10-21 18:00:14 Retweeting this tweet from four years ago because it perfectly anticipated @ylecun’s three stages of grief. https://t.co/lgoWfHjQBU

2022-10-21 17:59:06 RT @atldgtltrnsfnug: “In defense of skepticism about deep learning” by @GaryMarcus https://t.co/eYns9BWklh“All truth passes through three…

2022-10-21 17:57:40 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical and not the DNA itself?? just the reading process? what are you saying??

2022-10-21 17:56:54 @kohn_gregory @TonyZador @sanewman1 @PaulTopping @ehud @SpeciesTypical do you honestly think that i don’t that? or that is any different from an unplayed record? or unread book?

2022-10-21 17:54:59 @SpeciesTypical @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory i have never blocked someone simply for being obtuse before but I am feeling close…

2022-10-21 17:54:36 @kohn_gregory @PaulTopping @ehud @sanewman1 @SpeciesTypical “not relevant until it participates” doesn’t mean it doesn’t exist!same thing with my record player groove

2022-10-21 17:34:40 Pretty prescient, no? https://t.co/lgoWfHjQBU

2022-10-21 17:17:32 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical and more seriously: https://t.co/JS6U3kWl1v

2022-10-21 17:16:51 @TonyZador @sanewman1 @PaulTopping @ehud @kohn_gregory @SpeciesTypical nope! @kohn_gregory assures me there is no information in genes at all, &

2022-10-21 17:13:55 @WiringTheBrain @sanewman1 learning, clearly. because empiricism is all that is left!

2022-10-21 17:00:40 true, but fundamentally misleadinggenes aren’t blueprints. they don’t “directly” contain information for phenotypes—but they do contain an enormous amount of information that is relevant and causally influential. https://t.co/jH07H5Cjvh

2022-10-21 16:31:01 @math_dandy so now “deep” also encompasses a single layer, which was its historical point of contrast? and maybe literally all of statistics and computer science, too?

2022-10-21 15:53:02 so maybe you don’t need deep learning after all? just good statistics? 2nd hint of this in October… https://t.co/6Ks02OmNfr

2022-10-21 00:00:22 RT @ruchowdh: The 8 bit bias bounty is now live!! Thank you @Melissahei for the article on what the bounty program means in context of the…

2022-10-20 23:38:05 anyone want to organize a conference or conference session, red-teaming current AI, as suggest below by @nirsd? it’s a good idea. https://t.co/mTZGIrwqAi

2022-10-20 23:30:06 RT @kanjun: Today, AI systems can create stunning art &

2022-10-20 23:26:04 @nirsd i have actually seen such things advertised and honestly don’t know how you proceed if you don’t know what you are trying to fix

2022-10-20 23:24:01 RT @dcli: saw ted chiang give a talk yesterday where he basically told a bunch of ai bros, "we are nowhere near having real ai, and what we…

2022-10-20 23:14:20 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb and repeating since you missed it the first time: https://t.co/5MjfJ1h6k6

2022-10-20 23:13:41 @berent_iris for sure. no competition there

2022-10-20 23:09:55 @berent_iris also mouse wins on 3d understanding of where objects are

2022-10-20 23:07:52 @berent_iris mice have mediocre vision. so maybe not (of course they also use smell etc): but they beat models by a lot in knowing what to do with objects that they do recognize

2022-10-20 21:36:52 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb put differently, if a n-gram model (with very large n) got compositional syntax right 66% of the time, it wouldn't mean the n-gram model had compositionality, would it?

2022-10-20 21:35:34 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb whatever they get or don't get is parasitic on corpus stats. on the image synthesis stuff some colleagues and i have data we will report shortly and feel really confident that you are mistaking getting things right for the wrong reason with actually get them right.

2022-10-20 19:36:08 excellent work, offering a different and important take on compositionality

2022-10-20 19:30:56 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb another great example of the mappings are naive: https://t.co/d1bO9t8Spj

2022-10-20 19:28:50 @ihorgowda if herschel is our goal, we are doomed

2022-10-20 14:53:10 where is AI relative to mouse cognition? excellent thread! mouse experts please add to it! https://t.co/71a6kQJbqy

2022-10-20 14:48:26 @kohn_gregory @SpeciesTypical @sanewman1 that’s silly. honestly we have nothing to talk about

2022-10-20 14:43:41 @kohn_gregory @sanewman1 @SpeciesTypical not seeing the argument there, but yes i was trying to make a more sophisticated nativism.i am sure the @WiringTheBrain would say the subsequent 17 years of dev neuro have been kind to that position: eg https://t.co/bUgW8d0XTQ &

2022-10-20 14:40:08 @kohn_gregory @SpeciesTypical @sanewman1 this is confused. it is like saying there is no information in an LP, because you need to have a record player to play it.if you flatten the groove, you lose the info

2022-10-20 14:37:13 @kohn_gregory @SpeciesTypical umm… we were talking about the physical structure of the hand, and you have swapped to the motor programs controlling it. obviously the latter involves learning.

2022-10-20 14:33:19 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb @ZoubinGhahrama1 should give @TristanThrush access so Tristan can run a Winoground, then.

2022-10-20 14:31:17 and a pioneer in common sense reasoning! and now on Twitter: @ErnestSDavis https://t.co/3VWbdX2wTi

2022-10-20 08:09:29 RT @BarryOSullivan: I highly recommend this talk! And check out Ernie’s excellent book with @GaryMarcus: “Rebooting AI: Building Artificial…

2022-10-20 07:40:51 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb also note that @slatestarcodex initially declared victory in a bet re Imagen, and then retracted. i will believe that Imagen is real progress on compositionality only when it is carefully demonstrated.

2022-10-20 07:30:29 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i have asked them endlessly to make it available or test the Winoground benchmark. they have refused. i can speak w authority only re systems i (or others) can test. but remain skeptical about those hidden from testing.

2022-10-20 07:25:16 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb ie it is this claim above on image synthesis that i am questioning w.r.t. grammar: https://t.co/QLqpvsjqyL

2022-10-20 07:23:02 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb i was responding to a tweet that seemed to be about image synthesis

2022-10-20 07:02:11 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb best evidence that an image synthesis system captures some grammar? they seem to have some serious problems with binding.

2022-10-20 04:02:23 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb correction to last sentence: even when it gets it right it is bloviating. in no case does LLM do inference and reasoning over a world model

2022-10-20 03:59:45 @pmddomingos @bengoertzel @ch402 @AvilaGarcez @luislamb they composit but without proper variable binding they aren’t properly compositional. they also fail on part-whole relationships, function, etcin some ways the language side is akin to bag-of-words model, which represent language in a very impoverished way.

2022-10-20 03:45:25 RT @databu: Interesting discussion with @GaryMarcus and @bengoertzel there

2022-10-20 03:21:11 @pmddomingos @erikbryn without social media the riot probably wouldn’t have happened

2022-11-17 23:52:52 Tripling down, @yLeCun's contention is that people "abused" Galactica. But what safeguards were put in place to prevent bad actors from using the system to generate compelling disinformation around scientific topics like vaccines or racial differences? A: None. https://t.co/HlVjdHz1Sr

2022-11-17 23:39:33 @mmitchell_ai @trashcanmagic @timnitGebru But that is not me. I fully agree with Dr King's methods, and with direct action

2022-11-17 23:18:17 @andrewthesmart So the left is now pro-choice but anti-genetic screening? This is eugenics? https://t.co/gu4vD7v0en

2022-11-17 22:10:22 @emilymbender @timnitGebru and if what you say is true, you can simply provide a link. there certainly has not been anything from either of you in direct replies to me when i have asked.

2022-11-17 22:01:59 Why is this is any different at all than McCarthyism? I honestly don't see a difference between labeling someone a communist and labeling them a eugenicist, if you do so without either evidence or definition. I have asked, repeatedly, for both, and gotten neither. https://t.co/efvhRYCgEw

2022-11-17 21:54:48 @emilymbender @timnitGebru Nonsense. I have repeatedly asked a legitimate,straightforward question. Neither of you have answered: are you claiming eg that MacAskill (whom I have never met) is a eugenicist a la 19th century scientific racism? yes or no, and if yes, in what sense, and what is the argument?

2022-11-17 21:52:13 @athundt @timnitGebru The argument against Singer there is much more straightforward

2022-11-17 21:50:36 @athundt @timnitGebru I don't doubt that scientific racism still exists. But that label should not be applied to specific people without real evidence. That's all I am saying. &

2022-11-17 21:38:20 I have no shame about speaking out against McCarthyism.

2022-11-17 21:32:21 @timnitGebru no definitions still? i will carry on.

2022-11-17 21:26:23 @trashcanmagic @timnitGebru How dare you? My father was doing sit-ins for fair housing before you were born, and did discrimination law for years. I helped him (with statistics) as a teenager, and went to protests from when I was a newborn. I am with your struggle, but not your tactics.

2022-11-17 21:24:28 @timnitGebru We need definitions, @timnitGebru, if you want me and others take you seriously. I was once a huge fan of yours, but am really losing the faith.

2022-11-17 21:16:44 @timnitGebru Only quote I have seen was nearly fabricated, inserting words that Bostrom didn't use. I haven't seen any direct evidence on MacAskill. &

2022-11-17 21:07:19 @timnitGebru If I had been alive in the 1950s I hope that I would have called out McCarthy (white male) for doing what you do: naming names, without due process, without definitions, without evidence. It's not what groups you belong to Timnit, it's what you are saying that is problematic.

2022-11-17 20:41:16 RT @GaryMarcus: The reality is that large language models like GPT-3 annd Galactica are like bulls in a china shop, powerful but reckless.…

2022-11-17 19:56:24 @ylecun @Grady_Booch And they use lots of symbol-manipulation too. Funny that you never mention that.

2022-11-17 19:52:00 RT @GidMK: @GaryMarcus I'm worried about fake science. There are already hundreds of pay-to-play journals that publish openly plagiarised w…

2022-11-17 19:42:43 I doubt it. but Prince said it best, “Forever [is] a mighty long time” https://t.co/VXE5Up59tl

2022-11-17 19:32:50 RT @Plinz: Generative AI is incredibly useful and valuable, but it will force us to change our practices in science, journalism and social…

2022-11-17 19:31:30 @yLeCun: It’s all good, man https://t.co/PrX2PBawlc

2022-11-17 03:58:52 RT @frossi_t: On my way to the AAAI fall symposium on “thinking fast and slow and other cognitive theories in AI”. The program includes 12…

2022-11-17 01:37:42 RT @tytusm: 100% agree with Gary The threat posed to science by #Galactica is real Try it yourself using the gold standard…

2022-11-16 23:11:10 Flattery will get you nowhere, young battlestar. (image courtesy @UlrichJunker) https://t.co/d8GYGDCgTm

2022-11-16 22:59:03 @shokunin_studio @Grady_Booch @fchollet i replied to @tansuyegen's tweet

2022-11-16 22:58:06 @TansuYegen I think you have misunderstood wha the tech purports to do. I think that it is morphing actual human to look like a a different human. still cool, but not what you describe. see https://t.co/Lk5wipCh2c Face Shifting tab

2022-11-16 21:30:15 RT @karaswisher: Read these fine few words from ⁦@GaryMarcus⁩ https://t.co/Giqr1JI0by

2022-11-16 18:48:44 RT @ruthaylett: He is right. People laughing at this are mistaken, it is a serious problem in the making.

2022-11-16 18:40:34 RT @RemivanTrijp: I sincerely hope this doesn’t turn out to be prophetic

2022-11-16 17:56:14 @Grady_Booch @elonmusk I wasn’t assigning probabilities here

2022-11-16 17:52:44 @Grady_Booch @elonmusk Only two ways this can go - Musk can succeed, and rewrite everything we thought we knew about management and software engineering - He can fail, and become the textbook example of how not to do things.

2022-11-16 17:49:47 There’s a lot going on. Midterms, FTX, Twitter chaos. But in time we will see the release of #Galactica as an epochal event, a tipping point in a gigantic increase in the flow of misinformation. This is no joke. Galactica is funny, but the uses it will be put to are not. https://t.co/UPoe1q4BEi

2022-11-16 17:22:36 @MetaAI’s Galactica has made up my birthday (wrong decade), my education,my research interests, etc, all sounding plausible, 85% untrue. You know what’s really embarrassing about this? A decent AI could just look this stuff on the web &

2022-11-16 15:04:18 @ylecun @MetaAI @paperswithcode https://t.co/EvMKYUZcN0

2022-11-16 14:54:38 RT @pelimuc:

2022-11-16 14:41:15 never had this happened before - timeout loading my own Tweets. Things gonna get wild here in Bird-land! https://t.co/URBAdZpBhX

2022-11-16 14:38:05 For examples, see https://t.co/GUdfR6nR6y and this bonus fake bio of @ylecun here: https://t.co/gN1gvFkSjh

2022-11-16 14:38:03 AI Joke of the Day! Q: What do a CNC machine and MetaAI’s #Galactica have in common? .. .. .. A: They both help you fabricate! https://t.co/uEouaaO8iQ

2022-11-16 14:11:51 RT @dmonett: "Is this really what AI has come to, automatically mixing reality with bullshit so finely we can no longer recognize the diffe…

2022-11-16 14:11:23 RT @dougclow: Scientific bullshit generator, argh! Having spent many disheartening hours reviewing academic manuscripts that strongly resem…

2022-11-16 06:56:19 @kevinroose you don’t think the product placement for Storybook Brawl was a nice touch? https://t.co/15jySgxqR9 @TrungTPhan

2022-11-16 05:49:14 @mcuban just liked my new post

2022-11-16 05:35:37 @_jasonwei @colinraffel i need to think about @colinraffel’s specific tasks but i am willing to bet money that LLMs will continue to face serious challenges in factuality even as they scale.

2022-11-16 00:08:58 @ykilcher which benchmarks have you published in the last four weeks? I have posted two, one in a hot new ML journal, linked above, which you didn’t comment, the other on arXiv. You? the third was published in AI Magazine 2015. It’s not *my* fault that the field doesn’t rise to them.

2022-11-15 20:59:45 @ykilcher Not my job to decide what community disagrees about. But people ought look at the zero-shot version of Truthful QA and to test comprehension as I described in the 2014 New Yorker article Beyond The Turing Test, as well recent vision benchmark I helped w: https://t.co/LDPoiyTPNi

2022-11-15 19:02:22 OpenAI give me access to GPT-4? https://t.co/mN0OSkugcZ

2022-11-15 16:51:13 Rumor has it GPT-4 is coming. Here’s a prediction: it will still have serious issues with confabulation/misinformation, toxicity and bias. https://t.co/TKm6DA3Dvb

2022-11-15 16:24:59 . @sacca wasn’t wrong. https://t.co/2WVxcZAPMn

2022-11-15 15:23:55 "it was sobering to see how close we came to hitting a parked car after we rolled over a low curb separating the parking lot." — @cademetz on six hours in a Tesla:https://t.co/UUQXUBzE4q

2022-11-15 14:36:53 @_jasonwei @drjwrae @OwainEvans_UK post hoc.performance on original 0-shot task is not what you graphed. you are graphing only 10-shot results (and focusing on the multiple choice results which are the easier form of the task).On zero shot MC, 280B performance is <

2022-11-15 06:42:33 @_jasonwei @drjwrae @OwainEvans_UK If they used a different method you can’t tell how much is method change and how much is scale

2022-11-14 19:26:55 RT @QasimRashid: This is horrifying. The Iranian parliament votes overwhelmingly (227-63) to execute the 15,000 protesters they already arr…

2022-11-14 19:24:35 @_jasonwei @OwainEvans_UK your drawing doesn't match the leaderboard here: https://t.co/7HGyLImHHs and i think it's because Gopher changed the task...

2022-11-14 19:08:28 RT @svpino: In 2017, a team led by Andrew Ng published a paper showing off a Deep Learning model to detect pneumonia.Andrew is one of the…

2022-11-14 18:45:29 @_jasonwei Taking TruthfulQA as an example, your evidence for scaling seems (if I understand correctly) to come from a single data point from Gopher that was (a) selected post hoc, (b) has not been replicated and (c) relied on making the task multiple choice. cc @OwainEvans_UKAm I wrong?

2022-11-14 14:06:15 RT @federicobianchy: Text-to-image generation models (like Stable Diffusion and DALLE) are being used to generate millions of images a day.…

2022-11-14 13:10:22 RT @drewharwell: New: That fake verified Eli Lilly account sparked a panic inside the pharma giant, and senior execs decided they'd had eno…

2022-11-14 01:58:42 @karaswisher Will certainly be interesting to see what happens to advertising revenue in the Lockheed Martini era…

2022-11-17 23:52:52 Tripling down, @yLeCun's contention is that people "abused" Galactica. But what safeguards were put in place to prevent bad actors from using the system to generate compelling disinformation around scientific topics like vaccines or racial differences? A: None. https://t.co/HlVjdHz1Sr

2022-11-17 23:39:33 @mmitchell_ai @trashcanmagic @timnitGebru But that is not me. I fully agree with Dr King's methods, and with direct action

2022-11-17 23:18:17 @andrewthesmart So the left is now pro-choice but anti-genetic screening? This is eugenics? https://t.co/gu4vD7v0en

2022-11-17 22:10:22 @emilymbender @timnitGebru and if what you say is true, you can simply provide a link. there certainly has not been anything from either of you in direct replies to me when i have asked.

2022-11-17 22:01:59 Why is this is any different at all than McCarthyism? I honestly don't see a difference between labeling someone a communist and labeling them a eugenicist, if you do so without either evidence or definition. I have asked, repeatedly, for both, and gotten neither. https://t.co/efvhRYCgEw

2022-11-17 21:54:48 @emilymbender @timnitGebru Nonsense. I have repeatedly asked a legitimate,straightforward question. Neither of you have answered: are you claiming eg that MacAskill (whom I have never met) is a eugenicist a la 19th century scientific racism? yes or no, and if yes, in what sense, and what is the argument?

2022-11-17 21:52:13 @athundt @timnitGebru The argument against Singer there is much more straightforward

2022-11-17 21:50:36 @athundt @timnitGebru I don't doubt that scientific racism still exists. But that label should not be applied to specific people without real evidence. That's all I am saying. &

2022-11-17 21:38:20 I have no shame about speaking out against McCarthyism.

2022-11-17 21:32:21 @timnitGebru no definitions still? i will carry on.

2022-11-17 21:26:23 @trashcanmagic @timnitGebru How dare you? My father was doing sit-ins for fair housing before you were born, and did discrimination law for years. I helped him (with statistics) as a teenager, and went to protests from when I was a newborn. I am with your struggle, but not your tactics.

2022-11-17 21:24:28 @timnitGebru We need definitions, @timnitGebru, if you want me and others take you seriously. I was once a huge fan of yours, but am really losing the faith.

2022-11-17 21:16:44 @timnitGebru Only quote I have seen was nearly fabricated, inserting words that Bostrom didn't use. I haven't seen any direct evidence on MacAskill. &

2022-11-17 21:07:19 @timnitGebru If I had been alive in the 1950s I hope that I would have called out McCarthy (white male) for doing what you do: naming names, without due process, without definitions, without evidence. It's not what groups you belong to Timnit, it's what you are saying that is problematic.

2022-11-17 20:41:16 RT @GaryMarcus: The reality is that large language models like GPT-3 annd Galactica are like bulls in a china shop, powerful but reckless.…

2022-11-17 19:56:24 @ylecun @Grady_Booch And they use lots of symbol-manipulation too. Funny that you never mention that.

2022-11-17 19:52:00 RT @GidMK: @GaryMarcus I'm worried about fake science. There are already hundreds of pay-to-play journals that publish openly plagiarised w…

2022-11-17 19:42:43 I doubt it. but Prince said it best, “Forever [is] a mighty long time” https://t.co/VXE5Up59tl

2022-11-17 19:32:50 RT @Plinz: Generative AI is incredibly useful and valuable, but it will force us to change our practices in science, journalism and social…

2022-11-17 19:31:30 @yLeCun: It’s all good, man https://t.co/PrX2PBawlc

2022-11-17 03:58:52 RT @frossi_t: On my way to the AAAI fall symposium on “thinking fast and slow and other cognitive theories in AI”. The program includes 12…

2022-11-17 01:37:42 RT @tytusm: 100% agree with Gary The threat posed to science by #Galactica is real Try it yourself using the gold standard…

2022-11-16 23:11:10 Flattery will get you nowhere, young battlestar. (image courtesy @UlrichJunker) https://t.co/d8GYGDCgTm

2022-11-16 22:59:03 @shokunin_studio @Grady_Booch @fchollet i replied to @tansuyegen's tweet

2022-11-16 22:58:06 @TansuYegen I think you have misunderstood wha the tech purports to do. I think that it is morphing actual human to look like a a different human. still cool, but not what you describe. see https://t.co/Lk5wipCh2c Face Shifting tab

2022-11-16 21:30:15 RT @karaswisher: Read these fine few words from ⁦@GaryMarcus⁩ https://t.co/Giqr1JI0by

2022-11-16 18:48:44 RT @ruthaylett: He is right. People laughing at this are mistaken, it is a serious problem in the making.

2022-11-16 18:40:34 RT @RemivanTrijp: I sincerely hope this doesn’t turn out to be prophetic

2022-11-16 17:56:14 @Grady_Booch @elonmusk I wasn’t assigning probabilities here

2022-11-16 17:52:44 @Grady_Booch @elonmusk Only two ways this can go - Musk can succeed, and rewrite everything we thought we knew about management and software engineering - He can fail, and become the textbook example of how not to do things.

2022-11-16 17:49:47 There’s a lot going on. Midterms, FTX, Twitter chaos. But in time we will see the release of #Galactica as an epochal event, a tipping point in a gigantic increase in the flow of misinformation. This is no joke. Galactica is funny, but the uses it will be put to are not. https://t.co/UPoe1q4BEi

2022-11-16 17:22:36 @MetaAI’s Galactica has made up my birthday (wrong decade), my education,my research interests, etc, all sounding plausible, 85% untrue. You know what’s really embarrassing about this? A decent AI could just look this stuff on the web &

2022-11-16 15:04:18 @ylecun @MetaAI @paperswithcode https://t.co/EvMKYUZcN0

2022-11-16 14:54:38 RT @pelimuc:

2022-11-16 14:41:15 never had this happened before - timeout loading my own Tweets. Things gonna get wild here in Bird-land! https://t.co/URBAdZpBhX

2022-11-16 14:38:05 For examples, see https://t.co/GUdfR6nR6y and this bonus fake bio of @ylecun here: https://t.co/gN1gvFkSjh

2022-11-16 14:38:03 AI Joke of the Day! Q: What do a CNC machine and MetaAI’s #Galactica have in common? .. .. .. A: They both help you fabricate! https://t.co/uEouaaO8iQ

2022-11-16 14:11:51 RT @dmonett: "Is this really what AI has come to, automatically mixing reality with bullshit so finely we can no longer recognize the diffe…

2022-11-16 14:11:23 RT @dougclow: Scientific bullshit generator, argh! Having spent many disheartening hours reviewing academic manuscripts that strongly resem…

2022-11-16 06:56:19 @kevinroose you don’t think the product placement for Storybook Brawl was a nice touch? https://t.co/15jySgxqR9 @TrungTPhan

2022-11-16 05:49:14 @mcuban just liked my new post

2022-11-16 05:35:37 @_jasonwei @colinraffel i need to think about @colinraffel’s specific tasks but i am willing to bet money that LLMs will continue to face serious challenges in factuality even as they scale.

2022-11-16 00:08:58 @ykilcher which benchmarks have you published in the last four weeks? I have posted two, one in a hot new ML journal, linked above, which you didn’t comment, the other on arXiv. You? the third was published in AI Magazine 2015. It’s not *my* fault that the field doesn’t rise to them.

2022-11-15 20:59:45 @ykilcher Not my job to decide what community disagrees about. But people ought look at the zero-shot version of Truthful QA and to test comprehension as I described in the 2014 New Yorker article Beyond The Turing Test, as well recent vision benchmark I helped w: https://t.co/LDPoiyTPNi

2022-11-15 19:02:22 OpenAI give me access to GPT-4? https://t.co/mN0OSkugcZ

2022-11-15 16:51:13 Rumor has it GPT-4 is coming. Here’s a prediction: it will still have serious issues with confabulation/misinformation, toxicity and bias. https://t.co/TKm6DA3Dvb

2022-11-15 16:24:59 . @sacca wasn’t wrong. https://t.co/2WVxcZAPMn

2022-11-15 15:23:55 "it was sobering to see how close we came to hitting a parked car after we rolled over a low curb separating the parking lot." — @cademetz on six hours in a Tesla:https://t.co/UUQXUBzE4q

2022-11-15 14:36:53 @_jasonwei @drjwrae @OwainEvans_UK post hoc.performance on original 0-shot task is not what you graphed. you are graphing only 10-shot results (and focusing on the multiple choice results which are the easier form of the task).On zero shot MC, 280B performance is <

2022-11-15 06:42:33 @_jasonwei @drjwrae @OwainEvans_UK If they used a different method you can’t tell how much is method change and how much is scale

2022-11-14 19:26:55 RT @QasimRashid: This is horrifying. The Iranian parliament votes overwhelmingly (227-63) to execute the 15,000 protesters they already arr…

2022-11-14 19:24:35 @_jasonwei @OwainEvans_UK your drawing doesn't match the leaderboard here: https://t.co/7HGyLImHHs and i think it's because Gopher changed the task...

2022-11-14 19:08:28 RT @svpino: In 2017, a team led by Andrew Ng published a paper showing off a Deep Learning model to detect pneumonia.Andrew is one of the…

2022-11-14 18:45:29 @_jasonwei Taking TruthfulQA as an example, your evidence for scaling seems (if I understand correctly) to come from a single data point from Gopher that was (a) selected post hoc, (b) has not been replicated and (c) relied on making the task multiple choice. cc @OwainEvans_UKAm I wrong?

2022-11-14 14:06:15 RT @federicobianchy: Text-to-image generation models (like Stable Diffusion and DALLE) are being used to generate millions of images a day.…

2022-11-14 13:10:22 RT @drewharwell: New: That fake verified Eli Lilly account sparked a panic inside the pharma giant, and senior execs decided they'd had eno…

2022-11-14 01:58:42 @karaswisher Will certainly be interesting to see what happens to advertising revenue in the Lockheed Martini era…

2022-11-18 21:21:17 @ylecun @mrgreene1977 Potentially deadly TikTok challenges

2022-11-18 21:15:06 RT @michaelharriot: They won’t know what breaks until it breaks. It’s a matter of what breaks. “It’s not like McDonalds’ bro. You can’t ta…

2022-11-18 21:06:16 https://t.co/7aDMXp6G3g

2022-11-18 19:21:26 RT @GaryMarcus: @elonmusk What about misinformation?

2022-11-18 19:20:56 I Like Free Speech So Much I’ve Decided to Buy It - McSweeney’s https://t.co/Yw3n94HAQW

2022-11-18 19:18:59 @elonmusk What about misinformation?

2022-11-18 19:06:43 Galactica

2022-11-18 15:50:32 @yoavgo @ylecun @rasbt @manes @Abebab Troll farms cost money and presumably limited by budget. The more they can put out the more effective they are.

2022-11-18 15:36:47 @yoavgo @rasbt @manes @ylecun @Abebab On the scenario, secondary customers like journalists will have no longer have any trust in arXiv at all.

2022-11-18 15:34:22 @ylecun @rasbt @manes @Abebab It will reduce the cost hence increase the volume, hence make the problem worse.

2022-11-18 15:12:49 RT @GaryMarcus: @ylecun @Abebab Galactica *is* dangerous. And you have no idea how to eliminate that danger. large language models inherent…

2022-11-18 15:12:34 @ylecun @Abebab Galactica *is* dangerous. And you have no idea how to eliminate that danger. large language models inherently hallucinate. It weaponized the ability of these tools to spread fake science. The vitriol is proportionate to the risk. I am shocked that you are indifferent to this.

2022-11-18 15:06:40 “once perfected” is carrying a lot of weight here. there is no way Galactica will be perfected in next year or two, probably a lot longer. happy to publicly debate you about this, but recommend you start by rereading your own recent critique @noemamag of large language models. https://t.co/LYXD0E0jyI

2022-11-18 13:51:45 very rational fear https://t.co/bY0hUrnzcv

2022-11-18 13:41:25 follow friday may not be a thing much longer over here on Twitter but follow @abebab. And she’s on Mastodon now, too. https://t.co/29GfLksNFx

2022-11-18 13:38:38 RT @Abebab: @ylecun Pretty much exactly what happened was you overhyped and released a model. People tested it (you should be thankful for…

2022-11-18 13:15:06 Just saying, in case this ship goes down, @garymarcus@sigmoid.social

2022-11-18 12:55:18 @HeidyKhlaaf @AmandaAskell @andrewthesmart @Abebab 100% support thinking through the potential social consequences. There is a lot of room for important argument there.

2022-11-18 12:48:16 @HeidyKhlaaf @AmandaAskell @andrewthesmart @Abebab Thank you. and to be clear I am all for spelling out the latter part and will amplify the message if is is constructed carefully and respectfully. (If you want to write something longer, DM.)

2022-11-18 12:24:50 @HeidyKhlaaf @AmandaAskell @andrewthesmart Yes, there are lot of ideas within EA, some well-intentioned, some not. some briefly considered, some taken seriously, some not

2022-11-17 23:52:52 Tripling down, @yLeCun's contention is that people "abused" Galactica. But what safeguards were put in place to prevent bad actors from using the system to generate compelling disinformation around scientific topics like vaccines or racial differences? A: None. https://t.co/HlVjdHz1Sr

2022-11-17 23:39:33 @mmitchell_ai @trashcanmagic @timnitGebru But that is not me. I fully agree with Dr King's methods, and with direct action

2022-11-17 23:18:17 @andrewthesmart So the left is now pro-choice but anti-genetic screening? This is eugenics? https://t.co/gu4vD7v0en

2022-11-17 22:10:22 @emilymbender @timnitGebru and if what you say is true, you can simply provide a link. there certainly has not been anything from either of you in direct replies to me when i have asked.

2022-11-17 22:01:59 Why is this is any different at all than McCarthyism? I honestly don't see a difference between labeling someone a communist and labeling them a eugenicist, if you do so without either evidence or definition. I have asked, repeatedly, for both, and gotten neither. https://t.co/efvhRYCgEw

2022-11-17 21:54:48 @emilymbender @timnitGebru Nonsense. I have repeatedly asked a legitimate,straightforward question. Neither of you have answered: are you claiming eg that MacAskill (whom I have never met) is a eugenicist a la 19th century scientific racism? yes or no, and if yes, in what sense, and what is the argument?

2022-11-17 21:52:13 @athundt @timnitGebru The argument against Singer there is much more straightforward

2022-11-17 21:50:36 @athundt @timnitGebru I don't doubt that scientific racism still exists. But that label should not be applied to specific people without real evidence. That's all I am saying. &

2022-11-17 21:38:20 I have no shame about speaking out against McCarthyism.

2022-11-17 21:32:21 @timnitGebru no definitions still? i will carry on.

2022-11-17 21:26:23 @trashcanmagic @timnitGebru How dare you? My father was doing sit-ins for fair housing before you were born, and did discrimination law for years. I helped him (with statistics) as a teenager, and went to protests from when I was a newborn. I am with your struggle, but not your tactics.

2022-11-17 21:24:28 @timnitGebru We need definitions, @timnitGebru, if you want me and others take you seriously. I was once a huge fan of yours, but am really losing the faith.

2022-11-17 21:16:44 @timnitGebru Only quote I have seen was nearly fabricated, inserting words that Bostrom didn't use. I haven't seen any direct evidence on MacAskill. &

2022-11-17 21:07:19 @timnitGebru If I had been alive in the 1950s I hope that I would have called out McCarthy (white male) for doing what you do: naming names, without due process, without definitions, without evidence. It's not what groups you belong to Timnit, it's what you are saying that is problematic.

2022-11-17 20:41:16 RT @GaryMarcus: The reality is that large language models like GPT-3 annd Galactica are like bulls in a china shop, powerful but reckless.…

2022-11-17 19:56:24 @ylecun @Grady_Booch And they use lots of symbol-manipulation too. Funny that you never mention that.

2022-11-17 19:52:00 RT @GidMK: @GaryMarcus I'm worried about fake science. There are already hundreds of pay-to-play journals that publish openly plagiarised w…

2022-11-17 19:42:43 I doubt it. but Prince said it best, “Forever [is] a mighty long time” https://t.co/VXE5Up59tl

2022-11-17 19:32:50 RT @Plinz: Generative AI is incredibly useful and valuable, but it will force us to change our practices in science, journalism and social…

2022-11-17 19:31:30 @yLeCun: It’s all good, man https://t.co/PrX2PBawlc

2022-11-17 03:58:52 RT @frossi_t: On my way to the AAAI fall symposium on “thinking fast and slow and other cognitive theories in AI”. The program includes 12…

2022-11-17 01:37:42 RT @tytusm: 100% agree with Gary The threat posed to science by #Galactica is real Try it yourself using the gold standard…

2022-11-16 23:11:10 Flattery will get you nowhere, young battlestar. (image courtesy @UlrichJunker) https://t.co/d8GYGDCgTm

2022-11-16 22:59:03 @shokunin_studio @Grady_Booch @fchollet i replied to @tansuyegen's tweet

2022-11-16 22:58:06 @TansuYegen I think you have misunderstood wha the tech purports to do. I think that it is morphing actual human to look like a a different human. still cool, but not what you describe. see https://t.co/Lk5wipCh2c Face Shifting tab

2022-11-16 21:30:15 RT @karaswisher: Read these fine few words from ⁦@GaryMarcus⁩ https://t.co/Giqr1JI0by

2022-11-16 18:48:44 RT @ruthaylett: He is right. People laughing at this are mistaken, it is a serious problem in the making.

2022-11-16 18:40:34 RT @RemivanTrijp: I sincerely hope this doesn’t turn out to be prophetic

2022-11-16 17:56:14 @Grady_Booch @elonmusk I wasn’t assigning probabilities here

2022-11-16 17:52:44 @Grady_Booch @elonmusk Only two ways this can go - Musk can succeed, and rewrite everything we thought we knew about management and software engineering - He can fail, and become the textbook example of how not to do things.

2022-11-16 17:49:47 There’s a lot going on. Midterms, FTX, Twitter chaos. But in time we will see the release of #Galactica as an epochal event, a tipping point in a gigantic increase in the flow of misinformation. This is no joke. Galactica is funny, but the uses it will be put to are not. https://t.co/UPoe1q4BEi

2022-11-16 17:22:36 @MetaAI’s Galactica has made up my birthday (wrong decade), my education,my research interests, etc, all sounding plausible, 85% untrue. You know what’s really embarrassing about this? A decent AI could just look this stuff on the web &

2022-11-16 15:04:18 @ylecun @MetaAI @paperswithcode https://t.co/EvMKYUZcN0

2022-11-16 14:54:38 RT @pelimuc:

2022-11-16 14:41:15 never had this happened before - timeout loading my own Tweets. Things gonna get wild here in Bird-land! https://t.co/URBAdZpBhX

2022-11-16 14:38:05 For examples, see https://t.co/GUdfR6nR6y and this bonus fake bio of @ylecun here: https://t.co/gN1gvFkSjh

2022-11-16 14:38:03 AI Joke of the Day! Q: What do a CNC machine and MetaAI’s #Galactica have in common? .. .. .. A: They both help you fabricate! https://t.co/uEouaaO8iQ

2022-11-16 14:11:51 RT @dmonett: "Is this really what AI has come to, automatically mixing reality with bullshit so finely we can no longer recognize the diffe…

2022-11-16 14:11:23 RT @dougclow: Scientific bullshit generator, argh! Having spent many disheartening hours reviewing academic manuscripts that strongly resem…

2022-11-16 06:56:19 @kevinroose you don’t think the product placement for Storybook Brawl was a nice touch? https://t.co/15jySgxqR9 @TrungTPhan

2022-11-16 05:49:14 @mcuban just liked my new post

2022-11-16 05:35:37 @_jasonwei @colinraffel i need to think about @colinraffel’s specific tasks but i am willing to bet money that LLMs will continue to face serious challenges in factuality even as they scale.

2022-11-16 00:08:58 @ykilcher which benchmarks have you published in the last four weeks? I have posted two, one in a hot new ML journal, linked above, which you didn’t comment, the other on arXiv. You? the third was published in AI Magazine 2015. It’s not *my* fault that the field doesn’t rise to them.

2022-11-15 20:59:45 @ykilcher Not my job to decide what community disagrees about. But people ought look at the zero-shot version of Truthful QA and to test comprehension as I described in the 2014 New Yorker article Beyond The Turing Test, as well recent vision benchmark I helped w: https://t.co/LDPoiyTPNi

2022-11-15 19:02:22 OpenAI give me access to GPT-4? https://t.co/mN0OSkugcZ

2022-11-15 16:51:13 Rumor has it GPT-4 is coming. Here’s a prediction: it will still have serious issues with confabulation/misinformation, toxicity and bias. https://t.co/TKm6DA3Dvb

2022-11-15 16:24:59 . @sacca wasn’t wrong. https://t.co/2WVxcZAPMn

2022-11-15 15:23:55 "it was sobering to see how close we came to hitting a parked car after we rolled over a low curb separating the parking lot." — @cademetz on six hours in a Tesla:https://t.co/UUQXUBzE4q

2022-11-15 14:36:53 @_jasonwei @drjwrae @OwainEvans_UK post hoc.performance on original 0-shot task is not what you graphed. you are graphing only 10-shot results (and focusing on the multiple choice results which are the easier form of the task).On zero shot MC, 280B performance is <

2022-11-15 06:42:33 @_jasonwei @drjwrae @OwainEvans_UK If they used a different method you can’t tell how much is method change and how much is scale

2022-11-14 19:26:55 RT @QasimRashid: This is horrifying. The Iranian parliament votes overwhelmingly (227-63) to execute the 15,000 protesters they already arr…

2022-11-14 19:24:35 @_jasonwei @OwainEvans_UK your drawing doesn't match the leaderboard here: https://t.co/7HGyLImHHs and i think it's because Gopher changed the task...

2022-11-14 19:08:28 RT @svpino: In 2017, a team led by Andrew Ng published a paper showing off a Deep Learning model to detect pneumonia.Andrew is one of the…

2022-11-14 18:45:29 @_jasonwei Taking TruthfulQA as an example, your evidence for scaling seems (if I understand correctly) to come from a single data point from Gopher that was (a) selected post hoc, (b) has not been replicated and (c) relied on making the task multiple choice. cc @OwainEvans_UKAm I wrong?

2022-11-14 14:06:15 RT @federicobianchy: Text-to-image generation models (like Stable Diffusion and DALLE) are being used to generate millions of images a day.…

2022-11-14 13:10:22 RT @drewharwell: New: That fake verified Eli Lilly account sparked a panic inside the pharma giant, and senior execs decided they'd had eno…

2022-11-14 01:58:42 @karaswisher Will certainly be interesting to see what happens to advertising revenue in the Lockheed Martini era…

2022-11-20 03:29:28 @AI4Code @SumitGulwani

2022-11-20 01:02:03 October 28: No reinstatements without a council Nov 19: Never mind about the council. I asked my fans, mostly right of center. Trump’s back in. https://t.co/b3xrNznGcO

2022-11-20 00:45:06 @ylecun @drng @Grady_Booch @Jeff_Aronson @Abebab @mrgreene1977 Galactica has deeper problems: - it is unconstrained by the facts are in its database (eg it said that Elon Musk died in 2018 car crash but probably has data to the contrary) - it gives no indication when it makes something up that isn’t database

2022-11-20 00:33:06 @ylecun @drng @Grady_Booch @Jeff_Aronson @Abebab I think what he means that the model made false inferences eg about value in eating crushed glass (per @mrgreene1977) https://t.co/eOlWcTDTxL

2022-11-20 00:31:09 @mark_riedl @Sergei_Imaging @ylecun @mrgreene1977 no doubt that cat is out of the bag, and it’s easy enough to replicate. soon enough we find out empirically how much impact systems like these have. (though of course hard to trace since the quality of the mimicry is high and there are no watermarks)

2022-11-20 00:29:43 @srijankedia @ylecun @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian @emilio__ferrara @davidlazer @gianluca_string @CrashTheMod3 It’s quite worrisome that @metaAI’s Chief AI officer shows little apparent familiarity with such work.

2022-11-20 00:27:54 @tdietterich @ylecun @mrgreene1977 and they are in fact starting to be used, and as cost of doing them well decreases, we can expect them to be more widespread. some examples from 2022: https://t.co/4kxqUjHXXT 2024 election, who knows?

2022-11-20 00:22:17 #Galactica is fine, because if it makes stuff up, no problemo. @metaAI’s Chief AI Officer just told me so. And he’s *sure* nobody would ever misuse his tool. Yeah, right. https://t.co/9uk9YE56ad

2022-11-20 00:20:04 @noUpside @ylecun @mrgreene1977 @katestarbird @sinanaral @ProfNoahGian Ask Yann, who thinks it is actually hunky-dory, but my theory was this: https://t.co/tfjaWTQuHW

2022-11-20 00:10:39 @ylecun @srijankedia @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian Low probabilities ≠ zero probabilities. Low probabilities with low cost ≠ zero probabilities with zero cost. Every spammer, ever, knows this. Same principles will apply to misinfo

2022-11-20 00:06:40 Twitter on IOS just crashed midtweet and I lost my work. First time that happened in a decade of using it. also having some load time issues. And so it begins.

2022-11-20 00:02:31 @ylecun @jppesky @mrgreene1977 on latter point, better said: https://t.co/SLZ7lVHQ4f

2022-11-20 00:00:29 @ylecun @jppesky @mrgreene1977 You are coming across as obtuse. - You act as if misinfo has had no impact in the real world (despite success of anti-vax campaigns) - Act as if the only way in which scientific-sounding nonsense has impact is if it is in a peer review journal - Treat low prob as zero probability

2022-11-19 23:52:54 @TonyZador @ylecun @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian perhaps someone in 1985 said “email isn’t going to have any effect on scamming. it’s driven by how much people want to be taken.” the biggest application of LLMs is now? it’s apparently making nonsense for SEO optimization. New ways to do in volume, because cost is cheaper.

2022-11-19 23:14:23 @ylecun @mrgreene1977 Come on, it’s not just about journals. To take one example, vaccine misinformation, oft presented as “scientific” [eg https://t.co/B1JmJihJpl], likely had significant impact on vaccination rates. Tagging @noUpside @katestarbird @sinanaral @ProfNoahGian for more perspectives.

2022-11-19 22:13:07 RT @GaryMarcus: @ylecun @mrgreene1977 How could Galactica *not* make misinfo easier, @ylecun, when it makes writing faux scientific-like a…

2022-11-19 22:11:50 @ylecun @mrgreene1977 How could Galactica *not* make misinfo easier, @ylecun, when it makes writing faux scientific-like articles trivial rather than laborious? &

2022-11-19 19:43:40 @ylecun @Grady_Booch @Jeff_Aronson @Abebab The sand is in thinking that a system like this can stick to reality. It cannot.

2022-11-19 19:13:54 @ylecun @Abebab Ps repeating for convenience: https://t.co/BBE3moQFea

2022-11-18 21:21:17 @ylecun @mrgreene1977 Potentially deadly TikTok challenges

2022-11-18 21:15:06 RT @michaelharriot: They won’t know what breaks until it breaks. It’s a matter of what breaks. “It’s not like McDonalds’ bro. You can’t ta…

2022-11-18 21:06:16 https://t.co/7aDMXp6G3g

2022-11-18 19:21:26 RT @GaryMarcus: @elonmusk What about misinformation?

2022-11-18 19:20:56 I Like Free Speech So Much I’ve Decided to Buy It - McSweeney’s https://t.co/Yw3n94HAQW

2022-11-18 19:18:59 @elonmusk What about misinformation?

2022-11-18 19:06:43 Galactica

2022-11-18 15:50:32 @yoavgo @ylecun @rasbt @manes @Abebab Troll farms cost money and presumably limited by budget. The more they can put out the more effective they are.

2022-11-18 15:36:47 @yoavgo @rasbt @manes @ylecun @Abebab On the scenario, secondary customers like journalists will have no longer have any trust in arXiv at all.

2022-11-18 15:34:22 @ylecun @rasbt @manes @Abebab It will reduce the cost hence increase the volume, hence make the problem worse.

2022-11-18 15:12:49 RT @GaryMarcus: @ylecun @Abebab Galactica *is* dangerous. And you have no idea how to eliminate that danger. large language models inherent…

2022-11-18 15:12:34 @ylecun @Abebab Galactica *is* dangerous. And you have no idea how to eliminate that danger. large language models inherently hallucinate. It weaponized the ability of these tools to spread fake science. The vitriol is proportionate to the risk. I am shocked that you are indifferent to this.

2022-11-18 15:06:40 “once perfected” is carrying a lot of weight here. there is no way Galactica will be perfected in next year or two, probably a lot longer. happy to publicly debate you about this, but recommend you start by rereading your own recent critique @noemamag of large language models. https://t.co/LYXD0E0jyI

2022-11-18 13:51:45 very rational fear https://t.co/bY0hUrnzcv

2022-11-18 13:41:25 follow friday may not be a thing much longer over here on Twitter but follow @abebab. And she’s on Mastodon now, too. https://t.co/29GfLksNFx

2022-11-18 13:38:38 RT @Abebab: @ylecun Pretty much exactly what happened was you overhyped and released a model. People tested it (you should be thankful for…

2022-11-18 13:15:06 Just saying, in case this ship goes down, @garymarcus@sigmoid.social

2022-11-18 12:55:18 @HeidyKhlaaf @AmandaAskell @andrewthesmart @Abebab 100% support thinking through the potential social consequences. There is a lot of room for important argument there.

2022-11-18 12:48:16 @HeidyKhlaaf @AmandaAskell @andrewthesmart @Abebab Thank you. and to be clear I am all for spelling out the latter part and will amplify the message if is is constructed carefully and respectfully. (If you want to write something longer, DM.)

2022-11-18 12:24:50 @HeidyKhlaaf @AmandaAskell @andrewthesmart Yes, there are lot of ideas within EA, some well-intentioned, some not. some briefly considered, some taken seriously, some not

2022-11-17 23:52:52 Tripling down, @yLeCun's contention is that people "abused" Galactica. But what safeguards were put in place to prevent bad actors from using the system to generate compelling disinformation around scientific topics like vaccines or racial differences? A: None. https://t.co/HlVjdHz1Sr

2022-11-17 23:39:33 @mmitchell_ai @trashcanmagic @timnitGebru But that is not me. I fully agree with Dr King's methods, and with direct action

2022-11-17 23:18:17 @andrewthesmart So the left is now pro-choice but anti-genetic screening? This is eugenics? https://t.co/gu4vD7v0en

2022-11-17 22:10:22 @emilymbender @timnitGebru and if what you say is true, you can simply provide a link. there certainly has not been anything from either of you in direct replies to me when i have asked.

2022-11-17 22:01:59 Why is this is any different at all than McCarthyism? I honestly don't see a difference between labeling someone a communist and labeling them a eugenicist, if you do so without either evidence or definition. I have asked, repeatedly, for both, and gotten neither. https://t.co/efvhRYCgEw

2022-11-17 21:54:48 @emilymbender @timnitGebru Nonsense. I have repeatedly asked a legitimate,straightforward question. Neither of you have answered: are you claiming eg that MacAskill (whom I have never met) is a eugenicist a la 19th century scientific racism? yes or no, and if yes, in what sense, and what is the argument?

2022-11-17 21:52:13 @athundt @timnitGebru The argument against Singer there is much more straightforward

2022-11-17 21:50:36 @athundt @timnitGebru I don't doubt that scientific racism still exists. But that label should not be applied to specific people without real evidence. That's all I am saying. &

2022-11-17 21:38:20 I have no shame about speaking out against McCarthyism.

2022-11-17 21:32:21 @timnitGebru no definitions still? i will carry on.

2022-11-17 21:26:23 @trashcanmagic @timnitGebru How dare you? My father was doing sit-ins for fair housing before you were born, and did discrimination law for years. I helped him (with statistics) as a teenager, and went to protests from when I was a newborn. I am with your struggle, but not your tactics.

2022-11-17 21:24:28 @timnitGebru We need definitions, @timnitGebru, if you want me and others take you seriously. I was once a huge fan of yours, but am really losing the faith.

2022-11-17 21:16:44 @timnitGebru Only quote I have seen was nearly fabricated, inserting words that Bostrom didn't use. I haven't seen any direct evidence on MacAskill. &

2022-11-17 21:07:19 @timnitGebru If I had been alive in the 1950s I hope that I would have called out McCarthy (white male) for doing what you do: naming names, without due process, without definitions, without evidence. It's not what groups you belong to Timnit, it's what you are saying that is problematic.

2022-11-17 20:41:16 RT @GaryMarcus: The reality is that large language models like GPT-3 annd Galactica are like bulls in a china shop, powerful but reckless.…

2022-11-17 19:56:24 @ylecun @Grady_Booch And they use lots of symbol-manipulation too. Funny that you never mention that.

2022-11-17 19:52:00 RT @GidMK: @GaryMarcus I'm worried about fake science. There are already hundreds of pay-to-play journals that publish openly plagiarised w…

2022-11-17 19:42:43 I doubt it. but Prince said it best, “Forever [is] a mighty long time” https://t.co/VXE5Up59tl

2022-11-17 19:32:50 RT @Plinz: Generative AI is incredibly useful and valuable, but it will force us to change our practices in science, journalism and social…

2022-11-17 19:31:30 @yLeCun: It’s all good, man https://t.co/PrX2PBawlc

2022-11-17 03:58:52 RT @frossi_t: On my way to the AAAI fall symposium on “thinking fast and slow and other cognitive theories in AI”. The program includes 12…

2022-11-17 01:37:42 RT @tytusm: 100% agree with Gary The threat posed to science by #Galactica is real Try it yourself using the gold standard…

2022-11-16 23:11:10 Flattery will get you nowhere, young battlestar. (image courtesy @UlrichJunker) https://t.co/d8GYGDCgTm

2022-11-16 22:59:03 @shokunin_studio @Grady_Booch @fchollet i replied to @tansuyegen's tweet

2022-11-16 22:58:06 @TansuYegen I think you have misunderstood wha the tech purports to do. I think that it is morphing actual human to look like a a different human. still cool, but not what you describe. see https://t.co/Lk5wipCh2c Face Shifting tab

2022-11-16 21:30:15 RT @karaswisher: Read these fine few words from ⁦@GaryMarcus⁩ https://t.co/Giqr1JI0by

2022-11-16 18:48:44 RT @ruthaylett: He is right. People laughing at this are mistaken, it is a serious problem in the making.

2022-11-16 18:40:34 RT @RemivanTrijp: I sincerely hope this doesn’t turn out to be prophetic

2022-11-16 17:56:14 @Grady_Booch @elonmusk I wasn’t assigning probabilities here

2022-11-16 17:52:44 @Grady_Booch @elonmusk Only two ways this can go - Musk can succeed, and rewrite everything we thought we knew about management and software engineering - He can fail, and become the textbook example of how not to do things.

2022-11-16 17:49:47 There’s a lot going on. Midterms, FTX, Twitter chaos. But in time we will see the release of #Galactica as an epochal event, a tipping point in a gigantic increase in the flow of misinformation. This is no joke. Galactica is funny, but the uses it will be put to are not. https://t.co/UPoe1q4BEi

2022-11-16 17:22:36 @MetaAI’s Galactica has made up my birthday (wrong decade), my education,my research interests, etc, all sounding plausible, 85% untrue. You know what’s really embarrassing about this? A decent AI could just look this stuff on the web &

2022-11-16 15:04:18 @ylecun @MetaAI @paperswithcode https://t.co/EvMKYUZcN0

2022-11-16 14:54:38 RT @pelimuc:

2022-11-16 14:41:15 never had this happened before - timeout loading my own Tweets. Things gonna get wild here in Bird-land! https://t.co/URBAdZpBhX

2022-11-16 14:38:05 For examples, see https://t.co/GUdfR6nR6y and this bonus fake bio of @ylecun here: https://t.co/gN1gvFkSjh

2022-11-16 14:38:03 AI Joke of the Day! Q: What do a CNC machine and MetaAI’s #Galactica have in common? .. .. .. A: They both help you fabricate! https://t.co/uEouaaO8iQ

2022-11-16 14:11:51 RT @dmonett: "Is this really what AI has come to, automatically mixing reality with bullshit so finely we can no longer recognize the diffe…

2022-11-16 14:11:23 RT @dougclow: Scientific bullshit generator, argh! Having spent many disheartening hours reviewing academic manuscripts that strongly resem…

2022-11-16 06:56:19 @kevinroose you don’t think the product placement for Storybook Brawl was a nice touch? https://t.co/15jySgxqR9 @TrungTPhan

2022-11-16 05:49:14 @mcuban just liked my new post

2022-11-16 05:35:37 @_jasonwei @colinraffel i need to think about @colinraffel’s specific tasks but i am willing to bet money that LLMs will continue to face serious challenges in factuality even as they scale.

2022-11-16 00:08:58 @ykilcher which benchmarks have you published in the last four weeks? I have posted two, one in a hot new ML journal, linked above, which you didn’t comment, the other on arXiv. You? the third was published in AI Magazine 2015. It’s not *my* fault that the field doesn’t rise to them.

2022-11-15 20:59:45 @ykilcher Not my job to decide what community disagrees about. But people ought look at the zero-shot version of Truthful QA and to test comprehension as I described in the 2014 New Yorker article Beyond The Turing Test, as well recent vision benchmark I helped w: https://t.co/LDPoiyTPNi

2022-11-15 19:02:22 OpenAI give me access to GPT-4? https://t.co/mN0OSkugcZ

2022-11-15 16:51:13 Rumor has it GPT-4 is coming. Here’s a prediction: it will still have serious issues with confabulation/misinformation, toxicity and bias. https://t.co/TKm6DA3Dvb

2022-11-15 16:24:59 . @sacca wasn’t wrong. https://t.co/2WVxcZAPMn

2022-11-15 15:23:55 "it was sobering to see how close we came to hitting a parked car after we rolled over a low curb separating the parking lot." — @cademetz on six hours in a Tesla:https://t.co/UUQXUBzE4q

2022-11-15 14:36:53 @_jasonwei @drjwrae @OwainEvans_UK post hoc.performance on original 0-shot task is not what you graphed. you are graphing only 10-shot results (and focusing on the multiple choice results which are the easier form of the task).On zero shot MC, 280B performance is <

2022-11-15 06:42:33 @_jasonwei @drjwrae @OwainEvans_UK If they used a different method you can’t tell how much is method change and how much is scale

2022-11-14 19:26:55 RT @QasimRashid: This is horrifying. The Iranian parliament votes overwhelmingly (227-63) to execute the 15,000 protesters they already arr…

2022-11-14 19:24:35 @_jasonwei @OwainEvans_UK your drawing doesn't match the leaderboard here: https://t.co/7HGyLImHHs and i think it's because Gopher changed the task...

2022-11-14 19:08:28 RT @svpino: In 2017, a team led by Andrew Ng published a paper showing off a Deep Learning model to detect pneumonia.Andrew is one of the…

2022-11-14 18:45:29 @_jasonwei Taking TruthfulQA as an example, your evidence for scaling seems (if I understand correctly) to come from a single data point from Gopher that was (a) selected post hoc, (b) has not been replicated and (c) relied on making the task multiple choice. cc @OwainEvans_UKAm I wrong?

2022-11-14 14:06:15 RT @federicobianchy: Text-to-image generation models (like Stable Diffusion and DALLE) are being used to generate millions of images a day.…

2022-11-14 13:10:22 RT @drewharwell: New: That fake verified Eli Lilly account sparked a panic inside the pharma giant, and senior execs decided they'd had eno…

2022-11-14 01:58:42 @karaswisher Will certainly be interesting to see what happens to advertising revenue in the Lockheed Martini era…

2022-11-21 04:18:23 epic trolls from @rsalakhu (2 of 2, and so subtle the target isn’t even named). 100% factual, too. @soumithchintala’s response, which I initially missed, is indeed great and very professional. (He’s not the one being trolled.) https://t.co/ar5MIW5nFr

2022-11-21 04:17:18 @soumithchintala @moreisdifferent Appreciate your sober thread. I do think though that there is an inherent unsolved challenge though inasmuch as no current tech can really eliminate the widespread hallucinations.

2022-11-21 04:08:02 epic trolls from @rsalakhu (1 of 2) https://t.co/cMuq49z0Bo

2022-11-21 04:05:29 @pmddomingos @YejinChoinka in the limit of unlimited data, they are same, as I said. The difference is in the learning. But the learning isn’t real compositionality

2022-11-21 02:57:56 @pmddomingos @YejinChoinka And see also https://t.co/UWUg6kQ3Df by @EvelinaLeivada @ElliotMurphy91 and myself in different but related architecture

2022-11-21 02:55:39 @pmddomingos In the limit they are the same (universal function approximators). They are better at learning syntax but they still don’t have a compositional semantics. Which is why eg they fail at @YejinChoinka’s recent https://t.co/wsic38AdLR

2022-11-21 02:42:05 @pmddomingos Representing grammar is cheap - lots of architecture can do that in the limit. learning grammar from data is same problem it ever was. And you are still inaccurate in what you said about compositionality.

2022-11-21 02:22:41 @pmddomingos not quite right. They have a proxy that looks vaguely like compositionality. They don’t have meanings lawfully composed from parts, governed by syntax.

2022-11-21 02:18:55 amazing how many people missed that this was a joke https://t.co/myv5yd2xEU

2022-11-21 00:04:24 RT @MuhammedKambal: @GaryMarcus On a related note, this morning I watched your debate with Chnosky on debunking the big lie of AI. Quite ex…

2022-11-20 23:34:04 @Ted_Underwood @CriticalAI @true_mxp @JrKibs @_akpiper https://t.co/wsic38AdLR

2022-11-20 23:18:10 @ylecun @untitled01ipynb @rsalakhu @rsalakhu trolls @ylecun and I get blamed. (Does this count as a countercountertroll?)

2022-11-20 19:55:33 @ylecun @CriticalAI @TonyZador @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian Just answer this one question, Yann: https://t.co/BctdIM1PSW

2022-11-20 19:49:08 @PeterShor1 @ylecun @Grady_Booch @Jeff_Aronson @Abebab Were there any usability studies with actual users? You’d want to evaluate accuracy, and not just speed.

2022-11-20 18:10:34 Tweeted this day before Galactica came out https://t.co/p0cb9QUOW1

2022-11-20 18:09:27 @Ted_Underwood @true_mxp @JrKibs @_akpiper people can (and are) approaching it, but they can’t do it. with training and diet, i can “approach” running a 2 min mile, but it ain’t gonna happen with my body.

2022-11-20 18:07:18 @ykilcher has anyone done *any*? how? I just asked Yann re FB

2022-11-20 18:06:03 @Ted_Underwood @true_mxp @JrKibs @_akpiper the abject failure of Galactica to do so is btw an object lesson here. so are the woeful results on TruthfulQA across the whole field, and much harder tasks might be devised.

2022-11-20 18:03:37 @Ted_Underwood @true_mxp @JrKibs @_akpiper no, current software cannot do that remotely reliable

2022-11-20 03:29:28 @AI4Code @SumitGulwani

2022-11-20 01:02:03 October 28: No reinstatements without a council Nov 19: Never mind about the council. I asked my fans, mostly right of center. Trump’s back in. https://t.co/b3xrNznGcO

2022-11-20 00:45:06 @ylecun @drng @Grady_Booch @Jeff_Aronson @Abebab @mrgreene1977 Galactica has deeper problems: - it is unconstrained by the facts are in its database (eg it said that Elon Musk died in 2018 car crash but probably has data to the contrary) - it gives no indication when it makes something up that isn’t database

2022-11-20 00:33:06 @ylecun @drng @Grady_Booch @Jeff_Aronson @Abebab I think what he means that the model made false inferences eg about value in eating crushed glass (per @mrgreene1977) https://t.co/eOlWcTDTxL

2022-11-20 00:31:09 @mark_riedl @Sergei_Imaging @ylecun @mrgreene1977 no doubt that cat is out of the bag, and it’s easy enough to replicate. soon enough we find out empirically how much impact systems like these have. (though of course hard to trace since the quality of the mimicry is high and there are no watermarks)

2022-11-20 00:29:43 @srijankedia @ylecun @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian @emilio__ferrara @davidlazer @gianluca_string @CrashTheMod3 It’s quite worrisome that @metaAI’s Chief AI officer shows little apparent familiarity with such work.

2022-11-20 00:27:54 @tdietterich @ylecun @mrgreene1977 and they are in fact starting to be used, and as cost of doing them well decreases, we can expect them to be more widespread. some examples from 2022: https://t.co/4kxqUjHXXT 2024 election, who knows?

2022-11-20 00:22:17 #Galactica is fine, because if it makes stuff up, no problemo. @metaAI’s Chief AI Officer just told me so. And he’s *sure* nobody would ever misuse his tool. Yeah, right. https://t.co/9uk9YE56ad

2022-11-20 00:20:04 @noUpside @ylecun @mrgreene1977 @katestarbird @sinanaral @ProfNoahGian Ask Yann, who thinks it is actually hunky-dory, but my theory was this: https://t.co/tfjaWTQuHW

2022-11-20 00:10:39 @ylecun @srijankedia @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian Low probabilities ≠ zero probabilities. Low probabilities with low cost ≠ zero probabilities with zero cost. Every spammer, ever, knows this. Same principles will apply to misinfo

2022-11-20 00:06:40 Twitter on IOS just crashed midtweet and I lost my work. First time that happened in a decade of using it. also having some load time issues. And so it begins.

2022-11-20 00:02:31 @ylecun @jppesky @mrgreene1977 on latter point, better said: https://t.co/SLZ7lVHQ4f

2022-11-20 00:00:29 @ylecun @jppesky @mrgreene1977 You are coming across as obtuse. - You act as if misinfo has had no impact in the real world (despite success of anti-vax campaigns) - Act as if the only way in which scientific-sounding nonsense has impact is if it is in a peer review journal - Treat low prob as zero probability

2022-11-19 23:52:54 @TonyZador @ylecun @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian perhaps someone in 1985 said “email isn’t going to have any effect on scamming. it’s driven by how much people want to be taken.” the biggest application of LLMs is now? it’s apparently making nonsense for SEO optimization. New ways to do in volume, because cost is cheaper.

2022-11-19 23:14:23 @ylecun @mrgreene1977 Come on, it’s not just about journals. To take one example, vaccine misinformation, oft presented as “scientific” [eg https://t.co/B1JmJihJpl], likely had significant impact on vaccination rates. Tagging @noUpside @katestarbird @sinanaral @ProfNoahGian for more perspectives.

2022-11-19 22:13:07 RT @GaryMarcus: @ylecun @mrgreene1977 How could Galactica *not* make misinfo easier, @ylecun, when it makes writing faux scientific-like a…

2022-11-19 22:11:50 @ylecun @mrgreene1977 How could Galactica *not* make misinfo easier, @ylecun, when it makes writing faux scientific-like articles trivial rather than laborious? &

2022-11-19 19:43:40 @ylecun @Grady_Booch @Jeff_Aronson @Abebab The sand is in thinking that a system like this can stick to reality. It cannot.

2022-11-19 19:13:54 @ylecun @Abebab Ps repeating for convenience: https://t.co/BBE3moQFea

2022-11-18 21:21:17 @ylecun @mrgreene1977 Potentially deadly TikTok challenges

2022-11-18 21:15:06 RT @michaelharriot: They won’t know what breaks until it breaks. It’s a matter of what breaks. “It’s not like McDonalds’ bro. You can’t ta…

2022-11-18 21:06:16 https://t.co/7aDMXp6G3g

2022-11-18 19:21:26 RT @GaryMarcus: @elonmusk What about misinformation?

2022-11-18 19:20:56 I Like Free Speech So Much I’ve Decided to Buy It - McSweeney’s https://t.co/Yw3n94HAQW

2022-11-18 19:18:59 @elonmusk What about misinformation?

2022-11-18 19:06:43 Galactica

2022-11-18 15:50:32 @yoavgo @ylecun @rasbt @manes @Abebab Troll farms cost money and presumably limited by budget. The more they can put out the more effective they are.

2022-11-18 15:36:47 @yoavgo @rasbt @manes @ylecun @Abebab On the scenario, secondary customers like journalists will have no longer have any trust in arXiv at all.

2022-11-18 15:34:22 @ylecun @rasbt @manes @Abebab It will reduce the cost hence increase the volume, hence make the problem worse.

2022-11-18 15:12:49 RT @GaryMarcus: @ylecun @Abebab Galactica *is* dangerous. And you have no idea how to eliminate that danger. large language models inherent…

2022-11-18 15:12:34 @ylecun @Abebab Galactica *is* dangerous. And you have no idea how to eliminate that danger. large language models inherently hallucinate. It weaponized the ability of these tools to spread fake science. The vitriol is proportionate to the risk. I am shocked that you are indifferent to this.

2022-11-18 15:06:40 “once perfected” is carrying a lot of weight here. there is no way Galactica will be perfected in next year or two, probably a lot longer. happy to publicly debate you about this, but recommend you start by rereading your own recent critique @noemamag of large language models. https://t.co/LYXD0E0jyI

2022-11-18 13:51:45 very rational fear https://t.co/bY0hUrnzcv

2022-11-18 13:41:25 follow friday may not be a thing much longer over here on Twitter but follow @abebab. And she’s on Mastodon now, too. https://t.co/29GfLksNFx

2022-11-18 13:38:38 RT @Abebab: @ylecun Pretty much exactly what happened was you overhyped and released a model. People tested it (you should be thankful for…

2022-11-18 13:15:06 Just saying, in case this ship goes down, @garymarcus@sigmoid.social

2022-11-18 12:55:18 @HeidyKhlaaf @AmandaAskell @andrewthesmart @Abebab 100% support thinking through the potential social consequences. There is a lot of room for important argument there.

2022-11-18 12:48:16 @HeidyKhlaaf @AmandaAskell @andrewthesmart @Abebab Thank you. and to be clear I am all for spelling out the latter part and will amplify the message if is is constructed carefully and respectfully. (If you want to write something longer, DM.)

2022-11-18 12:24:50 @HeidyKhlaaf @AmandaAskell @andrewthesmart Yes, there are lot of ideas within EA, some well-intentioned, some not. some briefly considered, some taken seriously, some not

2022-11-17 23:52:52 Tripling down, @yLeCun's contention is that people "abused" Galactica. But what safeguards were put in place to prevent bad actors from using the system to generate compelling disinformation around scientific topics like vaccines or racial differences? A: None. https://t.co/HlVjdHz1Sr

2022-11-17 23:39:33 @mmitchell_ai @trashcanmagic @timnitGebru But that is not me. I fully agree with Dr King's methods, and with direct action

2022-11-17 23:18:17 @andrewthesmart So the left is now pro-choice but anti-genetic screening? This is eugenics? https://t.co/gu4vD7v0en

2022-11-17 22:10:22 @emilymbender @timnitGebru and if what you say is true, you can simply provide a link. there certainly has not been anything from either of you in direct replies to me when i have asked.

2022-11-17 22:01:59 Why is this is any different at all than McCarthyism? I honestly don't see a difference between labeling someone a communist and labeling them a eugenicist, if you do so without either evidence or definition. I have asked, repeatedly, for both, and gotten neither. https://t.co/efvhRYCgEw

2022-11-17 21:54:48 @emilymbender @timnitGebru Nonsense. I have repeatedly asked a legitimate,straightforward question. Neither of you have answered: are you claiming eg that MacAskill (whom I have never met) is a eugenicist a la 19th century scientific racism? yes or no, and if yes, in what sense, and what is the argument?

2022-11-17 21:52:13 @athundt @timnitGebru The argument against Singer there is much more straightforward

2022-11-17 21:50:36 @athundt @timnitGebru I don't doubt that scientific racism still exists. But that label should not be applied to specific people without real evidence. That's all I am saying. &

2022-11-17 21:38:20 I have no shame about speaking out against McCarthyism.

2022-11-17 21:32:21 @timnitGebru no definitions still? i will carry on.

2022-11-17 21:26:23 @trashcanmagic @timnitGebru How dare you? My father was doing sit-ins for fair housing before you were born, and did discrimination law for years. I helped him (with statistics) as a teenager, and went to protests from when I was a newborn. I am with your struggle, but not your tactics.

2022-11-17 21:24:28 @timnitGebru We need definitions, @timnitGebru, if you want me and others take you seriously. I was once a huge fan of yours, but am really losing the faith.

2022-11-17 21:16:44 @timnitGebru Only quote I have seen was nearly fabricated, inserting words that Bostrom didn't use. I haven't seen any direct evidence on MacAskill. &

2022-11-17 21:07:19 @timnitGebru If I had been alive in the 1950s I hope that I would have called out McCarthy (white male) for doing what you do: naming names, without due process, without definitions, without evidence. It's not what groups you belong to Timnit, it's what you are saying that is problematic.

2022-11-17 20:41:16 RT @GaryMarcus: The reality is that large language models like GPT-3 annd Galactica are like bulls in a china shop, powerful but reckless.…

2022-11-17 19:56:24 @ylecun @Grady_Booch And they use lots of symbol-manipulation too. Funny that you never mention that.

2022-11-17 19:52:00 RT @GidMK: @GaryMarcus I'm worried about fake science. There are already hundreds of pay-to-play journals that publish openly plagiarised w…

2022-11-17 19:42:43 I doubt it. but Prince said it best, “Forever [is] a mighty long time” https://t.co/VXE5Up59tl

2022-11-17 19:32:50 RT @Plinz: Generative AI is incredibly useful and valuable, but it will force us to change our practices in science, journalism and social…

2022-11-17 19:31:30 @yLeCun: It’s all good, man https://t.co/PrX2PBawlc

2022-11-17 03:58:52 RT @frossi_t: On my way to the AAAI fall symposium on “thinking fast and slow and other cognitive theories in AI”. The program includes 12…

2022-11-17 01:37:42 RT @tytusm: 100% agree with Gary The threat posed to science by #Galactica is real Try it yourself using the gold standard…

2022-11-16 23:11:10 Flattery will get you nowhere, young battlestar. (image courtesy @UlrichJunker) https://t.co/d8GYGDCgTm

2022-11-16 22:59:03 @shokunin_studio @Grady_Booch @fchollet i replied to @tansuyegen's tweet

2022-11-16 22:58:06 @TansuYegen I think you have misunderstood wha the tech purports to do. I think that it is morphing actual human to look like a a different human. still cool, but not what you describe. see https://t.co/Lk5wipCh2c Face Shifting tab

2022-11-16 21:30:15 RT @karaswisher: Read these fine few words from ⁦@GaryMarcus⁩ https://t.co/Giqr1JI0by

2022-11-16 18:48:44 RT @ruthaylett: He is right. People laughing at this are mistaken, it is a serious problem in the making.

2022-11-16 18:40:34 RT @RemivanTrijp: I sincerely hope this doesn’t turn out to be prophetic

2022-11-16 17:56:14 @Grady_Booch @elonmusk I wasn’t assigning probabilities here

2022-11-16 17:52:44 @Grady_Booch @elonmusk Only two ways this can go - Musk can succeed, and rewrite everything we thought we knew about management and software engineering - He can fail, and become the textbook example of how not to do things.

2022-11-16 17:49:47 There’s a lot going on. Midterms, FTX, Twitter chaos. But in time we will see the release of #Galactica as an epochal event, a tipping point in a gigantic increase in the flow of misinformation. This is no joke. Galactica is funny, but the uses it will be put to are not. https://t.co/UPoe1q4BEi

2022-11-16 17:22:36 @MetaAI’s Galactica has made up my birthday (wrong decade), my education,my research interests, etc, all sounding plausible, 85% untrue. You know what’s really embarrassing about this? A decent AI could just look this stuff on the web &

2022-11-16 15:04:18 @ylecun @MetaAI @paperswithcode https://t.co/EvMKYUZcN0

2022-11-16 14:54:38 RT @pelimuc:

2022-11-16 14:41:15 never had this happened before - timeout loading my own Tweets. Things gonna get wild here in Bird-land! https://t.co/URBAdZpBhX

2022-11-16 14:38:05 For examples, see https://t.co/GUdfR6nR6y and this bonus fake bio of @ylecun here: https://t.co/gN1gvFkSjh

2022-11-16 14:38:03 AI Joke of the Day! Q: What do a CNC machine and MetaAI’s #Galactica have in common? .. .. .. A: They both help you fabricate! https://t.co/uEouaaO8iQ

2022-11-16 14:11:51 RT @dmonett: "Is this really what AI has come to, automatically mixing reality with bullshit so finely we can no longer recognize the diffe…

2022-11-16 14:11:23 RT @dougclow: Scientific bullshit generator, argh! Having spent many disheartening hours reviewing academic manuscripts that strongly resem…

2022-11-16 06:56:19 @kevinroose you don’t think the product placement for Storybook Brawl was a nice touch? https://t.co/15jySgxqR9 @TrungTPhan

2022-11-16 05:49:14 @mcuban just liked my new post

2022-11-16 05:35:37 @_jasonwei @colinraffel i need to think about @colinraffel’s specific tasks but i am willing to bet money that LLMs will continue to face serious challenges in factuality even as they scale.

2022-11-16 00:08:58 @ykilcher which benchmarks have you published in the last four weeks? I have posted two, one in a hot new ML journal, linked above, which you didn’t comment, the other on arXiv. You? the third was published in AI Magazine 2015. It’s not *my* fault that the field doesn’t rise to them.

2022-11-15 20:59:45 @ykilcher Not my job to decide what community disagrees about. But people ought look at the zero-shot version of Truthful QA and to test comprehension as I described in the 2014 New Yorker article Beyond The Turing Test, as well recent vision benchmark I helped w: https://t.co/LDPoiyTPNi

2022-11-15 19:02:22 OpenAI give me access to GPT-4? https://t.co/mN0OSkugcZ

2022-11-15 16:51:13 Rumor has it GPT-4 is coming. Here’s a prediction: it will still have serious issues with confabulation/misinformation, toxicity and bias. https://t.co/TKm6DA3Dvb

2022-11-15 16:24:59 . @sacca wasn’t wrong. https://t.co/2WVxcZAPMn

2022-11-15 15:23:55 "it was sobering to see how close we came to hitting a parked car after we rolled over a low curb separating the parking lot." — @cademetz on six hours in a Tesla:https://t.co/UUQXUBzE4q

2022-11-15 14:36:53 @_jasonwei @drjwrae @OwainEvans_UK post hoc.performance on original 0-shot task is not what you graphed. you are graphing only 10-shot results (and focusing on the multiple choice results which are the easier form of the task).On zero shot MC, 280B performance is <

2022-11-15 06:42:33 @_jasonwei @drjwrae @OwainEvans_UK If they used a different method you can’t tell how much is method change and how much is scale

2022-11-14 19:26:55 RT @QasimRashid: This is horrifying. The Iranian parliament votes overwhelmingly (227-63) to execute the 15,000 protesters they already arr…

2022-11-14 19:24:35 @_jasonwei @OwainEvans_UK your drawing doesn't match the leaderboard here: https://t.co/7HGyLImHHs and i think it's because Gopher changed the task...

2022-11-14 19:08:28 RT @svpino: In 2017, a team led by Andrew Ng published a paper showing off a Deep Learning model to detect pneumonia.Andrew is one of the…

2022-11-14 18:45:29 @_jasonwei Taking TruthfulQA as an example, your evidence for scaling seems (if I understand correctly) to come from a single data point from Gopher that was (a) selected post hoc, (b) has not been replicated and (c) relied on making the task multiple choice. cc @OwainEvans_UKAm I wrong?

2022-11-14 14:06:15 RT @federicobianchy: Text-to-image generation models (like Stable Diffusion and DALLE) are being used to generate millions of images a day.…

2022-11-14 13:10:22 RT @drewharwell: New: That fake verified Eli Lilly account sparked a panic inside the pharma giant, and senior execs decided they'd had eno…

2022-11-14 01:58:42 @karaswisher Will certainly be interesting to see what happens to advertising revenue in the Lockheed Martini era…

2022-11-22 04:29:55 Interesting that none of these (including in the replies) get the QWERTY layout even though so many pictures of that exist. https://t.co/v1wMtG2O6c

2022-11-22 03:55:02 @Tkaraletsos @eerac Irene Pepperberg has written quite bit about how clever they are.

2022-11-22 03:00:45 RT @tribelaw: Richard Fierro saved so many lives. He deserves all our gratitude. Thanksgiving! https://t.co/00VYhHdDwd

2022-11-22 02:34:59 @MichaelTrazzi has there been any significant progress in the (mis)alignment of goals since she started research the topic? how serious a problem is it now?

2022-11-22 02:30:26 Elon Musk got paid about 10% of the market cap of Tesla to pay attention to … Tesla (which has fallen by roughly half since he decided to buy Twitter). &

2022-11-22 01:25:28 I Vancouver https://t.co/ifnquaT7Fw

2022-11-22 00:49:33 Did AI fail to meet expectations again? Personally, I would pay real money for a virtual assistant of the caliber of the OS in the film Her, but not for an AI with the narrowness of Alexa or the hallucinations of Galactica. https://t.co/G68u1XKWG8

2022-11-22 00:08:54 Via @danielmalmer on LinkedIn on some hate-filled ads that were purchased (but not published)

2022-11-22 00:04:45 @riedelcastro @soumithchintala @moreisdifferent and would welcome other pointers to relevant work!

2022-11-22 00:04:09 @riedelcastro @soumithchintala @moreisdifferent best evidence that this is near-term plausible? i’ve enjoyed some of your work (eg on Wiki verifiability), but wonder how it would manage around the kinds of things I have seen Galactica produce/that I posted here and on my Substack over last few days.

2022-11-21 23:08:15 @NaveenGRao I also wrote this earlier in same topic: https://t.co/DAKe19V1Dy

2022-11-21 22:55:25 @NaveenGRao “We’re thinking about it but haven’t figured out how to do it” might be an honest answer -- but also at odds with the absolutely certainty that there has been zero harm that Yann professed earlier.

2022-11-21 21:25:39 @ylecun I thought subtweeting without tagging was rude? Anyway, my reply, it’s not like I missed yours: https://t.co/QhudBxILW5

2022-11-21 21:23:31 I am that dude. You have answered at least a dozen of my recent queries. But not that one

2022-11-21 21:17:22 @pmddomingos 100% agree and yet already >

2022-11-21 21:04:42 @pmddomingos true! It’s only anti-wokes that can turn Twitter into a wild success like Parler or Truth Social.

2022-11-21 20:54:45 @Pehdrew_ Super hard. Which is why you can’t blandly assert that there has been no harm. And why meta ought put real work into the problem and report what they find.

2022-11-21 20:51:49 @lorakolodny @mer__edith

2022-11-21 20:42:26 RT @Grady_Booch: @GaryMarcus When <

2022-11-21 20:39:52 @prem_k @Grady_Booch @abeba @Michael_J_Black @MadamePratolung https://t.co/KlQJhILKgf

2022-11-21 14:11:40 https://t.co/kiviFxW0RM

2022-11-21 14:07:53 Galactica, summed up perfectly @futurism https://t.co/cRRQVQhZno

2022-11-21 13:44:42 RT @KDHungerford: #Artificialintelligence has mostly been focusing on a technique called #deeplearning. It might be time to reconsider. - @…

2022-11-21 04:18:23 epic trolls from @rsalakhu (2 of 2, and so subtle the target isn’t even named). 100% factual, too. @soumithchintala’s response, which I initially missed, is indeed great and very professional. (He’s not the one being trolled.) https://t.co/ar5MIW5nFr

2022-11-21 04:17:18 @soumithchintala @moreisdifferent Appreciate your sober thread. I do think though that there is an inherent unsolved challenge though inasmuch as no current tech can really eliminate the widespread hallucinations.

2022-11-21 04:08:02 epic trolls from @rsalakhu (1 of 2) https://t.co/cMuq49z0Bo

2022-11-21 04:05:29 @pmddomingos @YejinChoinka in the limit of unlimited data, they are same, as I said. The difference is in the learning. But the learning isn’t real compositionality

2022-11-21 02:57:56 @pmddomingos @YejinChoinka And see also https://t.co/UWUg6kQ3Df by @EvelinaLeivada @ElliotMurphy91 and myself in different but related architecture

2022-11-21 02:55:39 @pmddomingos In the limit they are the same (universal function approximators). They are better at learning syntax but they still don’t have a compositional semantics. Which is why eg they fail at @YejinChoinka’s recent https://t.co/wsic38AdLR

2022-11-21 02:42:05 @pmddomingos Representing grammar is cheap - lots of architecture can do that in the limit. learning grammar from data is same problem it ever was. And you are still inaccurate in what you said about compositionality.

2022-11-21 02:22:41 @pmddomingos not quite right. They have a proxy that looks vaguely like compositionality. They don’t have meanings lawfully composed from parts, governed by syntax.

2022-11-21 02:18:55 amazing how many people missed that this was a joke https://t.co/myv5yd2xEU

2022-11-21 00:04:24 RT @MuhammedKambal: @GaryMarcus On a related note, this morning I watched your debate with Chnosky on debunking the big lie of AI. Quite ex…

2022-11-20 23:34:04 @Ted_Underwood @CriticalAI @true_mxp @JrKibs @_akpiper https://t.co/wsic38AdLR

2022-11-20 23:18:10 @ylecun @untitled01ipynb @rsalakhu @rsalakhu trolls @ylecun and I get blamed. (Does this count as a countercountertroll?)

2022-11-20 19:55:33 @ylecun @CriticalAI @TonyZador @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian Just answer this one question, Yann: https://t.co/BctdIM1PSW

2022-11-20 19:49:08 @PeterShor1 @ylecun @Grady_Booch @Jeff_Aronson @Abebab Were there any usability studies with actual users? You’d want to evaluate accuracy, and not just speed.

2022-11-20 18:10:34 Tweeted this day before Galactica came out https://t.co/p0cb9QUOW1

2022-11-20 18:09:27 @Ted_Underwood @true_mxp @JrKibs @_akpiper people can (and are) approaching it, but they can’t do it. with training and diet, i can “approach” running a 2 min mile, but it ain’t gonna happen with my body.

2022-11-20 18:07:18 @ykilcher has anyone done *any*? how? I just asked Yann re FB

2022-11-20 18:06:03 @Ted_Underwood @true_mxp @JrKibs @_akpiper the abject failure of Galactica to do so is btw an object lesson here. so are the woeful results on TruthfulQA across the whole field, and much harder tasks might be devised.

2022-11-20 18:03:37 @Ted_Underwood @true_mxp @JrKibs @_akpiper no, current software cannot do that remotely reliable

2022-11-20 03:29:28 @AI4Code @SumitGulwani

2022-11-20 01:02:03 October 28: No reinstatements without a council Nov 19: Never mind about the council. I asked my fans, mostly right of center. Trump’s back in. https://t.co/b3xrNznGcO

2022-11-20 00:45:06 @ylecun @drng @Grady_Booch @Jeff_Aronson @Abebab @mrgreene1977 Galactica has deeper problems: - it is unconstrained by the facts are in its database (eg it said that Elon Musk died in 2018 car crash but probably has data to the contrary) - it gives no indication when it makes something up that isn’t database

2022-11-20 00:33:06 @ylecun @drng @Grady_Booch @Jeff_Aronson @Abebab I think what he means that the model made false inferences eg about value in eating crushed glass (per @mrgreene1977) https://t.co/eOlWcTDTxL

2022-11-20 00:31:09 @mark_riedl @Sergei_Imaging @ylecun @mrgreene1977 no doubt that cat is out of the bag, and it’s easy enough to replicate. soon enough we find out empirically how much impact systems like these have. (though of course hard to trace since the quality of the mimicry is high and there are no watermarks)

2022-11-20 00:29:43 @srijankedia @ylecun @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian @emilio__ferrara @davidlazer @gianluca_string @CrashTheMod3 It’s quite worrisome that @metaAI’s Chief AI officer shows little apparent familiarity with such work.

2022-11-20 00:27:54 @tdietterich @ylecun @mrgreene1977 and they are in fact starting to be used, and as cost of doing them well decreases, we can expect them to be more widespread. some examples from 2022: https://t.co/4kxqUjHXXT 2024 election, who knows?

2022-11-20 00:22:17 #Galactica is fine, because if it makes stuff up, no problemo. @metaAI’s Chief AI Officer just told me so. And he’s *sure* nobody would ever misuse his tool. Yeah, right. https://t.co/9uk9YE56ad

2022-11-20 00:20:04 @noUpside @ylecun @mrgreene1977 @katestarbird @sinanaral @ProfNoahGian Ask Yann, who thinks it is actually hunky-dory, but my theory was this: https://t.co/tfjaWTQuHW

2022-11-20 00:10:39 @ylecun @srijankedia @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian Low probabilities ≠ zero probabilities. Low probabilities with low cost ≠ zero probabilities with zero cost. Every spammer, ever, knows this. Same principles will apply to misinfo

2022-11-20 00:06:40 Twitter on IOS just crashed midtweet and I lost my work. First time that happened in a decade of using it. also having some load time issues. And so it begins.

2022-11-20 00:02:31 @ylecun @jppesky @mrgreene1977 on latter point, better said: https://t.co/SLZ7lVHQ4f

2022-11-20 00:00:29 @ylecun @jppesky @mrgreene1977 You are coming across as obtuse. - You act as if misinfo has had no impact in the real world (despite success of anti-vax campaigns) - Act as if the only way in which scientific-sounding nonsense has impact is if it is in a peer review journal - Treat low prob as zero probability

2022-11-19 23:52:54 @TonyZador @ylecun @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian perhaps someone in 1985 said “email isn’t going to have any effect on scamming. it’s driven by how much people want to be taken.” the biggest application of LLMs is now? it’s apparently making nonsense for SEO optimization. New ways to do in volume, because cost is cheaper.

2022-11-19 23:14:23 @ylecun @mrgreene1977 Come on, it’s not just about journals. To take one example, vaccine misinformation, oft presented as “scientific” [eg https://t.co/B1JmJihJpl], likely had significant impact on vaccination rates. Tagging @noUpside @katestarbird @sinanaral @ProfNoahGian for more perspectives.

2022-11-19 22:13:07 RT @GaryMarcus: @ylecun @mrgreene1977 How could Galactica *not* make misinfo easier, @ylecun, when it makes writing faux scientific-like a…

2022-11-19 22:11:50 @ylecun @mrgreene1977 How could Galactica *not* make misinfo easier, @ylecun, when it makes writing faux scientific-like articles trivial rather than laborious? &

2022-11-19 19:43:40 @ylecun @Grady_Booch @Jeff_Aronson @Abebab The sand is in thinking that a system like this can stick to reality. It cannot.

2022-11-19 19:13:54 @ylecun @Abebab Ps repeating for convenience: https://t.co/BBE3moQFea

2022-11-18 21:21:17 @ylecun @mrgreene1977 Potentially deadly TikTok challenges

2022-11-18 21:15:06 RT @michaelharriot: They won’t know what breaks until it breaks. It’s a matter of what breaks. “It’s not like McDonalds’ bro. You can’t ta…

2022-11-18 21:06:16 https://t.co/7aDMXp6G3g

2022-11-18 19:21:26 RT @GaryMarcus: @elonmusk What about misinformation?

2022-11-18 19:20:56 I Like Free Speech So Much I’ve Decided to Buy It - McSweeney’s https://t.co/Yw3n94HAQW

2022-11-18 19:18:59 @elonmusk What about misinformation?

2022-11-18 19:06:43 Galactica

2022-11-18 15:50:32 @yoavgo @ylecun @rasbt @manes @Abebab Troll farms cost money and presumably limited by budget. The more they can put out the more effective they are.

2022-11-18 15:36:47 @yoavgo @rasbt @manes @ylecun @Abebab On the scenario, secondary customers like journalists will have no longer have any trust in arXiv at all.

2022-11-18 15:34:22 @ylecun @rasbt @manes @Abebab It will reduce the cost hence increase the volume, hence make the problem worse.

2022-11-18 15:12:49 RT @GaryMarcus: @ylecun @Abebab Galactica *is* dangerous. And you have no idea how to eliminate that danger. large language models inherent…

2022-11-18 15:12:34 @ylecun @Abebab Galactica *is* dangerous. And you have no idea how to eliminate that danger. large language models inherently hallucinate. It weaponized the ability of these tools to spread fake science. The vitriol is proportionate to the risk. I am shocked that you are indifferent to this.

2022-11-18 15:06:40 “once perfected” is carrying a lot of weight here. there is no way Galactica will be perfected in next year or two, probably a lot longer. happy to publicly debate you about this, but recommend you start by rereading your own recent critique @noemamag of large language models. https://t.co/LYXD0E0jyI

2022-11-18 13:51:45 very rational fear https://t.co/bY0hUrnzcv

2022-11-18 13:41:25 follow friday may not be a thing much longer over here on Twitter but follow @abebab. And she’s on Mastodon now, too. https://t.co/29GfLksNFx

2022-11-18 13:38:38 RT @Abebab: @ylecun Pretty much exactly what happened was you overhyped and released a model. People tested it (you should be thankful for…

2022-11-18 13:15:06 Just saying, in case this ship goes down, @garymarcus@sigmoid.social

2022-11-18 12:55:18 @HeidyKhlaaf @AmandaAskell @andrewthesmart @Abebab 100% support thinking through the potential social consequences. There is a lot of room for important argument there.

2022-11-18 12:48:16 @HeidyKhlaaf @AmandaAskell @andrewthesmart @Abebab Thank you. and to be clear I am all for spelling out the latter part and will amplify the message if is is constructed carefully and respectfully. (If you want to write something longer, DM.)

2022-11-18 12:24:50 @HeidyKhlaaf @AmandaAskell @andrewthesmart Yes, there are lot of ideas within EA, some well-intentioned, some not. some briefly considered, some taken seriously, some not

2022-11-17 23:52:52 Tripling down, @yLeCun's contention is that people "abused" Galactica. But what safeguards were put in place to prevent bad actors from using the system to generate compelling disinformation around scientific topics like vaccines or racial differences? A: None. https://t.co/HlVjdHz1Sr

2022-11-17 23:39:33 @mmitchell_ai @trashcanmagic @timnitGebru But that is not me. I fully agree with Dr King's methods, and with direct action

2022-11-17 23:18:17 @andrewthesmart So the left is now pro-choice but anti-genetic screening? This is eugenics? https://t.co/gu4vD7v0en

2022-11-17 22:10:22 @emilymbender @timnitGebru and if what you say is true, you can simply provide a link. there certainly has not been anything from either of you in direct replies to me when i have asked.

2022-11-17 22:01:59 Why is this is any different at all than McCarthyism? I honestly don't see a difference between labeling someone a communist and labeling them a eugenicist, if you do so without either evidence or definition. I have asked, repeatedly, for both, and gotten neither. https://t.co/efvhRYCgEw

2022-11-17 21:54:48 @emilymbender @timnitGebru Nonsense. I have repeatedly asked a legitimate,straightforward question. Neither of you have answered: are you claiming eg that MacAskill (whom I have never met) is a eugenicist a la 19th century scientific racism? yes or no, and if yes, in what sense, and what is the argument?

2022-11-17 21:52:13 @athundt @timnitGebru The argument against Singer there is much more straightforward

2022-11-17 21:50:36 @athundt @timnitGebru I don't doubt that scientific racism still exists. But that label should not be applied to specific people without real evidence. That's all I am saying. &

2022-11-17 21:38:20 I have no shame about speaking out against McCarthyism.

2022-11-17 21:32:21 @timnitGebru no definitions still? i will carry on.

2022-11-17 21:26:23 @trashcanmagic @timnitGebru How dare you? My father was doing sit-ins for fair housing before you were born, and did discrimination law for years. I helped him (with statistics) as a teenager, and went to protests from when I was a newborn. I am with your struggle, but not your tactics.

2022-11-17 21:24:28 @timnitGebru We need definitions, @timnitGebru, if you want me and others take you seriously. I was once a huge fan of yours, but am really losing the faith.

2022-11-17 21:16:44 @timnitGebru Only quote I have seen was nearly fabricated, inserting words that Bostrom didn't use. I haven't seen any direct evidence on MacAskill. &

2022-11-17 21:07:19 @timnitGebru If I had been alive in the 1950s I hope that I would have called out McCarthy (white male) for doing what you do: naming names, without due process, without definitions, without evidence. It's not what groups you belong to Timnit, it's what you are saying that is problematic.

2022-11-17 20:41:16 RT @GaryMarcus: The reality is that large language models like GPT-3 annd Galactica are like bulls in a china shop, powerful but reckless.…

2022-11-17 19:56:24 @ylecun @Grady_Booch And they use lots of symbol-manipulation too. Funny that you never mention that.

2022-11-17 19:52:00 RT @GidMK: @GaryMarcus I'm worried about fake science. There are already hundreds of pay-to-play journals that publish openly plagiarised w…

2022-11-17 19:42:43 I doubt it. but Prince said it best, “Forever [is] a mighty long time” https://t.co/VXE5Up59tl

2022-11-17 19:32:50 RT @Plinz: Generative AI is incredibly useful and valuable, but it will force us to change our practices in science, journalism and social…

2022-11-17 19:31:30 @yLeCun: It’s all good, man https://t.co/PrX2PBawlc

2022-11-17 03:58:52 RT @frossi_t: On my way to the AAAI fall symposium on “thinking fast and slow and other cognitive theories in AI”. The program includes 12…

2022-11-17 01:37:42 RT @tytusm: 100% agree with Gary The threat posed to science by #Galactica is real Try it yourself using the gold standard…

2022-11-16 23:11:10 Flattery will get you nowhere, young battlestar. (image courtesy @UlrichJunker) https://t.co/d8GYGDCgTm

2022-11-16 22:59:03 @shokunin_studio @Grady_Booch @fchollet i replied to @tansuyegen's tweet

2022-11-16 22:58:06 @TansuYegen I think you have misunderstood wha the tech purports to do. I think that it is morphing actual human to look like a a different human. still cool, but not what you describe. see https://t.co/Lk5wipCh2c Face Shifting tab

2022-11-16 21:30:15 RT @karaswisher: Read these fine few words from ⁦@GaryMarcus⁩ https://t.co/Giqr1JI0by

2022-11-16 18:48:44 RT @ruthaylett: He is right. People laughing at this are mistaken, it is a serious problem in the making.

2022-11-16 18:40:34 RT @RemivanTrijp: I sincerely hope this doesn’t turn out to be prophetic

2022-11-16 17:56:14 @Grady_Booch @elonmusk I wasn’t assigning probabilities here

2022-11-16 17:52:44 @Grady_Booch @elonmusk Only two ways this can go - Musk can succeed, and rewrite everything we thought we knew about management and software engineering - He can fail, and become the textbook example of how not to do things.

2022-11-16 17:49:47 There’s a lot going on. Midterms, FTX, Twitter chaos. But in time we will see the release of #Galactica as an epochal event, a tipping point in a gigantic increase in the flow of misinformation. This is no joke. Galactica is funny, but the uses it will be put to are not. https://t.co/UPoe1q4BEi

2022-11-16 17:22:36 @MetaAI’s Galactica has made up my birthday (wrong decade), my education,my research interests, etc, all sounding plausible, 85% untrue. You know what’s really embarrassing about this? A decent AI could just look this stuff on the web &

2022-11-16 15:04:18 @ylecun @MetaAI @paperswithcode https://t.co/EvMKYUZcN0

2022-11-16 14:54:38 RT @pelimuc:

2022-11-16 14:41:15 never had this happened before - timeout loading my own Tweets. Things gonna get wild here in Bird-land! https://t.co/URBAdZpBhX

2022-11-16 14:38:05 For examples, see https://t.co/GUdfR6nR6y and this bonus fake bio of @ylecun here: https://t.co/gN1gvFkSjh

2022-11-16 14:38:03 AI Joke of the Day! Q: What do a CNC machine and MetaAI’s #Galactica have in common? .. .. .. A: They both help you fabricate! https://t.co/uEouaaO8iQ

2022-11-16 14:11:51 RT @dmonett: "Is this really what AI has come to, automatically mixing reality with bullshit so finely we can no longer recognize the diffe…

2022-11-16 14:11:23 RT @dougclow: Scientific bullshit generator, argh! Having spent many disheartening hours reviewing academic manuscripts that strongly resem…

2022-11-16 06:56:19 @kevinroose you don’t think the product placement for Storybook Brawl was a nice touch? https://t.co/15jySgxqR9 @TrungTPhan

2022-11-16 05:49:14 @mcuban just liked my new post

2022-11-16 05:35:37 @_jasonwei @colinraffel i need to think about @colinraffel’s specific tasks but i am willing to bet money that LLMs will continue to face serious challenges in factuality even as they scale.

2022-11-16 00:08:58 @ykilcher which benchmarks have you published in the last four weeks? I have posted two, one in a hot new ML journal, linked above, which you didn’t comment, the other on arXiv. You? the third was published in AI Magazine 2015. It’s not *my* fault that the field doesn’t rise to them.

2022-11-15 20:59:45 @ykilcher Not my job to decide what community disagrees about. But people ought look at the zero-shot version of Truthful QA and to test comprehension as I described in the 2014 New Yorker article Beyond The Turing Test, as well recent vision benchmark I helped w: https://t.co/LDPoiyTPNi

2022-11-15 19:02:22 OpenAI give me access to GPT-4? https://t.co/mN0OSkugcZ

2022-11-15 16:51:13 Rumor has it GPT-4 is coming. Here’s a prediction: it will still have serious issues with confabulation/misinformation, toxicity and bias. https://t.co/TKm6DA3Dvb

2022-11-15 16:24:59 . @sacca wasn’t wrong. https://t.co/2WVxcZAPMn

2022-11-15 15:23:55 "it was sobering to see how close we came to hitting a parked car after we rolled over a low curb separating the parking lot." — @cademetz on six hours in a Tesla:https://t.co/UUQXUBzE4q

2022-11-15 14:36:53 @_jasonwei @drjwrae @OwainEvans_UK post hoc.performance on original 0-shot task is not what you graphed. you are graphing only 10-shot results (and focusing on the multiple choice results which are the easier form of the task).On zero shot MC, 280B performance is <

2022-11-15 06:42:33 @_jasonwei @drjwrae @OwainEvans_UK If they used a different method you can’t tell how much is method change and how much is scale

2022-11-14 19:26:55 RT @QasimRashid: This is horrifying. The Iranian parliament votes overwhelmingly (227-63) to execute the 15,000 protesters they already arr…

2022-11-14 19:24:35 @_jasonwei @OwainEvans_UK your drawing doesn't match the leaderboard here: https://t.co/7HGyLImHHs and i think it's because Gopher changed the task...

2022-11-14 19:08:28 RT @svpino: In 2017, a team led by Andrew Ng published a paper showing off a Deep Learning model to detect pneumonia.Andrew is one of the…

2022-11-14 18:45:29 @_jasonwei Taking TruthfulQA as an example, your evidence for scaling seems (if I understand correctly) to come from a single data point from Gopher that was (a) selected post hoc, (b) has not been replicated and (c) relied on making the task multiple choice. cc @OwainEvans_UKAm I wrong?

2022-11-14 14:06:15 RT @federicobianchy: Text-to-image generation models (like Stable Diffusion and DALLE) are being used to generate millions of images a day.…

2022-11-14 13:10:22 RT @drewharwell: New: That fake verified Eli Lilly account sparked a panic inside the pharma giant, and senior execs decided they'd had eno…

2022-11-14 01:58:42 @karaswisher Will certainly be interesting to see what happens to advertising revenue in the Lockheed Martini era…

2022-11-23 00:31:29 RT @plibin: Truth is the Achilles' heel of current large language model AIs

2022-11-22 23:48:47 Really shocking how many platform owners and self-professed engineers can’t distinguish between vicious and viscous. Navier-Stokes for the win! https://t.co/fczz8lD4bl

2022-11-22 23:46:00 @MicahHoffmann @Grady_Booch I defer to @davidchalmers42 @anilkseth @De_dicto

2022-11-22 23:44:54 @leonpalafox @ylecun @yudapearl that was all true before the change of management

2022-11-22 22:52:47 “We failed to deliver our vision of AI-assisted programming because… tech is not ready yet," Smith explained. "You can see this in GitHub Copilot, which is built by GitHub in collaboration w OpenAI. As of late 2022, Copilot shows a lot of promise but still has a long way to go." https://t.co/uYnxlIw8z5

2022-11-22 21:48:41 At first I misread this as being about Twitter. My mistake. https://t.co/R6uaLOcLRu

2022-11-22 21:23:27 RT @Tweetermeyer: So, after Elon claimed that SBF "set off his bullshit meter," and implied he'd corrupted the justice system through polit…

2022-11-22 21:20:47 @voxbec @___merc___ Key word was: *reliably*

2022-11-22 20:44:56 I already publicly predicted that #GPT-4 will continue to hallucinate heavily

2022-11-22 20:38:08 RT @billbrown: Great essay! It is remarkable how little progress voice assistants have made over time. The ST:TNG computer seems as distan…

2022-11-22 20:34:46 RT @EthanVPorter: What were the effects of fact-checks/misinfo during 2020 election? New paper, with @aecoppock, Kim Gross, @emilythorson &

2022-11-22 20:30:35 @DG_Rand @AnnCC12 I got into a huge row with @ylecun @metaAI about whether large language models have generated misinfo in the wild. My view is that it would be hard to detect but easy to generate &

2022-11-22 19:47:25 @wayneholmes @an_open_mind @dmonett @Abebab @ylecun @emilymbender @MelMitchell1 Can you elaborate? Haven’t read those details carefully yet. Strong blitz play seems reasonable as a metric, no?

2022-11-22 19:31:04 @an_open_mind @dmonett @Abebab @wayneholmes @ylecun @emilymbender @MelMitchell1 my hot take

2022-11-22 19:28:54 @nealkhosla and NB I totally endorsed your idea! (just think LLMs per se are inadequate for implementing it)

2022-11-22 19:28:16 and, yes, we should be concerned about deception in machines.

2022-11-22 19:25:19 Very impressive Diplomacy model from @polynoamial &

2022-11-22 19:13:19 @nealkhosla of course not. (but what's your argument against the premise?)

2022-11-22 18:54:19 GPT-3 itself is utterly unqualified to do this, since it lacks interpretability and the ability to reason and dissect arguments with reliability

2022-11-22 18:50:30 RT @efarmer1: @GaryMarcus Gary you have summed it up in multiple nutshells. I wholeheartedly agree with your assessments. Now over to your…

2022-11-22 04:29:55 Interesting that none of these (including in the replies) get the QWERTY layout even though so many pictures of that exist. https://t.co/v1wMtG2O6c

2022-11-22 03:55:02 @Tkaraletsos @eerac Irene Pepperberg has written quite bit about how clever they are.

2022-11-22 03:00:45 RT @tribelaw: Richard Fierro saved so many lives. He deserves all our gratitude. Thanksgiving! https://t.co/00VYhHdDwd

2022-11-22 02:34:59 @MichaelTrazzi has there been any significant progress in the (mis)alignment of goals since she started research the topic? how serious a problem is it now?

2022-11-22 02:30:26 Elon Musk got paid about 10% of the market cap of Tesla to pay attention to … Tesla (which has fallen by roughly half since he decided to buy Twitter). &

2022-11-22 01:25:28 I Vancouver https://t.co/ifnquaT7Fw

2022-11-22 00:49:33 Did AI fail to meet expectations again? Personally, I would pay real money for a virtual assistant of the caliber of the OS in the film Her, but not for an AI with the narrowness of Alexa or the hallucinations of Galactica. https://t.co/G68u1XKWG8

2022-11-22 00:08:54 Via @danielmalmer on LinkedIn on some hate-filled ads that were purchased (but not published)

2022-11-22 00:04:45 @riedelcastro @soumithchintala @moreisdifferent and would welcome other pointers to relevant work!

2022-11-22 00:04:09 @riedelcastro @soumithchintala @moreisdifferent best evidence that this is near-term plausible? i’ve enjoyed some of your work (eg on Wiki verifiability), but wonder how it would manage around the kinds of things I have seen Galactica produce/that I posted here and on my Substack over last few days.

2022-11-21 23:08:15 @NaveenGRao I also wrote this earlier in same topic: https://t.co/DAKe19V1Dy

2022-11-21 22:55:25 @NaveenGRao “We’re thinking about it but haven’t figured out how to do it” might be an honest answer -- but also at odds with the absolutely certainty that there has been zero harm that Yann professed earlier.

2022-11-21 21:25:39 @ylecun I thought subtweeting without tagging was rude? Anyway, my reply, it’s not like I missed yours: https://t.co/QhudBxILW5

2022-11-21 21:23:31 I am that dude. You have answered at least a dozen of my recent queries. But not that one

2022-11-21 21:17:22 @pmddomingos 100% agree and yet already >

2022-11-21 21:04:42 @pmddomingos true! It’s only anti-wokes that can turn Twitter into a wild success like Parler or Truth Social.

2022-11-21 20:54:45 @Pehdrew_ Super hard. Which is why you can’t blandly assert that there has been no harm. And why meta ought put real work into the problem and report what they find.

2022-11-21 20:51:49 @lorakolodny @mer__edith

2022-11-21 20:42:26 RT @Grady_Booch: @GaryMarcus When <

2022-11-21 20:39:52 @prem_k @Grady_Booch @abeba @Michael_J_Black @MadamePratolung https://t.co/KlQJhILKgf

2022-11-21 14:11:40 https://t.co/kiviFxW0RM

2022-11-21 14:07:53 Galactica, summed up perfectly @futurism https://t.co/cRRQVQhZno

2022-11-21 13:44:42 RT @KDHungerford: #Artificialintelligence has mostly been focusing on a technique called #deeplearning. It might be time to reconsider. - @…

2022-11-21 04:18:23 epic trolls from @rsalakhu (2 of 2, and so subtle the target isn’t even named). 100% factual, too. @soumithchintala’s response, which I initially missed, is indeed great and very professional. (He’s not the one being trolled.) https://t.co/ar5MIW5nFr

2022-11-21 04:17:18 @soumithchintala @moreisdifferent Appreciate your sober thread. I do think though that there is an inherent unsolved challenge though inasmuch as no current tech can really eliminate the widespread hallucinations.

2022-11-21 04:08:02 epic trolls from @rsalakhu (1 of 2) https://t.co/cMuq49z0Bo

2022-11-21 04:05:29 @pmddomingos @YejinChoinka in the limit of unlimited data, they are same, as I said. The difference is in the learning. But the learning isn’t real compositionality

2022-11-21 02:57:56 @pmddomingos @YejinChoinka And see also https://t.co/UWUg6kQ3Df by @EvelinaLeivada @ElliotMurphy91 and myself in different but related architecture

2022-11-21 02:55:39 @pmddomingos In the limit they are the same (universal function approximators). They are better at learning syntax but they still don’t have a compositional semantics. Which is why eg they fail at @YejinChoinka’s recent https://t.co/wsic38AdLR

2022-11-21 02:42:05 @pmddomingos Representing grammar is cheap - lots of architecture can do that in the limit. learning grammar from data is same problem it ever was. And you are still inaccurate in what you said about compositionality.

2022-11-21 02:22:41 @pmddomingos not quite right. They have a proxy that looks vaguely like compositionality. They don’t have meanings lawfully composed from parts, governed by syntax.

2022-11-21 02:18:55 amazing how many people missed that this was a joke https://t.co/myv5yd2xEU

2022-11-21 00:04:24 RT @MuhammedKambal: @GaryMarcus On a related note, this morning I watched your debate with Chnosky on debunking the big lie of AI. Quite ex…

2022-11-20 23:34:04 @Ted_Underwood @CriticalAI @true_mxp @JrKibs @_akpiper https://t.co/wsic38AdLR

2022-11-20 23:18:10 @ylecun @untitled01ipynb @rsalakhu @rsalakhu trolls @ylecun and I get blamed. (Does this count as a countercountertroll?)

2022-11-20 19:55:33 @ylecun @CriticalAI @TonyZador @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian Just answer this one question, Yann: https://t.co/BctdIM1PSW

2022-11-20 19:49:08 @PeterShor1 @ylecun @Grady_Booch @Jeff_Aronson @Abebab Were there any usability studies with actual users? You’d want to evaluate accuracy, and not just speed.

2022-11-20 18:10:34 Tweeted this day before Galactica came out https://t.co/p0cb9QUOW1

2022-11-20 18:09:27 @Ted_Underwood @true_mxp @JrKibs @_akpiper people can (and are) approaching it, but they can’t do it. with training and diet, i can “approach” running a 2 min mile, but it ain’t gonna happen with my body.

2022-11-20 18:07:18 @ykilcher has anyone done *any*? how? I just asked Yann re FB

2022-11-20 18:06:03 @Ted_Underwood @true_mxp @JrKibs @_akpiper the abject failure of Galactica to do so is btw an object lesson here. so are the woeful results on TruthfulQA across the whole field, and much harder tasks might be devised.

2022-11-20 18:03:37 @Ted_Underwood @true_mxp @JrKibs @_akpiper no, current software cannot do that remotely reliable

2022-11-20 03:29:28 @AI4Code @SumitGulwani

2022-11-20 01:02:03 October 28: No reinstatements without a council Nov 19: Never mind about the council. I asked my fans, mostly right of center. Trump’s back in. https://t.co/b3xrNznGcO

2022-11-20 00:45:06 @ylecun @drng @Grady_Booch @Jeff_Aronson @Abebab @mrgreene1977 Galactica has deeper problems: - it is unconstrained by the facts are in its database (eg it said that Elon Musk died in 2018 car crash but probably has data to the contrary) - it gives no indication when it makes something up that isn’t database

2022-11-20 00:33:06 @ylecun @drng @Grady_Booch @Jeff_Aronson @Abebab I think what he means that the model made false inferences eg about value in eating crushed glass (per @mrgreene1977) https://t.co/eOlWcTDTxL

2022-11-20 00:31:09 @mark_riedl @Sergei_Imaging @ylecun @mrgreene1977 no doubt that cat is out of the bag, and it’s easy enough to replicate. soon enough we find out empirically how much impact systems like these have. (though of course hard to trace since the quality of the mimicry is high and there are no watermarks)

2022-11-20 00:29:43 @srijankedia @ylecun @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian @emilio__ferrara @davidlazer @gianluca_string @CrashTheMod3 It’s quite worrisome that @metaAI’s Chief AI officer shows little apparent familiarity with such work.

2022-11-20 00:27:54 @tdietterich @ylecun @mrgreene1977 and they are in fact starting to be used, and as cost of doing them well decreases, we can expect them to be more widespread. some examples from 2022: https://t.co/4kxqUjHXXT 2024 election, who knows?

2022-11-20 00:22:17 #Galactica is fine, because if it makes stuff up, no problemo. @metaAI’s Chief AI Officer just told me so. And he’s *sure* nobody would ever misuse his tool. Yeah, right. https://t.co/9uk9YE56ad

2022-11-20 00:20:04 @noUpside @ylecun @mrgreene1977 @katestarbird @sinanaral @ProfNoahGian Ask Yann, who thinks it is actually hunky-dory, but my theory was this: https://t.co/tfjaWTQuHW

2022-11-20 00:10:39 @ylecun @srijankedia @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian Low probabilities ≠ zero probabilities. Low probabilities with low cost ≠ zero probabilities with zero cost. Every spammer, ever, knows this. Same principles will apply to misinfo

2022-11-20 00:06:40 Twitter on IOS just crashed midtweet and I lost my work. First time that happened in a decade of using it. also having some load time issues. And so it begins.

2022-11-20 00:02:31 @ylecun @jppesky @mrgreene1977 on latter point, better said: https://t.co/SLZ7lVHQ4f

2022-11-20 00:00:29 @ylecun @jppesky @mrgreene1977 You are coming across as obtuse. - You act as if misinfo has had no impact in the real world (despite success of anti-vax campaigns) - Act as if the only way in which scientific-sounding nonsense has impact is if it is in a peer review journal - Treat low prob as zero probability

2022-11-19 23:52:54 @TonyZador @ylecun @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian perhaps someone in 1985 said “email isn’t going to have any effect on scamming. it’s driven by how much people want to be taken.” the biggest application of LLMs is now? it’s apparently making nonsense for SEO optimization. New ways to do in volume, because cost is cheaper.

2022-11-19 23:14:23 @ylecun @mrgreene1977 Come on, it’s not just about journals. To take one example, vaccine misinformation, oft presented as “scientific” [eg https://t.co/B1JmJihJpl], likely had significant impact on vaccination rates. Tagging @noUpside @katestarbird @sinanaral @ProfNoahGian for more perspectives.

2022-11-19 22:13:07 RT @GaryMarcus: @ylecun @mrgreene1977 How could Galactica *not* make misinfo easier, @ylecun, when it makes writing faux scientific-like a…

2022-11-19 22:11:50 @ylecun @mrgreene1977 How could Galactica *not* make misinfo easier, @ylecun, when it makes writing faux scientific-like articles trivial rather than laborious? &

2022-11-19 19:43:40 @ylecun @Grady_Booch @Jeff_Aronson @Abebab The sand is in thinking that a system like this can stick to reality. It cannot.

2022-11-19 19:13:54 @ylecun @Abebab Ps repeating for convenience: https://t.co/BBE3moQFea

2022-11-18 21:21:17 @ylecun @mrgreene1977 Potentially deadly TikTok challenges

2022-11-18 21:15:06 RT @michaelharriot: They won’t know what breaks until it breaks. It’s a matter of what breaks. “It’s not like McDonalds’ bro. You can’t ta…

2022-11-18 21:06:16 https://t.co/7aDMXp6G3g

2022-11-18 19:21:26 RT @GaryMarcus: @elonmusk What about misinformation?

2022-11-18 19:20:56 I Like Free Speech So Much I’ve Decided to Buy It - McSweeney’s https://t.co/Yw3n94HAQW

2022-11-18 19:18:59 @elonmusk What about misinformation?

2022-11-18 19:06:43 Galactica

2022-11-18 15:50:32 @yoavgo @ylecun @rasbt @manes @Abebab Troll farms cost money and presumably limited by budget. The more they can put out the more effective they are.

2022-11-18 15:36:47 @yoavgo @rasbt @manes @ylecun @Abebab On the scenario, secondary customers like journalists will have no longer have any trust in arXiv at all.

2022-11-18 15:34:22 @ylecun @rasbt @manes @Abebab It will reduce the cost hence increase the volume, hence make the problem worse.

2022-11-18 15:12:49 RT @GaryMarcus: @ylecun @Abebab Galactica *is* dangerous. And you have no idea how to eliminate that danger. large language models inherent…

2022-11-18 15:12:34 @ylecun @Abebab Galactica *is* dangerous. And you have no idea how to eliminate that danger. large language models inherently hallucinate. It weaponized the ability of these tools to spread fake science. The vitriol is proportionate to the risk. I am shocked that you are indifferent to this.

2022-11-18 15:06:40 “once perfected” is carrying a lot of weight here. there is no way Galactica will be perfected in next year or two, probably a lot longer. happy to publicly debate you about this, but recommend you start by rereading your own recent critique @noemamag of large language models. https://t.co/LYXD0E0jyI

2022-11-18 13:51:45 very rational fear https://t.co/bY0hUrnzcv

2022-11-18 13:41:25 follow friday may not be a thing much longer over here on Twitter but follow @abebab. And she’s on Mastodon now, too. https://t.co/29GfLksNFx

2022-11-18 13:38:38 RT @Abebab: @ylecun Pretty much exactly what happened was you overhyped and released a model. People tested it (you should be thankful for…

2022-11-18 13:15:06 Just saying, in case this ship goes down, @garymarcus@sigmoid.social

2022-11-18 12:55:18 @HeidyKhlaaf @AmandaAskell @andrewthesmart @Abebab 100% support thinking through the potential social consequences. There is a lot of room for important argument there.

2022-11-18 12:48:16 @HeidyKhlaaf @AmandaAskell @andrewthesmart @Abebab Thank you. and to be clear I am all for spelling out the latter part and will amplify the message if is is constructed carefully and respectfully. (If you want to write something longer, DM.)

2022-11-18 12:24:50 @HeidyKhlaaf @AmandaAskell @andrewthesmart Yes, there are lot of ideas within EA, some well-intentioned, some not. some briefly considered, some taken seriously, some not

2022-11-17 23:52:52 Tripling down, @yLeCun's contention is that people "abused" Galactica. But what safeguards were put in place to prevent bad actors from using the system to generate compelling disinformation around scientific topics like vaccines or racial differences? A: None. https://t.co/HlVjdHz1Sr

2022-11-17 23:39:33 @mmitchell_ai @trashcanmagic @timnitGebru But that is not me. I fully agree with Dr King's methods, and with direct action

2022-11-17 23:18:17 @andrewthesmart So the left is now pro-choice but anti-genetic screening? This is eugenics? https://t.co/gu4vD7v0en

2022-11-17 22:10:22 @emilymbender @timnitGebru and if what you say is true, you can simply provide a link. there certainly has not been anything from either of you in direct replies to me when i have asked.

2022-11-17 22:01:59 Why is this is any different at all than McCarthyism? I honestly don't see a difference between labeling someone a communist and labeling them a eugenicist, if you do so without either evidence or definition. I have asked, repeatedly, for both, and gotten neither. https://t.co/efvhRYCgEw

2022-11-17 21:54:48 @emilymbender @timnitGebru Nonsense. I have repeatedly asked a legitimate,straightforward question. Neither of you have answered: are you claiming eg that MacAskill (whom I have never met) is a eugenicist a la 19th century scientific racism? yes or no, and if yes, in what sense, and what is the argument?

2022-11-17 21:52:13 @athundt @timnitGebru The argument against Singer there is much more straightforward

2022-11-17 21:50:36 @athundt @timnitGebru I don't doubt that scientific racism still exists. But that label should not be applied to specific people without real evidence. That's all I am saying. &

2022-11-17 21:38:20 I have no shame about speaking out against McCarthyism.

2022-11-17 21:32:21 @timnitGebru no definitions still? i will carry on.

2022-11-17 21:26:23 @trashcanmagic @timnitGebru How dare you? My father was doing sit-ins for fair housing before you were born, and did discrimination law for years. I helped him (with statistics) as a teenager, and went to protests from when I was a newborn. I am with your struggle, but not your tactics.

2022-11-17 21:24:28 @timnitGebru We need definitions, @timnitGebru, if you want me and others take you seriously. I was once a huge fan of yours, but am really losing the faith.

2022-11-17 21:16:44 @timnitGebru Only quote I have seen was nearly fabricated, inserting words that Bostrom didn't use. I haven't seen any direct evidence on MacAskill. &

2022-11-17 21:07:19 @timnitGebru If I had been alive in the 1950s I hope that I would have called out McCarthy (white male) for doing what you do: naming names, without due process, without definitions, without evidence. It's not what groups you belong to Timnit, it's what you are saying that is problematic.

2022-11-17 20:41:16 RT @GaryMarcus: The reality is that large language models like GPT-3 annd Galactica are like bulls in a china shop, powerful but reckless.…

2022-11-17 19:56:24 @ylecun @Grady_Booch And they use lots of symbol-manipulation too. Funny that you never mention that.

2022-11-17 19:52:00 RT @GidMK: @GaryMarcus I'm worried about fake science. There are already hundreds of pay-to-play journals that publish openly plagiarised w…

2022-11-17 19:42:43 I doubt it. but Prince said it best, “Forever [is] a mighty long time” https://t.co/VXE5Up59tl

2022-11-17 19:32:50 RT @Plinz: Generative AI is incredibly useful and valuable, but it will force us to change our practices in science, journalism and social…

2022-11-17 19:31:30 @yLeCun: It’s all good, man https://t.co/PrX2PBawlc

2022-11-17 03:58:52 RT @frossi_t: On my way to the AAAI fall symposium on “thinking fast and slow and other cognitive theories in AI”. The program includes 12…

2022-11-17 01:37:42 RT @tytusm: 100% agree with Gary The threat posed to science by #Galactica is real Try it yourself using the gold standard…

2022-11-16 23:11:10 Flattery will get you nowhere, young battlestar. (image courtesy @UlrichJunker) https://t.co/d8GYGDCgTm

2022-11-16 22:59:03 @shokunin_studio @Grady_Booch @fchollet i replied to @tansuyegen's tweet

2022-11-16 22:58:06 @TansuYegen I think you have misunderstood wha the tech purports to do. I think that it is morphing actual human to look like a a different human. still cool, but not what you describe. see https://t.co/Lk5wipCh2c Face Shifting tab

2022-11-16 21:30:15 RT @karaswisher: Read these fine few words from ⁦@GaryMarcus⁩ https://t.co/Giqr1JI0by

2022-11-16 18:48:44 RT @ruthaylett: He is right. People laughing at this are mistaken, it is a serious problem in the making.

2022-11-16 18:40:34 RT @RemivanTrijp: I sincerely hope this doesn’t turn out to be prophetic

2022-11-16 17:56:14 @Grady_Booch @elonmusk I wasn’t assigning probabilities here

2022-11-16 17:52:44 @Grady_Booch @elonmusk Only two ways this can go - Musk can succeed, and rewrite everything we thought we knew about management and software engineering - He can fail, and become the textbook example of how not to do things.

2022-11-16 17:49:47 There’s a lot going on. Midterms, FTX, Twitter chaos. But in time we will see the release of #Galactica as an epochal event, a tipping point in a gigantic increase in the flow of misinformation. This is no joke. Galactica is funny, but the uses it will be put to are not. https://t.co/UPoe1q4BEi

2022-11-16 17:22:36 @MetaAI’s Galactica has made up my birthday (wrong decade), my education,my research interests, etc, all sounding plausible, 85% untrue. You know what’s really embarrassing about this? A decent AI could just look this stuff on the web &

2022-11-16 15:04:18 @ylecun @MetaAI @paperswithcode https://t.co/EvMKYUZcN0

2022-11-16 14:54:38 RT @pelimuc:

2022-11-16 14:41:15 never had this happened before - timeout loading my own Tweets. Things gonna get wild here in Bird-land! https://t.co/URBAdZpBhX

2022-11-16 14:38:05 For examples, see https://t.co/GUdfR6nR6y and this bonus fake bio of @ylecun here: https://t.co/gN1gvFkSjh

2022-11-16 14:38:03 AI Joke of the Day! Q: What do a CNC machine and MetaAI’s #Galactica have in common? .. .. .. A: They both help you fabricate! https://t.co/uEouaaO8iQ

2022-11-16 14:11:51 RT @dmonett: "Is this really what AI has come to, automatically mixing reality with bullshit so finely we can no longer recognize the diffe…

2022-11-16 14:11:23 RT @dougclow: Scientific bullshit generator, argh! Having spent many disheartening hours reviewing academic manuscripts that strongly resem…

2022-11-16 06:56:19 @kevinroose you don’t think the product placement for Storybook Brawl was a nice touch? https://t.co/15jySgxqR9 @TrungTPhan

2022-11-16 05:49:14 @mcuban just liked my new post

2022-11-16 05:35:37 @_jasonwei @colinraffel i need to think about @colinraffel’s specific tasks but i am willing to bet money that LLMs will continue to face serious challenges in factuality even as they scale.

2022-11-16 00:08:58 @ykilcher which benchmarks have you published in the last four weeks? I have posted two, one in a hot new ML journal, linked above, which you didn’t comment, the other on arXiv. You? the third was published in AI Magazine 2015. It’s not *my* fault that the field doesn’t rise to them.

2022-11-15 20:59:45 @ykilcher Not my job to decide what community disagrees about. But people ought look at the zero-shot version of Truthful QA and to test comprehension as I described in the 2014 New Yorker article Beyond The Turing Test, as well recent vision benchmark I helped w: https://t.co/LDPoiyTPNi

2022-11-15 19:02:22 OpenAI give me access to GPT-4? https://t.co/mN0OSkugcZ

2022-11-15 16:51:13 Rumor has it GPT-4 is coming. Here’s a prediction: it will still have serious issues with confabulation/misinformation, toxicity and bias. https://t.co/TKm6DA3Dvb

2022-11-15 16:24:59 . @sacca wasn’t wrong. https://t.co/2WVxcZAPMn

2022-11-15 15:23:55 "it was sobering to see how close we came to hitting a parked car after we rolled over a low curb separating the parking lot." — @cademetz on six hours in a Tesla:https://t.co/UUQXUBzE4q

2022-11-15 14:36:53 @_jasonwei @drjwrae @OwainEvans_UK post hoc.performance on original 0-shot task is not what you graphed. you are graphing only 10-shot results (and focusing on the multiple choice results which are the easier form of the task).On zero shot MC, 280B performance is <

2022-11-15 06:42:33 @_jasonwei @drjwrae @OwainEvans_UK If they used a different method you can’t tell how much is method change and how much is scale

2022-11-14 19:26:55 RT @QasimRashid: This is horrifying. The Iranian parliament votes overwhelmingly (227-63) to execute the 15,000 protesters they already arr…

2022-11-14 19:24:35 @_jasonwei @OwainEvans_UK your drawing doesn't match the leaderboard here: https://t.co/7HGyLImHHs and i think it's because Gopher changed the task...

2022-11-14 19:08:28 RT @svpino: In 2017, a team led by Andrew Ng published a paper showing off a Deep Learning model to detect pneumonia.Andrew is one of the…

2022-11-14 18:45:29 @_jasonwei Taking TruthfulQA as an example, your evidence for scaling seems (if I understand correctly) to come from a single data point from Gopher that was (a) selected post hoc, (b) has not been replicated and (c) relied on making the task multiple choice. cc @OwainEvans_UKAm I wrong?

2022-11-14 14:06:15 RT @federicobianchy: Text-to-image generation models (like Stable Diffusion and DALLE) are being used to generate millions of images a day.…

2022-11-14 13:10:22 RT @drewharwell: New: That fake verified Eli Lilly account sparked a panic inside the pharma giant, and senior execs decided they'd had eno…

2022-11-14 01:58:42 @karaswisher Will certainly be interesting to see what happens to advertising revenue in the Lockheed Martini era…

2022-11-29 05:02:39 @BobbyAlter @ylecun @mrgreene1977 @rhinigtas @ykilcher But absent any data, please excuse us for being skeptical

2022-11-29 04:01:45 @srikumarks @ylecun @mrgreene1977 @rhinigtas @ykilcher @JeanDreze @AltNews Different harms and different processes but yes, just illustrative. It would be great to have real data here.

2022-11-29 03:56:15 RT @GlobalSight: At #SightTechGlobal2022, @neddesmond will chat with Dave Ferrucci &

2022-11-29 03:48:08 @ylecun @srikumarks @mrgreene1977 @rhinigtas @ykilcher @JeanDreze @AltNews @theliverdr Again do you have *data*? Also: * You said above "almost all of it is harmless“

2022-11-29 03:06:44 @ylecun @BobbyAlter @mrgreene1977 @rhinigtas @ykilcher surely not 100% You have any data on efficacy? if even .01% of LLM generated misinfo makes it through, we may be in trouble Cc @ProfNoahGian @noUpside @davidlazer

2022-11-29 01:06:39 @ylecun @mrgreene1977 @rhinigtas @ykilcher Folks like @davidlazer, @ethanporter, and @ProfNoahGian have done both correlational and experimental work that suggests otherwise. As an example, from Lazer: https://t.co/u0ORdBeL53

2022-11-28 20:06:52 RT @CNNPR: @elonmusk This headline never appeared on CNN. Be better. https://t.co/n8xGp2Z6qm

2022-11-28 19:56:18 RT @ykilcher: Ad hominems, classy

2022-11-28 19:28:17 RT @KHCourage: AI is really good at spinning up misinformation (including "scientific" papers about the benefits of eating crushed glass).…

2022-11-28 15:14:19 RT @CaulfieldTim: "#Misinformation is soon going to come at a pace never seen before." "Anybody who is not worried, should be." AI Is Te…

2022-11-28 14:46:44 RT @CriticalAI: This is happening tomorrow! We invite you to join us with @GaryMarcus for a webinar on “foundations” for AI, 11/29 at 5pm E…

2022-11-28 13:43:31 RT @DorotheaBaur: An #LLM is like "a turbocharged version of autocomplete". It only considers "the states of the world on which it is train…

2022-11-27 23:24:52 @Grady_Booch @_ashawndabney me, too

2022-11-27 23:00:10 RT @olliewaters: Great breakdown of new AI that can win at Diplomacy by @GaryMarcus. I'd say an agent that can't handle a slight variation…

2022-11-27 22:54:31 @ProfNoahGian as discussed here: https://t.co/4PQpe24ISL cc @ylecun

2022-11-27 22:53:49 RT @ProfNoahGian: No, Yann, AI is part of the solution AND part of the problem here.

2022-11-27 22:27:43 @realGeorgeHotz @elonmusk @NautilusMag the trick is do that at scale

2022-11-27 22:25:52 RT @casparhenderson: "AI is unlikely to be perfect at challenge of misinformation, but years of trying with largely human content moderatio…

2022-11-27 20:21:25 @elonmusk Right now, AI is terrible at detecting misinformation. It doesn’t have to be. my look at your predicament @elonmusk and how AI might help, @nautilusmag https://t.co/4PQpe24ISL

2022-11-27 20:18:56 @realGeorgeHotz FYI: Right now, AI is terrible at detecting misinformation. It doesn’t have to be. my look at @elonmusk's predicament and how AI might help, @nautilusmag https://t.co/4PQpe24ISL

2022-11-27 20:09:34 AI Is Terrible at Detecting Misinformation. It Doesn’t Have to Be. My look at @elonmusk's predicament and how AI might help https://t.co/4PQpe24ISL

2022-11-27 19:55:48 RT @psowinski_: Great analysis of Meta's Cicero AI! It's a good example of how neurosymbolics &

2022-11-27 19:02:17 Excellent thread by @drjimfan summarizing all 15 Outstanding Papers at NeurIPS (including his own, which I tweeted a few days ago). https://t.co/Bpc0iTA8WS

2022-11-27 18:55:56 later today @NautilusMag, a memo to @elonmusk

2022-11-27 18:20:15 @davidmanheim @polynoamial Fair enough, butI was raising AZ as a thought experiment/point of comparison. not sure how much Noam can speak freely about how Cicero evolved or will evolve, but I find the modular structure that they used to be interesting, esp since DM’s related work was well known to them

2022-11-27 16:52:55 RT @BallouxFrancois: The Covid situation in China is not looking good right now. The authorities have trapped themselves into a situation…

2022-11-27 16:44:40 Sorta sucks to feel intellectually obliged to retweet all @polynoamial’s smart takes on AlphaZero and Cicero, because now I think nobody’s gonna take my bet

2022-11-27 16:42:55 RT @polynoamial: @ykilcher @GaryMarcus A big challenge of Diplomacy (and the real world) is dealing with the fact that not all humans behav…

2022-11-27 16:42:46 RT @polynoamial: @ykilcher @GaryMarcus It's worth mentioning that we *did* train a bot in the style of AZ for no-press Diplomacy (a non-lan…

2022-11-27 16:42:29 RT @polynoamial: @ykilcher @GaryMarcus AZ can't play Diplomacy. Aside from the imperfect info (which AZ can't handle), using human data is…

2022-11-27 16:42:15 @ykilcher @polynoamial not sure who you are asking

2022-11-27 16:35:46 @ykilcher @polynoamial (just to be clear, re the bet I was referring to the language version, though the no-press version is an interesting question in its own right)

2022-11-27 05:28:21 @natanielruizg @MatthewJBar @MetaAI @ErnestSDavis @NYUScience @polynoamial Knock yourself out :)

2022-11-27 05:17:03 @asyncmind Cicero is open source so it would be easy for @openai to try if they wanted to

2022-11-27 05:09:20 Anyone want to be against me on this? #gpt4 vs Cicero @ blitz Diplomacy. https://t.co/N3hWwtrVZ4

2022-11-27 03:47:28 RT @Fayeexplore: I highly recommend subscribing to Gary  Marcus' column, which focuses on the latest research and analysis in AI, thank him…

2022-11-27 00:58:01 RT @AvilaGarcez: "It is worth noting that some aspects of Cicero use a neurosymbolic approach to AI"

2022-11-26 22:07:26 RT @NathanpmYoung: Should @elonmusk give $ 1billion directly to the poorest people in the world via @GiveDirectly vox populi, vox dei

2022-11-26 16:47:42 @bsgallagher @elonmusk @WholeMarsBlog sounds like city driving remains problematic, given reports like these

2022-11-26 16:45:28 @bsgallagher @elonmusk @WholeMarsBlog Unless they have solved the outlier problem, I seriously doubt it.

2022-11-26 16:29:38 what's Noam Chomsky think about GPT-3? why doesn't Alexa use GPT-3? what does Meta's new Diplomacy AI bot means for AI? what can linguistics teach AI? what's the debate over “sentience” in AI really about? https://t.co/8ir1xKenr6 has got you covered. subscribe for free

2022-11-26 16:12:04 Way too soon. RIP, Irene Cara. https://t.co/upCVb75lKq

2022-11-26 15:26:46 RT @JianqiangM: “If Cicero is any guide, machine learning may ultimately prove to be even more valuable if it is embedded in highly structu…

2022-11-26 15:20:26 RT @Kuzeko: «Cicero’s overall architecture Is not something that simply emerged spontaneously from the basic data, but it is rather an exqu…

2022-11-26 15:17:45 @anderssandberg @dw2 True, but perhaps more around emotion than hallucination

2022-11-26 03:50:55 Seems like, I dunno, maybe, Twitter’s trending algorithms could use some small tweaks? https://t.co/SRR6O30Eoi

2022-11-26 03:43:56 @Noahpinion @lenlayton @0xfbifemboy @ylecun 1. there is no standard measure 2. closest thing is TruthfulQA on which LLMs are very poor 3. Galactica &

2022-11-26 02:38:51 RT @polynoamial: Branching factor of Chess: ~30 Branching factor of Go: ~300 Branching factor of Diplomacy: ~10^20 But that's even before…

2022-11-26 02:21:44 @FelixHill84 @ylecun @Grady_Booch @Meta I genuinely think that LeCun needs perspective here, and that he is embarrassing himself, as a great number of people in many fields have pointed out

2022-11-26 02:14:24 @pmddomingos one mechanism among several, eg see also https://t.co/wd7SDrII6c

2022-11-26 01:39:23 At this rate, you won’t have to pay the $8 to have an ad-free Twitter https://t.co/PTUnbdhL3P

2022-11-26 01:03:14 https://t.co/iGrkbGxI0T

2022-11-25 23:42:19 @ylecun @BobbyAlter https://t.co/vnADWFV31r

2022-11-25 23:32:24 RT @Jim_Salmons: 2of2 To my #DigitalHumanities &

2022-11-25 22:40:47 I feel like I am living in a real-life version of Don’t Look Up. https://t.co/iwf86WGFW0

2022-11-25 21:52:14 @BobbyAlter @ylecun thanks

2022-11-25 21:50:55 Example of medical misinfo in the wild, for purposes of ad sales hoax, that looks LLM-generated to me, sample text. [ 3/3] cc @noUpside @ylecun @ProfNoahGian @EthanVPorter @katestarbird @sinanaral @justinhendrix @DG_Rand @lorakolodny @chafkin https://t.co/ydTxFClrzI

2022-11-25 21:50:54 Using fake medical info to sell ad clicks: Disturbing, plausible, and perhaps inevitable. [ 1/3] #gpt3r #misinformation #possiblesightinginthewild https://t.co/yl2c63P3Kf

2022-11-25 21:09:38 I too would love to hear from the integrity teams @meta @twitter @youtube @tiktok_us etc. Not clear that Twitter still has one, but would love to hear from former employees. My DM’s are open, and happy to continue conversations elsewhere. https://t.co/KzVuig1dkP

2022-11-25 21:05:15 @ylecun @Grady_Booch @Meta You, my former friend, are burning your reputation to the ground. Everyone is telling you to lie down and go home. Listen to what they are saying. Not for me

2022-11-25 20:51:10 RT @dmonett: "We have known for some time that machine learning is valuable

2022-11-25 20:50:52 @Pehdrew_ No, but by default it gives too little consideration to ethics and consequences. what’s happened lately is next level.

2022-11-25 20:45:51 @ylecun @Grady_Booch @Meta Easy there, Yann. Maybe you might need a few days off? @Grady_Booch is an IBM Fellow, and one of the world’s leading software architects, eminent in his field, fully capable of reaching his own conclusions.

2022-11-25 19:56:42 Good science, poor ethical leadership. https://t.co/ByWLWisowq

2022-11-25 19:48:43 This is EXACTLY why you need to answer my misinfo question with candor. By ducking it, you make it impossible for anyone outside of @meta to properly consider those cost-benefit tradeoffs. and most people therefore naturally infer that you are hiding something damaging. https://t.co/bqoD97OWcY

2022-11-25 19:47:46 @AlexCEngler @noUpside @erikbryn @ylecun There are multiple potential serious harms that need serious attention

2022-11-25 18:30:31 RT @MMikeMMa: Been waiting for the @GaryMarcus take, here w/ Dr. Ernest Davis! “Cicero is in many ways a marvel.. achieved.. the deepest…

2022-11-25 18:16:23 @MMikeMMa (it’s coauthored, by @ErnestSDavis and me, to just to be clear, so with not via)

2022-11-25 18:15:37 @AlexCEngler @jmourabarbosa Consider that Birdwatch / Community Notes @twitter relies on heavy human annotation

2022-11-25 02:49:41 @leecronin i have written several essays about this at https://t.co/8ir1xKenr6, if you want to find out more.

2022-11-25 00:21:53 My first post on Post, https://t.co/ENuPLSpsQc

2022-11-24 21:18:51 It is sad. This receipt-laden thread by @dmonett yesterday on same topic is startling. https://t.co/GNUkGW4PhJ https://t.co/WCVbRc7y8E

2022-11-24 21:14:19 RT @plibin: @GaryMarcus I think dissemination, influence, and production are all important. Especially when influence is so heavily determi…

2022-11-24 21:10:00 ps one key passage is here

2022-11-24 21:08:25 This is what you call a “failure of imagination” argument: I, personally, sitting here in my comfy chair, can’t figure out how the bad guys would ever [steal the jewels/deliberately crash a stolen jet/disseminate misinfo] so it could never happen. https://t.co/yfXwUJJpks

2022-11-24 21:07:54 @vayuvegula @ylecun or maybe c. doing a really long comedy bit and not breaking character? Colbert would be proud!

2022-11-24 21:02:58 @MMikeMMa @plibin @sgourley that’s my whole point to @ylecun, which he insists cannot be true.

2022-11-24 20:47:31 The amount of automatically generated misinformation in wild (in this mild case, to drive ad sales) is clearly not zero. No way a human wrote the linked page in its entirety. Bet there are many more like it, and as cost drops to zero, amount goes up. Your move, @ylecun. https://t.co/bzb3TBdInT

2022-11-24 20:36:19 RT @plibin: I have such mixed feelings about this unedited AI playground output from GPT-3. On the one hand: holy shit! On the other hand:…

2022-11-24 20:35:30 @dlowd url? we have all the best training data, the best

2022-11-24 19:51:55 @dromanocpm @cajundiscordian it was an amazing scene in the show

2022-11-24 19:39:12 The Wire fans, you remember the scene where Bunk McNulty finally get to investigate Stringer Bell's apartment? "Who in the F... was I chasing"? @cajundiscordian ain't the guy I thought I was writing about. Can't shake that scene. https://t.co/cmpbVkafCa

2022-11-24 19:33:03 @cajundiscordian @Inframethod @ch402 that's interesting

2022-11-24 19:31:47 @Willyintheworld @cajundiscordian @Inframethod @ch402 sure, although eg people under anesthesia may or may not be sentient while under anesthesia. it'd be really nice to have an external measurement that we were fully confident in.

2022-11-24 19:30:29 @cajundiscordian @Inframethod @ch402 i allow that i don't have a full understanding of what is inside nor of the behavior. what do you think is most salient about its differences with GPT-3 as a baseline?

2022-11-24 19:28:26 @cajundiscordian @Inframethod @ch402 if you had no autobiographical memory, short term or long, what is the *you* that you'd be experiencing? (&

2022-11-24 19:22:21 @cajundiscordian @Inframethod @ch402 all i know is the paper and your own examples

2022-11-24 19:21:39 @cajundiscordian @Inframethod @ch402 when you say something, and you are a bot, what you say is your history, &

2022-11-23 00:31:29 RT @plibin: Truth is the Achilles' heel of current large language model AIs

2022-11-22 23:48:47 Really shocking how many platform owners and self-professed engineers can’t distinguish between vicious and viscous. Navier-Stokes for the win! https://t.co/fczz8lD4bl

2022-11-22 23:46:00 @MicahHoffmann @Grady_Booch I defer to @davidchalmers42 @anilkseth @De_dicto

2022-11-22 23:44:54 @leonpalafox @ylecun @yudapearl that was all true before the change of management

2022-11-22 22:52:47 “We failed to deliver our vision of AI-assisted programming because… tech is not ready yet," Smith explained. "You can see this in GitHub Copilot, which is built by GitHub in collaboration w OpenAI. As of late 2022, Copilot shows a lot of promise but still has a long way to go." https://t.co/uYnxlIw8z5

2022-11-22 21:48:41 At first I misread this as being about Twitter. My mistake. https://t.co/R6uaLOcLRu

2022-11-22 21:23:27 RT @Tweetermeyer: So, after Elon claimed that SBF "set off his bullshit meter," and implied he'd corrupted the justice system through polit…

2022-11-22 21:20:47 @voxbec @___merc___ Key word was: *reliably*

2022-11-22 20:44:56 I already publicly predicted that #GPT-4 will continue to hallucinate heavily

2022-11-22 20:38:08 RT @billbrown: Great essay! It is remarkable how little progress voice assistants have made over time. The ST:TNG computer seems as distan…

2022-11-22 20:34:46 RT @EthanVPorter: What were the effects of fact-checks/misinfo during 2020 election? New paper, with @aecoppock, Kim Gross, @emilythorson &

2022-11-22 20:30:35 @DG_Rand @AnnCC12 I got into a huge row with @ylecun @metaAI about whether large language models have generated misinfo in the wild. My view is that it would be hard to detect but easy to generate &

2022-11-22 19:47:25 @wayneholmes @an_open_mind @dmonett @Abebab @ylecun @emilymbender @MelMitchell1 Can you elaborate? Haven’t read those details carefully yet. Strong blitz play seems reasonable as a metric, no?

2022-11-22 19:31:04 @an_open_mind @dmonett @Abebab @wayneholmes @ylecun @emilymbender @MelMitchell1 my hot take

2022-11-22 19:28:54 @nealkhosla and NB I totally endorsed your idea! (just think LLMs per se are inadequate for implementing it)

2022-11-22 19:28:16 and, yes, we should be concerned about deception in machines.

2022-11-22 19:25:19 Very impressive Diplomacy model from @polynoamial &

2022-11-22 19:13:19 @nealkhosla of course not. (but what's your argument against the premise?)

2022-11-22 18:54:19 GPT-3 itself is utterly unqualified to do this, since it lacks interpretability and the ability to reason and dissect arguments with reliability

2022-11-22 18:50:30 RT @efarmer1: @GaryMarcus Gary you have summed it up in multiple nutshells. I wholeheartedly agree with your assessments. Now over to your…

2022-11-22 04:29:55 Interesting that none of these (including in the replies) get the QWERTY layout even though so many pictures of that exist. https://t.co/v1wMtG2O6c

2022-11-22 03:55:02 @Tkaraletsos @eerac Irene Pepperberg has written quite bit about how clever they are.

2022-11-22 03:00:45 RT @tribelaw: Richard Fierro saved so many lives. He deserves all our gratitude. Thanksgiving! https://t.co/00VYhHdDwd

2022-11-22 02:34:59 @MichaelTrazzi has there been any significant progress in the (mis)alignment of goals since she started research the topic? how serious a problem is it now?

2022-11-22 02:30:26 Elon Musk got paid about 10% of the market cap of Tesla to pay attention to … Tesla (which has fallen by roughly half since he decided to buy Twitter). &

2022-11-22 01:25:28 I Vancouver https://t.co/ifnquaT7Fw

2022-11-22 00:49:33 Did AI fail to meet expectations again? Personally, I would pay real money for a virtual assistant of the caliber of the OS in the film Her, but not for an AI with the narrowness of Alexa or the hallucinations of Galactica. https://t.co/G68u1XKWG8

2022-11-22 00:08:54 Via @danielmalmer on LinkedIn on some hate-filled ads that were purchased (but not published)

2022-11-22 00:04:45 @riedelcastro @soumithchintala @moreisdifferent and would welcome other pointers to relevant work!

2022-11-22 00:04:09 @riedelcastro @soumithchintala @moreisdifferent best evidence that this is near-term plausible? i’ve enjoyed some of your work (eg on Wiki verifiability), but wonder how it would manage around the kinds of things I have seen Galactica produce/that I posted here and on my Substack over last few days.

2022-11-21 23:08:15 @NaveenGRao I also wrote this earlier in same topic: https://t.co/DAKe19V1Dy

2022-11-21 22:55:25 @NaveenGRao “We’re thinking about it but haven’t figured out how to do it” might be an honest answer -- but also at odds with the absolutely certainty that there has been zero harm that Yann professed earlier.

2022-11-21 21:25:39 @ylecun I thought subtweeting without tagging was rude? Anyway, my reply, it’s not like I missed yours: https://t.co/QhudBxILW5

2022-11-21 21:23:31 I am that dude. You have answered at least a dozen of my recent queries. But not that one

2022-11-21 21:17:22 @pmddomingos 100% agree and yet already >

2022-11-21 21:04:42 @pmddomingos true! It’s only anti-wokes that can turn Twitter into a wild success like Parler or Truth Social.

2022-11-21 20:54:45 @Pehdrew_ Super hard. Which is why you can’t blandly assert that there has been no harm. And why meta ought put real work into the problem and report what they find.

2022-11-21 20:51:49 @lorakolodny @mer__edith

2022-11-21 20:42:26 RT @Grady_Booch: @GaryMarcus When <

2022-11-21 20:39:52 @prem_k @Grady_Booch @abeba @Michael_J_Black @MadamePratolung https://t.co/KlQJhILKgf

2022-11-21 14:11:40 https://t.co/kiviFxW0RM

2022-11-21 14:07:53 Galactica, summed up perfectly @futurism https://t.co/cRRQVQhZno

2022-11-21 13:44:42 RT @KDHungerford: #Artificialintelligence has mostly been focusing on a technique called #deeplearning. It might be time to reconsider. - @…

2022-11-21 04:18:23 epic trolls from @rsalakhu (2 of 2, and so subtle the target isn’t even named). 100% factual, too. @soumithchintala’s response, which I initially missed, is indeed great and very professional. (He’s not the one being trolled.) https://t.co/ar5MIW5nFr

2022-11-21 04:17:18 @soumithchintala @moreisdifferent Appreciate your sober thread. I do think though that there is an inherent unsolved challenge though inasmuch as no current tech can really eliminate the widespread hallucinations.

2022-11-21 04:08:02 epic trolls from @rsalakhu (1 of 2) https://t.co/cMuq49z0Bo

2022-11-21 04:05:29 @pmddomingos @YejinChoinka in the limit of unlimited data, they are same, as I said. The difference is in the learning. But the learning isn’t real compositionality

2022-11-21 02:57:56 @pmddomingos @YejinChoinka And see also https://t.co/UWUg6kQ3Df by @EvelinaLeivada @ElliotMurphy91 and myself in different but related architecture

2022-11-21 02:55:39 @pmddomingos In the limit they are the same (universal function approximators). They are better at learning syntax but they still don’t have a compositional semantics. Which is why eg they fail at @YejinChoinka’s recent https://t.co/wsic38AdLR

2022-11-21 02:42:05 @pmddomingos Representing grammar is cheap - lots of architecture can do that in the limit. learning grammar from data is same problem it ever was. And you are still inaccurate in what you said about compositionality.

2022-11-21 02:22:41 @pmddomingos not quite right. They have a proxy that looks vaguely like compositionality. They don’t have meanings lawfully composed from parts, governed by syntax.

2022-11-21 02:18:55 amazing how many people missed that this was a joke https://t.co/myv5yd2xEU

2022-11-21 00:04:24 RT @MuhammedKambal: @GaryMarcus On a related note, this morning I watched your debate with Chnosky on debunking the big lie of AI. Quite ex…

2022-11-20 23:34:04 @Ted_Underwood @CriticalAI @true_mxp @JrKibs @_akpiper https://t.co/wsic38AdLR

2022-11-20 23:18:10 @ylecun @untitled01ipynb @rsalakhu @rsalakhu trolls @ylecun and I get blamed. (Does this count as a countercountertroll?)

2022-11-20 19:55:33 @ylecun @CriticalAI @TonyZador @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian Just answer this one question, Yann: https://t.co/BctdIM1PSW

2022-11-20 19:49:08 @PeterShor1 @ylecun @Grady_Booch @Jeff_Aronson @Abebab Were there any usability studies with actual users? You’d want to evaluate accuracy, and not just speed.

2022-11-20 18:10:34 Tweeted this day before Galactica came out https://t.co/p0cb9QUOW1

2022-11-20 18:09:27 @Ted_Underwood @true_mxp @JrKibs @_akpiper people can (and are) approaching it, but they can’t do it. with training and diet, i can “approach” running a 2 min mile, but it ain’t gonna happen with my body.

2022-11-20 18:07:18 @ykilcher has anyone done *any*? how? I just asked Yann re FB

2022-11-20 18:06:03 @Ted_Underwood @true_mxp @JrKibs @_akpiper the abject failure of Galactica to do so is btw an object lesson here. so are the woeful results on TruthfulQA across the whole field, and much harder tasks might be devised.

2022-11-20 18:03:37 @Ted_Underwood @true_mxp @JrKibs @_akpiper no, current software cannot do that remotely reliable

2022-11-20 03:29:28 @AI4Code @SumitGulwani

2022-11-20 01:02:03 October 28: No reinstatements without a council Nov 19: Never mind about the council. I asked my fans, mostly right of center. Trump’s back in. https://t.co/b3xrNznGcO

2022-11-20 00:45:06 @ylecun @drng @Grady_Booch @Jeff_Aronson @Abebab @mrgreene1977 Galactica has deeper problems: - it is unconstrained by the facts are in its database (eg it said that Elon Musk died in 2018 car crash but probably has data to the contrary) - it gives no indication when it makes something up that isn’t database

2022-11-20 00:33:06 @ylecun @drng @Grady_Booch @Jeff_Aronson @Abebab I think what he means that the model made false inferences eg about value in eating crushed glass (per @mrgreene1977) https://t.co/eOlWcTDTxL

2022-11-20 00:31:09 @mark_riedl @Sergei_Imaging @ylecun @mrgreene1977 no doubt that cat is out of the bag, and it’s easy enough to replicate. soon enough we find out empirically how much impact systems like these have. (though of course hard to trace since the quality of the mimicry is high and there are no watermarks)

2022-11-20 00:29:43 @srijankedia @ylecun @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian @emilio__ferrara @davidlazer @gianluca_string @CrashTheMod3 It’s quite worrisome that @metaAI’s Chief AI officer shows little apparent familiarity with such work.

2022-11-20 00:27:54 @tdietterich @ylecun @mrgreene1977 and they are in fact starting to be used, and as cost of doing them well decreases, we can expect them to be more widespread. some examples from 2022: https://t.co/4kxqUjHXXT 2024 election, who knows?

2022-11-20 00:22:17 #Galactica is fine, because if it makes stuff up, no problemo. @metaAI’s Chief AI Officer just told me so. And he’s *sure* nobody would ever misuse his tool. Yeah, right. https://t.co/9uk9YE56ad

2022-11-20 00:20:04 @noUpside @ylecun @mrgreene1977 @katestarbird @sinanaral @ProfNoahGian Ask Yann, who thinks it is actually hunky-dory, but my theory was this: https://t.co/tfjaWTQuHW

2022-11-20 00:10:39 @ylecun @srijankedia @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian Low probabilities ≠ zero probabilities. Low probabilities with low cost ≠ zero probabilities with zero cost. Every spammer, ever, knows this. Same principles will apply to misinfo

2022-11-20 00:06:40 Twitter on IOS just crashed midtweet and I lost my work. First time that happened in a decade of using it. also having some load time issues. And so it begins.

2022-11-20 00:02:31 @ylecun @jppesky @mrgreene1977 on latter point, better said: https://t.co/SLZ7lVHQ4f

2022-11-20 00:00:29 @ylecun @jppesky @mrgreene1977 You are coming across as obtuse. - You act as if misinfo has had no impact in the real world (despite success of anti-vax campaigns) - Act as if the only way in which scientific-sounding nonsense has impact is if it is in a peer review journal - Treat low prob as zero probability

2022-11-19 23:52:54 @TonyZador @ylecun @mrgreene1977 @noUpside @katestarbird @sinanaral @ProfNoahGian perhaps someone in 1985 said “email isn’t going to have any effect on scamming. it’s driven by how much people want to be taken.” the biggest application of LLMs is now? it’s apparently making nonsense for SEO optimization. New ways to do in volume, because cost is cheaper.

2022-11-19 23:14:23 @ylecun @mrgreene1977 Come on, it’s not just about journals. To take one example, vaccine misinformation, oft presented as “scientific” [eg https://t.co/B1JmJihJpl], likely had significant impact on vaccination rates. Tagging @noUpside @katestarbird @sinanaral @ProfNoahGian for more perspectives.

2022-11-19 22:13:07 RT @GaryMarcus: @ylecun @mrgreene1977 How could Galactica *not* make misinfo easier, @ylecun, when it makes writing faux scientific-like a…

2022-11-19 22:11:50 @ylecun @mrgreene1977 How could Galactica *not* make misinfo easier, @ylecun, when it makes writing faux scientific-like articles trivial rather than laborious? &

2022-11-19 19:43:40 @ylecun @Grady_Booch @Jeff_Aronson @Abebab The sand is in thinking that a system like this can stick to reality. It cannot.

2022-11-19 19:13:54 @ylecun @Abebab Ps repeating for convenience: https://t.co/BBE3moQFea

2022-11-18 21:21:17 @ylecun @mrgreene1977 Potentially deadly TikTok challenges

2022-11-18 21:15:06 RT @michaelharriot: They won’t know what breaks until it breaks. It’s a matter of what breaks. “It’s not like McDonalds’ bro. You can’t ta…

2022-11-18 21:06:16 https://t.co/7aDMXp6G3g

2022-11-18 19:21:26 RT @GaryMarcus: @elonmusk What about misinformation?

2022-11-18 19:20:56 I Like Free Speech So Much I’ve Decided to Buy It - McSweeney’s https://t.co/Yw3n94HAQW

2022-11-18 19:18:59 @elonmusk What about misinformation?

2022-11-18 19:06:43 Galactica

2022-11-18 15:50:32 @yoavgo @ylecun @rasbt @manes @Abebab Troll farms cost money and presumably limited by budget. The more they can put out the more effective they are.

2022-11-18 15:36:47 @yoavgo @rasbt @manes @ylecun @Abebab On the scenario, secondary customers like journalists will have no longer have any trust in arXiv at all.

2022-11-18 15:34:22 @ylecun @rasbt @manes @Abebab It will reduce the cost hence increase the volume, hence make the problem worse.

2022-11-18 15:12:49 RT @GaryMarcus: @ylecun @Abebab Galactica *is* dangerous. And you have no idea how to eliminate that danger. large language models inherent…

2022-11-18 15:12:34 @ylecun @Abebab Galactica *is* dangerous. And you have no idea how to eliminate that danger. large language models inherently hallucinate. It weaponized the ability of these tools to spread fake science. The vitriol is proportionate to the risk. I am shocked that you are indifferent to this.

2022-11-18 15:06:40 “once perfected” is carrying a lot of weight here. there is no way Galactica will be perfected in next year or two, probably a lot longer. happy to publicly debate you about this, but recommend you start by rereading your own recent critique @noemamag of large language models. https://t.co/LYXD0E0jyI

2022-11-18 13:51:45 very rational fear https://t.co/bY0hUrnzcv

2022-11-18 13:41:25 follow friday may not be a thing much longer over here on Twitter but follow @abebab. And she’s on Mastodon now, too. https://t.co/29GfLksNFx

2022-11-18 13:38:38 RT @Abebab: @ylecun Pretty much exactly what happened was you overhyped and released a model. People tested it (you should be thankful for…

2022-11-18 13:15:06 Just saying, in case this ship goes down, @garymarcus@sigmoid.social

2022-11-18 12:55:18 @HeidyKhlaaf @AmandaAskell @andrewthesmart @Abebab 100% support thinking through the potential social consequences. There is a lot of room for important argument there.

2022-11-18 12:48:16 @HeidyKhlaaf @AmandaAskell @andrewthesmart @Abebab Thank you. and to be clear I am all for spelling out the latter part and will amplify the message if is is constructed carefully and respectfully. (If you want to write something longer, DM.)

2022-11-18 12:24:50 @HeidyKhlaaf @AmandaAskell @andrewthesmart Yes, there are lot of ideas within EA, some well-intentioned, some not. some briefly considered, some taken seriously, some not

2022-11-17 23:52:52 Tripling down, @yLeCun's contention is that people "abused" Galactica. But what safeguards were put in place to prevent bad actors from using the system to generate compelling disinformation around scientific topics like vaccines or racial differences? A: None. https://t.co/HlVjdHz1Sr

2022-11-17 23:39:33 @mmitchell_ai @trashcanmagic @timnitGebru But that is not me. I fully agree with Dr King's methods, and with direct action

2022-11-17 23:18:17 @andrewthesmart So the left is now pro-choice but anti-genetic screening? This is eugenics? https://t.co/gu4vD7v0en

2022-11-17 22:10:22 @emilymbender @timnitGebru and if what you say is true, you can simply provide a link. there certainly has not been anything from either of you in direct replies to me when i have asked.

2022-11-17 22:01:59 Why is this is any different at all than McCarthyism? I honestly don't see a difference between labeling someone a communist and labeling them a eugenicist, if you do so without either evidence or definition. I have asked, repeatedly, for both, and gotten neither. https://t.co/efvhRYCgEw

2022-11-17 21:54:48 @emilymbender @timnitGebru Nonsense. I have repeatedly asked a legitimate,straightforward question. Neither of you have answered: are you claiming eg that MacAskill (whom I have never met) is a eugenicist a la 19th century scientific racism? yes or no, and if yes, in what sense, and what is the argument?

2022-11-17 21:52:13 @athundt @timnitGebru The argument against Singer there is much more straightforward

2022-11-17 21:50:36 @athundt @timnitGebru I don't doubt that scientific racism still exists. But that label should not be applied to specific people without real evidence. That's all I am saying. &

2022-11-17 21:38:20 I have no shame about speaking out against McCarthyism.

2022-11-17 21:32:21 @timnitGebru no definitions still? i will carry on.

2022-11-17 21:26:23 @trashcanmagic @timnitGebru How dare you? My father was doing sit-ins for fair housing before you were born, and did discrimination law for years. I helped him (with statistics) as a teenager, and went to protests from when I was a newborn. I am with your struggle, but not your tactics.

2022-11-17 21:24:28 @timnitGebru We need definitions, @timnitGebru, if you want me and others take you seriously. I was once a huge fan of yours, but am really losing the faith.

2022-11-17 21:16:44 @timnitGebru Only quote I have seen was nearly fabricated, inserting words that Bostrom didn't use. I haven't seen any direct evidence on MacAskill. &

2022-11-17 21:07:19 @timnitGebru If I had been alive in the 1950s I hope that I would have called out McCarthy (white male) for doing what you do: naming names, without due process, without definitions, without evidence. It's not what groups you belong to Timnit, it's what you are saying that is problematic.

2022-11-17 20:41:16 RT @GaryMarcus: The reality is that large language models like GPT-3 annd Galactica are like bulls in a china shop, powerful but reckless.…

2022-11-17 19:56:24 @ylecun @Grady_Booch And they use lots of symbol-manipulation too. Funny that you never mention that.

2022-11-17 19:52:00 RT @GidMK: @GaryMarcus I'm worried about fake science. There are already hundreds of pay-to-play journals that publish openly plagiarised w…

2022-11-17 19:42:43 I doubt it. but Prince said it best, “Forever [is] a mighty long time” https://t.co/VXE5Up59tl

2022-11-17 19:32:50 RT @Plinz: Generative AI is incredibly useful and valuable, but it will force us to change our practices in science, journalism and social…

2022-11-17 19:31:30 @yLeCun: It’s all good, man https://t.co/PrX2PBawlc

2022-11-17 03:58:52 RT @frossi_t: On my way to the AAAI fall symposium on “thinking fast and slow and other cognitive theories in AI”. The program includes 12…

2022-11-17 01:37:42 RT @tytusm: 100% agree with Gary The threat posed to science by #Galactica is real Try it yourself using the gold standard…

2022-11-16 23:11:10 Flattery will get you nowhere, young battlestar. (image courtesy @UlrichJunker) https://t.co/d8GYGDCgTm

2022-11-16 22:59:03 @shokunin_studio @Grady_Booch @fchollet i replied to @tansuyegen's tweet

2022-11-16 22:58:06 @TansuYegen I think you have misunderstood wha the tech purports to do. I think that it is morphing actual human to look like a a different human. still cool, but not what you describe. see https://t.co/Lk5wipCh2c Face Shifting tab

2022-11-16 21:30:15 RT @karaswisher: Read these fine few words from ⁦@GaryMarcus⁩ https://t.co/Giqr1JI0by

2022-11-16 18:48:44 RT @ruthaylett: He is right. People laughing at this are mistaken, it is a serious problem in the making.

2022-11-16 18:40:34 RT @RemivanTrijp: I sincerely hope this doesn’t turn out to be prophetic

2022-11-16 17:56:14 @Grady_Booch @elonmusk I wasn’t assigning probabilities here

2022-11-16 17:52:44 @Grady_Booch @elonmusk Only two ways this can go - Musk can succeed, and rewrite everything we thought we knew about management and software engineering - He can fail, and become the textbook example of how not to do things.

2022-11-16 17:49:47 There’s a lot going on. Midterms, FTX, Twitter chaos. But in time we will see the release of #Galactica as an epochal event, a tipping point in a gigantic increase in the flow of misinformation. This is no joke. Galactica is funny, but the uses it will be put to are not. https://t.co/UPoe1q4BEi

2022-11-16 17:22:36 @MetaAI’s Galactica has made up my birthday (wrong decade), my education,my research interests, etc, all sounding plausible, 85% untrue. You know what’s really embarrassing about this? A decent AI could just look this stuff on the web &

2022-11-16 15:04:18 @ylecun @MetaAI @paperswithcode https://t.co/EvMKYUZcN0

2022-11-16 14:54:38 RT @pelimuc:

2022-11-16 14:41:15 never had this happened before - timeout loading my own Tweets. Things gonna get wild here in Bird-land! https://t.co/URBAdZpBhX

2022-11-16 14:38:05 For examples, see https://t.co/GUdfR6nR6y and this bonus fake bio of @ylecun here: https://t.co/gN1gvFkSjh

2022-11-16 14:38:03 AI Joke of the Day! Q: What do a CNC machine and MetaAI’s #Galactica have in common? .. .. .. A: They both help you fabricate! https://t.co/uEouaaO8iQ

2022-11-16 14:11:51 RT @dmonett: "Is this really what AI has come to, automatically mixing reality with bullshit so finely we can no longer recognize the diffe…

2022-11-16 14:11:23 RT @dougclow: Scientific bullshit generator, argh! Having spent many disheartening hours reviewing academic manuscripts that strongly resem…

2022-11-16 06:56:19 @kevinroose you don’t think the product placement for Storybook Brawl was a nice touch? https://t.co/15jySgxqR9 @TrungTPhan

2022-11-16 05:49:14 @mcuban just liked my new post

2022-11-16 05:35:37 @_jasonwei @colinraffel i need to think about @colinraffel’s specific tasks but i am willing to bet money that LLMs will continue to face serious challenges in factuality even as they scale.

2022-11-16 00:08:58 @ykilcher which benchmarks have you published in the last four weeks? I have posted two, one in a hot new ML journal, linked above, which you didn’t comment, the other on arXiv. You? the third was published in AI Magazine 2015. It’s not *my* fault that the field doesn’t rise to them.

2022-11-15 20:59:45 @ykilcher Not my job to decide what community disagrees about. But people ought look at the zero-shot version of Truthful QA and to test comprehension as I described in the 2014 New Yorker article Beyond The Turing Test, as well recent vision benchmark I helped w: https://t.co/LDPoiyTPNi

2022-11-15 19:02:22 OpenAI give me access to GPT-4? https://t.co/mN0OSkugcZ

2022-11-15 16:51:13 Rumor has it GPT-4 is coming. Here’s a prediction: it will still have serious issues with confabulation/misinformation, toxicity and bias. https://t.co/TKm6DA3Dvb

2022-11-15 16:24:59 . @sacca wasn’t wrong. https://t.co/2WVxcZAPMn

2022-11-15 15:23:55 "it was sobering to see how close we came to hitting a parked car after we rolled over a low curb separating the parking lot." — @cademetz on six hours in a Tesla:https://t.co/UUQXUBzE4q

2022-11-15 14:36:53 @_jasonwei @drjwrae @OwainEvans_UK post hoc.performance on original 0-shot task is not what you graphed. you are graphing only 10-shot results (and focusing on the multiple choice results which are the easier form of the task).On zero shot MC, 280B performance is <

2022-11-15 06:42:33 @_jasonwei @drjwrae @OwainEvans_UK If they used a different method you can’t tell how much is method change and how much is scale

2022-11-14 19:26:55 RT @QasimRashid: This is horrifying. The Iranian parliament votes overwhelmingly (227-63) to execute the 15,000 protesters they already arr…

2022-11-14 19:24:35 @_jasonwei @OwainEvans_UK your drawing doesn't match the leaderboard here: https://t.co/7HGyLImHHs and i think it's because Gopher changed the task...

2022-11-14 19:08:28 RT @svpino: In 2017, a team led by Andrew Ng published a paper showing off a Deep Learning model to detect pneumonia.Andrew is one of the…

2022-11-14 18:45:29 @_jasonwei Taking TruthfulQA as an example, your evidence for scaling seems (if I understand correctly) to come from a single data point from Gopher that was (a) selected post hoc, (b) has not been replicated and (c) relied on making the task multiple choice. cc @OwainEvans_UKAm I wrong?

2022-11-14 14:06:15 RT @federicobianchy: Text-to-image generation models (like Stable Diffusion and DALLE) are being used to generate millions of images a day.…

2022-11-14 13:10:22 RT @drewharwell: New: That fake verified Eli Lilly account sparked a panic inside the pharma giant, and senior execs decided they'd had eno…

2022-11-14 01:58:42 @karaswisher Will certainly be interesting to see what happens to advertising revenue in the Lockheed Martini era…

2022-12-07 21:23:10 RT @ibogost: I talked to ChatGPT about poetry, behavioral psychology, computer programming, hamburgers, and more. It's not a tool that will…

2022-12-07 20:50:20 @Michael_J_Black you might be interested in examples like these (I noticed you followed back on post but not here btw)

2022-12-07 20:34:22 More utterly bogus references from the alleged Future of Search https://t.co/y6ahzULYqO

2022-12-07 20:23:18 @RadioFreeTom He’s 94! And bemused by the fake GPTChat news I passed along, reporting his alleged 2020 death.

2022-12-07 20:04:33 @jjarangoes @ylecun @percyliang actually, @ylecun and agree on a lot: I started out 2917 debate with multiple points of agreement &

2022-12-07 19:40:19 @Agz5885 @ahatamiz1 @Amine57240266 stochasticity + authoritative presentation is not a winning formula

2022-12-07 19:39:08 @erikbryn Destroying trust in the information economy, on the Russian “firehouse of falsehood” model. Stack Overflow took the first blows. there will be more.

2022-12-07 19:34:43 @RadioFreeTom If it’s today, you also share it with Noam Chomsky

2022-12-07 19:32:34 @EricRWeinstein @elonmusk @ESYudkowsky I offered Elon some money about this topic, along with a definition (below) vetted by two of the people who invented the term. offer still stands: https://t.co/yfXycYXxb9 @wadhwa and others over to pentuple the bet to $500k.

2022-12-07 19:28:31 @ChristophMolnar I gave the clearest example

2022-12-07 19:25:38 Yes! But they include discussion of laws and explicit values in that education. They don’t leave it all to implicit reinforcement learning. See my pinned tweet, @ylecun cc: @percyliang https://t.co/5EykwGQGGA

2022-12-07 18:48:46 @MelMitchell1 I think the Turing Test is a terrible test, but two hours gives a good amount of time for depth. Cc @mkapor What Comes After the Turing Test? https://t.co/1N4qeEn6uH

2022-12-07 18:42:13 @ChristophMolnar Look forward to lots of novices puzzling over stuff like this https://t.co/7MsqFd2knj

2022-12-07 18:40:56 @ahatamiz1 @Amine57240266 @Agz5885 have you been able to replicate? Can you elaborate?

2022-12-07 18:09:27 RT @innovaTutor: A *must listen* interview with @GaryMarcus on the @BigTechnology Podcast https://t.co/pO0qQPMZ08 Is AI dangerously overh…

2022-12-07 18:09:05 @srikumarks The question of course not whether these systems sometimes create correct answers, but how often, and whether users (or downstream processors) can tell the difference when they instead generate wrong answers

2022-12-07 17:48:11 RT @percyliang: RL from human feedback seems to be the main tool for alignment. Given reward hacking and the falliability of humans, this s…

2022-12-07 16:42:21 @Ricardo_Joseh_L He took it well. (But shares my concern about the imminent deluge)

2022-12-07 16:37:37 Happy 94th birthday, Noam Chomsky! And don’t believe everything chatGPT tells you! (Spoiler alert: he doesn’t) https://t.co/VcdCVopBwx

2022-12-07 14:26:48 Absolutely. But it makes them up. Eg Noam Chomsky is very much alive (we emailed a few days ago), but check out the fake yet plausible looking sources below (&

2022-12-08 00:48:15 @Plinz @pwlot @seanmcbride @lexfridman Ok but I was on his show, he should have had thicker skin than that

2022-12-08 00:23:59 Ha ha Lex blocked me (after our episode!) because I didn’t agree that AI was going to relocate to Austin. Cc @Grady_Booch https://t.co/A5MaEtxDGH

2022-12-07 23:47:10 @seanmcbride @Plinz What takes tremendous intellectual effort is humility—and recognizing where improvements are needed.

2022-12-07 23:03:42 RT @emilkeyder: this is going to be a limiting factor for this kind of approach for a long time. there is no understanding or knowledge or…

2022-12-07 23:02:17 if you don’t understand this point, you don’t understand how ChatGPT works. and if you don’t understand this point, you are very likely overestimating GPT’s capabilities. https://t.co/0byiYi1orb

2022-12-07 22:54:50 No! even when there is structured data, chatGPT is limited in its ability to leverage that information. example: Wiki says Noam Chomsky is alive (which I can confirm)

2022-12-07 22:42:18 @John68Richmond The facts are clear: https://t.co/RIsKhp3yzX

2022-12-07 22:40:21 @kenneth0stanley @ylecun @hayabhay @ZoubinGhahrama1 @jeffclune @Tkaraletsos I hope @ylecun that @kenneth0stanley’s response will put an end to this line of discussion.

2022-12-07 22:38:08 @ylecun @hayabhay @ZDNET @TiernanRayTech So no counterargument then to the fact that - you publicly lied about my credentials - are resting a further attack on my credentials by conflating whether I helped develop an algorithm with whether I was permitted by its buyer to publish it. Classy!

2022-12-07 22:27:27 @ylecun @hayabhay @ZDNET @TiernanRayTech You know perfectly well that that work is under NDA and wasn’t published. That wasn’t my choice (since I no longer own it) but I have told you that before. You are just trying to extend your efforts to deceive.

2022-12-07 22:18:59 @fchollet Counterpoint: robotics (esp domestic humanoid robots) is one of the domains that most needs major advances in physical psychological and causal reasoning, and is one of the least tolerant of the current approximative approach. So could actually be the catalyst.

2022-12-07 22:15:19 @johnsontoddr4 @TonyD993 @yudapearl lovely example of how little gpt understands causality

2022-12-07 22:13:20 @ylecun @hayabhay Ps you also recently lied about my credentials to @zdnet, forcing @TiernanRayTech to publish a correction. And you said crazy things at end of 2019 about history that whole field knew was not true. Please stop attacking my credentials with false &

2022-12-07 22:06:23 @ylecun @hayabhay Uber bought that algorithm &

2022-12-07 21:23:10 RT @ibogost: I talked to ChatGPT about poetry, behavioral psychology, computer programming, hamburgers, and more. It's not a tool that will…

2022-12-07 20:50:20 @Michael_J_Black you might be interested in examples like these (I noticed you followed back on post but not here btw)

2022-12-07 20:34:22 More utterly bogus references from the alleged Future of Search https://t.co/y6ahzULYqO

2022-12-07 20:23:18 @RadioFreeTom He’s 94! And bemused by the fake GPTChat news I passed along, reporting his alleged 2020 death.

2022-12-07 20:04:33 @jjarangoes @ylecun @percyliang actually, @ylecun and agree on a lot: I started out 2917 debate with multiple points of agreement &

2022-12-07 19:40:19 @Agz5885 @ahatamiz1 @Amine57240266 stochasticity + authoritative presentation is not a winning formula

2022-12-07 19:39:08 @erikbryn Destroying trust in the information economy, on the Russian “firehouse of falsehood” model. Stack Overflow took the first blows. there will be more.

2022-12-07 19:34:43 @RadioFreeTom If it’s today, you also share it with Noam Chomsky

2022-12-07 19:32:34 @EricRWeinstein @elonmusk @ESYudkowsky I offered Elon some money about this topic, along with a definition (below) vetted by two of the people who invented the term. offer still stands: https://t.co/yfXycYXxb9 @wadhwa and others over to pentuple the bet to $500k.

2022-12-07 19:28:31 @ChristophMolnar I gave the clearest example

2022-12-07 19:25:38 Yes! But they include discussion of laws and explicit values in that education. They don’t leave it all to implicit reinforcement learning. See my pinned tweet, @ylecun cc: @percyliang https://t.co/5EykwGQGGA

2022-12-07 18:48:46 @MelMitchell1 I think the Turing Test is a terrible test, but two hours gives a good amount of time for depth. Cc @mkapor What Comes After the Turing Test? https://t.co/1N4qeEn6uH

2022-12-07 18:42:13 @ChristophMolnar Look forward to lots of novices puzzling over stuff like this https://t.co/7MsqFd2knj

2022-12-07 18:40:56 @ahatamiz1 @Amine57240266 @Agz5885 have you been able to replicate? Can you elaborate?

2022-12-07 18:09:27 RT @innovaTutor: A *must listen* interview with @GaryMarcus on the @BigTechnology Podcast https://t.co/pO0qQPMZ08 Is AI dangerously overh…

2022-12-07 18:09:05 @srikumarks The question of course not whether these systems sometimes create correct answers, but how often, and whether users (or downstream processors) can tell the difference when they instead generate wrong answers

2022-12-07 17:48:11 RT @percyliang: RL from human feedback seems to be the main tool for alignment. Given reward hacking and the falliability of humans, this s…

2022-12-07 16:42:21 @Ricardo_Joseh_L He took it well. (But shares my concern about the imminent deluge)

2022-12-07 16:37:37 Happy 94th birthday, Noam Chomsky! And don’t believe everything chatGPT tells you! (Spoiler alert: he doesn’t) https://t.co/VcdCVopBwx

2022-12-07 14:26:48 Absolutely. But it makes them up. Eg Noam Chomsky is very much alive (we emailed a few days ago), but check out the fake yet plausible looking sources below (&

2022-12-08 21:04:55 RT @GaryMarcus: @blattnerma @Plinz Yes &

2022-12-08 21:04:15 @tsushil_thapa @ShizaCharania @ylecun not friends with people who repeatedly lie about me

2022-12-08 21:03:50 absolutely - the edge cases in open-ended chatbots are endless

2022-12-08 21:03:00 @info_sprinkles @ylecun sorry but lying is not cool

2022-12-08 21:02:51 RT @MMikeMMa: @pgolding @plibin @GaryMarcus Isn’t a concern that these models are striving towards “general” capabilities across many use c…

2022-12-08 19:52:42 RT @GaryMarcus: @ylecun . @yLeCun the last time you said this to @zdnet, I presented multiple AI publications in top journals (sample scree…

2022-12-08 19:52:39 @ylecun . @yLeCun the last time you said this to @zdnet, I presented multiple AI publications in top journals (sample screenshots below). @zdnet corrected you, because you lied. @kenneth0stanley corrected you yesterday re our algorithm. *why* are you lying about my credentials again? https://t.co/uG07u8sYsi

2022-12-08 19:27:20 RT @GaryMarcus: @ShizaCharania @ylecun That’ll be $8 :)

2022-12-08 19:25:54 @ShizaCharania @ylecun That’ll be $8 :)

2022-12-08 19:18:02 @wadhwa @Shrikantha5 AFAIK all LLMs are vulnerable to hallucination. The thing you sent was made by a different recent OpenAI model (not ChatGPT) one from @Shrikantha5 is indeed ChatGPT. General problem is very real! Ps: LLMs are stochastic &

2022-12-08 18:37:38 @MuhammedKambal Sure would like to know what was in the training corpus

2022-12-08 18:31:46 Eerily prescient paper from 2021: language models &

2022-12-08 18:12:21 @pwlot @jeblad @StolfiAlberto I got blocked (I believe) for doubting his claim that Austin would become the AI capital of the world. And if you read the thread carefully someone else brought the wiki to my attention. Something didn’t feel right tbh when he interviewed me in the MIT library.

2022-12-08 18:09:38 @__delas__ @Plinz Love it when Plinz publicly attacks me and I respond and then the fanboys pile on without reading the time stamps or understanding the chronology.

2022-12-08 17:44:14 @pwlot @jeblad @StolfiAlberto he's representing himself as an MIT research scientist, not as a Drexel PhD.

2022-12-08 17:27:49 @Tweetermeyer wow. not a great look, especially given how he styles himself.

2022-12-08 17:26:03 @pwlot @jeblad @StolfiAlberto *did* he graduate? from the answers I am getting, he taught some January classes, but was not on faculty.

2022-12-08 17:21:33 RT @ChloeSavignac: Excellent line-up

2022-12-08 16:36:51 I suspect that this is inaccurate, but am curious if anyone can confirm or deny it. https://t.co/qlDi5cyH9c https://t.co/myflXl9IAr

2022-12-08 15:28:07 RT @wadhwa: And I did not found "Telesoft Partners". As well, Google clearly says that I authored 5 books, not 3 I am beginning to agree w…

2022-12-08 00:48:15 @Plinz @pwlot @seanmcbride @lexfridman Ok but I was on his show, he should have had thicker skin than that

2022-12-08 00:23:59 Ha ha Lex blocked me (after our episode!) because I didn’t agree that AI was going to relocate to Austin. Cc @Grady_Booch https://t.co/A5MaEtxDGH

2022-12-07 23:47:10 @seanmcbride @Plinz What takes tremendous intellectual effort is humility—and recognizing where improvements are needed.

2022-12-07 23:03:42 RT @emilkeyder: this is going to be a limiting factor for this kind of approach for a long time. there is no understanding or knowledge or…

2022-12-07 23:02:17 if you don’t understand this point, you don’t understand how ChatGPT works. and if you don’t understand this point, you are very likely overestimating GPT’s capabilities. https://t.co/0byiYi1orb

2022-12-07 22:54:50 No! even when there is structured data, chatGPT is limited in its ability to leverage that information. example: Wiki says Noam Chomsky is alive (which I can confirm)

2022-12-07 22:42:18 @John68Richmond The facts are clear: https://t.co/RIsKhp3yzX

2022-12-07 22:40:21 @kenneth0stanley @ylecun @hayabhay @ZoubinGhahrama1 @jeffclune @Tkaraletsos I hope @ylecun that @kenneth0stanley’s response will put an end to this line of discussion.

2022-12-07 22:38:08 @ylecun @hayabhay @ZDNET @TiernanRayTech So no counterargument then to the fact that - you publicly lied about my credentials - are resting a further attack on my credentials by conflating whether I helped develop an algorithm with whether I was permitted by its buyer to publish it. Classy!

2022-12-07 22:27:27 @ylecun @hayabhay @ZDNET @TiernanRayTech You know perfectly well that that work is under NDA and wasn’t published. That wasn’t my choice (since I no longer own it) but I have told you that before. You are just trying to extend your efforts to deceive.

2022-12-07 22:18:59 @fchollet Counterpoint: robotics (esp domestic humanoid robots) is one of the domains that most needs major advances in physical psychological and causal reasoning, and is one of the least tolerant of the current approximative approach. So could actually be the catalyst.

2022-12-07 22:15:19 @johnsontoddr4 @TonyD993 @yudapearl lovely example of how little gpt understands causality

2022-12-07 22:13:20 @ylecun @hayabhay Ps you also recently lied about my credentials to @zdnet, forcing @TiernanRayTech to publish a correction. And you said crazy things at end of 2019 about history that whole field knew was not true. Please stop attacking my credentials with false &

2022-12-07 22:06:23 @ylecun @hayabhay Uber bought that algorithm &

2022-12-07 21:23:10 RT @ibogost: I talked to ChatGPT about poetry, behavioral psychology, computer programming, hamburgers, and more. It's not a tool that will…

2022-12-07 20:50:20 @Michael_J_Black you might be interested in examples like these (I noticed you followed back on post but not here btw)

2022-12-07 20:34:22 More utterly bogus references from the alleged Future of Search https://t.co/y6ahzULYqO

2022-12-07 20:23:18 @RadioFreeTom He’s 94! And bemused by the fake GPTChat news I passed along, reporting his alleged 2020 death.

2022-12-07 20:04:33 @jjarangoes @ylecun @percyliang actually, @ylecun and agree on a lot: I started out 2917 debate with multiple points of agreement &

2022-12-07 19:40:19 @Agz5885 @ahatamiz1 @Amine57240266 stochasticity + authoritative presentation is not a winning formula

2022-12-07 19:39:08 @erikbryn Destroying trust in the information economy, on the Russian “firehouse of falsehood” model. Stack Overflow took the first blows. there will be more.

2022-12-07 19:34:43 @RadioFreeTom If it’s today, you also share it with Noam Chomsky

2022-12-07 19:32:34 @EricRWeinstein @elonmusk @ESYudkowsky I offered Elon some money about this topic, along with a definition (below) vetted by two of the people who invented the term. offer still stands: https://t.co/yfXycYXxb9 @wadhwa and others over to pentuple the bet to $500k.

2022-12-07 19:28:31 @ChristophMolnar I gave the clearest example

2022-12-07 19:25:38 Yes! But they include discussion of laws and explicit values in that education. They don’t leave it all to implicit reinforcement learning. See my pinned tweet, @ylecun cc: @percyliang https://t.co/5EykwGQGGA

2022-12-07 18:48:46 @MelMitchell1 I think the Turing Test is a terrible test, but two hours gives a good amount of time for depth. Cc @mkapor What Comes After the Turing Test? https://t.co/1N4qeEn6uH

2022-12-07 18:42:13 @ChristophMolnar Look forward to lots of novices puzzling over stuff like this https://t.co/7MsqFd2knj

2022-12-07 18:40:56 @ahatamiz1 @Amine57240266 @Agz5885 have you been able to replicate? Can you elaborate?

2022-12-07 18:09:27 RT @innovaTutor: A *must listen* interview with @GaryMarcus on the @BigTechnology Podcast https://t.co/pO0qQPMZ08 Is AI dangerously overh…

2022-12-07 18:09:05 @srikumarks The question of course not whether these systems sometimes create correct answers, but how often, and whether users (or downstream processors) can tell the difference when they instead generate wrong answers

2022-12-07 17:48:11 RT @percyliang: RL from human feedback seems to be the main tool for alignment. Given reward hacking and the falliability of humans, this s…

2022-12-07 16:42:21 @Ricardo_Joseh_L He took it well. (But shares my concern about the imminent deluge)

2022-12-07 16:37:37 Happy 94th birthday, Noam Chomsky! And don’t believe everything chatGPT tells you! (Spoiler alert: he doesn’t) https://t.co/VcdCVopBwx

2022-12-07 14:26:48 Absolutely. But it makes them up. Eg Noam Chomsky is very much alive (we emailed a few days ago), but check out the fake yet plausible looking sources below (&

2022-12-08 21:04:55 RT @GaryMarcus: @blattnerma @Plinz Yes &

2022-12-08 21:04:15 @tsushil_thapa @ShizaCharania @ylecun not friends with people who repeatedly lie about me

2022-12-08 21:03:50 absolutely - the edge cases in open-ended chatbots are endless

2022-12-08 21:03:00 @info_sprinkles @ylecun sorry but lying is not cool

2022-12-08 21:02:51 RT @MMikeMMa: @pgolding @plibin @GaryMarcus Isn’t a concern that these models are striving towards “general” capabilities across many use c…

2022-12-08 19:52:42 RT @GaryMarcus: @ylecun . @yLeCun the last time you said this to @zdnet, I presented multiple AI publications in top journals (sample scree…

2022-12-08 19:52:39 @ylecun . @yLeCun the last time you said this to @zdnet, I presented multiple AI publications in top journals (sample screenshots below). @zdnet corrected you, because you lied. @kenneth0stanley corrected you yesterday re our algorithm. *why* are you lying about my credentials again? https://t.co/uG07u8sYsi

2022-12-08 19:27:20 RT @GaryMarcus: @ShizaCharania @ylecun That’ll be $8 :)

2022-12-08 19:25:54 @ShizaCharania @ylecun That’ll be $8 :)

2022-12-08 19:18:02 @wadhwa @Shrikantha5 AFAIK all LLMs are vulnerable to hallucination. The thing you sent was made by a different recent OpenAI model (not ChatGPT) one from @Shrikantha5 is indeed ChatGPT. General problem is very real! Ps: LLMs are stochastic &

2022-12-08 18:37:38 @MuhammedKambal Sure would like to know what was in the training corpus

2022-12-08 18:31:46 Eerily prescient paper from 2021: language models &

2022-12-08 18:12:21 @pwlot @jeblad @StolfiAlberto I got blocked (I believe) for doubting his claim that Austin would become the AI capital of the world. And if you read the thread carefully someone else brought the wiki to my attention. Something didn’t feel right tbh when he interviewed me in the MIT library.

2022-12-08 18:09:38 @__delas__ @Plinz Love it when Plinz publicly attacks me and I respond and then the fanboys pile on without reading the time stamps or understanding the chronology.

2022-12-08 17:44:14 @pwlot @jeblad @StolfiAlberto he's representing himself as an MIT research scientist, not as a Drexel PhD.

2022-12-08 17:27:49 @Tweetermeyer wow. not a great look, especially given how he styles himself.

2022-12-08 17:26:03 @pwlot @jeblad @StolfiAlberto *did* he graduate? from the answers I am getting, he taught some January classes, but was not on faculty.

2022-12-08 17:21:33 RT @ChloeSavignac: Excellent line-up

2022-12-08 16:36:51 I suspect that this is inaccurate, but am curious if anyone can confirm or deny it. https://t.co/qlDi5cyH9c https://t.co/myflXl9IAr

2022-12-08 15:28:07 RT @wadhwa: And I did not found "Telesoft Partners". As well, Google clearly says that I authored 5 books, not 3 I am beginning to agree w…

2022-12-08 00:48:15 @Plinz @pwlot @seanmcbride @lexfridman Ok but I was on his show, he should have had thicker skin than that

2022-12-08 00:23:59 Ha ha Lex blocked me (after our episode!) because I didn’t agree that AI was going to relocate to Austin. Cc @Grady_Booch https://t.co/A5MaEtxDGH

2022-12-07 23:47:10 @seanmcbride @Plinz What takes tremendous intellectual effort is humility—and recognizing where improvements are needed.

2022-12-07 23:03:42 RT @emilkeyder: this is going to be a limiting factor for this kind of approach for a long time. there is no understanding or knowledge or…

2022-12-07 23:02:17 if you don’t understand this point, you don’t understand how ChatGPT works. and if you don’t understand this point, you are very likely overestimating GPT’s capabilities. https://t.co/0byiYi1orb

2022-12-07 22:54:50 No! even when there is structured data, chatGPT is limited in its ability to leverage that information. example: Wiki says Noam Chomsky is alive (which I can confirm)

2022-12-07 22:42:18 @John68Richmond The facts are clear: https://t.co/RIsKhp3yzX

2022-12-07 22:40:21 @kenneth0stanley @ylecun @hayabhay @ZoubinGhahrama1 @jeffclune @Tkaraletsos I hope @ylecun that @kenneth0stanley’s response will put an end to this line of discussion.

2022-12-07 22:38:08 @ylecun @hayabhay @ZDNET @TiernanRayTech So no counterargument then to the fact that - you publicly lied about my credentials - are resting a further attack on my credentials by conflating whether I helped develop an algorithm with whether I was permitted by its buyer to publish it. Classy!

2022-12-07 22:27:27 @ylecun @hayabhay @ZDNET @TiernanRayTech You know perfectly well that that work is under NDA and wasn’t published. That wasn’t my choice (since I no longer own it) but I have told you that before. You are just trying to extend your efforts to deceive.

2022-12-07 22:18:59 @fchollet Counterpoint: robotics (esp domestic humanoid robots) is one of the domains that most needs major advances in physical psychological and causal reasoning, and is one of the least tolerant of the current approximative approach. So could actually be the catalyst.

2022-12-07 22:15:19 @johnsontoddr4 @TonyD993 @yudapearl lovely example of how little gpt understands causality

2022-12-07 22:13:20 @ylecun @hayabhay Ps you also recently lied about my credentials to @zdnet, forcing @TiernanRayTech to publish a correction. And you said crazy things at end of 2019 about history that whole field knew was not true. Please stop attacking my credentials with false &

2022-12-07 22:06:23 @ylecun @hayabhay Uber bought that algorithm &

2022-12-07 21:23:10 RT @ibogost: I talked to ChatGPT about poetry, behavioral psychology, computer programming, hamburgers, and more. It's not a tool that will…

2022-12-07 20:50:20 @Michael_J_Black you might be interested in examples like these (I noticed you followed back on post but not here btw)

2022-12-07 20:34:22 More utterly bogus references from the alleged Future of Search https://t.co/y6ahzULYqO

2022-12-07 20:23:18 @RadioFreeTom He’s 94! And bemused by the fake GPTChat news I passed along, reporting his alleged 2020 death.

2022-12-07 20:04:33 @jjarangoes @ylecun @percyliang actually, @ylecun and agree on a lot: I started out 2917 debate with multiple points of agreement &

2022-12-07 19:40:19 @Agz5885 @ahatamiz1 @Amine57240266 stochasticity + authoritative presentation is not a winning formula

2022-12-07 19:39:08 @erikbryn Destroying trust in the information economy, on the Russian “firehouse of falsehood” model. Stack Overflow took the first blows. there will be more.

2022-12-07 19:34:43 @RadioFreeTom If it’s today, you also share it with Noam Chomsky

2022-12-07 19:32:34 @EricRWeinstein @elonmusk @ESYudkowsky I offered Elon some money about this topic, along with a definition (below) vetted by two of the people who invented the term. offer still stands: https://t.co/yfXycYXxb9 @wadhwa and others over to pentuple the bet to $500k.

2022-12-07 19:28:31 @ChristophMolnar I gave the clearest example

2022-12-07 19:25:38 Yes! But they include discussion of laws and explicit values in that education. They don’t leave it all to implicit reinforcement learning. See my pinned tweet, @ylecun cc: @percyliang https://t.co/5EykwGQGGA

2022-12-07 18:48:46 @MelMitchell1 I think the Turing Test is a terrible test, but two hours gives a good amount of time for depth. Cc @mkapor What Comes After the Turing Test? https://t.co/1N4qeEn6uH

2022-12-07 18:42:13 @ChristophMolnar Look forward to lots of novices puzzling over stuff like this https://t.co/7MsqFd2knj

2022-12-07 18:40:56 @ahatamiz1 @Amine57240266 @Agz5885 have you been able to replicate? Can you elaborate?

2022-12-07 18:09:27 RT @innovaTutor: A *must listen* interview with @GaryMarcus on the @BigTechnology Podcast https://t.co/pO0qQPMZ08 Is AI dangerously overh…

2022-12-07 18:09:05 @srikumarks The question of course not whether these systems sometimes create correct answers, but how often, and whether users (or downstream processors) can tell the difference when they instead generate wrong answers

2022-12-07 17:48:11 RT @percyliang: RL from human feedback seems to be the main tool for alignment. Given reward hacking and the falliability of humans, this s…

2022-12-07 16:42:21 @Ricardo_Joseh_L He took it well. (But shares my concern about the imminent deluge)

2022-12-07 16:37:37 Happy 94th birthday, Noam Chomsky! And don’t believe everything chatGPT tells you! (Spoiler alert: he doesn’t) https://t.co/VcdCVopBwx

2022-12-07 14:26:48 Absolutely. But it makes them up. Eg Noam Chomsky is very much alive (we emailed a few days ago), but check out the fake yet plausible looking sources below (&

2022-12-08 21:04:55 RT @GaryMarcus: @blattnerma @Plinz Yes &

2022-12-08 21:04:15 @tsushil_thapa @ShizaCharania @ylecun not friends with people who repeatedly lie about me

2022-12-08 21:03:50 absolutely - the edge cases in open-ended chatbots are endless

2022-12-08 21:03:00 @info_sprinkles @ylecun sorry but lying is not cool

2022-12-08 21:02:51 RT @MMikeMMa: @pgolding @plibin @GaryMarcus Isn’t a concern that these models are striving towards “general” capabilities across many use c…

2022-12-08 19:52:42 RT @GaryMarcus: @ylecun . @yLeCun the last time you said this to @zdnet, I presented multiple AI publications in top journals (sample scree…

2022-12-08 19:52:39 @ylecun . @yLeCun the last time you said this to @zdnet, I presented multiple AI publications in top journals (sample screenshots below). @zdnet corrected you, because you lied. @kenneth0stanley corrected you yesterday re our algorithm. *why* are you lying about my credentials again? https://t.co/uG07u8sYsi

2022-12-08 19:27:20 RT @GaryMarcus: @ShizaCharania @ylecun That’ll be $8 :)

2022-12-08 19:25:54 @ShizaCharania @ylecun That’ll be $8 :)

2022-12-08 19:18:02 @wadhwa @Shrikantha5 AFAIK all LLMs are vulnerable to hallucination. The thing you sent was made by a different recent OpenAI model (not ChatGPT) one from @Shrikantha5 is indeed ChatGPT. General problem is very real! Ps: LLMs are stochastic &

2022-12-08 18:37:38 @MuhammedKambal Sure would like to know what was in the training corpus

2022-12-08 18:31:46 Eerily prescient paper from 2021: language models &

2022-12-08 18:12:21 @pwlot @jeblad @StolfiAlberto I got blocked (I believe) for doubting his claim that Austin would become the AI capital of the world. And if you read the thread carefully someone else brought the wiki to my attention. Something didn’t feel right tbh when he interviewed me in the MIT library.

2022-12-08 18:09:38 @__delas__ @Plinz Love it when Plinz publicly attacks me and I respond and then the fanboys pile on without reading the time stamps or understanding the chronology.

2022-12-08 17:44:14 @pwlot @jeblad @StolfiAlberto he's representing himself as an MIT research scientist, not as a Drexel PhD.

2022-12-08 17:27:49 @Tweetermeyer wow. not a great look, especially given how he styles himself.

2022-12-08 17:26:03 @pwlot @jeblad @StolfiAlberto *did* he graduate? from the answers I am getting, he taught some January classes, but was not on faculty.

2022-12-08 17:21:33 RT @ChloeSavignac: Excellent line-up

2022-12-08 16:36:51 I suspect that this is inaccurate, but am curious if anyone can confirm or deny it. https://t.co/qlDi5cyH9c https://t.co/myflXl9IAr

2022-12-08 15:28:07 RT @wadhwa: And I did not found "Telesoft Partners". As well, Google clearly says that I authored 5 books, not 3 I am beginning to agree w…