Découvrez Les IA Experts
 |
Nando de Freitas | Chercheur chez Deepind |
 |
Nige Willson | Conférencier |
 |
Ria Pratyusha Kalluri | Chercheur, MIT |
 |
Ifeoma Ozoma | Directrice, Earthseed |
 |
Will Knight | Journaliste, Wired |
Profil AI Expert
Non Disponible
Les derniers messages de l'Expert:
2024-03-01 00:00:00 CAFIAC FIX
2024-03-11 00:00:00 CAFIAC FIX
2023-05-22 23:26:19 @ChiefScientist @yoavgo Or at least have cross-disciplinary teams addressing these questions
2023-05-22 22:34:37 @yoavgo Then there is all of the work on natural language interfaces to systems, which falls clearly under HCI. It all sounds like CS to me
2023-05-22 22:33:01 @yoavgo Probing and explanation are debugging tools for engineering these models and understanding what they’ve learned. AFAICT, prompting is just hacking and shows that we need better models.
2023-05-22 22:26:33 @yoavgo Why?
2023-05-22 15:15:45 @ArmandDoma I’m a 95% PowerPoint user
2023-05-22 14:12:56 RT @emilymbender: So the problem I see with the "FDA for AI" model of regulation is that it posits that AI needs to be regulated *separatel…
2023-05-22 05:14:51 RT @Miles_Brundage: The cluster of issues around: - Use of AI in influence operations + scams - Watermarking/provenance for AI outputs - O…
2023-05-22 05:14:45 @LeonDerczynski @Miles_Brundage I'd love to hear some details about what has already been implemented and what peoples' experience has been.
2023-05-22 02:12:30 This was the paper I was trying to find last week related to the question of whether the RLHF training data collected by companies gave them a moat. Looks like the answer is probably no. https://t.co/8oM7FyRODV
2023-05-22 00:47:56 @vardi and anonymized it!
2023-05-21 18:07:46 @mustafasuleymn You might also consider Beijing and Moscow. They aren't tech visions, but they are proposing to change the world.
2023-05-20 15:50:21 RT @ben_golub: *Laughs in Foster and Vohra*
2023-05-20 06:00:46 @TaliaRinger My Italian friends have an espresso before bed for this purpose
2023-05-19 22:20:44 @mrgemy95 Thanks!
2023-05-19 22:19:19 @TobyWalsh Well, $600 to fine tune an open weights LLM, not to train from zero.
2023-05-19 22:10:22 RT @rharang: So let's talk about prompt injection in large language models: the interesting, the dangerous, and the silly. 1/
2023-05-19 22:03:51 @AlvinGrissomII That looks very interesting, thanks!
2023-05-19 22:01:08 @mrgemy95 Yes, this is the canonical reference. Do you know of work that replicates their results or explains why ResNets exhibit poor calibration?
2023-05-19 19:00:00 CAFIAC FIX
2023-05-21 19:00:00 CAFIAC FIX
2023-04-21 00:00:01 CAFIAC FIX
2023-04-14 20:44:21 @cafharvard What an amazing photograph.
2023-04-14 16:38:10 @balazskegl @GaryMarcus In other words, if I create an AI system and program it at the level of goals, it has some agency in the sense that it’s behavior is causally related to its goals and can be predicted in terms of them.
2023-04-14 16:34:36 @balazskegl @GaryMarcus All code has bugs
2023-04-14 04:19:16 @GaryMarcus In other words, much ado about nothing.
2023-04-14 00:07:56 @GaryMarcus The other thing generative AI systems offer is improved persuasiveness for highly selective highly effective spearphishing. If only a few people are attacked, detection probability goes way down. end/
2023-04-14 00:05:46 @GaryMarcus Phishing (e.g., for ransomware) faces a similar detection risk. My understanding is that one effective strategy is to make the phish so obvious that 99% of people just ignore it. And a single person falling for it is sufficient. 2/
2023-04-14 00:03:38 @GaryMarcus People can already do all of these things (or write code to do them). The main thing that AI systems offer is scale. But a large scale attack of these kinds is also much more likely to be detected. 1/
2023-04-14 00:00:06 @corecursion @tlbtlbtlb @JeffLadish Any appeal to an unlimitedly-smart mind is not a serious argument.
2023-04-13 20:32:27 @tech_queen Math is important for algorithm design and XAI, but there are parts of ML that are less math-intensive such as MLOps, an area that needs a lot more skilled people. This mixes distributed systems, software engineering, and ML.
2023-04-13 19:08:22 @norabelrose @sindero Could this also be used to delete individual facts? Eg for compliance with GDPR right to be forgotten? @yoavgo . Pointers to other relevant work?
2023-04-13 17:38:05 RT @ddwoods2: w/ explosion of unexpected effects from deployed AI, time to remind all of (Robin) Murphy's Law: Any deployment of AI systems…
2023-04-13 01:00:18 @marek_rosa At least anyone paying for LLM queries via an API will be strongly incentivized to control their agents! That won't stop a nation state actor, however.
2023-04-13 00:58:50 Deterrence is one approach: Robert Morris was convicted of a felony for releasing the Morris Worm. https://t.co/Mx7cJehsBt https://t.co/wtejYgjf75
2023-04-12 21:52:59 @shortstein Back when many citations were to books, the page number was very important. Back when physical journals were bound into hardcover volumes, the page number of important to know which volume to pull off the shelf. Still need them to request inter-library loan
2023-04-12 20:42:58 @daniel_kraft What safety measures are you taking to prevent abuse of these capabilities? Seems like this could take DDoS attacks to a new, semantically-powerful level.
2023-04-11 22:00:51 @mattturck How well do prompts for GPT-2 work for GPT-3? How well do prompts for GPT-3 work for GPT-4? Is there any concern that prompting skills will need constant revision or do they generalize well?
2023-04-11 12:15:28 RT @aihuborg: Forthcoming machine learning and AI seminars: April 2023 edition - https://t.co/nGYVUvpQXS https://t.co/ykszs0B73t
2023-04-11 12:06:02 RT @mcxfrank: What does it mean for a large language model (LLM) to "have" a particular ability? Developmental psychologists argue about th…
2023-04-10 20:40:32 @ziv_ravid Well, it knows the appropriate words to output. As it has never had children or gone on spring break, it has only a limited understanding of what they mean.
2023-04-10 15:11:30 @rao2z But of course they should instead be expressed in a formal language. Anyone working on that?
2023-04-10 15:02:02 @andyroyle_pwrc Hang in there
2023-04-10 01:09:43 @ShannonVallor A revealing example of how anything unusual will tend to disappear when LLMs are used. Machine learning is good at the common case and bad in the low probability tails of the distribution.
2023-04-10 01:05:38 @mishadavinci I would say 0% understand ChatGPT deeply. I also don’t see how this recital of AI history helps us understand ChatGPT. It is nice though.
2023-04-09 21:07:04 @davidmanheim @ylecun This workshop is a reasonable starting point for folks who are interested: https://t.co/7qyorlsJva. Also check out Zico Kolter's work @zicokolter
2023-04-09 21:02:31 @davidmanheim @ylecun But for LLMs, we don't have a good definition of correct performance, so that makes them impossible to validate *in general*. I think it is feasible (but challenging) to study specific criteria such detecting (and censoring) porn, racism, etc.
2023-04-09 21:00:36 @davidmanheim @ylecun There is a community of researchers looking at verification and validation of ML-based systems in safety-critical applications. But those systems are not based on LLMs. It would be great to see more people working on problems such as confirming *systematic* behavior
2023-04-09 19:19:11 @davidmanheim @ylecun *end user
2023-04-09 19:06:35 @davidmanheim @ylecun The analogy seems a bit broken. We require certification of aircraft (including software) before it can fly. But running simulations during the design process is not regulated. Is the end used analogous to the pilot or to the passenger?
2023-04-09 15:45:07 @MoarPart @dvassallo Yes. And if you look at the GPT-4 report, you will see that those probabilities are well calibrated prior to RLHF fine tuning. But the fine tuning damages that, and the model becomes quite overconfident. This can probably be fixed, at least in targeted applications
2023-04-08 23:29:32 @dvassallo The key difference is that a database query will return the empty set if there are no matching records. The LLM will generate plausible tuples instead. It’s a probabilistic model of a database rather than a database itself.
2023-04-08 23:25:35 @ajcwebdev @dvassallo How will it get these reasoning capabilities?
2023-04-08 22:24:35 @GaryMarcus @zakkohane “Judgment”: yet another inappropriate anthropomorphism! Yikes
2023-04-08 15:20:30 @rasbt @EMostaque Hmm, arxiv has plenty of junk food, including quite a few papers on AI.
2023-04-08 03:42:27 RT @suchisaria: I’ve worked in AI for 20+ yrs &
2023-04-07 20:53:05 @IDoTheThinking Depends on how much they charge. It needs to be much cheaper than a blue check!
2023-04-07 20:33:00 @fchollet Do you consider in-context learning to be an example of rapid skill acquisition?
2023-04-07 15:42:58 RT @ichbinilya: This is hilarious. Elon comes in trumpeting "free speech." Thousands of his biggest fans have been parroting that rhetoric…
2023-04-07 03:46:43 @karpathy I suppose retrieval-augmented models can then be viewed as having external read-only-memory from which they can retrieve. This allows the "GPT CPU" to be 10x smaller (e.g., in RETRO https://t.co/Hc9OeXpH8A)
2023-04-06 19:35:03 @BallouxFrancois Isn't the question at hand whether Pfizer would have submitted their EUA application earlier if Topol had not intervened? And hence, the EUAs might have been granted a month earlier?
2023-04-06 19:08:35 @BallouxFrancois This article doesn't claim that Topol actually was successful in slowing EUA, only that he tried very hard. Is there any analysis of the decision making process inside Pfizer? It would be very interesting to know if they took a political decision
2023-04-06 19:03:08 @BallouxFrancois Especially as none of them were actual lockdowns as they were in the UK, Spain, etc.
2023-04-06 06:15:14 @TobyWalsh I had in mind certificates that also handled image manipulations. I think there are some efforts in this direction for scientific imagery. The camera signs the image, then each tool documents the changes and signs the result, etc.
2023-04-06 05:48:52 It is such a pleasure to read papers from Girshick's team (@inkynumbers). The Segment Anything paper is wonderful. https://t.co/ECgijKaMcc
2023-04-06 05:05:02 @TobyWalsh Toby, do you know of any efforts to develop digital certificates of authenticity for real photos, videos, and documents? If these were widely available, we could assume everything else should not be trusted.
2023-04-06 03:58:01 @hamishmckenzie Who are the writers and creators? If I post, does that make me one of them?
2023-04-05 20:46:07 AAAI presidents seek to create a community focused on beneficial AI. We call on researchers, governments, corporations, and the public to join this effort. https://t.co/hY5OksPP98
2023-04-05 15:56:33 @bitcloud @random_walker Can you point me to that work? I’m mostly seeing work like https://t.co/wPjFQdQX8E that mixes retrieved passages into the context buffer
2023-04-05 14:28:12 @random_walker It has to be one of the oldest design flaws in communications engineering: mixing commands and data in the same stream like the old long distance telephone network did. https://t.co/34aeK5PRQ2
2023-04-04 22:32:59 @erichorvitz @WHOSTP Thank you for your service to the AI community and the country
2023-04-04 20:58:03 @bleepbeepbzzz @compthink I am aware of that. What do you think is the best work in this direction?
2023-04-04 20:26:31 @katecrawford @mjnblack Or, tbh, maybe the h-index was always a bit of a hallucination...
2023-04-04 20:26:02 @katecrawford @mjnblack Gives a whole new meaning to the "h-index" based on hallucitation counts
2023-04-04 20:23:47 @katecrawford @mjnblack This is the word we've been needing
2023-04-04 18:25:32 @compthink LLMs have short-term memory in the context buffer, and the buffers are getting quite large, so one must be cautious with this argument. It is still basically correct, but we may be surprised how much useful memory can fit in the context buffer
2023-04-04 18:23:21 RT @mcxfrank: People are testing large language models (LLMs) on their "cognitive" abilities - theory of mind, causality, syllogistic reaso…
2023-04-04 18:22:04 @ylecun Everyone is a stakeholder, so everyone has a legitimate role to play in determining policy. We have an obligation to educate the public in the clearest and most unbiased way we can. I don't think this diagram is helpful in that regard. @savvyRL
2023-04-04 18:17:35 @BallouxFrancois @spinfocl Is there a benefit to having Covid sweep rapidly through the population? India seems to have had very low numbers after their horrible wave, and maybe China is benefiting similarly?
2023-04-04 05:43:14 @alexolshevsky1 @aryehazan I agree that these models understand sometimes. But the understanding has a point-wise character. It is fragile. Hence, we can't rely on them. We say a person understands something when they understand it systematically and robustly. Then we can trust them.
2023-04-03 20:55:26 @annajobin @ovoss @TspBackgroundDi @random_w @sayashk I thought this article captured the various points of view very well.
2023-04-03 19:22:15 @GaryMarcus Has anyone confirmed that GPT-5 is being trained?
2023-04-02 23:32:50 I like this insight from @sarahookr https://t.co/7akgSngbK0
2023-04-02 04:50:06 @GaryMarcus @a11enhw I agree that it is good practice to adopt a charitable interpretation of things our colleagues post. I extend that to @ylecun's posts too.
2023-04-02 04:23:52 @GaryMarcus @a11enhw This seems very different from concerns about, say, autonomous weapons or political destabilization
2023-04-02 04:22:03 @GaryMarcus @a11enhw But the letter you signed tried to link legitimate concerns about safety and trust with weird statements about being replaced by AIs or having our culture destroyed.
2023-04-01 22:50:05 @dingywhiskers @ylecun @erikbryn We could ask the LLM whether its proposed output would violate a rule. There are lots of things to try (and maybe some of them have already been tried and didn't work). Much more research is needed.
2023-04-01 22:41:28 @dingywhiskers @ylecun @erikbryn It doesn't work very well. We can do better, for example, by actually reasoning about the generated output and societal norms rather than just biasing the weights. We know RLHF damages calibration, and it is entirely opaque.
2023-04-01 22:33:33 @GaryMarcus One minute Gary is writing a book about how to improve AI and the next he is signing a letter that warns "we risk loss of control of our civilization". I'm puzzled...
2023-04-01 21:23:31 @mark_riedl It would be nice to have a trainable agent that could serve as my "personal recommender".
2023-04-01 21:22:02 @ylecun @erikbryn I like to say, only partially in jest, that our systems need a superego
2023-04-01 21:21:15 @ylecun @erikbryn We are already incorporating guardrails of various kinds. Building a meta-cognitive layer that can modulate system outputs based on societal norms and organizational rules should be a high priority. It is clearly not straightforward, but I think it is doable.
2023-04-01 20:18:58 @erikbryn I think there is broad agreement that safety is essential. The disagreement is about the nature of the danger. Application of deep fakes in fraud and defamation are the dangers to focus on.
2023-04-01 20:13:53 @wellingmax @ylecun I was particularly puzzled by what "we risk loss of control of our civilization" could mean.
2023-04-01 20:13:20 @wellingmax @ylecun But I wonder if they were signing because "we develop nonhuman minds that might eventually outnumber, outsmart, obsolete and replace us" or "we let machines flood our information channels with propaganda and untruth". I suspect it was much more the latter than the former
2023-04-01 17:14:18 It is notable that while we are willing to use an opaque language model to generate text, we are not happy with an opaque mechanism for preventing the model from violating social and ethical norms. We want the self-censorship to be interpretable to ensure fairness
2023-04-01 17:07:34 @geomblog @sharongoldman I think it is also important to analyze the rhetoric. The FLI letter uses "control" in ways that can have multiple interpretations. Technically, OpenAI has trouble controlling ChatGPT's output, but that is different from AI being out of control &
2023-04-01 15:52:18 @davisblalock @PierreAlquier Bias in the training data is likely as well, of course.
2023-04-01 15:51:02 @davisblalock @PierreAlquier I wonder if this results from spotty and inconsistent generalization of RLHF. It would seem difficult to inspect the learned reward function of RLHF as well as to understand what gets changed inside the LLM when trained. A more transparent technique is needed.
2023-04-01 04:17:59 @roydanroy Certainly not me
2023-03-31 22:12:33 I also support their call for standards and regulations that reduce the risk that users will be deceived by an AI system. Interaction design should discourage anthropomorphization, clearly identify non-human systems, and communicate the limitations of these systems.
2023-03-31 21:56:46 This paper presents an interesting perspective. Nearly everything missing from current generative AI models corresponds to functions in the pre-frontal cortex including self-regulation. "Deep Learning Needs a Pre-Frontal Cortex". https://t.co/Ia3JOmJz6Z
2023-03-31 21:56:45 Many good points in this statement. On the technical side, we need much more research on how to constrain the output of generative AI to comply with societal norms. On the corporate side, companies need to adopt and enforce safe and fair employment practices. https://t.co/P1MXCILzJi
2023-03-31 18:55:15 @Sandcommando201 Which country is yours?
2023-03-31 05:33:12 @GaryMarcus @bleepbeepbzzz @Noahpinion @OpenAI Thank you
2023-03-31 05:26:07 @bleepbeepbzzz @GaryMarcus @Noahpinion Is there evidence that GPT-4 is more dangerous than GPT-3 or ChatGPT? It seems more truthful, which should make it less dangerous. [https://t.co/ARWzqZAwL9] https://t.co/aZgoFzk4LL
2023-03-31 03:08:19 @GaryMarcus They will also be more accurate, interpretable, and robust. In short, achieving the goals laid out in the letter requires continued model development
2023-03-31 03:06:35 @GaryMarcus Those will be "more powerful than GPT-4" if you and I are right about hybrid systems.
2023-03-31 03:06:15 @GaryMarcus Yes, and I'm seeing lots of work on hybrid systems. But some of those systems--if they are to successfully separate, say, factual knowledge from linguistic knowledge, will need to train new generations of LLMs in concert with knowledge bases and reasoning engines.
2023-03-31 03:00:22 @GaryMarcus And who is to say how long such a moratorium would actually last? Will fear mongering end up where it did with GMOs, nuclear power, and vaccines?
2023-03-31 02:57:55 @GaryMarcus If you read the newspaper coverage, you will see that this is not what people (or congresspeople) are hearing. Furthermore, addressing many of the shortcomings of current models will require training new ones. Not allowed by the moratorium.
2023-03-30 22:38:03 @bleepbeepbzzz @Noahpinion Fixing these would make it easier to attain the benefits without a lot of ad hoc prompt engineering and guardrails. end/
2023-03-30 22:37:14 @bleepbeepbzzz @Noahpinion Our current models have many shortcomings--ones that I don't think will be fixed by scaling but that might fall under the proposed moratorium. Examples: hallucination, poor ability to plan and reason, inability to construct situation models (for dialogue/narrative), etc. 1/
2023-03-30 18:09:03 @ArmandDoma The key word is “it”. There is no one home. The system, as far as we know, has no privileged access to its own internals, so it is just making up answers about itself based upon what it has read.
2023-03-30 16:40:34 @ArmandDoma The whole premise that you can interview a system is out-of-control anthropomorphism
2023-03-30 05:34:54 A couple of notes: 1. This project is led by the Indian government. I believe Microsoft worked on parts of the software stack. 2. The demo only shows Hindi. I couldn't find a demo in a regional language, but I believe the goal is to support all 22 constitutional languages
2023-03-30 05:12:25 @shettyprajwal_ @insign02 @AndrewYNg That's not clear. On most web sites, I don't even try to get any customer service. It is the one thing that currently doesn't scale to millions and billions of customers. I think there is huge unserved demand
2023-03-30 05:07:49 @shettyprajwal_ @insign02 @AndrewYNg I agree that call center jobs will change. But we know that existing customer service does not come close to scaling with demand. If each call center worker can supervise dozens of chatbots, each worker will become more valuable to their employer.
2023-03-30 05:03:49 @nutanc Yes, I should have been more precise. It is a collaboration led by MeitY. My understanding is that Microsoft has contributed to the design of the software stack. But you are totally correct.
2023-03-30 05:00:21 @insign02 @shettyprajwal_ @AndrewYNg I recommend you read this post for some economic thinking about jobs and AI. https://t.co/MCxQvRNEfb
2023-03-30 04:58:07 @insign02 @shettyprajwal_ @AndrewYNg What is an example of a job that you think will be lost? I'd be very interested in specific examples. Maybe the typists who fill out forms for low-literate folks? Other ideas?
2023-03-30 04:12:04 This app combines speech recognition in local language, translation to Hindi, ChatGPT interaction with govt documents to get answers in Hindi, translation back to the local language, and speech synthesis.
2023-03-30 04:12:03 Important post from @Noahpinion. LLM-based tools have the potential to make all of us more efficient. Don't let the fearmongers rob us of the benefits of good AI tools. https://t.co/emFJjmaZT8
2023-03-30 00:00:58 @shettyprajwal_ @AndrewYNg What are the threats that you see? I don't see how better writing tools, for example, are going to lead to mass unemployment in India. But maybe this is just a failure of imagination on my part. I'd love to know more.
2023-03-29 19:25:37 @yoavgo @boazbaraktcs Everyone will end up on street corners
2023-03-29 17:00:28 @DrJimFan (including my tweets, I'm afraid)
2023-03-29 16:58:40 @DrJimFan I suspect the humility is hard-coded. Difficult to believe it could be acquired from reading the internet
2023-03-29 16:54:46 @RishiBommasani This is great! Thank you
2023-03-29 16:51:15 @MelMitchell1 I also did not sign. The letter is such a mess of scary rhetoric and ineffective/non-existent policy prescriptions. There are important technical and policy issues, and many of us are working on them.
2023-03-29 16:02:16 RT @random_walker: This open letter — ironically but unsurprisingly — further fuels AI hype and makes it harder to tackle real, already occ…
2023-03-29 15:48:15 @MLJ_Social On the other hand, if we knew who you were, the opinions might carry some weight.
2023-03-29 14:50:26 @MLJ_Social It is odd for a journal social account to express opinions about papers. I’d prefer that you just post the titles and links.
2023-03-29 06:00:11 @BenFMaier @BauwensRobin I do find that ChatGPT can improve paragraphs, particularly by writers who are not very fluent in English. I haven't tried using it to make something more concise though.
2023-03-29 02:41:39 RT @boazbaraktcs: I think it’s more important for safety and understanding of AI that a model matching GPT4 is trained in the academic doma…
2023-03-28 23:47:46 @TDeryugina You qualify to write promotion and tenure letters for a wider range of people. This is both a cost but also an increase in power to help good people advance.
2023-03-28 22:24:00 @squidjerk Thank you!
2023-03-28 21:59:57 Your front page currently plays 4 videos simultaneously. My computer is overheating.
2023-03-28 21:55:20 @nytimes How do I disable video autoplay in Firefox? It is ironic that you have published multiple how-to articles on disabling autoplay but then you force it on your paying customers anyway.
2023-03-28 20:04:02 RT @melgirm: “We have not loved children enough in this country to change our relationships to violence.” - @KieseLaymon https://t.co/6kVKf…
2023-03-28 14:24:39 @MLJ_Social Who is writing these tweets?
2023-03-28 05:27:21 RT @thegautamkamath: TMLR (@TmlrOrg) just put out a report on their first (almost) year of operations. Here's a with some of the highligh…
2023-03-28 04:42:16 @LeonDerczynski Thanks! I can imagine a system that looks not only at the content of a message but also at the broader situation. For example, my university recently rolled out the Microsoft email quarantine system. After a week, I got a phishing email designed to imitate it.
2023-03-28 04:04:14 @emilymbender @TonyHoWasHere I'm skeptical of the sentiment machines as well
2023-03-28 03:41:49 The main short-term risks of LLMs is that they will improve the effectiveness of phishing, cyber attacks, and disinformation. But LLM-enabled systems have the potential to help defend against these as well. A priority should be to invest in such defensive systems. 1/
2023-03-28 03:36:06 @timpplex @Noahpinion When I was growing up, all my nerdy friends wanted to work either for NASA or for Bell Labs. As @Noahpinion has written, the place was incredibly well-managed to promote great work. You don't need to offer big salaries to get top people if you make everything else easy.
2023-03-28 03:17:10 @Noahpinion One might ask why our existing national labs fail to be as effective as Bell Labs. end/
2023-03-28 03:16:38 @Noahpinion My understanding was that it was the loss of the telephone monopoly that led to the loss in surplus available for research. Bell Labs was essentially a national lab supported by that surplus. The USG could create another lab using tax dollars if they wanted to. 1/
2023-03-27 19:10:49 @shortstein @arxiv Maybe it doesn't, but the variety of latex packages and different layouts suggests to me that what would be hard to code with a parser would be trivial with an LLM.
2023-03-27 18:47:58 A potential LLM application: Extract all of the metadata required by @arxiv from the latex or pdf file. As a moderator, I see so many copy/paste errors that result from a tedious and error-prone task
2023-03-27 16:42:29 RT @mcxfrank: How do we compare the scale of language learning input for large language models vs. humans? I've been trying to come to grip…
2023-03-27 16:37:49 @ChrisMurphyCT @rao2z “It” didn’t “decide” anything. Programmers fit the model to many many chemistry papers.
2023-03-27 05:57:37 @AlexHBruns @Noahpinion As an AI researcher, I agree it is possible, but only up to a point. To the extent that the requirements gathering is a political process that gains legitimacy from the process itself, we will insist on people leading it.
2023-03-27 01:39:09 @Noahpinion I know @google is reporting big productivity improvements with their in-house system similar to Copilot. https://t.co/N8bOaMLaxt
2023-03-27 01:27:36 Machine tools didn't eliminate mechanical engineers. I think the same will apply to software engineering.
2023-03-27 01:26:37 @Noahpinion There is also going to be a huge effort to apply LLM-based methods to understand and replace legacy systems. Maybe they can help us retire decades of technical debt?
2023-03-27 01:25:13 @Noahpinion The second most important software engineering problem is keeping the software running as the requirements and the surrounding environment change. Perhaps LLMs will have their biggest impact here?
2023-03-27 01:24:08 @Noahpinion The central problem in software engineering is to figure out what the software is supposed to do and then express that precisely. How much of this can be easily and accurately expressed in natural language?
2023-03-26 22:54:06 @mapto @emilymbender True enough, but I do have a tendency to think out loud. I apologize to my followers (and @emilymbender's) for adding noise to their feeds.
2023-03-26 21:50:15 @raphaelmilliere Please thank the sound and camera people. The live stream was great!
2023-03-26 04:38:28 @ccanonne_ Brilliant!
2023-03-25 20:10:48 @emilymbender Indeed. Thank you for explaining twitter to me.
2023-03-25 20:01:34 @emilymbender Sorry
2023-03-25 20:00:42 RT @emilymbender: Carl Bergstrom has a banger thread over on Mastodon about some serious #privacy problems with #Bing #GPT. You can find th…
2023-03-25 20:00:28 @emilymbender Of course, this is exactly the point you made in your paper. I just had not fully appreciated the consequences.
2023-03-25 19:59:40 @emilymbender Excellent example of the failure of these systems as dialogue agents. A dialogue is more than a Markov chain of alternating utterances. I suspect this problem will be hard to solve without much more work
2023-03-25 19:35:03 @katebevan Google: "organize the world's information and make it universally accessible and useful" Us: We didn't mean *that* accessible!
2023-03-25 18:57:10 RT @ProfNoahGian: I found the recent @nytimes opinion piece on AI by @harari_yuval @tristanharris @aza very interesting and I agree with so…
2023-03-25 17:20:49 @GaryMarcus @miguelisolano If you want deterministic results, set temperature to zero. If you want to replicate stochastic results, you will need a large sample size.
2023-03-25 17:16:15 RT @ruchowdh: I kinda hate that responsible AI has come full circle to the uneducated yet highly opinionated pontificating on topics they k…
2023-03-25 17:13:46 I'm really enjoying this symposium on Deep Learning and Philosophy. Thank you to the organizers for the live stream! https://t.co/FjoZFc4hel
2023-03-25 16:28:50 @GaryMarcus @miguelisolano He is using non-zero temperature. Doesn't that guarantee a high probability that answers will not replicate?
2023-03-25 16:26:59 @AndrewLampinen Yes. Smarter systems that can marshal multiple software tools to carry out tasks must necessarily have access to our private information in those tools. Privacy needs to be designed in from the very start.
2023-03-25 16:21:12 @random_walker @ccanonne_ In a separate thread, the discussion has focused on the aspects of latex (programmability) that make it difficult to automatically remove comments with 100% success. @arxiv has good reasons to collect and distribute the latex source code, but these introduce risks as well.
2023-03-25 16:14:50 RT @random_walker: The YOLO attitude to security is baffling. I see a pattern: OpenAI overplays hypothetical risks arising from the models…
2023-03-25 03:47:48 @shortstein @ccanonne_ @arxiv Yes. In fact, in the paper I just uploaded this week (https://t.co/St45pZNxhq), I forgot to delete a big section that had been commented out using \begin{comment}\end{comment}. I had only searched for %. (sigh)
2023-03-25 00:52:06 @solvay_1927 @dustinvtran Given that I don't think AGI is a well-defined thing, I predict progress will look a lot like Moore's law--a staircase of sigmoids each reflecting a new advance. The LLM sigmoid is a big step in generality, but there is lots of head room to go beyond LLMs.
2023-03-24 23:42:37 @ccanonne_ @arxiv Good point
2023-03-24 23:33:25 @ccanonne_ @arxiv I think it would be nice for @arxiv to offer to automatically delete comments from the source files. This is a common error.
2023-03-24 22:16:05 @eclecticleaps @MelMitchell1 @info_sprinkles @erikbryn I recommend @ylecun's paper https://t.co/7RjwYz23Uy for a thorough review of non-linguistic knowledge.
2023-03-24 19:14:34 @rasbt Aren't you a big fan of tree boosting? I am, and I use the R implementation. In fact, I do most of my ML research in R except for deep learning work
2023-03-24 19:12:15 @conways_law @Grady_Booch @davetroy https://t.co/ddtCqZtzoe
2023-03-24 16:08:31 @peteratmsr Pursuing the analogy, will most of these pre-Cambrian organisms rapidly go extinct?
2023-03-24 05:38:30 https://t.co/ek3ozG57af
2023-03-24 04:39:17 @dustinvtran We certainly have a lot of new capabilities to understand. But if the history of tech is any guide, this wave will have a sigmoidal shape. In any case, the second derivative sure feels positive right now.
2023-03-24 04:37:06 @eclecticleaps @MelMitchell1 @info_sprinkles @erikbryn How about all of physical interaction? How about the smell of rotten food? The sound of a bearing about to fail? The feeling in your hand when you realize you are going to drop a knife and you can't prevent it from hitting your toe?
2023-03-24 01:23:52 @Dilip_Arumugam It might be a limiting case of RDT where the conditional mutual information is driven to zero. Interesting...
2023-03-24 00:56:22 This plot shows our methods (GRDS, SRAS) match the oracle (purple) and do much better than the PPO baseline that uses the original reward function. 6/
2023-03-24 00:56:21 If the reward decomposes additively as R_exo(X) + R_end(X,R,A), then the MDP can be decomposed into a Markov Reward Process that involves only R_exo and an MDP that only maximizes R_end (the endogenous reward). 3/
2023-03-24 00:56:20 George Trimponias and I are pleased to announce our paper on Exogenous State Variables and Rewards in reinforcement learning. The paper develops a mathematical theory of exogenous state variables based on Pearl's causality theory. 1/ https://t.co/St45pZNxhq
2023-03-23 23:50:52 @ben_levinstein "intelligence" is not a boolean but a relative measure. GPT-4 is more intelligent than many previous AI systems and its breadth of knowledge is probably wider than any previous AI system
2023-03-23 20:04:29 @rasbt Of course like all quizzes the results cut both ways. Maybe the reviewers couldn't understand the paper (or the quiz) because it was poorly written
2023-03-23 19:06:15 Nice blog post by the always-insightful @AaronHertzmann https://t.co/eHkm7fT2Iz
2023-03-23 17:30:59 @MelMitchell1 @info_sprinkles @erikbryn Huge amounts are missing, and current large models don't capture things correctly, either (as you have pointed out). But for people worrying about intelligence explosions, encoding humanity's existing knowledge isn't going to cause one.
2023-03-23 17:20:48 @info_sprinkles @erikbryn Of course #2 could also lead to a huge expansion in our understanding of the world. So much knowledge is siloed, and LLMs++ can span all silos and uncover surprising connections. That could be a one-time "explosion", but not a "chain reaction".
2023-03-23 17:17:51 RT @tdietterich: @info_sprinkles @erikbryn It might be useful to view our progress in AI in three main phases: 1. Encoding existing humanit…
2023-03-23 17:16:37 @info_sprinkles @erikbryn It might be useful to view our progress in AI in three main phases: 1. Encoding existing humanity's knowledge 2. Working out the consequences of that knowledge 3. Adding new knowledge by making new discoveries #3 might lead to an explosion. We are working on #2 right now
2023-03-23 16:12:19 @erikbryn But passing the Turing Test turns out to just be a parlor trick (with lots of downsides). It doesn't lead to a singularity
2023-03-23 04:52:12 @GaryMarcus @TobyWalsh I must have replied to the wrong thread. Sorry
2023-03-23 04:17:24 @rasbt Yes. Furthermore all prior versions of the paper are available. The arXiv paper can also point to the published paper (which is highly encouraged and presumably helps @SemanticScholar and @Google scholar detect duplicates).
2023-03-22 22:38:38 @GaryMarcus @TobyWalsh Here https://t.co/cICTS4Aj41
2023-03-22 19:47:31 @erikbryn @McKinsey_MGI What explains Miami, Las Vegas, and Orlando? Wasn't Miami supposed to be the Next Great Place? The data end in 2019
2023-03-20 23:02:00 @thegautamkamath When AI conferences first considered giving awards the goal was to boost people applying for jobs, for tenure, etc. Other fields were giving awards, and this led to an "arms race".
2023-03-20 22:00:46 @yoavgo In particular, while it is reasonable to expect people who do well on these exams would also do well across the domain, we lack any basis for such an expectation with LLMs
2023-03-20 20:53:26 @amcafee @matt_levine I’d delete the comma instead
2023-03-20 18:41:44 @MazMHussain IMO The strategic error was to invade Iraq at all. It had little strategic value for the US. And it led to multiple strategic setbacks for the US.
2023-03-20 05:28:50 @rao2z @SCAI_ASU @ASUEngineering @iiit_hyderabad Is this like in Spain, where "profesor" means teacher, and "real" professors are called "catedratico" (i.e., they have a chair)?
2023-03-20 03:13:12 @BobMcIntyre53 Carlin is only right if the median and the mean coincide. Seems plausible...
2023-03-19 23:54:28 @MuzafferKal_ Exactly! I didn't mean to give the impression I was defending him. I think it is a disingenuous statement, and he has been incredibly irresponsible testing his software on the general public.
2023-03-19 23:08:22 Notice how carefully @elonmusk worded his claim: "we’ll have over a million Tesla cars on the road with full self-driving hardware". He's said the same thing since 2015. No mention of software. https://t.co/ulEcWZNIZA
2023-03-19 22:19:08 RT @AndrewLampinen: The recent discussions of what language models can and can't accomplish highlight some important issues in how cognitiv…
2023-03-19 20:22:02 @Pajuhaan @davidchalmers42 I don't think McCarthy ever understood machine learning or its potential. He occasionally used the word "learn", but it usually mean "acquire a logical formula from a programmer".
2023-03-19 17:06:55 @cloudquistador @Abebab @Grady_Booch @OpenAI My point is that it is the task that is the problem, not the wages. You could pay $1000 an hour and it would still be unethical
2023-03-19 17:03:49 @cloudquistador @Abebab @Grady_Booch @OpenAI Agreed
2023-03-19 16:58:05 @cloudquistador @Abebab @Grady_Booch @OpenAI No, but @abebab said "uncompensated". Perhaps she meant "undercompensated"? My view is that some of these labeling tasks are unethical regardless of the pay rate. For some tasks, supervised learning is impossible because obtaining the labels is unethical.
2023-03-19 16:09:26 @yoavgo The basic use case for citation is when you use a technique or result without providing all of the details. Your want to point the reader to the best place to get those details. I agree this could be a handbook or subsequent paper rather than the original
2023-03-19 16:05:23 @Abebab @Grady_Booch Are you saying @OpenAI doesn’t pay their crowd workers?
2023-03-19 05:22:10 @stephxsher @taylorswift13 @karpathy Swift clearly has a larger out-degree in the influence graph, but the edge from Karpathy to me has much higher weight than the edge from Swift to me.
2023-03-19 05:19:10 @mgruchy @DrJimFan @GaryMarcus @random_walker I'd be curious to see if you provided not only the standard chess move notation but also the board positions (as images). Properly generalized, this would give the LM the correct model of the world and its dynamics.
2023-03-18 20:10:41 @rao2z Some representation probing experiments might falsify this claim.
2023-03-17 20:46:34 @danbri @iruletheworldmo @GaryMarcus @ESYudkowsky @sama "pretend" implies that the system is aware that it is not sentient, which is almost an impossibility
2023-03-17 18:09:58 @fadeyifemi @iruletheworldmo @GaryMarcus @ESYudkowsky @sama But many such humans will be imprisoned and held responsible for the damage they cause.
2023-03-17 18:08:01 @iruletheworldmo @GaryMarcus @ESYudkowsky @sama Please don't repeat the mistakes of the past. https://t.co/Mx7cJei0r1 Morris also just wanted to see if it could be done.
2023-03-17 17:49:35 @iruletheworldmo @GaryMarcus @ESYudkowsky @sama If I had to guess, I'd bet you are enforcing a narrow scope of potential actions the agent can perform. But I'd love to learn more and understand what is easy/hard to achieve.
2023-03-17 17:45:13 @RamaswmySridhar @vivek7ue @OpenAI @verge The research community needs full access to at least some models (including the largest ones) so that we can study why they work, how to improve them, etc. But a few well-documented, well-supported ones is more important than many, poorly-supported ones
2023-03-17 17:42:06 @iruletheworldmo @GaryMarcus @ESYudkowsky @sama What steps are you taking to prevent your agent from causing damage?
2023-03-17 15:05:35 @GaryMarcus @ESYudkowsky @sama Converting an LLM into an agent with unrestricted access to the internet (or even to the local shell) is asking for trouble.
2023-03-16 21:40:49 @taexalon @AlbertoCairo tbh That is the first thing everyone should try
2023-03-16 18:46:04 @LeonDerczynski @fchollet Nicer fonts, though
2023-03-16 16:52:08 RT @geomblog: Roko's* five stages of AI grief: 1. Denial: "there is no bias in AI systems. Math is not racist." I feel like we have lar…
2023-03-16 15:19:46 @beenwrekt Yes. Making it easier to generate "communication" does nothing to increase the supply of human attention. Seems like this will favor the creation of bots communicating on our behalf. This might work for some routine things, but I think it will quickly lead to trouble.
2023-03-16 15:10:53 Bulleted lists in gmail when received by Outlook are rendered with lots of extra bullets. Does anyone know why? @gmail @outlook https://t.co/pNwoQBOpOC
2023-03-16 04:48:02 @fchollet Evidence in favor: the popularity of powerpoint Evidence against: powerpoint presentations
2023-03-16 03:28:46 @IDoTheThinking Also the pizza places near college campuses (which is consistent with your foot traffic point)
2023-03-15 20:57:16 Great story in @nytimes on technological risks. Sobering reading for all of us. https://t.co/A2039AD9yW
2023-03-15 18:48:24 RT @etzioni: Commentary: OpenAI’s GPT-4 has some limitations that are fixable — and some that are not https://t.co/Ni2KImcfwj via @GeekWire
2023-03-15 18:36:57 RT @GWilliamThomas: @NateSilver538 "American AI Strategy" (2019) https://t.co/3pRxexxiQQ "National AI Initiative" (2020) https://t.co/5DcZ…
2023-03-15 18:29:30 @Noahpinion Should they have bought a mix of maturities, or would that expose them to other risks?
2023-03-15 15:30:06 @DhruvBatraDB @soumithchintala @OpenAI @MetaAI @Stanford @StabilityAI @huggingface Yes, which makes this a meta signaling tweet, I guess. I ll see myself out now…
2023-03-15 15:28:35 @soumithchintala @OpenAI @MetaAI @Stanford @StabilityAI @huggingface When OpenAI was created it was clear that to the extent they succeeded, their work would create risks of misuse that would undermine the name “Open”. I completely understand their reasons for being careful.
2023-03-15 15:09:41 @soumithchintala @OpenAI @MetaAI @Stanford @StabilityAI @huggingface I don’t think people are “mad”. It’s the hypocrisy of calling yourself “open” that has us choking on our drinks
2023-03-15 03:24:27 @ProfLHunter Yes
2023-03-15 02:17:58 @vardi If we asked GPT-4 to improve it, I wonder what it would do
2023-03-14 21:37:07 @geoffreyhinton Or maybe GPT-4 is the soup of nutrients, but more training is required to make a butterfly that can actually fly
2023-03-14 21:29:17 If the #1 merit of your paper is that it is the first to do X, maybe you should make sure it does X well. In other words, delete "first to show" and see if your paper is still any good. [arXiv subtweet]
2023-03-14 21:16:29 RT @boazbaraktcs: The new scaling laws https://t.co/Rwq2H79oc8
2023-03-14 16:39:17 @emilymbender @SemanticScholar They have been generating these since 2020 (https://t.co/9PQduvPHP8). I agree they should be clearly marked.
2023-03-14 14:20:39 @Arfness @iamharaldur @sindero I was viewing them abstractly. They are both mechanisms for spreading risk. But are they really different? Federal disaster funds are precisely a form of taxpayer-paid insurance. The key difference is that taxes are involuntary, private insurance is voluntary.
2023-03-14 13:57:50 @Arfness @iamharaldur @sindero The "other party" in insurance is the other people who pay premiums to the insurance company. This spreads the risk. Otherwise, you could just put your premiums in a bank account and withdraw them when you experienced a loss.
2023-03-14 04:07:08 @grbradsk @NickKristof Ah the joys (and sorrows) of security theater
2023-03-14 03:32:46 @NickKristof But you do need to know the rules: take every last bit of electronics out of your bags. And even a scissor with rounded edges will be confiscated. Nail clippers will usually cause your bag to be pulled for a double-check.
2023-03-14 02:52:58 @KiltedRef @MichaelDAmbro17 @zeynep The risk to me, as a 68 year old male, is very high. And there is hard data (not "lies") behind that risk assessment. That's why I continue to mask carefully.
2023-03-14 02:50:20 @ArmandDoma Diversity in bond maturity dates would have helped...
2023-03-14 02:47:52 @KiltedRef @MichaelDAmbro17 @zeynep I am glad you are doing well. As I like to say, "Life is an uncontrolled experiment."
2023-03-14 00:51:19 "Mr. Becker of S.V.B. took home $9.9 million in compensation last year, including a $1.5 million bonus [...]. Joseph DePaolo of Signature got $8.6 million. We should claw all of that back, along with bonuses for other executives at these banks. " https://t.co/K7QOXiOjrf
2023-03-14 00:05:55 @MichaelDAmbro17 @zeynep We've seen mask mandates fail. I've been traveling extensively world-wide wearing well-fitting N95s, and I have never gotten Covid19. So that's one anecdotal data point in favor of masking. Have you gotten it?
2023-03-13 23:24:35 @GaryMarcus @KatjaGrace @primalpoly The biggest risk I see is millions of chatGPT-enabled bots bringing down the internet by conducting random experiments on every web site. People, please don't give your learning agent unrestricted internet access!
2023-03-13 23:21:27 @GaryMarcus @KatjaGrace @primalpoly The kinds of risks I worry about were not well-captured in the survey questions. Perhaps other folks had the same reaction?
2023-03-13 21:44:12 @emollick If may approximate a short-term theory of mind, but it obviously will fail when the relevant tokens fall off the end of the context buffer. Without durable episodic memory, one can easily construct ToM tests that it will fail.
2023-03-13 15:24:04 @PGewgle @iamharaldur @sindero But no taxpayer dollars are being spent in the case of SVB. The bank's owners and shareholders are taking huge financial losses. But viewed abstractly, both bailouts and insurance are cases where one person's loss is covered by another group of people (i.e., socialized).
2023-03-13 05:39:28 Here's hoping the libertarians in the valley will listen. https://t.co/BVxNiSPtbn
2023-03-13 05:35:03 @JeffDean @Khipu_AI Likewise! I hope to make the next one.
2023-03-13 04:45:36 @rndmcnlly @docmilanfar @kchonyc See @HastieTrevor et al: https://t.co/EaVqG8eMrN. As soon as you introduce regularization (or Bayesian priors "with teeth"), you start to move away from a purely parametric model. This can make it hard to interpret the fitted parameters
2023-03-13 04:38:57 @rndmcnlly @docmilanfar @kchonyc Not necessarily. DNNs are non-parametric, but they don't store the training data. Instead, they have many more weights than strictly needed ("over-parameterized"), and we combine stochastic gradient descent + regularization to use only some of that representational capacity
2023-03-13 04:14:16 @andrewgwils @kchonyc To (approximately) quote Larry Wasserman: Bayesian questions require Bayesian answers
2023-03-13 04:05:22 @docmilanfar @kchonyc While this gives an example of a good nonparametric method, it doesn’t explain what we mean by nonparametric. The key, in my view, is that the model does not have a fixed number of degrees of freedom. Do you agree?
2023-03-13 01:50:05 @PGewgle @iamharaldur @sindero As I replied elsewhere in this tree, I was using "socialize" in the sense of spreading the risk among multiple people (not necessarily just government). Premiums and tax dollars are both paid in advance, and insurance payouts and govt disaster aid are paid later.
2023-03-13 00:25:45 @yudapearl My pet hypothesis is that when we deliberate, we experience this as an intervention in our own thought processes, and we view our decisions as being caused by those deliberations. We can perform experiments on this to confirm. "Look before you leap", "Think before you act".
2023-03-12 22:51:25 @chrislintott That's what @Katiebouman_AN did in her black hole visualizations.
2023-03-12 22:47:46 @chrislintott Ideally we should treat super-resolution as Bayesian inference based on a prior distribution over the "hidden" content. Then we could clarify our assumptions and quantify our uncertainty.
2023-03-12 22:39:47 @bleepbeepbzzz @iamharaldur @sindero I do agree there can be bad incentives. People should not be able to rebuild houses on flood plains or fire-prone landscapes after being bailed out by the govt. But one of the points of insurance is to enable people to take risks, e.g., on starting a business.
2023-03-12 22:37:32 @bleepbeepbzzz @iamharaldur @sindero I am using "socialize" in the sense of spreading risk across multiple people, "society" in general. Private insurance does this, and so do our tax dollars that we spend on disaster assistance, public health, etc. Both are paid in advance (mostly).
2023-03-12 15:24:11 @Grady_Booch One of the benefits of living in the tropics
2023-03-12 15:19:02 RT @TobyWalsh: https://t.co/yQaHv4A4Xp
2023-03-12 15:02:01 RT @suchisaria: Friends, I’m putting together an invited year in review for conference on health, inference &
2023-03-12 05:02:14 @beenwrekt @drvictoriafox Would you really walk into a Covid-19 ward in a hospital without a well-fitting N95?
2023-03-11 21:12:38 RT @AlanPaulFern1: I've very excited that we will be getting a Digit humanoid robot in our lab any day now. So I'm combing over all the wor…
2023-03-11 20:44:54 @qatarairways being super weird! https://t.co/PApa5f1oQO
2023-03-11 20:43:01 @flypdx this bathroom design violates the principle that the user should be able to see the things they are trying to use. Totally unusable! https://t.co/lbKAGwhROR
2023-03-11 20:20:46 @matloff @iamharaldur @sindero Yes, but there are limits to this individualistic view. When an event threatens the wider society, government needs to step in. There is a moral hazard, and perhaps these owners should be banned from banking. Same with people who rebuild on a floodplain post flood.
2023-03-11 18:59:02 @rasbt @iamharaldur @sindero The big losers here will be the bank’s shareholders. That presumably includes the bank’s managers, who are exactly the people who should be held responsible. There is no bailout afaik
2023-03-11 18:24:55 @GaryMarcus Still waiting for a transcript
2023-03-11 18:23:55 @blu_dechkin @fchollet They can find a buyer who can afford to hold those treasuries to maturity.
2023-03-11 16:54:31 @cdp_physics @iamharaldur @sindero We do this for natural disasters. But I suppose you can consider some part of your taxes as insurance premiums. In any case, it looks like no tax dollars will go into SVB
2023-03-11 16:11:06 @iamharaldur @sindero I get your point. But… Privatizing profit and socializing losses is called “insurance”. Everyone of us who buys insurance does this.
2023-03-11 15:52:48 Another black eye for @Airbnb https://t.co/y5RLtY1TEo
2023-03-11 03:01:06 @Noahpinion Great post!
2023-03-10 23:59:31 @MazMHussain Of course, it would have been even better if the US had not invaded Iraq. That was a violation of the rules-based order that we have tried to hard to create and uphold. Maybe the Syrian disaster would have been averted entirely?
2023-03-10 23:55:48 @MazMHussain The world is interconnected. Suppose the US and Europe had intervened in Syria. This could have prevented the huge inflow of refugees to Europe. That would have resulted in a much weaker Far Right. It might also have discouraged Russian from invading Ukraine.
2023-03-10 18:56:57 @SashaMTL If our goal is to decarbonize AI, then we should be asking questions both about the generic energy cost (regardless of energy source) and strategies for acquiring that energy from low-carbon sources. Looking back historically doesn't directly address either of these questions.
2023-03-10 14:53:04 @oyvinht @Noahpinion Yes, I expect we will see orders of magnitude reductions in energy and data requirements over the next 6-10 years
2023-03-10 04:50:40 @srchvrs @YejinChoinka The Constitutional Learning mechanism (@anthropicAI) imposes a constraint that the outputs of the model should be consistent with the model's own assessment of consistency with its "constitution". end/
2023-03-10 04:48:06 @srchvrs I wonder if you can get better results by imposing a consistency requirement? @YejinChoinka does this for extracting common sense by asking a set of related queries and applying a MAX SAT solver to find the best-supported answer. 1/
2023-03-10 00:53:00 @deliprao You can’t even get a list of the performers or composers on @SpotifyUSA
2023-03-09 23:42:23 @deliprao The listening experience could be so much richer if they integrated music history, analysis, comparison of different performances, etc etc. with the listening experience. Would work for other genres as well, e.g., jazz, flamenco, Indian classical music, Chinese classical, etc.
2023-03-09 22:36:13 @yaroslavvb Safari in Tanzania or Kenya. I loved this place near Mt. Kilimanjaro: https://t.co/SrF4MdPm26
2023-03-09 19:01:41 @conor_muldoon @ylecun What I was trying to say is that by giving the bot a name and a persona, we are creating the expectation that the user is engaged in a conversation with a socially-competent agent.
2023-03-09 18:03:07 @Grady_Booch I think he will regret this thread. People are coming up with thousands of creative ways of using LLMs. I share his skepticism that search is one of them, though.
2023-03-09 15:26:34 @TDeryugina Ah, I see your point. I hate that kind of spin!
2023-03-09 14:56:58 @TDeryugina It seems tautological. If you communicate poorly, you communicate poorly
2023-03-09 14:52:48 @chelseabfinn @hseas @boazbaraktcs @ShamKakade6 Thank you for the very nice talk, and thank you @hseas for sharing it
2023-03-09 04:54:03 @MakingSenseHQ It there a transcript? I started listening, but the data rate in podcasts is too low
2023-03-09 04:39:49 @olcan @sinanaral @GaryMarcus @nytimes @yudapearl I'm not familiar with the task. Pointers?
2023-03-09 03:46:11 @QRDL @yudapearl @sinanaral @GaryMarcus @nytimes Indeed, we often learn a trick (e.g., how to open a jar) before we understand the causal model behind it. But the great thing about causal models is that we can bring them to bear to reason about situations we have never seen "during training". end/
2023-03-09 03:44:36 @QRDL @yudapearl @sinanaral @GaryMarcus @nytimes Of course we also memorize the consequences of causal reasoning to create the automatic routines and inference paths that underlie our subconscious, automatic performance in both the physical and social worlds. 3/
2023-03-09 03:43:01 @QRDL @yudapearl @sinanaral @GaryMarcus @nytimes During university, we are taught causal models in virtually all branches of knowledge, especially science and social science. Causal reasoning draws on this causal knowledge to explain, predict, and control the world around us. 2/
2023-03-09 03:40:59 @QRDL @yudapearl @sinanaral @GaryMarcus @nytimes We usually draw a distinction between causal discovery and causal reasoning. Infants do quite a lot of causal discovery about the physical world
2023-03-09 01:09:37 Nice discussion in this article. The whole concept of a chat bot seems broken. We expect a "bot" to be an agent with the kinds of social intelligence described in this article. This is very different from how we treat, e.g., google translate or google search. https://t.co/E8YfkiH1SR
2023-03-08 21:35:41 @LoneDruid_1 @sinanaral @GaryMarcus @nytimes @yudapearl Interesting. Moral intelligence could also include persuading people to behave well, building consensus and coalitions. So I think it can be creative too. I agree with Chomsky that a chatbot needs social intelligence (including moral thinking)
2023-03-08 20:42:16 @LoneDruid_1 @sinanaral @GaryMarcus @nytimes @yudapearl My understanding is that they were very concerned about the moral implications (quite appropriately). Note that they used computers to help with the design, but those computers did not need to do moral reasoning.
2023-03-08 20:15:22 @Noahpinion AI won't be economically practical without an energy revolution
2023-03-08 20:11:59 RT @kaggle: New ML competition is up! The 3rd annual BirdClef competition w/ @CornellBirds, @Google, &
2023-03-08 20:11:09 RT @tdietterich: @sinanaral @GaryMarcus @nytimes Nice to see such dispassionate writing from Chomsky. AI systems and "their ilk", "lumberin…
2023-03-08 20:10:53 @sinanaral @GaryMarcus @nytimes @yudapearl Hence, I agree with Chomsky that current LLMs lack this component of social intelligence. But I think that "True intelligence is also capable of moral thinking" is an overstatement. end/
2023-03-08 20:08:57 @sinanaral @GaryMarcus @nytimes @yudapearl However, a system that interacts with people in ways that concern their role as members of society cannot be considered socially intelligent unless it can reason about the moral dimension of its actions. 4/
2023-03-08 20:06:58 @sinanaral @GaryMarcus @nytimes @yudapearl But I don't see why a system cannot be "intelligent" without being able to perform moral thinking. One form of intelligence is to prove theorems from a set of assumptions, and that does not involve moral thinking. Mastering Go did not require moral thinking. 3/
2023-03-08 20:05:25 @sinanaral @GaryMarcus @nytimes I only see two substantive arguments: lack of causal reasoning and lack of moral thinking. The first is correct: @yudapearl has proved that statistical ML without causal assumptions cannot draw causal inferences. 2/
2023-03-08 20:01:46 @sinanaral @GaryMarcus @nytimes Nice to see such dispassionate writing from Chomsky. AI systems and "their ilk", "lumbering", "conventional", "mere probabilities" whereas grammar "almost mathematical elegance". What are the core arguments when you strip away the rhetoric? 1/
2023-03-08 15:20:28 @emilymbender I wonder if that same doctor types queries about his own medical symptoms into a search engine (without first using a VPN + incognito browsing)? We've made it so easy to leak information
2023-03-08 04:35:45 RT @KieranSnyder: There may be nothing I’ve seen wreck the careers of high-performing, hardworking people more commonly than stepping into…
2023-03-08 04:26:26 @monsoon0 SYD?
2023-03-07 22:45:36 @MiTiBennett Thanks
2023-03-07 20:32:04 RT @AaronHertzmann: Spoiler: the real problem is unregulated Profit Maximizers, which have been around for a few decades at least
2023-03-07 19:14:12 @J_K_Chesterton @jlindy77 @wanyeburkett Fascinating
2023-03-07 06:00:53 @zacharylipton @Uber @lyft It must vary hugely depending on the city. Pittsburgh?
2023-03-07 05:13:20 @GaryMarcus @JSEllenberg Gary, what is your operational definition of a representation? PaLM-E sure behaves as if it has an internal world model
2023-03-06 20:16:09 @VeronicaSaron @Neeva I don't want search, I want assistance completing desktop workflows. Every time I search and every time I do copy/paste, this reveals a failure of the UI to understand what I'm doing and give me a one-click way to do it. I'd love to see an AI-powered workflow assistant
2023-03-06 04:38:57 @kchonyc What airline?
2023-03-06 01:46:48 @kchonyc Also on the subway in DC. Maybe their earbuds need recharging?
2023-03-05 23:55:09 @amcafee @superwuster *dime, but maybe nickel-and-time is accurate too :-)
2023-03-05 23:54:20 @amcafee @superwuster The latter model has much less transaction overhead and results in a much more pleasant user experience. The former leads to nickel-and-time (or, today, $5-$10 charges) for everything.
2023-03-05 23:52:49 @amcafee @superwuster One view of services is that each service should be supported only by the individuals it benefits. The other is that we consider the services jointly so that every individual benefits from the same fraction of such services and we fund them jointly. 1/
2023-03-05 22:24:35 @davidmanheim @GaryMarcus @jacyanthis @rgblong @lukeprog @yudapearl We model ourselves as agents who choose actions determined by our knowledge, goals, and beliefs. Our belief in free will is based on a causal model of our own behavior. end/
2023-03-05 22:22:36 @davidmanheim @GaryMarcus @jacyanthis @rgblong @lukeprog @yudapearl My naive understanding of "theory of mind" in dialogue is that we maintain a mental model of the knowledge, goals, and beliefs of our dialogue counterparts. These account for their utterances as actions taken by agents to achieve goals. 1/
2023-03-05 20:33:44 @sir_deenicus @sarahookr I agree that regulating uses is much better than regulating technology. The challenge is to discover and limit the bad uses while advancing the technology and making it accessible to everyone for good uses. What are your thoughts about how to achieve that? @sir_deenicus?
2023-03-05 19:20:22 @jacyanthis @rgblong @lukeprog @davidmanheim @GaryMarcus For applications where we can live with a noisy approximation, statistical learning methods suffice. But the word "solve" would seem to require extreme reliability, which I do not believe statistical learning alone can achieve.
2023-03-05 19:16:39 @sarahookr @sir_deenicus I think we need to develop and adopt such guidelines promptly. We don't want our field to contribute another notorious incident to future classes on the Responsible Conduct of Research.
2023-03-05 19:12:39 RT @sarahookr: We need more nuanced discussions around the risk of open sourcing models. Open source brings valuable access, but it is ab…
2023-03-05 18:21:15 @burkov Is it really a killer app for search? Im not convinced yet
2023-03-05 18:07:31 @rgblong @lukeprog @davidmanheim We have methods that can do all of these things, but they are fragile and unpredictable. They demonstrate a lack of systematic understanding. I suspect this is because statistics alone can’t give us causal understanding
2023-03-05 18:04:11 @amcafee @superwuster How individualistic! Isn’t there some social value in welcoming international visitors to the US with generosity?
2023-03-05 16:22:50 @lukeprog @davidmanheim Still not solved
2023-03-05 16:20:29 @amcafee @superwuster Even if it is, it is really rude to arriving passengers. I’d rather include the cost in the facilities fees, like we do with the toilets, floors, walls, and roof
2023-03-05 15:59:34 @MatthewJBar Do you think LLMs can also help formulate interesting conjectures?
2023-03-05 15:58:50 @MatthewJBar I think it was R W Hamming who said “Our theorems are correct but the proofs are wrong”. But the reality is that we must debug theorems and proofs against each other. LLM-based tools will really help with this.
2023-03-05 15:49:12 @unsorsodicorda @giffmana @roydanroy I had in mind, as Dan Roy was suggesting, creating an ensemble instead of selecting one model. Obviously not always computationally feasible…
2023-03-05 10:00:00 CAFIAC FIX
2023-03-02 22:00:00 CAFIAC FIX
2023-02-27 01:00:00 CAFIAC FIX
2023-02-20 15:18:45 @GaryMarcus Check out Cohen &
2023-02-20 15:17:08 @GaryMarcus The fundamental difficulty of ToM is that *we* researchers need to find a good representation for encoding the beliefs of the system about the beliefs of other agents. I haven't seen any evidence that such representations are learned by LLMs.
2023-02-20 04:23:53 @fixedpointfae Basic, Fortran, Assembly, COBOL, Pascal, C, Lisp, Prolog, BCPL, S/R, C++, Python, Pytorch. I’m probably forgetting something…
2023-02-20 02:37:15 @hackylawyER @ledell I don’t see email as very promising, however. Maybe saves a few keystrokes, but I don’t have to Google the pytorch documentation to answer routine email (unlike coding, where I probably do one search per line of code).
2023-02-20 02:34:36 @hackylawyER @ledell Programming tools like github copilot have produced substantial productivity improvements. I’m sure there will be other successes (and failures).
2023-02-19 21:49:51 I am very grateful to FAIR for leading the way and supporting the world-wide deep learning research community. https://t.co/cX5onibrIj
2023-02-19 20:38:58 RT @zicokolter: Generative models and P vs. NP: A clickbaity thread An important point that seems missing (as far as I've seen) in the d…
2023-02-18 16:09:20 @sirkodnap @Grady_Booch A repeated lesson in security is that you need to develop methods that are (a) secure and (b) highly usable and then you need to mandate their use. I know I hate to be told what to do, but this is a case where I need the push
2023-02-18 16:07:13 @sirkodnap @Grady_Booch Fair enough
2023-02-18 15:28:14 @JohnHMcWhorter I would like to learn more about what restorative policies have been tried and how well they succeeded. A good class topic would ask how to measure the success of such policies. How to define restorative justice?
2023-02-18 08:47:10 @lacisoft @Grady_Booch But why defend SMS 2FA when it is no longer the best technology for anyone with a smartphone?
2023-02-18 08:46:11 @lacisoft @Grady_Booch Yes, roaming in EU is wonderful. Would love to see similar agreements worldwide.
2023-02-18 07:24:26 @Grady_Booch So you are saying that companies should ignore the needs of a subset of their customers? Furthermore, SMS 2FA has vulnerabilities: https://t.co/FSvck9LoZY
2023-02-18 06:45:07 RT @etzioni: Well, at least your acquisition of Twitter is going as planned. https://t.co/vQ2Ci4FFgy
2023-02-18 06:43:15 @mosheroperandi @Grady_Booch Verizon’s international plans are a total ripoff. I buy a local SIM card instead.
2023-02-18 06:22:38 @mosheroperandi @Grady_Booch Yes, I noticed that for Twitter when I read their announcement. But a large number of companies only support SMS.
2023-02-18 03:33:20 @Grady_Booch SMS 2FA should be deprecated. It doesn't work when you are traveling internationally unless you also have an (expensive) international roaming plan for your phone. At a minimum, they should support WhatsApp, but I believe authentication apps are better.
2023-02-17 09:30:01 @rgblong Maybe step-by-step reasoning gone bad?
2023-02-16 04:23:12 @erichorvitz @yudapearl @aperianez Yes. In the original tweet, they wrote “foundational”, and that made me flinch.
2023-02-16 04:20:44 RT @ESYudkowsky: The thing about the Future is that it's made up of the same people and same sort of people who are implementing the presen…
2023-02-16 04:16:43 @fchollet A tool should show its capabilities through interpretable affordances.
2023-02-16 04:06:38 @yudapearl @aperianez @erichorvitz Agreed. The term “model” has always bothered me. It would be nice to have a shorter term than “fitted function”.
2023-02-16 04:02:26 RT @Noahpinion: Good (long) post about the cultural problems within Google: https://t.co/VoP668q4aA
2023-02-16 03:19:29 @aperianez @erichorvitz “Foundation”. There is nothing foundational about these models. In fact, we should probably call them “marginal models”, because they just give the marginal distribution of language after integrating out meaning, intent, etc.
2023-02-15 04:14:32 @rasbt @KyleCranmer Or any optimizer. Optimal solutions can be weird!
2023-02-14 17:17:50 @ross_mc_donnell @BallouxFrancois The (male) public was highly resistant to using seatbelts. Making seatbelts a legal requirement made a big difference in compliance. Same with motorcycle helmets. The objection to vax requirements (e.g., in military) is the same old story, but now it risks troop readiness.
2023-02-13 09:31:26 An excellent discussion of the economics of improving worker productivity with AI by analogy with machine tools. I always learn a lot from @Noahpinion's economics posts https://t.co/SMcTYGX0QJ
2023-02-13 09:29:33 @Bee_Y_s @Noahpinion @ElijahRavitz @Google The answers are currently unreliable, and it seems very challenging to fix this. Perhaps the way forward is to use interactive chat to refine the query but then traditional retrieval to produce reliable answers? Meanwhile, I expect major errors from rushed deployments
2023-02-13 08:07:13 “I’m not even 30, and I’ve already done the most significant thing I’ll ever do.” https://t.co/0MfIYWwtmO
2023-02-13 05:34:26 @Noahpinion @ElijahRavitz If (when?) chat-based search turns out to be a bust, @google may turn out to be the winner
2023-02-12 16:10:57 RT @tdietterich: @beenwrekt @zacharylipton That’s a great article! When I read Superintelligence, I had a similar reaction. The book made p…
2023-02-12 16:10:52 @beenwrekt @zacharylipton That’s a great article! When I read Superintelligence, I had a similar reaction. The book made perfect sense if you replaced the word “superintelligence” by the word “capitalism”. But I love how Chiang depicts this as a self-model of the tech industry! Thanks for posting it
2023-02-12 00:41:44 @jackclarkSF These are all ways of computing the transitive closure of existing knowledge as opposed to acquiring new knowledge. How much can be gained by combining what we already know?
2023-02-10 10:38:07 Congratulations to my colleague and collaborator, Kiri Wagstaff, for this recognition of her contributions to ML research, applications (esp in space exploration), and to AI education. https://t.co/1VlDJQrSAA
2023-02-10 08:12:43 RT @Noahpinion: Wow, an essay about "neoliberalism" that is actually interesting, novel, and useful! It must be American, not British. ht…
2023-02-08 05:13:04 RT @SarahKachovich: This is how the global ocean should be seen, with #Antarctic &
2023-02-07 03:59:13 @Noahpinion very good + very bad = neutral
2023-02-07 03:58:11 @rgblong ChatGPT’s answer would not be based on self knowledge (as I understand it)
2023-02-06 13:31:08 @americakaran Which states do you think will be most/least successful?
2023-02-05 12:15:49 @roydanroy The ML conferences have been using the Toronto paper matching system and I believe it has many shortcomings. But can we use the latest LLMs to build better expertise models? @andrewmccallum ?
2023-02-05 08:24:17 @roydanroy I agree. With our own tech, we should be able to assign reviewers automatically
2023-02-04 15:50:29 @RyanMcbeth @PhillipsPOBrien I am wrong
2023-02-04 13:36:39 @RyanMcbeth @PhillipsPOBrien Maybe this was just an editing error?
2023-02-04 13:16:02 @RyanMcbeth @PhillipsPOBrien This seemed pretty nonsensical. Spy satellites are not geosynchronous but low earth orbit.
2023-02-02 15:34:19 RT @aihuborg: Science communication for AI researchers: our tutorial at #AAAI2023 - https://t.co/mRCwTK0aMG https://t.co/gg0AtqV5Zn
2023-02-02 13:38:04 @HaydnBelfield This assumes that we continue down the brute force trajectory. I predict that within a decade, we will have figured out how to replace LLMs with a constellation of well-understood, efficient, engineered components.
2023-02-02 01:30:45 @BallouxFrancois Block and ignore
2023-02-01 10:24:02 @ridelfin @cloudquistador @Grady_Booch Yes, and there will be a need to train/fine-tune/augment large pretrained models with private data. Those models will need to be run on-premises (or in private clouds) to avoid leaking information.
2023-02-01 10:10:41 @ridelfin @cloudquistador @Grady_Booch Yes: search engines and social media. I'm not sure about ecommerce sites like Amazon or paid services like Office 365.
2023-02-01 09:53:37 @ridelfin @cloudquistador @Grady_Booch There are several costs: capital expense, operating expense of the data centers, and latency cost to the user. My understanding is that for most applications these costs are too expensive for ad-supported products
2023-02-01 09:16:07 @cloudquistador @Grady_Booch They won't deploy these models unless it makes business sense. My impression is that the models are too expensive today, but people are optimistic that huge cost reductions are possible.
2023-02-01 09:09:53 @Grady_Booch But will end users be able to meet their end-user software engineering needs without writing in DSLs like Excel? LLMs might help the computer understand the user's context and goals better and prevent many end-user programming errors
2023-02-01 07:01:48 @LauraALibby @daniela_witten Citation needed.
2023-01-31 12:04:41 RT @togelius: Saying "an AI" is definitely a red flag, but an even bigger red flag these days is arguably talking as if ChatGPT was all of…
2023-01-30 01:00:00 CAFIAC FIX
2023-01-17 00:59:59 RT @vardi: https://t.co/OeTJpKOZ0C?
2023-01-16 22:59:54 @rcalo Why should I bother to tweet anything when @rcalo has already tweeted it?
2023-01-16 22:50:59 I'm excited to see this new book. I was just rereading @martypute's wonderful 1994 book. It has long been an RL Bible https://t.co/LYcz4Ur591
2023-01-16 21:52:34 @paul_scharre @jackclarkSF @nmaslej @Miles_Brundage @jjding99 @mattsheehan88 @ylecun @indexingai @etzioni @CSETGeorgetown @DataInnovation @Noahpinion Did the report discuss papers with both Chinese and non-Chinese authors? These defy the US-vs-China narrative. It seems that there is still a healthy exchange of ideas across the global research community. That said, at least on the corporate side, secrecy seems to be growing
2023-01-15 21:55:55 @LechMazur @mezaoptimizer Yes, and now it has become a marketing term itself. I guess this was inevitable.
2023-01-15 19:47:05 @grbradsk @mezaoptimizer Yes. If I understand correctly, causality theory shows that, while you can learn a lot from observational data + causal assumptions, you can learn even more from interventional data @yudapearl @eliasbareinboim
2023-01-15 19:41:24 @PaulTopping @mezaoptimizer Very true
2023-01-15 08:54:10 @mezaoptimizer If it turns out to be easy, the joke will be on us.
2023-01-15 08:53:26 @mezaoptimizer Old time AI people like me have always been working on AGI. We just called it AI. To me AGI is just a marketing term. We old timers think we know how hard the problem is, and we are lolling as the younger generation discovers this
2023-01-15 08:48:25 @Noahpinion It is still a good place to post something where other people can find it. But it could be so much better
2023-01-15 01:33:36 @mark_riedl No, and not for any of these reasons but because it is not a moral agent who can be held accountable for errors in the paper.
2023-01-14 20:33:14 @rasbt @ljbuturovic @MuhammadAnas707 This same reasoning may apply to XGBoost, gbm, and so on. I haven't looked to see what their margin distributions are like. See also "Large Margin Deep Networks for Classification" https://t.co/bjyAvQHaFc
2023-01-14 20:30:27 @rasbt @ljbuturovic @MuhammadAnas707 In the soft margin case, you only get guarantees on the points that are outside the margin region. That said, I can imagine a system that says "Here is the predicted class, but it is not a robust answer" for points that don't have the desired margin.
2023-01-14 19:52:00 @rasbt @ljbuturovic @MuhammadAnas707 I think it depends on your evaluation criteria. For in-distribution accuracy, I suspect you are correct in general Sebastian. But for various forms of out-of-distribution performance, I think the max margin and margin distribution ideas have much to teach us about robustness.
2023-01-14 06:20:07 @RWerpachowski It seems to have worked (inadvertently) with abortion, but of course that didn't trash the economy or abandon Ukraine and Taiwan.
2023-01-14 06:05:00 Should the Democrats go ahead and cut everything the Republicans want to cut but make sure the cuts go into effect immediately? Then use the resulting outrage to soundly defeat the Republicans in the next election.
2023-01-14 05:45:13 House Republicans prepare their Thelma and Louise plan to drive the US off a cliff. Why do they want to damage America so much? https://t.co/mWdsIWuEHM
2023-01-14 04:58:48 @mark_riedl Chronological has now disappeared from my web client. I only have a choice of "Following" or "For You".
2023-01-14 04:49:45 @rao2z @edward_the6 I'm often perplexed by your sentences
2023-01-14 03:57:51 @paul_rietschka @rasbt @MuhammadAnas707 I'd be interested in citations to robustness guarantees for XGBoost.
2023-01-14 03:57:10 @rasbt @MuhammadAnas707 I know that Adaboost tends to produce a large margin classifier. Does gradient boosting do this? The SVM solution may not perform as well in-distribution, but it is guaranteed to be robust to changes in features and in class frequencies.
2023-01-14 02:41:27 @deliprao @SashaMTL Computer vision challenge: measure petal and stamen lengths so we can use Fisher’s linear discriminant for classification. Reduce to a previously solved problem:-)
2023-01-14 02:37:51 @rasbt @MuhammadAnas707 SVMs find very robust solutions. I wouldn’t say they are outdated at all.
2023-01-14 01:53:13 @Noahpinion Wow, you are really trolling us today, Noah. Reagan permitted laissez faire capitalism to run rampant, and that eventually damaged the fabric of society
2023-01-13 03:53:11 @mark_riedl The browser version HTML title is now "Following", so that's consistent
2023-01-13 03:42:51 @mark_riedl Chronological is still working for me in the browser
2023-01-12 02:05:09 @Poyonoz @zdeborova I’m not sure we need gatekeepers. In my experience, peer review improves my papers but my rejected papers did not contain errors and could have been safely published
2023-01-11 22:34:38 @Noahpinion *mRNA vaccines
2023-01-11 22:20:41 @Noahpinion I don't think web surveillance contributed to the development of mRNA viruses or solar power or battery tech, @Noahpinion, which is what you've mentioned in the past when talking about tech saving the world.
2023-01-11 19:30:01 RT @VC31415: @tdietterich @AOC Before I say anything silly, I wanted to mention that @ProfEmilyOster has a nice write up on the issue. htt…
2023-01-11 19:29:21 @VC31415 @AOC @ProfEmilyOster This is very helpful, thank you!
2023-01-11 18:21:49 @VC31415 @AOC So you think reduced cognitive performance causes people to buy gas stoves? Or that there is a confounder?
2023-01-11 05:15:07 @miguelisolano No. Solve factuality and you solve factuality. Why would you believe this is the only remaining roadblock?
2023-01-10 22:54:11 @rasbt Can hardly wait to have an LLM system that can do this
2023-01-10 21:57:49 @roxannedarling @DavidDeutschOxf As an example: The author suggests that we stop learning when we perform a task automatically. But a famous study of cigar rollers in Cuba showed that workers continue to improve even after decades of practice. In short: There is evidence that bears on this discussion!
2023-01-10 21:17:00 @DavidDeutschOxf I didn't see a single reference to relevant research on these questions. This is just an exercise in good old introspection--a notoriously unreliable source of evidence.
2023-01-10 19:59:05 @zdeborova Absolutely ridiculous. This should not be permitted. Let's stop thinking that acceptance at a "top" ML conference means anything about the quality of the work.
2023-01-10 19:48:23 @y0b1byte @shimon8282 I learned a lot from your dissertation
2023-01-10 16:20:58 RT @zeynep: It’s true Scott is on Pfizer’s board and ex head of FDA. I was *very* wary. I didn’t expect to find him to be so informative a…
2023-01-10 16:13:44 @BenKobren @MaxWinebach @Neeva @Google In my queries, I wasn't satisfied with the non-AI results. The queries were about taxi services in India
2023-01-10 04:16:27 @rao2z Maybe you can get the classroom declared to be a SCIF (Secure Facility).
2023-01-10 03:12:51 @omarsar0 Very nice
2023-01-10 01:46:02 @matloff @kareem_carr Love the Y axis position of the feet
2023-01-10 01:13:27 @jwoodgett @mattsclancy I deeply and sincerely agree!!
2023-01-10 01:12:49 @jwoodgett @mattsclancy But papers are not sampled iid but are written relative to a set of competitors. In my experience, when the field is small, all 100 papers are mediocre and would not rank in the above 10% after the field has grown.
2023-01-10 01:02:38 @jwoodgett @mattsclancy It depends on your model of competition. When a field has 100 people in it, you only need to write a paper better than 99 other papers. When it has 10,000 people in it, you have to write a paper better than 9,900 other papers (for a top 1% paper)
2023-01-09 22:19:04 @Raamana_ @GaryMarcus Additional progress: * ML theory to understand double descent * self-supervised learning (which enabled LLMs) Biggest challenges * continual learning (esp for LLMs) * representation learning that can generalize out-of-distribution
2023-01-09 22:16:17 @Raamana_ @GaryMarcus In #MachineLearning, * improvements in uncertainty quantification using conformal prediction * deeper understanding of out-of-distribution detection * methods for responding to distribution shifts * progress in causal ML
2023-01-09 22:12:36 Is anyone developing multi-modal LLMs that model and manipulate user interfaces of existing systems? Example: MS Word, MS Excel, Photoshop, etc. Seems like this would require understanding language, UI actions (clicks, typing), and screen contents.
2023-01-09 21:02:22 @MaxWinebach @Neeva I tried it for some tourism queries and it was a total fail. I suspect it is indexing less than @google.
2023-01-09 19:48:16 @mattsclancy Another hypothesis is that the internet has increased connectivity among researchers and this leads to less diversity in the problems being attacked (everyone works on the most popular/fashionable problems) which leads to overall less exploration of the innovation space
2023-01-09 19:45:44 @mattsclancy The increase in the size of a field seems like it would explain a lot of these metrics. E.g., it is obvious that with more people in a field, it would be harder to write a top 0.1% paper. More people would work on the same problems, so diversity of words drops. etc.
2023-01-09 16:38:23 @YiMaTweets I don’t allow my students to submit a paper unless I believe it is ready for publication. Otherwise you are wasting reviewer’s time
2023-01-09 01:18:05 @GaryMarcus ChatGPT seems to adopt the presuppositions of the prompt and tries to find an answer consistent with them. I would have expected the training data to contain arguments where the presuppositions were challenged. Another example of the difficulty of negation/contradiction?
2023-01-09 01:14:30 @adawan919 @NikkeiAsia 1. What determines the risk of long covid? 2. Are there medications (yet to be discovered) that can prevent long covid? 3. Do we have therapies for recovery from long covid?
2023-01-09 00:52:34 @NikkeiAsia That won't happen for many of us until long covid is understood and the risk reduced
2023-01-08 23:05:19 @adawan919 @ylecun I'm not arguing that teaching grammar rules is a good way to learn second languages. It is just a method that I personally find useful. I know nothing about the pedagogy of language learning. I just wanted an example of skill learning that involved pure practice.
2023-01-08 22:41:38 @adawan919 @ylecun You are arguing against a position that I do not hold.
2023-01-08 21:14:10 @adawan919 @ylecun I don't view this as an 'ethical' question
2023-01-08 21:12:41 @adawan919 @ylecun Yes, but they present a unified theory of chunking as a form of knowledge compilation that applies to any skill learning task.
2023-01-08 21:11:52 @adawan919 @ylecun I do a lot of language learning (as an adult, learning second languages). I realize this is different from child language acquisition!! As an adult, it is helpful to learn the conjugation rules (e.g., in Spanish, Hindi, etc.) and then practice speaking until you can forget them.
2023-01-08 19:21:49 @RitaKonaev Best tweet of 2023 so far
2023-01-08 18:40:27 @loretoparisi What do you mean by “already incorporated”?
2023-01-08 18:35:34 @pwlot @davidmanheim Strongly disagree. Successful prediction is one outcome of a intelligence, but other forms include rapid learning by asking the right questions
2023-01-08 18:31:56 @davidmanheim We may not need necessary and sufficient conditions for AGI, but we do need measurement methodologies and metrics.
2023-01-08 07:11:04 @Noahpinion I'm still seeing good stuff on Ukraine and COVID + your posts. Fewer posts from journalists, I think.
2023-01-07 19:46:33 @ylecun I agree. But practice can serve a knowledge compilation function that does not involve acquiring new knowledge. Example: By practicing conjugating verbs, I convert the conjugation rule into "muscle memory". Rosenbloom&
2023-01-07 07:21:49 Interesting insights into the dynamics of SARS-CoV-2 variants (and a lesson for ML folks) https://t.co/ZkKLOV8BLt
2023-01-07 06:58:43 RT @pierre_azoulay: OK. I have spent much of the week finally sinking my teeth into this paper. It's really well written and convincing. Bu…
2023-01-07 06:54:23 RT @luisceze: It is funny to see people on one end using ChatGPT to expand bullets into longer text prose to then send it to other people,…
2023-01-07 01:40:21 @ylecun Thank you CIFAR for placing a bet on this area!!
2023-01-07 01:38:52 @Noahpinion I'm not saying you have to plot zero, but just pointing out that for all of these questions, more than half of the 12th graders have answered YES. So a majority are still taking risks.
2023-01-07 01:24:50 @Noahpinion Check out where 0 is on the vertical axis. These charts exaggerate the effect, real as it may be
2023-01-06 20:04:54 @EvelinaLeivada @analisereal @GaryMarcus Then we need to know how these scores would compare to the scores assigned to " 'Colorless green ideas sleep furiously' is not grammatical " and to other sentences like "Colorless ideas green sleep furiously", etc. Interesting questions... end/
2023-01-06 20:01:36 @EvelinaLeivada @analisereal @GaryMarcus I'm not sure this addresses my concern. ChatGPT will assign one internal score to the sentence " 'Colorless green ideas sleep furiously' is grammatical " and another score to the sentence "Colorless green ideas sleep furiously". 1/
2023-01-06 18:37:13 @EvelinaLeivada @analisereal @GaryMarcus When you ask ChatCPT whether it thinks a sentence is grammatical, the answer it gives is its summary of what the utterances in the training corpus would say about the sentence. This is one-step removed from its internal assessment. Like its other answers, it can be a fabrication
2023-01-06 18:35:18 @EvelinaLeivada @analisereal @GaryMarcus As I understand LLMs, they are sampling from a distribution over continuations of the prompt much like n-gram models. The product of these one-step probabilities would give a score for the probability of a sentence.
2023-01-06 18:33:03 @emilymbender @mmitchell_ai @IrisVanRooij If ChatGPT is automated plagiarism, then so is human language learning. I'm currently studying Hindi, and this involves exact memorization of phrases. There is a difference in scale, obviously.
2023-01-06 18:02:54 RT @betanalpha: Anyways, to end on a more productive note let me mention my personal strategy for prior modeling of one-dimensional paramet…
2023-01-06 17:48:35 @EvelinaLeivada @analisereal @GaryMarcus Are you measuring its internal scores on sentences? I would never trust its natural language statements
2023-01-06 05:34:41 @sir_deenicus @csabaveres @rgblong @GaryMarcus @EvelinaLeivada And maybe this is enough for natural languages? What’s known about this?
2023-01-06 05:33:09 @sir_deenicus @csabaveres @rgblong @GaryMarcus @EvelinaLeivada Yes, I was being sloppy when I said “context sensitive”. It is of course impossible to generate all context free and context sensitive languages without a stack. But in an informal sense, attention over the huge input buffer lets transformers consider huge context
2023-01-06 04:06:07 @analisereal @GaryMarcus It can tell us about the grammar it read about, but that is not the “grammar” that produces its performance, of course
2023-01-06 00:56:53 @Noahpinion Marie Curie
2023-01-06 00:19:12 @GaryMarcus @s_r_constantin As the AI field attempts to build integrated systems, it is natural that it will spread into terrain that was previously explored by other fields. When Machine Learning discovered Statistics (around 1985?), we had to go educate ourselves. Time to (re)read these classics!
2023-01-06 00:16:24 @j_q_balter @GaryMarcus There was the small matter of massive statistical learning involved...I don't know about you, but I couldn't hand code these billions of weights
2023-01-05 23:29:43 @csabaveres @rgblong @GaryMarcus @EvelinaLeivada Yes, this is an interesting line of research to explore.
2023-01-05 23:28:32 @mazuretsky @IDoTheThinking Depends on how you measure "land". If you measure soil, we are in trouble. https://t.co/c8pcZVABfQ
2023-01-05 22:19:00 @csabaveres @rgblong @GaryMarcus @EvelinaLeivada Obviously it is "cheating" in some way, and that might be interesting.
2023-01-05 22:18:15 @csabaveres @rgblong @GaryMarcus @EvelinaLeivada For me as a computer scientist, it is interesting that it can produce context-sensitive language without a pushdown stack. Chomsky showed us that there are important relationships between the implementation and the grammar.
2023-01-05 22:16:20 @csabaveres @rgblong @GaryMarcus @EvelinaLeivada The question at hand is whether LLMs give us a new framework that might give us additional insights beyond those of Dr. Chomsky. I don't think the answer is clearly YES or NO. I do think it is worth studying.
2023-01-05 22:13:35 @csabaveres @rgblong @GaryMarcus @EvelinaLeivada How do you know? It will take some time for the research community to develop the tools for understanding these models. This is an active area of research. I like your idea of looking at LLMs for formal languages, because we know their grammars.
2023-01-05 22:09:53 @csabaveres @rgblong @GaryMarcus @EvelinaLeivada If people could explain how we understand language, there would be no mysteries for linguists to study. At least with an LLM, we can examine the weights and activations and their patterns. We can intervene and ablate, etc. to gain a causal understanding of the LLM.
2023-01-05 21:30:03 @csabaveres @rgblong @GaryMarcus @EvelinaLeivada What is the point of this line of argument? Python is much simpler than human natural language. That said, like natural language, there are long-distance dependencies in Python, and I'm impressed that these models can handle them well.
2023-01-05 19:03:42 @GaryMarcus @EvelinaLeivada As for semantics, I think LLMs show that distributional semantics is broken. That's a negative finding, but I think it is pretty interesting.
2023-01-05 19:02:02 @GaryMarcus @EvelinaLeivada Of course I agree that there are many more questions about language than just fluent production. But if we want to study fluent production (something of interest to Dr. Chomsky in the past), then I think we might learn a thing or two from studying LLMs.
2023-01-05 18:56:58 @EvelinaLeivada @GaryMarcus I conjecture that we can prune or distill transformer models to make them much more compact than the extremely high-order Markov models of equal performance.
2023-01-05 18:54:44 @EvelinaLeivada @GaryMarcus Obviously they learn language -- they are constructed by learning. They can generate novel sentences combinatorially at least as well as any human-written grammar (and certainly better than a voice recorder). I would bet they can assess grammaticality better too.
2023-01-05 18:39:39 @GaryMarcus They tell us that fluent language can be generated by models like these. Hence, these models are worth studying to understand how they do this. I wouldn’t expect anyone to have deep answers yet.
2023-01-05 02:17:33 I like this clarification. @icmlconf https://t.co/PXa0Na91At
2023-01-04 22:58:08 @emilymbender If we hold constant the capabilities of the human organism, then we can replace "evolution" with "optimization" to side-step the issues of teleology in evolution. This concludes my interpretation of @ylecun's tweet.
2023-01-04 22:56:37 @emilymbender Similar considerations should apply to the meanings of words. These should be well-separated in some conceptual space so that language is interpreted reliably. This of course reflects an information-theoretic view of language, which is impoverished, but perhaps still useful.
2023-01-04 22:54:04 @emilymbender Signal detection theory suggests that different signals need to be reliably distinguished by the sensor. So I interpret this as suggesting that phonemes should be reliably distinguished by the human auditory system.
2023-01-04 21:32:53 @Noahpinion Important application for AR glasses: round all prices up to the next dollar value :-)
2023-01-04 21:26:53 @ToddTheLinguist @mmitchell_ai You can't get anything interesting out of an LLM without giving it a rich input prompt such as a list of points to include in the abstract, etc. The LLM could also be applied to write short summaries of previous work. But of course ALL generated text will need checking
2023-01-04 20:14:07 @mmitchell_ai Formally, it seems the same in the sense that even for simple autocorrect, I need to check the correctness of the result. However, with LLMs, that checking process is much more difficult, and I think we need additional tools to help. Does this match your thinking @mmitchell_ai?
2023-01-03 23:08:28 @stolyassa @ylecun @Raamana_ This is why it would be so important for future LLMs to be able to cite their sources. But it also suggests that we could use LLMs to summarize previously published work in ways that could permit detection of "semantic plagiarism" and not just literal text copying.
2023-01-03 23:06:56 @stolyassa @ylecun @Raamana_ If ChatGPT just regurgitates someone else's work, that would be plagiarism. So the big risk to an author using ChatGPT is to ensure that ChatGPT isn't plagiarizing. That will require searching the prior work. Maybe a plagiarism detector can help?
2023-01-03 23:04:59 @stolyassa @ylecun @Raamana_ The paper will need to make a research contribution. That could result either from ChatGPT combining existing work in a novel way (by drawing inferences) or from the human authors including novelty in the prompt and asking ChatGPT to write it out fluently 1/
2023-01-03 21:40:25 @Raamana_ @ylecun Why are you so eager to relinquish your humanity to an algorithm? You do realize that this leads to treating people as "just algorithms", right? That doesn't strike me as a good basis for a thriving society.
2023-01-03 21:37:48 @Raamana_ @ylecun Well, those criteria are designed for humans. Humans are moral agents who can be held accountable for their mistakes. How will you hold ChatGPT accountable for its mistakes?
2023-01-03 20:33:06 @Raamana_ @ylecun When we use automl, pytorch, and compilers, they don't get acknowledged as co-authors. The people are the authors, and they are assisted by good tools.
2023-01-03 20:13:07 @ylecun Odd decision, especially for an AI conference. We should welcome papers that make a technical/scientific contribution regardless of whether they are AI-assisted. (Rejecting, of course, plagiarism and other unethical practices.)
2023-01-03 18:54:16 @peterbhase Yes, this is important work by my colleagues at @OregonState
2023-01-03 18:42:46 It was fun to hear @mkearnsupenn on @twimlai
2023-01-03 06:15:17 RT @tdietterich: @yoavgo Great writeup on the power and (current) limitations of large language models.
2023-01-03 06:15:13 @yoavgo Great writeup on the power and (current) limitations of large language models.
2023-01-03 05:46:04 Shame on @airbnb. Time for a rethink of your systems! https://t.co/HnORoHVt3d
2023-01-03 05:19:48 RT @Grady_Booch: Forrest is the Tom Lehrer of tech. @forrestbrazeal https://t.co/RtuBikidkp
2023-01-02 00:15:32 What’s with the promoted requests to follow? Is this a new blue feature? Makes me want to block instead
2023-01-01 20:12:12 @plain_simon @AndrewKemendo @EzraJNewman Yes, I agree with these, @plainsimon
2023-01-01 20:00:23 RT @rodneyabrooks: Posted my fifth annual review of my dated predictions from Jan 1, 2018, about self driving cars, AI/ML &
2023-01-01 19:23:56 @plain_simon @AndrewKemendo @EzraJNewman Global coherence is probably not necessary, but local coherence is essential. Example: code won’t work if it makes contradictory assumptions about its input or environment
2023-01-01 17:49:33 @plain_simon @AndrewKemendo @EzraJNewman In my view, it is you who is shifting the goal posts toward what DL can do today. But there is plenty of scope for people to disagree about research goals. Maybe we should rephrase the OP as “What should we work on next after ChatGPT?”
2023-01-01 06:05:10 @IDoTheThinking However, I view them as a form of gentrification. Outsiders with more money occupy housing and displace the locals without contributing socially to the neighborhood. I think they should be tightly regulated.
2023-01-01 06:03:24 @IDoTheThinking I like Airbnbs for the same reason that I don't live in a hotel room. Space, kitchen, laundry, free parking. Apartments and houses are nicer than hotel rooms
2023-01-01 06:01:14 @tinaintheburbs @IDoTheThinking In Spain and Australia, I have stayed in "apartment hotels". Some of these are in buildings that mix long-term residents with transients like me
2023-01-01 01:46:36 @mbrasneves @EzraJNewman But all of this might not be enough
2023-01-01 01:44:51 @mbrasneves @EzraJNewman I don't think we know. We aren't sure what AGI means, and we aren't sure what else might be missing. I think we want a system that combines massive knowledge with inference and consistency maintenance. ChatGPT seems to have massive knowledge, but lacks reasoning and consistency
2023-01-01 01:38:59 RT @paul_scharre: 1/ This new article by @sarahbauerle &
2023-01-01 01:37:25 @yudapearl Thank you for gently correcting our errors and helping us deepen our understanding of causality!
2022-12-31 23:56:29 @AndrewKemendo @EzraJNewman So what? I don't think the goal of AI is to replicate a human. It is to empower humans (and society) so that they can be more productive and happy
2022-12-31 22:24:24 @EzraJNewman “Seems”, but it has no memory once the tokens fall off the end of the input buffer. It lacks coherent beliefs and will emit direct contradictions in succession
2022-12-31 21:40:21 RT @tdietterich: @EzraJNewman 1. A model of the world, both the current state (the things it believes are true) and the dynamics (the thing…
2022-12-31 21:40:06 @EzraJNewman 6. In general, any AI system exists within a complex web of organizations. Most thinking in AI has been about an AI "individual", which perhaps reflects our individualistic society. I suggest we devote some attention to intelligent institutions and societies
2022-12-31 21:37:48 @EzraJNewman 5. A purpose/mission/set of goals. How is it supposed to behave? I like Stuart Russell's approach that the system has the goal of helping its user achieve the user's goals, which means it also seeks to understand what those goals might be and how they might change. 5/
2022-12-31 21:35:40 @EzraJNewman 4. An ability to reason and plan (a) to change the state of the world (including other peoples' beliefs), (b) to gather information to improve its own beliefs 4/
2022-12-31 21:34:40 @EzraJNewman 3. A model of larger entities (corporations, clubs, institutions, countries) and their beliefs, interests, and dynamics. 3/
2022-12-31 21:33:48 @EzraJNewman 2. A model of itself and of other individual people (theory of mind). What they believe right now
2022-12-31 21:32:40 @EzraJNewman 1. A model of the world, both the current state (the things it believes are true) and the dynamics (the things it believes will be true in the future, conditioned on its actions, the actions of other agents, and existing active processes) 1/
2022-12-31 16:51:53 @Noahpinion If you are investing in Riverside, make sure to use a different map
2022-12-31 16:37:25 RT @Aaroth: A good fairness question: https://t.co/fUcQKKkgb5 Should we include population base rates when making inferences for decisions…
2022-12-31 05:25:53 @matloff @rajiinio Time for me to go read the papers on causal notions of fairness.
2022-12-31 05:25:21 @matloff @rajiinio I think that is the causal argument. If the causality is x->
2022-12-31 04:36:17 @matloff @rajiinio I guess I'm making a causal argument rather than a probability argument.
2022-12-31 04:35:12 @matloff @rajiinio To me, it feels like "guilt by association". Suppose I apply for a loan in january and again in june
2022-12-31 04:32:00 @CsabaSzepesvari @rajiinio Another example might be loan defaults. My individual risk of default depends on my individual features, but in a down economy (which shifts p(y)) there may be unobserved factors that increase the risk of default. These might be reflected in p(y). end/
2022-12-31 04:30:33 @CsabaSzepesvari @rajiinio On the other hand, if the disease is an infectious one, then there is a causal link between the population p(y) and my personal p(y). In short, there may be a shared, unobserved confounder, and p(y) tells us something about it. 2/
2022-12-31 04:28:08 @CsabaSzepesvari @rajiinio I think the situation can be very complicated. If p(y) is computed on a whole population, why should the population prevalence affect my individual beliefs? I should be estimating my own p(y=0) for me. It is an instance of the reference class problem, I think. 1/
2022-12-31 01:30:24 @Aaroth Cool
2022-12-31 00:38:18 (I'm leaving out the normalizing terms, of course) end/
2022-12-31 00:38:17 Fairness question: When we apply a classifier, we combine P(x|y) and P(y) to get P(y|x). Why do we include the P(y) term? 1/
2022-12-30 23:57:46 @Grady_Booch Exams designed for people are rarely good tests for software systems. It is interesting to ask how we would want to test a software system for a legal application.
2022-12-30 19:55:03 @indoriitweeter @RickABright I’m just pointing out that there are some hygiene practices that I have always done and now will always do. No big deal
2022-12-30 19:04:07 @indoriitweeter @RickABright So what? I’m also stuck on washing hands
2022-12-30 03:36:55 @erikbryn @DigEconLab If we aren’t careful, the current batch of flawed methods will poison the well for a long time. I hope people are careful to focus on error-tolerant applications.
2022-12-30 02:41:55 @deliprao Differential privacy?
2022-12-30 02:38:46 @soboleffspaces @GaryMarcus @sethlazar That’s because we knew that Newton’s laws extended far beyond the moon, based on data that was gathered at long distances by astronomers
2022-12-30 01:50:59 @soboleffspaces @GaryMarcus @sethlazar I was focusing on ethical disagreements. While evidence from the world is important in ethical discussions, some disagreements (e.g., between utilitarian and duty-based frameworks) cannot be resolved via standard notions of evidence.
2022-12-29 20:50:25 @npparikh @yoavgo Understanding can always go deeper, it seems.
2022-12-29 16:56:57 @yoavgo There is a sub field of symbolic regression that tries to learn models like these.
2022-12-28 17:48:48 @tunguz Keep in mind that @ylecun is leading an industrial lab and working to keep it at the cutting edge. This lab has not indulged in least-publishable units
2022-12-28 07:25:21 @a_m_mastroianni I guess so. I’m excited about @TmlrOrg , which reviews for correctness but not for significance (similar to PLOS One)
2022-12-28 02:16:48 @a_m_mastroianni I have found anonymous peer reviews of my papers to be very helpful in improving them. The gate keeping function is not useful, however.
2022-12-27 21:13:36 @athundt They have a limited context buffer
2022-12-27 21:11:44 @vardi He was such a kind person
2022-12-27 20:18:47 @EpiEllie I always mask in indoor public spaces (e.g., supermarkets, theaters). But with people who know me, my masking (or lack thereof) can be correctly interpreted as a signal of my self-assessment of risk of infecting them.
2022-12-27 19:19:58 @yuvalpi Great Quantiles was kind of ok
2022-12-27 19:17:22 @EpiEllie Surely this policy can vary depending on your recent exposure risk. If I have been working from home all week, that’s different than if I just spent a week traveling internationally. I mask much more in the latter case.
2022-12-27 18:52:18 @crazyredranting I see a wide range of estimates online from 42% to 70% but nothing near 97%
2022-12-27 06:44:05 @aryehazan Robustness: max margin solves a robust optimization problem when your features are potentially noisy. It also improves robustness to class probability shift.
2022-12-27 06:02:36 @rao2z @francoisfleuret My mistake. I thought everything was accessible through the API, but I guess that is only true if you can guess the URI of each tweet. My original question still hasn't been answered: Just because someone "had access" doesn't mean they actually accessed and trained on the data
2022-12-27 02:11:46 @hartliss_chris @averylargebear @MazMHussain + the army and the church, esp Opus Dei, which provided technocrats
2022-12-27 00:27:46 @beenwrekt I guess it is a form of AI boosterism
2022-12-27 00:26:22 @beenwrekt Yes, we want to avert our eyes from the human labor behind this.
2022-12-27 00:18:11 @beenwrekt I agree with you there!
2022-12-27 00:17:40 @beenwrekt One would expect that the optimal combination would use ML for the easy cases and reserve human review for the more challenging ones. Escalation of the most difficult cases to the policymakers themselves. (Of course this requires good uncertainty quantification for the ML.)
2022-12-27 00:15:26 @beenwrekt From reading that thread, it looks like it is a combination of ML and people executing a decision tree. So there is *some* automation, yes?
2022-12-26 21:32:56 @FroudeNum @dogsnotbarking I'm not an expert on uncertainty quantification for language models, but in computer vision we have a toolbox of methods including post-hoc calibration, ensembles (including Bayesian posteriors), anomaly detection, and conformal prediction.
2022-12-26 21:15:47 @davidmanheim @TylerAlterman Judaism emphasizes the fate of the collective rather than the individual. Perhaps due to its pastoralist origins? Individuals can be ostracized, but God smites the community when it forgets the path of righteousness
2022-12-26 19:11:56 @kchonyc What exactly is your null? If these are two Bernoulli's with parameters theta1 and theta2 and you want to test H0: theta1 = theta2, you will need a lot more than 4 trials of each.
2022-12-26 17:23:27 RT @NickKristof: This is one of the great failings in Portland and up and down the West Coast: the impediments to home construction, raisin…
2022-12-26 06:55:12 @francoisfleuret Does “have access” mean “included in the training data”? Don’t we all have access to all public tweets?
2022-12-26 06:49:07 @yoavgo @Google Yes, but not at inference time, and that’s the problem, imo
2022-12-26 06:40:33 @zacharylipton Discovery seems best on Twitter
2022-12-26 02:40:38 @yoavgo Yes, I don't know anything about how search autocomplete works. But search results clearly benefit from human-curated knowledge. This has always been at the heart of @Google, starting with PageRank.
2022-12-26 01:08:58 @yoavgo Don’t they also make heavy use of knowledge graphs? How are google info boxes generated? Don’t the engines also know about events (sports, elections, movie premiers)?
2022-12-25 20:57:09 @danielleboccell @harryeskin Uncover important ideas that risk being lost. Embody critical expertise and teach it to others?
2022-12-25 20:54:11 @Joshua_B_Smith @dogsnotbarking It would be great for language translation as well
2022-12-25 19:22:57 What an excellent essay! Christmas gift to me: having time to catch up on articles like this https://t.co/Os5DzIphos
2022-12-25 19:07:01 RT @tdietterich: @dogsnotbarking I wonder if we could modulate fluency in generated language and speech to signal the level of confidence o…
2022-12-25 19:06:56 @dogsnotbarking I wonder if we could modulate fluency in generated language and speech to signal the level of confidence of the model? I assume this has been studied. Pointers?
2022-12-25 02:03:25 @rajivpant @OpenAI @YouSearchEngine Neither answer is correct.
2022-12-24 19:21:49 RT @ylecun: @egrefen Second-hand Dunning?
2022-12-24 17:22:56 RT @sh_reya: IMO the chatgpt discourse exposed just about how many people believe writing and communication is only about adhering to some…
2022-12-24 17:20:21 @AndrewKemendo @pfau Specifying what we want is the hard part. This is why we have languages with precise semantics—programming languages.
2022-12-24 02:52:40 @DavidDeutschOxf Let’s hope our AGI systems can learn from near misses rather than mistakes
2022-12-23 18:34:07 @faoliehoek @CharlotteHase The central problem in software engineering is to figure out what system we are trying to build. Coding is the easy part. It will be great to have LLMs helping with coding. Maybe they can also help us reduce specification and design error too?
2022-12-23 17:04:02 @soumithchintala @typedfemale Well-informed heuristic search is better than illl-informed. But the greedy algorithm fails when the heuristic is buggy
2022-12-23 17:01:31 @roydanroy @DamienQuerlioz @NatureElectron @C2N_com @INSIS_CNRS @UnivParisSaclay It’s a popular topic under https://t.co/rJUrSf3jAd on arxiv
2022-12-23 06:07:35 @RichardMCNgo @amt_shrma @ChenhaoTan I think LLMs need to be coupled with reasoning systems that can check the truth of LLM guesses against external resources such as knowledge graphs, data bases, simulators, causal models, and physical/social interaction with the world.
2022-12-23 06:03:13 @deliprao It’s not very different from building a company but success is measured in impact (intellectual, social, organizational, financial) rather than only financial
2022-12-23 05:58:44 @deliprao Success in academia requires combining long term focus with entrepreneurship. The people who are named to the National Academy are people who did something substantial through sustained focus rather than chasing conference deadlines. It usually requires building a great team
2022-12-23 05:54:47 @thanhnguyentang @JmlrOrg is better than the general ML conferences. +1 for COLT
2022-12-23 05:52:41 @ylecun @tobias_rees Successful technology disappears behind functionality. Failed technology shows through.
2022-12-22 16:36:57 @fchollet Extraordinary replies. I stand with you @fchollet and the Ukrainian people, and with everyone worldwide who is suffering and sacrificing to oppose Russian aggression.
2022-12-22 16:32:15 @aryehazan @pfau @roydanroy @aryehazan I think there are cases where actual harm is avoided. Example: A paper that will reveal a new security vulnerability in a deployed system without following responsible disclosure procedures is a clear case where ethics review is appropriate.
2022-12-22 07:21:13 @debarghya_das Google is essentially continually cleaning its web corpus (scraping and evaluating pages). A replacement will have to duplicate all of that and then figure out how to update the LLM.
2022-12-22 06:06:30 @matloff @vardi I have never heard of this. Citation?
2022-12-22 00:47:17 @ylecun I like to joke that the PHD is for people who love ideas more than money. Otherwise, an MS is much faster and usually leads to higher lifetime earnings. Of course either way, you need to keep learning new stuff all the time.
2022-12-22 00:43:51 @jimm_y_oung @ylecun And having both would be very strong
2022-12-21 21:21:34 @davidbau How much of this would be covered by the #NAIRR? I've lost track of its status.
2022-12-21 21:21:04 @davidbau I strongly support a public National Deep Inference service. It needs to be designed with flexibility in mind, as we will want to supports many different things: fine tuning, access to the training data, access to external resources (for LLMs that query them), etc.
2022-12-21 03:46:28 @amt_shrma @ChenhaoTan I would say that it can reproduce learned causal links and hypothesize new links. That is a kind of inference. However, it has no way to confirm whether those hypothesized links are real. These models generate, but they don't test.
2022-12-20 21:10:07 @DeepIndaba Congratulations on your huge impact!
2022-12-20 17:19:25 RT @geomblog: But this isn't about surveillance. It's about the weaponization of tech by those in power. It keeps happening and will keep h…
2022-12-20 02:40:40 RT @mmitchell_ai: In order to understand why ChatGPT can't replace Google Search, it's useful to understand the early days of web search an…
2022-12-19 21:41:45 @Noahpinion I do worry that the rabbit lobby, with its consistent delivery of happiness, is influencing your work
2022-12-19 21:37:15 @TobyWalsh @physorg_com Is it actually hard to design toxic chemicals? If compounds designed to have therapeutic value still have a high probability of being toxic, how much more likely is it for random compounds to be toxic? Or do molecules with therapeutic value usually have toxic neighbors?
2022-12-19 07:45:07 @docmilanfar Number of citations depends on the number of papers written in a field. Many more papers have been written in ML than in information retrieval
2022-12-19 07:42:55 @sdbaral But you know that, of course. I don’t see what the problem is. I need yellow fever vax to travel to places with yellow fever.
2022-12-19 07:41:03 @sdbaral Given the low vax rates in the US, it is very wise to get vaxxed before coming here.
2022-12-19 06:50:40 @peterwildeford Which one involves people? That's the hard one
2022-12-19 05:56:50 Let's give this lots of likes! https://t.co/xGIphLrhu1
2022-12-19 05:22:31 RT @glupyan: @FelixHill84 @spiantado @IbanDlank @tmalsburg @AndrewLampinen As a cognitive scientist (trained in the connectionism), I find…
2022-12-19 04:13:32 @erikbryn @DigEconLab A lot of Google’s service is to filter out spam websites, seo junk, etc. This is an adversarial game. How will the LLMs play it?
2022-12-19 01:50:21 RT @Bob_Wachter: Covid (@UCSF) Chronicles, Day 1006 (!) There’s an awful lot of Covid confusion &
2022-12-19 01:25:25 @matloff @overleaf There should be a little script you can run to fix the environment/registry after you were forced to restore files from backups after a disk crash.
2022-12-19 01:22:33 @matloff @overleaf I'm tired of reinstalling latex. On windows, the installation procedure is quite sensitive. I can have copied all of the files from another machine and updated all of the registry and path variables I can find, and yet it still won't run. My time is not free
2022-12-19 00:52:40 @matloff Ah, but using @overleaf you can (a) easily collaborate across multiple authors and (b) avoid having to install anything on your local machine. Overleaf is always up to date, and my impression is that they've worked hard on error message analysis as well.
2022-12-19 00:37:50 @schmangee Apparently there are still many millions of people in the US who have not had covid, but determining the exact percentage is very difficult.
2022-12-19 00:29:42 @schmangee As with every disease, some of us will die first of something else.
2022-12-19 00:09:28 @yudapearl You chant so beautifully. May we all be blessed by his memory
2022-12-18 23:03:10 @schmangee It can be a reminder that this virus is extremely contagious and very easy to catch. I think it helps remind folks to redouble their efforts. But I guess it could also lead to despair
2022-12-18 21:42:05 @tejuafonja This sounds like an appropriate use of a statistical test (== p-values)
2022-12-18 16:36:31 @yudapearl @DanielGordis @isaacdecastrog The dynamics of politics on both sides has favored the hardliners and rejectionists. Progress will require courage, but all I see is cowardice. It is heartbreaking
2022-12-18 16:33:08 @tejuafonja Is there a reason that you need to make a yes/no decision about independence? Or can you carry the uncertainty forward in your reasoning? If you need to make a decision, you will have to set a threshold
2022-12-18 01:24:19 @tejuafonja Confidence intervals, on the other hand, are very useful. The key question should be: What are the sources of variation, and am I quantifying them and understanding their impact on my inferences?
2022-12-17 19:52:39 @beenwrekt @RWerpachowski @rajiinio We used our physics-based mechanistic models to design safety mechanisms with huge margins. We also understood the threats to safety and were able to test against them.
2022-12-17 19:01:34 @beenwrekt @rajiinio Yes! And I don't see how statistical methods alone can model the tails of the distribution with sufficient accuracy to get five 9s. I think we need mechanistic models to achieve this. Is there any alternative?
2022-12-17 18:35:08 @predict_addict At a deep level, it is related to Vovk's conformal predictive distributions &
2022-12-17 18:33:46 @predict_addict This gives the same coverages as CQR but is easier to extend to multi-variate conformal prediction and conformal bounds on RL trajectories. So it offers an improvement over our previous work: https://t.co/eZiSSzqZt5.
2022-12-17 18:31:46 @predict_addict Thanks! We kind of buried the lede in this paper. The key idea is to fit conditional CDFs rather than quantile regressions as in Conformalized Quantile Regression (https://t.co/gCYaxY6PMI). Then we can conformalize in "probability space" without having to compute residuals. 1/
2022-12-17 18:25:31 @fchollet Do the ISPs have some responsibility? Perhaps blocking them will encourage them to know their customers better. Probably even more relevant to domain registrars who are willing to register all of these fly-by-night domains
2022-12-17 17:42:14 @rao2z @washingtonpost It’s like most political work: no fingerprints
2022-12-16 03:20:02 Time to quiet quit. I’ll check back in a few days after they’ve been restored https://t.co/v4pHL7I6C0
2022-12-16 03:13:54 @KLdivergence Multi-talented!!
2022-12-16 03:06:20 @unsorsodicorda @roydanroy @erikphoel Isn’t it just that statistical models are best at modeling the dense regions. So when sampling, we will mostly get draws from the typical set? ML optimizes for boring!
2022-12-16 02:51:53 RT @AlexGDimakis: After fine-tuning (Dreambooth) Stable diffusion with my face, I become everyone and get multiple copies even when I try t…
2022-12-16 02:50:42 @maosbot Wow!
2022-12-15 06:37:23 @vitavonni @shortstein I totally agree about the benchmark chase. It rarely leads to any insight.
2022-12-15 06:29:09 @vitavonni @shortstein You will need theory to understand why the scaling laws exist. My bet is that we will need a theory of the scale of different patterns that are being learned. Larger patterns require larger architectures, more data, and longer training. Why are you so negative about theory?
2022-12-15 05:42:40 @gchrupala @FelixHill84 I watch the wild turkeys in my neighborhood form leks of typically 3-4 males. Only one of those males will mate successfully, it seems. The others assist and "succeed" through kin selection.
2022-12-15 05:41:04 @gchrupala @FelixHill84 Evolution selects for reproductive success, but that doesn't mean that every individual maximizes the number of descendants. Check out https://t.co/rXhuE5oezn for some weird examples.
2022-12-15 04:51:50 @FelixHill84 Keep in mind that there is a fundamental difference between knowledgeable performance and intelligence. Performance exploits learned patterns, intelligence adapts to novel challenges. They are distinct, but they require each other.
2022-12-15 04:49:54 @FelixHill84 Biological intelligence is likely to be much more complex, much less elegant, and much more robust. Focus on finding a theory underlying intelligence
2022-12-14 21:57:04 @vitavonni @shortstein The theory community is hard at work understanding SGD, double descent, benign overfitting, and so on. Important work on privacy and its relationship to generalization.
2022-12-14 20:33:39 RT @outland_art: | From the desk of #OutlandGuestEditor @thesarahshow~ @NeuralBricolage’s work is best understood…
2022-12-14 20:10:38 @vitavonni @shortstein It's true that ML is currently in a phase where heuristics dominate, but that has not always been true. ML theory will rise again!! I'm not sure what you mean by "brute force all that seems reasonable"
2022-12-14 19:29:59 @boazbaraktcs @emilymbender @Azure In the medium term, I'm optimistic the carbon cost of ML will go to zero through a combination of carbon-free energy production, better computer architecture, and better algorithms. Many people are working on this. Carbon is the least of the problems of LLMs.
2022-12-14 19:22:15 @boazbaraktcs @emilymbender @Azure Or multiple queries could be batched
2022-12-14 19:21:49 @boazbaraktcs @emilymbender @Azure But if there is a security benefit, it is not pointless.
2022-12-14 19:17:38 @boazbaraktcs @emilymbender @Azure I do this! One advantage is the added security of knowing that google will correct my typos instead of sending me to a bogus site trying to mimic a legit site by having domain name substitutions (zero for `o').
2022-12-14 19:13:58 @vitavonni @shortstein I had in mind a more positive outcome. If an AI system can write a paper that adds to the knowledge of the field, then it should be accepted.
2022-12-14 19:12:22 @emilymbender @Azure If machines are powered down, then total demand has dropped. This REQUIRES some generation to decrease or some additional storage to increase (pumped water storage
2022-12-14 19:09:31 @SashaMTL @emilymbender @Azure Yes, they will continue using offsets until 2025, evidently. They plan to continue purchasing offsets beyond 2025, because they want to remove CO2 equivalent to all of their historical emissions since the company was founded.
2022-12-14 18:49:53 @emilymbender @Azure I don't agree. Cloud resources are powered down if there is no demand.
2022-12-14 18:32:32 @shortstein Your paper must include machine-checkable proofs and automatically reproducible experiments. These checks are performed prior to human review.
2022-12-14 18:30:16 @emilymbender ChatGPT is presumably running on @Azure which claims to have been carbon neutral since 2012 and plans to be operating on 100% renewable energy by 2025.
2022-12-14 18:20:35 Looking forward to this panel Friday. Do robustness and fairness trade off against each other? https://t.co/8w1uENKYWU
2022-12-14 04:31:35 Painful article in El Pais on the failed auto-golpe in Peru. “Nations don’t fall down like buildings. They simply pretend to exist.” https://t.co/w9HhIXxox6
2022-12-14 03:20:24 @doyouknowchamp So the system just makes up an image of a robot in a Tesla factory? It's all fake
2022-12-13 21:53:49 @Sergei_Imaging @zga_aaa @arxiv My reading of these rules is that you are not allowed to register a computer program as an author on the @arxiv site. Hence, it cannot be a coauthor on an arXiv submission. Journals and conferences may have other rules, I suppose.
2022-12-13 20:57:15 @Sergei_Imaging @zga_aaa @arxiv Misrepresentation of identity or affiliation, for any reason, is possible grounds for immediate and permanent suspension. 2/
2022-12-13 20:57:05 @Sergei_Imaging @zga_aaa @arxiv rules: "It is a violation of our policies to misrepresent your identity or organizational affiliation. Claimed affiliation should be current in the conventional sense: e.g., physical presence, funding, e-mail address, etc. 1/
2022-12-13 20:49:35 @Sergei_Imaging @zga_aaa Sure, but not as coauthor.
2022-12-13 20:45:01 @roydanroy Yes, if we get hit with a large number of bogus submissions, either we will need to start checking them in much more detail or arXiv will become much less useful.
2022-12-13 20:24:21 @roydanroy I believe we released a paper yesterday that, while it did not list ChatGPT as a coauthor, did claim that an LLM had written large parts of the paper. It is under review at some venue, presumably. It should be judged on whether its content is accurate and scientifically valuable
2022-12-13 20:21:41 @Sergei_Imaging Not being a moral agent, an AI system cannot be a coauthor because it cannot take responsibility for the content of the paper. Do you acknowledge spellcheck and grammarcheck as coauthors already?
2022-12-13 19:03:10 @Noahpinion Similarly, having a single platform for customer service requests that is public has been great for forcing companies to be more responsive when things go wrong. (But it has been weird seeing customer service complaints mixed in with technical and political discussions.)
2022-12-13 19:01:49 @Noahpinion It would be good to have one place where government agencies and companies can make announcements. On twitter, I can subscribe to my power company, local government, weather service, etc. and get alerts from all of them.
2022-12-13 16:46:42 @emilymbender This will be the longest ever tweet thread. Judging from my spam invitations, I’m an expert in materials science…
2022-12-13 02:45:16 @dan_s_becker My colleagues Alan Fern and Jonathan Hurst are having success with sim2real transfer teaching Cassie to walk and run. The key is high fidelity simulation and domain randomization. Sims are improving, and this will unlock RL applications https://t.co/AiD76mN1yN
2022-12-13 00:45:43 RT @Bob_Wachter: Current case rate in CA is 29/100K/d, 3-4x early November. SF's rate is 27. Asymptomatic test + rate @UCSFHospitals is 5%.…
2022-12-13 00:03:58 @jbseltzer @Plinz @ezraklein I thought google+ worked pretty well, and it allowed longer posts. But I'd also like some way to have structured discussions, not just chains of replies. And easy ways to add links/citations. Maybe overlay highlighting on news articles and papers, etc.
2022-12-12 23:52:27 RT @bigmlcom: Google finally launched a #MachineLearning add-on for #GoogleSheets this month. OTOH BigML's Add-on for Google Sheets keeps b…
2022-12-12 23:19:51 @Plinz @ezraklein @ezraklein's point is that the low cost of responding leads to thought-free, emotional reply-tweets. These in turn are attacks on our attention. Shortness in itself is not the issue except as it encourages poorly-thought-out responses.
2022-12-12 21:31:07 @rao2z Writing is a tool for reasoning--for checking the coherence of one's arguments and beliefs. This is totally different from being a "computer" who just executes specified calculations. (But I assume you aren't responding seriously, yes?)
2022-12-12 21:08:28 RT @yangyang_cheng: For @ChinaFile, I write about the recent protests in China and solidarity rallies overseas, including the one I attende…
2022-12-12 20:58:13 "Twitter’s value is how easy it makes it to talk. Its cost is how hard it makes it to listen." Brilliant essay by @ezraklein https://t.co/KquyqAxwjA
2022-12-12 20:30:13 @rao2z Human computers didn't decide what to compute
2022-12-11 23:06:40 @paul_scharre Thanks
2022-12-11 21:27:15 @paul_scharre I haven't had a chance to read your book yet, but I wonder how the numbers of changed in the last 3 years. Anecdotally, I would suspect they have crashed.
2022-12-11 17:40:11 RT @a_derfelGazette: 1) Hi, everyone. I wanted to share with you the scary recent experience I had with my original Twitter account, @Aaron…
2022-12-11 17:35:18 @ntoxeg @GaryMarcus @Noahpinion I don’t know. There are published papers about SQL. Is autocorrect in Google docs, gmail, and Outlook based on LLMs? The others are speculations on my part
2022-12-11 05:50:57 @Noahpinion And Google @deepmind
2022-12-11 05:49:31 @GaryMarcus @Noahpinion Pay attention to my last condition: Anything that can be validated. Obviously the LLM can't be trusted on its own. But there are plenty of settings where plans can be checked for correctness against a model (e.g., in PDDL+). I agree household robot plans are NOT such a setting
2022-12-11 03:35:34 RT @tdietterich: @GaryMarcus Auto-complete for everything. See @Noahpinion's recent substack. Mapping English specifications into first dr…
2022-12-11 03:35:24 @GaryMarcus Auto-complete for everything. See @Noahpinion's recent substack. Mapping English specifications into first draft designs: English->
2022-12-11 01:08:56 @IDoTheThinking @johnwhitehead81 In my psychology methods course, they said if R^2 >
2022-12-10 23:50:10 RT @BDehbozorgi83: Spot on @GaryMarcus!
2022-12-10 19:46:34 @Hello_World @SashaMTL We will make progress in ML much faster if we have high quality open LLMs to study. That said, @OpenAI has no obligation to open their particular model.
2022-12-10 18:41:07 @j_foerst How about MLDR (machine learning done right)…. (ducks)
2022-12-10 15:53:24 @Hello_World @SashaMTL She is doing exactly this: she determines that this is problematic. There is no fear involved. The goal is to advance scientific understanding
2022-12-10 05:09:37 @jefrankle @gkdziugaite @roydanroy @MITIBMLab @GoogleAI Congratulations! Looking forward to the revised LTH paper
2022-12-10 04:38:40 RT @WildInsights: Protecting biodiversity is essential to address climate change, but we need reliable data. The #WildlifeInsights communit…
2022-12-09 17:49:14 @rasbt @pfau But for most purposes the final fold is the key, yes?
2022-12-09 00:19:25 @rasbt @pfau People are doing de novo design, so it is not just refinement. It is an amazing success!
2022-12-08 21:45:21 @ai__pub It was not the only reason for renaming to NeurIPS, but it was a contributing factor. The name was a source of discontent in prior years.
2022-12-08 14:08:19 @adjiboussodieng I agree, this requires some care to design it well and to establish the validity of the methodology. Even the simplest ideas turn into research projects
2022-12-08 13:34:14 @adjiboussodieng Maybe we could write descriptions of three students and then ask volunteers from different countries to write letters? I could do US letters, maybe as a seed example. Anyone out there interested in giving this a try? Or maybe there is a better approach using example phrases?
2022-12-08 13:02:15 @adjiboussodieng Yes! It would be great to have some examples of "the same" letter as it might be written by people from different cultures as a reference for those of us reading these letters. Maybe also at 3 levels of excitement?
2022-12-08 13:00:00 CAFIAC FIX
2022-12-07 08:00:00 CAFIAC FIX
2022-11-13 15:52:25 @micahgoldblum Isn’t it a simple corollary of information theory that I need more bits from the data to discriminate among more models? You are controlling the model class via the prior. This hinges on what is meant by “can represent”
2022-11-13 06:21:00 @timnitGebru Relationships with funders require care. Tom Mitchell likes to say "Never let your funding interfere with your research". Easy to say, but it can be hard to achieve
2022-11-13 06:18:16 @timnitGebru What a researcher does with the money is at least as important as the source. Am I "complicit" with the US Government when I accept funds from NSF? What about from the Ford Foundation? Google? It's a complex issue
2022-11-12 23:30:20 @jdcmedlock Furthermore, in the long run, CTC might reduce gun violence as a side effect.
2022-11-12 21:15:53 @GaryMarcus @ylecun Yes, there is no disagreement here. That ibex has very little data available, so it needs super-strong priors. A lyrebird can learn an immense range of songs from data, so must have weaker priors. It takes them a fair amount of data to learn those songs.
2022-11-12 20:42:14 @IBJIYONGI @FractalEcho What I was trying to say is that people in the majority (i.e., white) don't encounter racism on Mastodon, so they are not aware that the host matters. Call it the "ignorance of the majority".
2022-11-12 06:52:33 @arayavalani @SarahCBlunt @katiedimartin I don’t believe that could possibly be true. How many applications do they get?
2022-11-12 05:24:15 RT @tdietterich: @twimlai @bohannon_bot @etzioni @allen_ai I greatly enjoyed this episode. I always learn a lot listening to @etzioni
2022-11-12 05:24:11 @twimlai @bohannon_bot @etzioni @allen_ai I greatly enjoyed this episode. I always learn a lot listening to @etzioni
2022-11-12 05:15:00 @sgourley What do you imagine the user experience to be? How would the AI know what I wanted to say? I usually figure that out as a write and revise.
2022-11-12 00:28:52 "There are some things it might be a mistake to grow out of." https://t.co/gohkfIHvwz
2022-11-11 04:07:04 @karpathy I've often wondered if taxation can be usefully viewed as regularization of the economy
2022-11-10 23:53:39 RT @rao2z: Soon the only reliable indicator that a twitter account is real may be that it is *NOT* a verified account..(just like the onl…
2022-11-10 23:53:09 @Sim_Startup_ @FelixHill84 @ai__pub I worry that the word "emergence" is encouraging us to think of this as some magical phenomenon. I look forward to additional research that demystifies these behaviors by showing that they result from discovering abstract patterns that are present in the training data.
2022-11-10 23:50:55 @ai__pub @FelixHill84 Maybe other folks can provide more insight?
2022-11-10 23:50:08 @ai__pub @FelixHill84 According to this hypothesis, the discovery of these abstract patterns emerges with sufficient data and sufficient training. Which is why I would call it "emergent". Does that make sense? end/
2022-11-10 23:48:43 @ai__pub @FelixHill84 The main example given in the Foundation Models papers is in-context few-shot learning. My hypothesis is that this capability results from discovering abstract patterns in the training data that can then be instantiated in the prompt and extended by the LLM. 2/
2022-11-10 23:47:29 @ai__pub @FelixHill84 The example they give is in-context few-shot learning, which I suspect is quite different than writing poetry or doing arithmetic. My guess is that poetry and arithmetic are the result of training on poetry and arithmetic, and the model is just generalizing from its training. 1/
2022-11-10 20:16:35 @ai__pub Is this definition of "emergence" generally accepted? By this definition, all forms of self-supervised learning on "pre-text" tasks exhibit emergence. I thought emergence is the phenomenon that qualitatively different things are learned as training continues
2022-11-10 05:10:00 @ccanonne_ I should have thought of this!
2022-11-10 00:55:00 @Noahpinion Tigray estimated at half a million or more. https://t.co/ieBnXi0g0j
2022-11-09 23:59:54 Oregon State EECS is recruiting (including in AI/ML). We will soon have the most powerful academic computing facility in the US thanks to a gift from Jen-Hsun and Lori Huang. https://t.co/8SR7WHPF1G
2022-11-09 23:51:21 RT @ipfo: **NEW** eBird Trend maps for N. America - Thanks to >
2022-11-09 23:48:00 RT @TorstenVolk: @bigmlcom definitely deserves a big shoutout for what these guys have achieved with just $4.4M in venture funding. 16 pate…
2022-11-09 23:07:31 RT @noUpside: Gonna be evergreen content for a while
2022-11-09 18:40:02 I can never remember what "shambolic" means. It is neither a sham nor related to symbolic. How did we get such a confusing word? Maybe from "shambles"?
2022-11-09 06:17:57 @utopiah @erikbryn They can be relied upon to operate for thousands of hours autonomously. All autonomy is limited. I myself am likely to fail completely in the next 30 years. In contrast, existing LLM tools can't operate autonomously for more than a few seconds.
2022-11-08 20:53:35 @utopiah @erikbryn The operating system on my PC is highly autonomous. Same with the ABS braking system in my car.
2022-11-08 20:11:54 @erikbryn Are any of these systems fully autonomous? Is it more accurate to say that a person + an AI tool co-create a chart-topping song?
2022-11-08 19:47:44 "a well funded problem": what a great line! https://t.co/YNLNhQQOrl
2022-11-08 19:31:27 @NWSPortland Why so much warmer in Portland? Usually, being near the Gorge makes Portland colder
2022-11-08 06:01:39 I would love to see these reforms adopted: https://t.co/KNOSC5zL5n
2022-11-08 02:52:12 @FractalEcho @IBJIYONGI Well, I guess we disagree. That’s fine. Cheers
2022-11-08 02:26:54 @FractalEcho @IBJIYONGI I’d guess ignorance rather than conceit
2022-11-07 20:26:03 @mapto @boazbaraktcs @AvivTamar1 @MannorShie Right
2022-11-07 19:59:14 RT @aihuborg: Free seminars and talks available in November!There are over 30 seminars and talks across the globe this month, here are…
2022-11-07 04:21:45 @boazbaraktcs I believe Stuart Russell has been looking at uncertain rewards. @AvivTamar1 @MannorShie
2022-11-07 04:20:09 @boazbaraktcs What is known formally about whether backing off from the optimum brings us closer to our true objective? Chow, et al. showed that in RL optimizing CVaR == robust optimization against uncertain dynamics, but they say nothing about uncertain rewardhttps://t.co/KBALehI5BD
2022-11-06 19:32:51 @RunPolisRun The protests in Portland contained a nihilistic element that was purely destructive. And they succeeded in destroying a lot of business activity. I wish I understood why the police were unable to stop them.
2022-11-06 00:23:56 @jdcmedlock Easily confused in my experience
2022-11-05 18:56:58 @MazMHussain @JohnHMcWhorter Beautiful sentence: "People succeed or fail based on a mixture of circumstance, character, and fortune. "
2022-11-05 16:54:19 @shai_s_shwartz @boazbaraktcs Very nice. The keys are how Y is represented and what loss we use
2022-11-05 01:18:25 @MazMHussain I've always wanted to learn Persian
2022-11-05 01:12:11 @MazMHussain How many languages do you speak?
2022-11-05 01:01:09 @MFordFuture @vardi Done a long time ago for dish washing and espresso making, not to mention mixing, blending, and kneading. The trend continues!
2022-11-04 23:13:42 @roydanroy anonymity leak
2022-11-04 18:06:46 The best thing Elon could do for Twitter is to get off twitter
2022-11-04 04:15:18 @mark_riedl I agree. @realaaai now that we have moved to electronic proceedings, can we lift the limit on the bibliography?
2022-11-04 02:47:02 @roydanroy @colinraffel @NeurIPSConf @kchonyc I know I’m deeply confused about it!
2022-11-04 02:08:05 @GeorgeMonbiot Isn’t most maize grown without irrigation? Can you show the numbers for pesticides?
2022-11-02 21:06:07 @MazMHussain I'd pay for a site without disinformation and twitter mobs. Can @elonmusk provide a Quality of Service guarantee?
2022-11-02 21:01:59 @CACMmag @vardi @chloexiang @motherboard Actually, what is increasing is our ability to understand how DL works and why it has the outputs it has. We are making progress on this problem.
2022-11-02 05:01:43 @__SimonCoste__ Wonderful! And it gives me much better intuition about what's going on.
2022-11-02 04:56:54 @DavidBrin Very common practice in Australia
2022-11-02 03:42:18 @noctusacinonyx @bennpeifert I think you have overfit on series that are too simple. There is no reason to believe that all stationary time series problems can be solved with traditional techniques. Certainly they are the first thing to try, though.
2022-11-01 19:49:38 @Rosejackson3 @geomblog @united Yes, I want to know how my biometrics are transmitted, verified, and whether they are stored and retained.
2022-11-01 04:39:32 @Rosejackson3 @geomblog @united When I recently asked to be checked by an agent instead, the United agent complied immediately.
2022-11-01 04:30:49 @dvassallo @Kal2Karthik It is a form of smoothing. It makes the UX more consistent
2022-11-01 04:28:37 Wise words from @Bob_Wachter https://t.co/i1L4gNoIrZ
2022-11-01 04:10:14 @LordOfTheAlts @scienceisstrat1 @Noahpinion @ylecun @erikbryn @nfergus There’s a huge amount of talent being wasted in China by the CCPs policies. Wouldn’t it be nice to free that talent to benefit the world?
2022-10-31 04:51:47 @matloff @JennyChachan How does someone attempting to murder you increase your empathy with the poor?
2022-10-31 04:39:40 @matloff @JennyChachan What evidence do you have that Paul Pelosi lacks empathy for ordinary folks?’ And somehow needs this lesson. Your tweet came across as nasty, and I think your pre-apology shows that you knew it was in poor taste
2022-10-30 17:18:41 RT @chowleen: Praise the mutilated world,and the gray feather a thrush lost,and the gentle light that strays and vanishesand returns.-…
2022-10-30 16:31:16 @scottastevenson I also think chess is not a good case to think about. If we look at protein folding, for example, we can check the energy of the folded protein to understand it. The explanation is not in terms of network weights but in terms of the properties of the solution
2022-10-30 14:25:39 @scottastevenson Why can’t AI systems also become better at explaining themselves?
2022-10-30 14:24:21 @scottastevenson Isn’t most illegibility in society the result of adversaries deliberately increasing noise and distractions?
2022-10-30 05:27:59 @Miles_Brundage @Lee_Morgan7 In a previous Twitter discussion, someone suggested "Large Corpus Models"
2022-10-29 22:32:04 @adawan919 @ylecun Evolution has no goal. It is stochastic forward search followed by natural selection (eg reproductive success or failure).
2022-10-29 22:30:42 @adawan919 @ylecun “Planning”, as we use it in AI, involves assembling a sequence of actions to achieve a goal. Efficient planning algorithms incorporate the goal into the search.
2022-10-29 16:15:21 @adawan919 @ylecun Natural selection does not plan
2022-10-29 15:48:28 @rao2z @SheilaMcIlraith That paper shows how to automatically extend the state space to convert non-Markovian reward specifications into Markovian reward functions
2022-10-29 15:47:38 @rao2z Yes, when teaching MDPs, I tend to focus too much on the Markovian transitions and not enough on the Markovian rewards. I love @SheilaMcIlraith's work on more expressive languages for expressing reward functions: https://t.co/7jIG8mvc6Y
2022-10-29 05:45:04 @matloff Yeah
2022-10-29 05:37:23 @ylecun Is there any evidence that this is more than a preprogrammed procedure?
2022-10-29 05:34:08 @boazbaraktcs Academic Twitter includes Academic #metoo (as it should). Then come the trolls. But presumably the trolls will come to mastodon too.
2022-10-29 05:30:09 @DKThomp @Noahpinion We always want to find a scapegoat. That’s why my prior is for a wild origin. The evidence continues to be ambiguous
2022-10-29 05:23:11 @deliprao This is why I liked teaching intro algorithms. You can teach TRUTH (eg sorting).
2022-10-29 05:05:04 Great thread by my colleague @taaltree on why being a professor is a great job https://t.co/sh1idUdZP7
2022-10-28 04:56:24 I am now @tdietterich@mastodon.social. Will cross-tweet for now
2022-10-27 23:06:50 @aaronharris @Noahpinion I guess the courts will decide that question: https://t.co/k1ROaMBVs9
2022-10-27 21:34:36 @aaronharris @Noahpinion Monopoly over the Apple App store and all financial transactions that go through it.
2022-10-27 16:09:35 RT @predict_addict: The long (17 years!) wait is over.The second edition of the ALRW book (Conformal Prediction Bible) written by the c…
2022-10-27 05:40:36 @roydanroy @KevinKaichuang Me too. Entrapment effort?
2022-10-27 04:59:32 @Noahpinion The harms caused by Apple are the standard harms of a monopoly. The harms of Facebook (and social media in general) threaten societies around the world through the toxic behavior that they enable. I agree that anti-trust is more relevant to the former.
2022-10-27 02:13:32 @Abebab Just being true to its history…
2022-10-26 03:05:48 @mhar4 @yangyang_cheng I think the two generations are very different. The first was triumphal, but today we are all extremely mindful of the fragility of freedom, health, and peace. end/
2022-10-26 03:03:52 @mhar4 @yangyang_cheng She writes "The new cold warriors are not so different from their counterparts a generation ago, who preached that marketisation and free trade would usher in political liberalisation. Both narratives stem from an unabashed belief in the supremacy of their own system..." 2/
2022-10-26 03:03:27 @mhar4 @yangyang_cheng I am always moved by @yangyang_cheng's writing, and this piece is no exception. For a whole host of reasons, we have all entered a state of perpetual mourning. But I disagree with one statement. 1/
2022-10-25 15:27:22 @Raza_Habib496 @yoavgo Interesting, thanks.
2022-10-25 00:11:52 @yoavgo I'm not sure what it would mean to be "about language", but existing models are certainly not "about communication" because an LLM is not an agent and has no communicative goals
2022-10-25 00:08:41 @schwellenbach But it isn't exactly a surprise is it? Anyone paying attention knows that Eric Schmidt has been devoting a lot of his time to AI. (I have no special insight
2022-10-24 17:33:48 RT @GaelVaroquaux: #NeurIPS2022 paper: https://t.co/qft4LgZTQ7An easily reusable benchmark that shows that deep learning underperforms…
2022-10-24 15:52:43 @MelMitchell1 @arxiv You might consider withdrawing and resubmitting the paper with the correct categories. Long delays are usu caused by discussions among moderators about the best category. These can get stalled because the discussion UI is clunky.
2022-10-24 15:49:40 @tinaeliassi @davecliff @MelMitchell1 @arxiv We clear the cs.LG category every day. (If we didn't, we would drown in submissions and never recover.) That should never be the category that is holding up a release.
2022-10-24 04:17:28 Interesting housing story in the @latimes. I knew the racist history of Los Angeles was bad, but it was much worse than I had realized. https://t.co/uHRFNnjro5
2022-10-24 01:27:17 @cjmaddison I would rephrase this as" It is by the grace of God AND the measures you are taking (vaccination, masking)". Those measures are not useless, but they provide no guarantees
2022-10-24 01:23:40 @rasbt I'd like to see metrics for ease of tuning or insensitivity to hyperparameters. More radical: It is not an algorithm until all tuning is automated. I guess that would be required for third-party evaluation.
2022-10-23 22:02:41 @cloudquistador @emilymbender Excellent point! Thanks for clarifying
2022-10-23 21:25:39 @AnthroPunk @emilymbender Yes, and I take very seriously the idea that human intelligence may have evolved primarily to support social interaction. I think that is why so many of our AI benchmarks seem to be mis-targeted.
2022-10-23 21:23:10 @cloudquistador @emilymbender Yes. I'm not sure what you mean by "aiming in error". Are we focusing on the wrong target behaviors?
2022-10-23 20:17:48 @cloudquistador @emilymbender The achievement of aerodynamics is that we (mostly) understand what material properties are required to mimic birds. It would be wonderful to have similar insights into the computational/biological substrates required to mimic human cognition in detail.
2022-10-23 19:50:25 @emilymbender At some level of detail, they fly for the same reasons.
2022-10-23 19:49:38 @emilymbender They fly.
2022-10-23 19:44:17 @emilymbender But I agree it is difficult to define the "functional behavior" of cognition, and that seems to be a major stumbling block in these discussions
2022-10-23 19:43:22 @emilymbender Hmm. If airplanes are not models of avian flight behavior, comparisons to avian flight behavior are only marketing/hype? I think there can be global functional comparisons without any claim that the underlying detailed structure is a "model"
2022-10-23 04:19:29 @deliprao I do love the feeling that the computer is going to work all night while I get to sleep.
2022-10-23 03:08:42 @deliprao This is an opportunity to get in touch with the early years of computing. In every generation, our ambitions outstrip the hardware and we must wait for computations to finish.
2022-10-22 22:47:02 @earnmyturns @Noahpinion And Spain was having enough trouble holding on to Catalonia, which continues to this day.
2022-10-22 21:21:07 @GaryMarcus I don't know how to quantify progress, so I don't know if it is exponential. But we have been making progress on dealing with change/novelty. e.g., Bryan Loyall at CRA leads an exciting project: https://t.co/dggKFIm3Cf
2022-10-22 21:15:51 @tejuafonja Instead, it is the one that is most clearly written and that clarifies the fundamentals. The paper that explains what is happening and why.
2022-10-22 21:14:53 @tejuafonja A reason not to wait is if you are in a race with some other research groups. But it is easy to worry too much about this. Good, mature work is a more valuable contribution than rushed, immature work. Usually the most cited/appreciated work is not the very first paper on a topic
2022-10-22 21:13:12 @tejuafonja One reason to wait is that if you publicize the unimproved version of your paper, that is the version people will read. And they may dismiss it or misunderstand it. After improvement via peer review, they will hopefully understand and appreciate it much more
2022-10-22 21:11:47 @tejuafonja It depends on how confident they are that the work is ready for public examination. I usually prefer to have a round of peer review before going public with my work. That can be internal peer review if you are part of a large enough organization.
2022-10-22 19:29:36 RT @tdietterich: @denny_zhou Nonsense. Each deployed app is a choice. I hope this attitude is not representative of Google brain as an orga…
2022-10-22 19:29:24 @denny_zhou Nonsense. Each deployed app is a choice. I hope this attitude is not representative of Google brain as an organization
2022-10-21 23:23:23 @RRamirezMorales @mbeisen *Most of the areas within these fields
2022-10-21 23:21:26 @RRamirezMorales @mbeisen But most of these fields don't need gate keeping.
2022-10-21 23:20:42 @RRamirezMorales @mbeisen The physicists have to deal with variations on perpetual motion and theories of everything.
2022-10-21 23:20:17 @RRamirezMorales @mbeisen In computer science, the sensitive areas are (a) proofs that P=NP, (b) papers claiming to diagnose COVID using chest x-rays and deep learning, (c) papers claiming to forecast COVID using fancy time series analysis.
2022-10-21 19:47:20 @RRamirezMorales @mbeisen Most areas of science and engineering don’t draw the attention of the media or the trolls. But I agree that there are some areas that require careful filtering
2022-10-21 00:15:06 @scottjshapiro I assume this particular tweet was aimed at philosophers, but it certainly applies generally. Technologists need to have a basic understanding of political theory, sociology, psychology, philosophy, and economics to analyze the broader context of their work.
2022-10-21 00:13:00 @scottjshapiro Whenever you cross disciplinary borders, it is best to have someone from that discipline as an escort. Interdisciplinary work requires a big investment to understand the concerns and conceptual structures of each discipline.
2022-10-20 20:33:05 @randall_balestr @sindero Tree ensembles create empty partitions, too, by intersecting no empty partitions. It is not obvious that this is a good thing
2022-10-20 16:59:00 @rasbt There is usually one spiking net paper per day on ArXiv
2022-10-20 14:53:59 @MazMHussain I always figured South Asian was adopted to remind people in the West that there are lots more Asians than just Chinese, Japanese, and Koreans.
2022-10-20 00:39:13 @_joaogui1 @FelixHill84 @erikbryn @gdb @OpenAI @scaleai The DL frameworks make neural net assignments more reliable, but tuning is still a black hole for students. end/
2022-10-20 00:37:22 @_joaogui1 @FelixHill84 @erikbryn @gdb @OpenAI @scaleai Back in the 90s, a colleague of mine said "If I give a neural network assignment, only 10% of the students can get it to work. If I assign an SVM, 90% get it to work, and they all get the same answer." 2/
2022-10-20 00:35:48 @_joaogui1 @FelixHill84 @erikbryn @gdb @OpenAI @scaleai Even in the 1990s, shallow neural networks could beat SVMs in the right hands. But they were much more difficult to tune and slower to train. The big advantage of SVMs was that the optimization problem was convex and they are very insensitive to the margin parameter C. 1/
2022-10-19 22:43:20 @davidthewid And one junior faculty member stepped up and did a lot of informal advising too
2022-10-19 22:42:19 @davidthewid In countries such as Canada, where funding goes directly to students, the power issues are reduced. But that still doesn’t fix the single advisor problem. My fellow students in grad school ended up advising each other during the 1980s AI boom
2022-10-19 19:10:33 @erikbryn @gdb @OpenAI @scaleai What has increased monotonically is the amount of data available for training
2022-10-19 19:10:02 @erikbryn @gdb @OpenAI @scaleai This is a simplification. There were intermediate points along the way where compute loads decreased. For example, when we switched to SVMs from 1980s neural networks.
2022-10-19 16:54:03 @davidmanheim @Noahpinion As long as the US continues to send $3B per year in military aid to Israel, can't it expect cooperation in military affairs?
2022-10-18 22:36:32 @robert_k_walter @JordanJacobs10 @geoffreyhinton Routine design problems will be much easier to automate than novel systems. But I do believe there is plenty of scope to create better tools for novel systems as well. Just don't expect full automation
2022-10-18 22:35:39 @robert_k_walter @JordanJacobs10 @geoffreyhinton The core problem in software engineering is figuring out what system we wish to build -- getting the requirements and specifications correct, even when everyone involved has excellent NLP skills.
2022-10-18 08:35:44 @nickchk Yes! Being an outlier is just a statistical indicator. You need to ask what underlying process generated that outlier and whether you want to include that process in your model.
2022-10-18 05:55:21 @robert_k_walter @JordanJacobs10 @geoffreyhinton It certainly helps with programming in the small. But it isn't going to design software architectures or manage CI/CD and MLOps. It gives programmers a power tool, but we need much more to achieve a revolution in software systems.
2022-10-18 05:07:37 @TheRealAdamG @JordanJacobs10 @geoffreyhinton Yes, there are exciting directions to pursue!
2022-10-18 01:05:10 @JordanJacobs10 @geoffreyhinton A word of caution. Machine learning introduces software change, not always improvement. That creates risk and requires careful testing. AI is not eating software engineering, at least not yet.
2022-10-18 00:26:47 @ylecun @Grady_Booch @synoptase If you can directly ask the learned model to successfully reproduce a particular artwork, then this is image retrieval, and copyright law should apply, yes?
2022-10-18 00:25:29 @ylecun @Grady_Booch @synoptase It is one thing to copy a single work by an artist (e.g., in order to learn). There is a long tradition of this in art. But to copy the entire lifetime of work is no longer fair use in my non-legal opinion.
2022-10-17 23:47:05 @ylecun @Grady_Booch @synoptase Yes, because digital copying is so much more accurate.
2022-10-17 21:01:04 @davidmanheim @Noahpinion We need to drive all of those curves to zero. Any country that doesn't will "make the difference"
2022-10-17 06:47:42 @balajis To which textbooks are you referring? Citations and quotes please.
2022-10-16 21:05:16 @geomblog I was wondering about this as a US Global Entry station scanned my face recently. Do you know how that is implemented?
2022-10-15 22:57:18 @leonpalafox @totlmstr But his companies build rockets and cars. True, they contain code, but if anything, the software in a Tesla is one of its weaker points
2022-10-15 20:09:24 @totlmstr Yes, I don't say this as a criticism. He is an excellent entrepreneur.
2022-10-15 20:08:32 @rao2z Newell &
2022-10-15 20:07:08 @rao2z Simulator-based #RL is also just inference by a different name. But sometimes it is easier to do inference by sampling than by intensional reasoning.
2022-10-15 19:59:42 @totlmstr Is Mr. Musk an inventor? I thought he was an entrepreneur and marketer
2022-10-14 12:40:02 @EigenGender One thing about software is that you can run it and obtain its operational semantics. So one can imagine reinforcement learning combined with domain adaptation to transfer from one programming language to another
2022-10-14 07:26:05 @nebul19926248 @AriDavidPaul @Austen As soon as the customer clicks, they can link and deanonymize the random id, right? Amounts to the same thing
2022-10-12 19:51:26 @kchonyc What is the geographic distribution of these mistakes?
2022-10-12 08:11:42 @IDoTheThinking It does mean that they have been to a place where they are surrounded by people who don't speak the same language and who have different customs. This is both humbling and enlightening
2022-10-10 05:08:14 @minimeadow @Noahpinion Don't you think there will be millions of climate migrants over the next 20 years? I think countries in temperate and subpolar regions need to develop a coordinated plan for assimilating the coming flood.
2022-10-08 00:41:33 @NLPnorth We don’t care about your paper count. Tell us some interesting content, some reason to read the papers
2022-10-07 21:47:29 @Grady_Booch We need to test the whole pipeline, because so many bugs enter during data preparation. We need to test the optimizer. We need to test the gradient computations. @karpathy had a nice post awhile back, IIRC
2022-10-07 21:45:20 @Grady_Booch I agree that ML needs good software engineering practice, but I think we need to DISCOVER those practices. ML does plenty of random testing on separate test data, for example, which is basically unit testing, but that is not at all sufficient.
2022-10-07 16:16:06 @unsorsodicorda @USGS_Quakes Yeah, I'm kind of disappointed I didn't feel it.
2022-10-07 13:32:42 @andrewgwils @tetraduzione @nsaphra If I'm training with a proper loss function, won't my model be calibrated in the large sample limit? If I've made poor modeling choices, accuracy may be terrible, of course. Are there other conditions that must be satisfied?
2022-10-07 13:27:51 @USGS_Quakes Didn't feel anything in Corvallis. I was awake and sitting in a chair
2022-10-06 14:01:06 @deliprao Keeping up with the literature? Speaking for myself, it is more like barely hanging on and faking the rest
2022-10-06 12:56:24 Inferring lava channel depth from standing waves in lava. My daughter Hannah Dietterich's latest paper. https://t.co/HqIz5qzoYG
2022-10-06 03:03:59 @rctatman The “safety” term covers a diverse range of research including safe exploration and execution in reinforcement learning, robustness to violations of assumptions (iid assumption, Markov assumption), uncertainty quantification, etc. It’s not all AGI stuff
2022-10-06 00:20:50 @_ashawndabney Me
2022-10-05 21:32:46 @emilymbender For me, it depends on the content. If I’m just giving an academic talk on my research, I wouldn’t expect an honorarium. But if I’m giving a longer tutorial, I charge my consulting rate plus travel expenses
2022-10-05 01:14:54 @ShriramKMurthi Does this reduce endogamy or increase it?
2022-10-05 01:08:50 @Alber_RomGar The field is mining a couple of ideas here, and because there are so many researchers, results flow quickly. Hard to know when this vein will play out. There is still a lot to do building on top of the large corpus models and diffusion.
2022-10-04 20:38:17 @SashaMTL @StephanieBrandl DM me the submission ID and I can look it up
2022-10-04 15:26:30 @kwekubediako_ @ghdatascience Love this photo!
2022-10-04 12:55:45 RT @geomblog: I'm very proud to see the release of the AI Bill of Rights (BoR) today. It all started with a vision articulated by @AlondraN…
2022-10-04 10:27:46 @acherm @ChristophMolnar Learning against an adversary is hard because of nonstationarity. It has been studied in the context of cyber security and email spam. @dlowd has great work in this area
2022-10-04 09:59:13 @yudapearl @Grady_Booch @ylecun @AlexTensor @pmddomingos @KordingLab @GaryMarcus Yes, these accounts are mutually compatible. The desire for explanation and control of danger at the individual level can be combined to build cohesive communities and institutions
2022-10-04 09:40:03 @acherm @ChristophMolnar But there are some applications, such as Google flu trends, which are obviously not going to work. The signal (symptom searches) is weakly correlated with the disease. It is based on human behavior that can rapidly change. It is confounded with advertising.
2022-10-04 09:37:56 @acherm @ChristophMolnar Yes, but not relevant to these cases. In many of these cases, an ML/statistical approach might make sense, but data need to be collected carefully to capture or control potential confounders and ensure validity.
2022-10-04 09:35:16 @ChristophMolnar @arxiv There are careful studies of ML in medicine. See, for example, https://t.co/yDjKiVpPLWIt is not an RCT, but of course there are many medical questions that cannot be answered via RCTs. @suchisaria
2022-10-04 09:33:47 @ChristophMolnar @arxiv Absolutely!
2022-10-04 09:28:41 @ChristophMolnar I certainly agree with you there. ML and statistical tools in the hands of poorly-trained researchers lead to useless/dangerous publications. I moderate https://t.co/zTxHX0xsHV on @arxiv, and we tried to identify and reject as many of these ridiculous COVID CX papers as we could
2022-10-04 09:15:38 @yudapearl @Grady_Booch @ylecun @AlexTensor @pmddomingos @KordingLab @GaryMarcus I believe the current theory about the origin of organized religion is that it arose at the same time as larger human settlements arose. So it may play a role in building and maintaining social cohesion
2022-10-04 09:10:48 @ChristophMolnar Interesting for a statistician to make a claim based on anecdote. These are high profile failures, but I can match your anecdata with successes too. Let’s start with Google translate, an application no statistician would ever have contemplated
2022-10-03 17:56:03 @tejuafonja That is the iron test. Wonderful! (And best wishes for your defense!)
2022-10-01 07:14:31 RT @PangWeiKoh: Reminder: Submissions to the NeurIPS DistShift workshop are due next Monday (Oct 3), AoE!
2022-09-30 21:26:10 @scottbudman Is this caused by the annual switch to winter fuel blend in California?
2022-09-30 21:23:51 @thegautamkamath In fact, I usually read the reviews first to decide whether the paper is worth my time
2022-09-30 16:10:15 @etzioni @allen_ai Congratulations on building such an amazing research organization!
2022-09-30 16:05:26 @GaryMarcus @vyodaiken @swarat @ylecun @wellingmax I'm a big fan of @swarat's work.
2022-09-30 13:16:52 @SemiCynic @svpino It would still be useful to write out the software contract. What size of errors are you willing to accept? Do you prefer overestimates to underestimates?
2022-09-30 12:11:15 @svpino The algorithm should implement a contract with the user that can be fully understood. (And that raises a huge challenge for ML.)
2022-09-30 07:15:53 @Carnage4Life I pack this way. Main benefit is shirt’s don’t get wrinkled. But both packing time and unpacking time increase.
2022-09-29 20:33:54 @masonfidino Also check the clock time on your camera
2022-09-29 16:48:57 @jackclarkSF I'd love to read a transcript of your testimony. Will you post it?
2022-09-29 16:18:01 @etzioni @jackhidary Yes, the full implications of this take a while to comprehend. Frightening!
2022-09-29 09:36:28 @Plinz @pwlot @fchollet This is a great idea @Plinz. Someone could analyze Cyc (at least the OpenCyc knowledge base) and assemble a corpus of interesting abstractions. Then researchers could probe some large corpus models to assess how well they have learned those abstractions.
2022-09-28 21:16:58 @pwlot @fchollet Maybe, but I would bet it represented thousands of regularities that no deep net can ever learn from language corpora
2022-09-28 21:13:35 @roydanroy MWWA: Make Workshops Workshops Again
2022-09-28 16:00:53 @RWJE_BA @eliasbareinboim @ylecun @GaryMarcus @yudapearl @TiernanRayTech Well, it may be philosophical, but it is backed by theorems. And those theorems tell us what inferences we can draw from the forms of evidence that you list.
2022-09-28 15:21:11 RT @aihuborg: We have too many highlights from September: Interviews with @lilyxu0 and @PaulaHarder8Papers and invited talks from @IJC…
2022-09-28 14:27:19 @boazbaraktcs @GaryMarcus @davidchalmers42 @ylecun Instead of debates, let's do research on these questions.
2022-09-28 14:25:57 @JulianLBaasch @GaryMarcus @davidchalmers42 @ylecun Judging by the "likes", many people. I care deeply about the science, and I look forward to seeing strong progress in both connectionist and hybrid systems. But much of the current discussion is a meta conversation that doesn't advance the science IMO
2022-09-28 14:23:45 @thegautamkamath I love papers that list the contributions of each author. Can we move to this as a default?
2022-09-28 13:39:13 @news4wombats @cassidycody @JohnHolbein1 I don't think I've ever written "as such" in a paper myself. So maybe I do just hate it full stop.
2022-09-28 13:17:47 @GaryMarcus @davidchalmers42 @ylecun We just don’t care.
2022-09-28 13:17:14 @JohnHolbein1 “Use” and its variants. A wimpy boring word that says nothing.
2022-09-28 13:15:23 @cassidycody @JohnHolbein1 strong agree! I hate it because it is hardly ever used correctly
2022-09-28 08:49:55 @shiwalimohan @ylecun @GaryMarcus @yudapearl @TiernanRayTech ML-based (including DL-based) next-state prediction can learn from interacting with the physical world too, not just a sim. But yes, large samples are usually required. This allows them to learn from interventions, which takes us to Rung 2 of the Ladder of Causality.
2022-09-28 08:45:34 RT @rao2z: My take on the religious wars of #DL vs. #symbolic #AI--that repeatedly get fought from the same polarized positions: IMHO, they…
2022-09-27 18:03:31 @GaryMarcus @Zergylord @ylecun @rao2z @MITCoCoSci @guyvdb @HenaffMikael @alfcnz @anirudhg9119 To what extent do you think @yudapearl 's causality theory is the solution? This is certainly where Yoshua Bengio is placing his bets.
2022-09-27 16:33:29 @hectorpal @vardi It might feel more Popperian in the online setting where the linear extrapolation leads to making a new measurement, but it is true even in the batch setting. 2/
2022-09-27 16:32:43 @hectorpal @vardi Can't you view gradient descent as a process of falsifying theories? Each gradient step is a linear extrapolation from the current theory, and each computation of the loss is identifying and responding to errors in that prediction (aka falsifications). 1/
2022-09-27 16:08:49 @TDeryugina When do you need the money?
2022-09-27 15:41:20 @heatherknight Is this because of the computational cost of reproducing other peoples' work?
2022-09-27 14:17:43 @ylecun @GaryMarcus @yudapearl @TiernanRayTech But maybe I don't understand the full scope of your approach.
2022-09-27 14:17:17 @ylecun @GaryMarcus @yudapearl @TiernanRayTech The third rung allows you to reason about a specific observed trajectory and predict what would have happened if different actions had been taken---based on what is learned (e.g., random events such as dice rolls) from that specific trajectory. 2/
2022-09-25 16:13:15 @DanHendrycks @DavidSKrueger Well, there does need to be an underlying theoretical basis for believing that there are risks, just as we have strong biological theory about viruses. There is no evidence for zombies, for example, but no reason to expect they exist either.
2022-09-25 15:28:50 @rgblong @DavidSKrueger A conceptual problem is that, unlike integrated circuits, it is unclear how we should be measuring intelligence.
2022-09-25 15:26:14 @rgblong @DavidSKrueger Good point, "rapid growth" need not be exponential. As most technologies exhibit saturating growth, the question is "Where are we located on the sigmoidal curve?". Perhaps "AI" will be a whole sequence of smaller sigmoids just like Moore's Law?
2022-09-25 15:21:34 @NickKristof @nytimes Maybe the Oregon Public Employees Benefit Board #PEBB should drop Providence from its network?
2022-09-25 13:58:04 @DavidSKrueger For the founders and the X-risk folks, there is an assumption that just beyond human-level intelligence, we will observe an exponential intelligence chain reaction. But there is no evidence for this. Meanwhile, near term risks need urgent attention.
2022-09-25 10:47:57 @UofTCompSci So sorry to learn of this. As so many have said, Fahiem Bacchus was a wonderful human being and a brilliant teacher and researcher
2022-09-25 09:05:19 @memotv @AaronHertzmann Yes, but when people use "An AI", it seems that they want to talk about an "entity" not a "stuff".
2022-09-25 08:02:44 @memotv @AaronHertzmann But you can’t say “I built a software to translate languages”. Wrong kind of noun.
2022-09-25 07:42:27 @AaronHertzmann We need a short, snappy word for “computer program”. 5 syllables is evidently too many
2022-09-24 21:30:56 @ttzteei @fchollet I believe the alternative phrase was “Complex information processing”
2022-09-24 20:37:28 @unsorsodicorda So the question is not “Is the brain a computer?” But rather “How are the brain’s computations organized?”
2022-09-24 20:36:09 @unsorsodicorda Yes. But I agree with his point that we do not store eidetic memories but rather construct memories generatively. His knowledge of AI seems seriously out of date. He needs to play with stable diffusion!
2022-09-24 12:34:41 @unsorsodicorda He claims we will need to understand the detailed chemistry of every cell. But surely there are more abstract levels of description that will provide better insights
2022-09-24 12:32:48 @unsorsodicorda One of his arguments is that the brain is a reinforcement learning system. And of course that is a kind of computer. He is mostly arguing against naive application of the von Neumann architecture. ML has also moved beyond von Neumann
2022-09-24 08:13:17 RT @tdietterich: @JulienMouchnino @rasbt There are many advantages to working with calibrated classifiers. First, suppose you are going to…
2022-09-24 08:13:01 @JulienMouchnino @rasbt and make the minimum cost decision y'.This carries over to any downstream probabilistic inference where the classifier is part of a larger system and the probabilities assigned by the classifier will be combined with other probabilities to drive system behavior. 4/
2022-09-24 08:12:31 @JulienMouchnino @rasbt The second advantage of calibrated classifiers is for cost-sensitive classification. Using calibrated probabilities P(y|x) and a cost matrix C(y,y') (cost of predicting y' when the true answer is y), I can compute the expected cost of predicting y' = sum_y P(y|x) C(y,y') 3/
2022-09-24 08:12:03 @JulienMouchnino @rasbt For any fixed threshold, We could get the same information using a validation set. Calibration analyzes the validation set and gives us this for all possible thresholds. 2/
2022-09-24 08:11:54 @JulienMouchnino @rasbt There are many advantages to working with calibrated classifiers. First, suppose you are going to make decisions by taking the output confidence of the classifier and comparing against a threshold. If the classifier is calibrated, then the threshold == prob of being correct 1/
2022-09-23 18:58:36 RT @JaneLubchenco46: DYK: If shipping were a ‘country’ it would be the 8th largest emitter? That’s why green shipping is a priority for @PO…
2022-09-23 16:40:18 @emilymbender Yes, CHR have always made me a bit uncomfortable, and I think you've put your finger on the reasons why.
2022-09-23 16:39:01 @emilymbender Maybe we should develop different rules for such cases? For example, I've been in meetings where you are not permitted to talk about anything person X presents without first getting permission from person X.
2022-09-23 16:33:39 @emilymbender This conflicts with the goal of getting intellectual credit for your ideas. The expectation in DC seems to be that getting credit is impossible. Maybe if you want credit, you should only talk about ideas you have already published?
2022-09-23 16:30:50 @emilymbender You raise interesting questions. When I took the CRA's policy course, they emphasized that in Washington DC, there is a "no fingerprints" phenomenon: So many people contribute to a policy outcome that you will not be able to tell whether your input had any effect. 1/
2022-09-23 16:27:10 @Julie188 I think it depends on the circumstances. My undergraduate college invested in me through financial aid. Sending money to them is a way of paying that forward.
2022-09-23 05:14:20 @simonw Load it into @bigmlcom
2022-09-23 05:06:21 @davidthewid And email me the table so I can visualize it in various ways
2022-09-22 13:18:01 @pfau Hierarchical paper structure is good. Progressive refinement: abstract, intro, body, appendix tell the same story at different levels of detail
2022-09-22 12:57:15 @Abel_TorresM @davidchalmers42 @ylecun This goes beyond statistical regularities
2022-09-22 12:55:42 @Abel_TorresM @davidchalmers42 @ylecun A good world model captures modular invariants. The interaction of two gears should only depend on their contact, not their names, etc.
2022-09-21 18:49:34 @AymericPM @neurips Ban that one too. If you don't think your work is novel, you shouldn't be writing a paper. Exceptions of course for work that replicates, critiques, or synthesizes existing work.
2022-09-21 18:38:56 Let's ban "for the first time" from @NeurIPS paper abstracts. Given the explosion in the literature, no one can be sure they are doing something "for the first time".
2022-09-21 15:01:19 RT @aihuborg: A huge thank you to @lilyxu0 for the great chat and sharing your inspiring work in using AI for conservation in the latest ep…
2022-09-21 11:31:44 @aryehazan Because of power asymmetry. Because of life long psychological damage.
2022-09-20 08:18:55 @aidybarnett @PLOSONE How much of this is the result of underpowered studies that tell us nothing?
2022-09-20 06:53:46 @SebastienBubeck Debugging an adaptive system will always be hard. While SGD raises specific challenges, all systems based on ML or optimization present versions of this problem.
2022-09-19 19:34:02 @ishaantharoor @MazMHussain the fastest way to fix the problem is to end poverty. Ending racism would help in the long run, but I despair of ever abolishing it.
2022-09-19 19:32:28 @ishaantharoor @MazMHussain I see a lot of news stories where poverty is the underlying cause of a problem, but instead of mentioning this, the problem is cast in terms of race. Now obviously poverty is often a result of racism, but the stories should at least mention poverty, because ...
2022-09-18 17:35:05 @deliprao Yes. If I recall correctly, the final version was a complete rewrite from the earlier one that did showed promise in the CASP competition. Diffusion methods have been through a similar amount of iterative improvement. Both transitioned much faster than the usual 20 years!
2022-09-18 17:28:50 @deliprao AlphaFold was very quick as well, and much more useful and trustworthy
2022-09-18 15:16:34 @rasbt @roydanroy @tetraduzione @NeurIPSConf Make poster sessions much longer. Arrange them spatially by topic. Specify times when the poster is manned but keep them up much longer.
2022-09-18 15:14:19 @rasbt @roydanroy @tetraduzione @NeurIPSConf Whats wrong with having thousands of attendees and talks? AGU makes it work every year.
2022-09-18 13:32:16 @abhas_rewcie @RWerpachowski @roydanroy @tetraduzione @NeurIPSConf Both are possible. One use case: everyone is listening to an invited speaker who mentions an arxiv paper. All 5000 people try to download it simultaneously. Both the network and arxiv crash.
2022-09-18 12:12:55 @RWerpachowski @abhas_rewcie @roydanroy @tetraduzione @NeurIPSConf When 5000 people simultaneously tey to use this phonesin and conference, what do you this will happen?
2022-09-18 00:39:32 @AvivTamar1 @roeiherzig There has been quite a lot f progress in offline RL. Distributional RL is also a nice advance
2022-09-17 22:00:01 @etzioni Of course, performance on downstream tasks suggests that these models are learning interesting abstractions for some things, but apparently not for arithmetic.
2022-09-17 21:58:25 @etzioni Without the capability to show me what training data are contributing to the output, I can't tell whether GPT-3 has learned any generalizations or just memorized facts. At least a rote memory database can answer provenance questions!
2022-09-17 21:47:17 @etzioni I don't understand why this is tantalizing? Even a 175B parameter lookup table can't master a task that should be possible with a 1M-element table.
2022-09-17 21:21:47 @RWJE_BA @smueller @yudapearl @stevekrouse @trunghlt @YiMaTweets But you must make assumptions to draw generalizations from data even in non-causal statistical reasoning. Even RCTs require some assumptions (that Pearl's work elucidates).
2022-09-17 16:39:36 @rasbt @matloff @marktenenholtz When I see "design matrix", I think about fractional factorial designs, split-plot designs, etc.
2022-09-17 14:34:33 @leonpalafox @tetraduzione @NeurIPSConf I agree!!
2022-09-17 14:14:24 @rasbt @matloff @marktenenholtz It isn't usually "designed" (at least in observational data)
2022-09-17 14:13:27 @tetraduzione @NeurIPSConf One thing that would help enormously would be to encourage employers to sponsor the conferences. Sponsorships translate directly into lower prices and travel grants for attendees.
2022-09-17 14:12:16 @tetraduzione @NeurIPSConf You know that all of the people expenses at conference venues have greatly increased post-pandemic just like all service-industry salaries. The added uncertainty in number of attendees means registration costs must be higher to hedge against worst-case outcomes
2022-09-17 14:10:32 @tetraduzione @NeurIPSConf I don't know the details for NeurIPS, but ICML lost a lot of money on the Baltimore conference. It turns out a hybrid conference has huge internet costs plus stuff like slideslive. I think the online registration fees are not covering expenses at all.
2022-09-17 14:09:49 @tetraduzione @NeurIPSConf Emergency funds make sense for emergencies, but if the new normal is to lose money on every meeting, organizations like NeurIPS and ICML won't be functioning very long
2022-09-17 13:57:19 @tetraduzione @NeurIPSConf Attendance is down and expenses are up. Therefore, prices must increase.
2022-09-17 02:52:32 @guicho271828 @YiMaTweets @yudapearl As with other forms of machine learning and reasoning, the implementation of causal reasoning may be approximate and automatic as well as exact and explicit
2022-09-17 02:51:14 @guicho271828 @YiMaTweets @yudapearl I view the existence of superstitions and conspiracy theories as extremely strong evidence that people are constantly looking for causes.
2022-09-16 17:36:18 @NateAFischer @myth_pilot In Europe, you can have bars and restaurants surrounding an area where children can play. In the US, we keep playgrounds and bars far apart. (Our Puritan legacy?)
2022-09-16 17:30:37 RT @amaldorai: I'm not surprised that Adobe is acquiring Figma for $20B, nor that Wall Street doesn't understand it and $ADBE stock is down…
2022-09-15 23:58:01 @adjiboussodieng Autodiff at scale is having an impact across many fields, not just ML. I'm so grateful to the developers who have built these amazing tools (and the mathematics underlying them)
2022-09-15 23:56:06 @adjiboussodieng I agree the frameworks are wonderful! But Chalmers is a scientist rather than an engineer, so I assumed he was asking about scientific rather than engineering breakthroughs.
2022-09-15 23:22:20 RT @tdietterich: @KenGardner11 I just gave to mvcommunityservices. You can too: https://t.co/6dn08pZMQV
2022-09-13 03:56:38 @zxul767 @isbellHFh @irinarish I have subjective opinions, but I suspect this has been studied rigorously. Morphine stops pain (in certain ways) by binding to certain receptors. Does it work on both kinds of pain?
2022-09-13 01:25:51 @Sheikheddy @DrMJoyner @DavidEpstein @nxthompson @GaryMarcus @cragcrest @geoffreyhinton @ACasadevall1 I do report missed spam using the Outlook reporting feature. I'll try to report false spam too. Thanks for your efforts!
2022-09-13 01:24:23 @Sheikheddy @DrMJoyner @DavidEpstein @nxthompson @GaryMarcus @cragcrest @geoffreyhinton @ACasadevall1 My group has been running an email classifier as an Outlook addin for more than a decade. We classify mail into user-defined categories. It is fully based on ML and still makes lots of mistakes. I agree that speed and memory footprint are serious engineering challenges.
2022-09-13 01:09:05 @andyroyle_pwrc So pretty!
2022-09-13 00:37:01 RT @nikitabier: Putin is a few days away from a Medium post about his team's amazing journey.
2022-09-13 00:35:24 @deliprao How about "progressively"?
2022-09-13 00:33:16 @noUpside @Noahpinion Same here, but I'm about to do a whole month of travel, so we'll see if I can dodge the virions successfully. Got my bivalent booster today. In the family, 2/4 have had it
2022-09-13 00:23:09 @isbellHFh @irinarish Yes, I suppose it is. But it is a different kind of suffering than physical pain, hunger, cold, etc.
2022-09-12 19:18:48 @geomblog “The size of Wales” is another standard unit of this type
2022-09-12 19:16:36 @irinarish I think social approval might be more powerful
2022-09-11 20:29:49 @marcinpaprzycki @zacharylipton @AnimaAnandkumar There are pros and cons to both. https://t.co/akAbm3HjnO
2022-09-11 19:37:47 @marcinpaprzycki @zacharylipton @AnimaAnandkumar Way
2022-09-11 17:10:47 @unsorsodicorda @ian_soboroff @emilymbender @ml_angelopoulos Agreed. I have a paper on CP for MDP trajectories, but we only handle a scalar value at each step along the trajectory. But maybe CP can be applied in the embedding space? https://t.co/eZiSSzqZt5
2022-09-11 04:06:34 @jdcmedlock Is this really true? Certainly as I get older, I take fewer physical risks, because I'm physically less reliable (balance, strength, etc.). But a standard trope is the old guy who sacrifices his life to save the village, because his life is almost over anyway
2022-09-10 23:44:58 @DrMJoyner @DavidEpstein @nxthompson @GaryMarcus @cragcrest @geoffreyhinton @ACasadevall1 This does seem doable. What mail system does the best job? In my experience, definitely not Microsoft 365
2022-09-10 21:57:11 @oren_sa As others have pointed out, need to take into consideration taking of political prisoners, hostages for exchange purposes, rights for women, LGBT, etc. as well
2022-09-10 21:55:30 @oren_sa Of course, Iran might decide you can't return home after the conference.
2022-09-10 19:13:55 @unsorsodicorda @ian_soboroff @emilymbender Conformal prediction could be applied to either of these two tasks: sequence extension or supervised probability prediction. end/
2022-09-10 19:12:33 @unsorsodicorda @ian_soboroff @emilymbender Unsurprisingly, when evaluated on the accuracy of the content, its probabilities are the most poorly calibrated. 7/
2022-09-10 19:11:37 @unsorsodicorda @ian_soboroff @emilymbender The authors also compare against the normalized log probability that the non-fine-tuned GPT-3 assigns to its output token sequence. This is an uncalibrated assessment of the original sequence extension task and has not been trained on the propositional content. 6/
2022-09-10 19:07:02 @unsorsodicorda @ian_soboroff @emilymbender The claim is that the fine-tuned GPT-3 makes accurate uncertainty statements for other tasks that were not in the calibration set. These other tasks also involve arithmetic, so one can imagine that this could work, and they show experimental evidence that it does. 5/
2022-09-10 19:04:59 @unsorsodicorda @ian_soboroff @emilymbender For this special case, therefore, GPT-3 does have a (statistical) basis for estimating the accuracy of the propositional content of these 1-token sequences. 4/
2022-09-10 19:03:32 @unsorsodicorda @ian_soboroff @emilymbender In this paper, they created a supervised calibration set of true propositions and measured GPT-3's accuracy. They then fine-tuned GPT-3 to output an uncertainty assessment (in five levels) of its accuracy. 3/
2022-09-10 19:01:08 @unsorsodicorda @ian_soboroff @emilymbender This is very different from the task of judging whether the propositions appearing in the token sequence are true in the world. GPT-3 has no way of knowing which token sequences contain true vs false propositions. 2/
2022-09-10 19:00:08 @unsorsodicorda @ian_soboroff @emilymbender The task of GPT-3 is to find a high-probability sequence of tokens to extend a prompt. Speaking informally, a well-calibrated GPT-3 would output the probability that the token sequence would be generated by a human drawn from the same set of humans that generated the corpus. 1/
2022-09-10 02:59:21 @sarahookr Do you think hardware designers should be exploring other architectures? Do you have any suggestions about architectural directions?
2022-09-09 17:44:59 @AlecStapp Perhaps it would be easier to pass a temporary emergency moratorium so that we don't have a huge battle about the long term?
2022-09-09 06:29:02 RT @mbeisen: King Charles III is an inspiration for postdocs everywhere
2022-09-08 23:50:09 RT @MazMHussain: … so naturally the tour dealt heavily with the crushing oppression of the Spanish, including mass enslavement and murder o…
2022-09-08 23:19:45 @XAI_Research @caro__lawrence @umangsbhatt @NickKroeger1 I think that is a very important use case! I also think that when CS people design explanation facilities, we implicitly design them for ourselves. But I know some groups start with end-user engagement to design their explanation facilities.
2022-09-08 23:17:52 @XAI_Research @caro__lawrence @umangsbhatt @NickKroeger1 That is consistent with our experience back in the expert system era (1980s). The explanation facilities in those systems were primarily used as a debugging tool similar to a stack backtrace.
2022-09-08 22:19:58 RT @paul_scharre: How might artificial intelligence change warfare? @jordanschnyc and I discuss on ChinaTalk [THREAD] https://t.co/oWX8Ss…
2022-09-08 22:18:25 @caro__lawrence This is nice. Is there a similar analysis of the kinds of queries that users of explanations have? In applications, I prefer to start with the user. For example, if the user is an ML engineer, what questions do they have about a model, an example, or the fitting process?
2022-09-08 03:03:40 RT @erichorvitz: On readying for new forms of #AI-powered manipulation &
2022-09-08 03:02:45 "the whole point of art is to make human interventions in the world for humans. I can imagine that robot generated art might be great for robots, but not for us. "
2022-09-08 03:02:44 Loved this about AI art from Ernest Edmonds (@eaedmonds) quoted in @benbendc's latest email: "It is...very possible to imitate a style using an AI program. But such imitation is exactly what a student artist does BEFORE they develop. Exactly what the human artist would discard."
2022-09-08 02:15:44 @PacificPower_OR This is a useless tweet, and it links to a useless page, and your corporate web page says nothing about this.
2022-09-07 21:58:33 @RanthonyEdmonds Oregon State offers a PhD in math (or CS) with a minor in Ecosystem Informatics. We have world-class programs in ecology, earth sciences, atmospheric sciences, and environmental engineering.
2022-09-07 21:22:05 @ian_soboroff @emilymbender The problem in this paper is the authors are confusing UQ on their downstream supervised task with both UQ of the language model and UQ of the propositional content (right word?) of the sentences.
2022-09-07 21:20:24 @ian_soboroff @emilymbender As long as you give me a separate validation data set, I can get black box prediction intervals using, for example, conformal prediction. Introspection is not required for in-distribution uncertainty quantification.
2022-09-07 20:27:31 @emilymbender The authors claim that this downstream confidence model also gives reasonable confidences for other, closely-related tasks. I don't see any theoretical justification for such a claim although it may be true for the tasks they have tested.
2022-09-07 20:26:04 @emilymbender Yes, the uncertainty that they are quantifying is the uncertainty of a second (supervised) model trained on data for which ground truth is known but using features from GPT-3. This is a down-stream task, not the "language modeling" task.
2022-09-07 20:05:43 @ian_soboroff @emilymbender Isn't it a question of what uncertainties are being quantified? With regression I can make a point prediction or predict a distribution or an interval. I might also assess the probability that the test data is from the same distribution as the training data, etc.
2022-09-07 03:15:48 RT @mauromycota: Have you ever wondered if the dried porcini products you are buying at the supermarket are truly what it says in the label…
2022-09-06 23:36:04 @TobyWalsh [To be precise, it doesn't contain any images labeled as people. There are people in some images, and they are sometimes very prominent, but they have not been labeled.]
2022-09-06 23:34:24 @TobyWalsh ImageNet-1K doesn't even contain any people. Does that invalidate what we've learned about ML and CV from training on ImageNet?
2022-09-06 23:31:18 @TobyWalsh In short, NO ONE CLAIMED THAT IMAGENET REPRESENTED BIODIVERSITY ACCURATELY, so this paper is attacking a claim that no one ever made.
2022-09-06 23:30:22 @TobyWalsh ImageNet has served quite well for this purpose, and the methods that work well on ImageNet also work well when trained on higher-quality data sets either de novo or by fine tuning or other transfer methods. 4/
2022-09-06 23:29:17 @TobyWalsh Instead, the goal was to develop better computer vision ML methods and learned representations. Ideally, methods that do not require 100% correctly-labeled data, as such data is rarely available in large volumes. 3/
2022-09-06 23:28:14 @TobyWalsh So the goal is not to be 100% correct on the categories
2022-09-06 23:27:26 @TobyWalsh I didn't mean to be dismissive. It depends on the purpose of creating and training in ImageNet. ImageNet doesn't correspond to any specific object recognition goal. The categories are at many different levels of specificity. 1/
2022-09-06 23:05:25 @TobyWalsh This is why most ImageNet evaluations measure top-5 accuracy. We know the labels are wrong and incomplete.
2022-09-06 23:02:29 @deliprao But what seems to be missing is the beliefs and goals of the speaker (i.e., the causal latent variables) that produced this behavior. If we could identify those, we could do much more. Do any latent variables in current LLMs correspond to these?
2022-09-06 23:01:03 @deliprao To the extent that sentences are records of human behavior, an LLM is a compact description of a giant corpus of human behavior. Given a "behavior prefix", we can predict the "behavior suffix" (and vice versa). 1/
2022-09-05 20:04:00 @mark_riedl Looking at this (esp. in peripheral vision), does anyone else experience the heart "pulsating" with each visual saccade?
2022-09-05 18:50:56 I retweeted this whole thread, but I wanted to call attention to this absolutely critical point. Our hardest societal problems (including, but not just, #natsec) require strong social science + humanities expertise coupled with technical expertise. https://t.co/EvOrZ0apNw
2022-09-05 17:50:56 RT @nadyabliss: Interesting analysis.TL/DR more humanities than #STEM majors regret their choices of major.STEM allows for effectively…
2022-09-05 17:28:08 @gileshooker @predict_addict @roydanroy @gilesh To me, what “hurts” are the missed/delayed opportunities. We need to ask ourselves what interesting research area is being ignored today
2022-09-05 17:10:26 @predict_addict @gileshooker @roydanroy @gilesh It is interesting that the early IJCAI conferences (late 1960s
2022-09-05 15:48:45 @predict_addict @roydanroy @gileshooker @gilesh Which Vapnik papers? Who is "They"? I haven't heard this history
2022-09-05 00:35:39 @roydanroy @gileshooker @gilesh What do you mean?
2022-09-05 00:22:54 @natfriedman This is just speech transcription. Dialogue involves much much more.
2022-09-04 22:09:16 @jjhikes @StevenGlinert @thebadfeeling_ Mechanics doesn't suffice. AlphaFold makes critical use of phylogenetic variation to infer candidate contact points
2022-09-04 21:30:40 @gileshooker Statisticians invented all of these terms: Jackknife, Bootstrap, Bagging, Lasso, Garrote. Although I'm happy to claim the Stanford and Berkeley stat departments as part of ML if @GilesH agrees
2022-09-04 21:27:28 @MoralOutrage1 @websterkaroon @MazMHussain It is automated, right?
2022-09-04 21:20:25 @jjhikes @thebadfeeling_ @StevenGlinert I love this use case. I think it is achievable, but it will require much deeper scene understanding than current models attain
2022-09-04 17:13:29 @mark_riedl Now do one for connectivity out of the Perceptrons book.
2022-09-04 17:11:16 @MazMHussain Yes, it was astonishing. This from March 2003 (start of war) https://t.co/vxvxpdxikW https://t.co/ecW47p0bSs
2022-09-04 05:31:05 @shortstein We’ve all been there! Good luck
2022-09-04 05:24:14 @feelsdesperate Wow. I thought DEI was about behaving like grownups and treating everyone with kindness.
2022-09-04 03:21:08 @TaliaRinger You are absolutely correct that I didn't stray from the traditional path. But I think what you are doing is much more significant.
2022-09-04 03:19:00 @TaliaRinger One way to say "no" would be something like "I'm sorry, but all of my service time is currently devoted to running the Computing Connections Fellowship and the SIGPLAN mentoring program. This leaves me no time to serve on X." This also reminds colleagues about your service work.
2022-09-04 03:15:39 @TaliaRinger Wow! Running an international mentoring program sounds much more impactful than conference/journal service! I hope your department takes this view. 1/
2022-09-03 20:11:31 @TaliaRinger Even those of us with long track records of service need to say no frequently. Also, tenure evaluation includes a service component, yes?
2022-09-03 03:09:27 @Grady_Booch or .fraud
2022-09-02 22:36:49 RT @oneunderscore__: I've been covering bad parts of the internet for long time now.For years, there was one site extremist researchers w…
2022-09-02 18:50:24 RT @geomblog: 3. Relying on the law is a little backwards in emerging areas. We have to figure out collectively what we believe is importan…
2022-09-02 17:07:17 @Patronsaintofd1 @Noahpinion This will greatly accelerate our understanding of disease processes and how those processes can be disrupted by novel drugs. It is revolutionary. Determining a rough protein structure used to take months-to-years, now it takes minutes.
2022-09-02 05:40:57 @sethlazar “Precipitates of the past”. Wow, this guy could write!
2022-09-02 03:54:50 @catherineols @kenarchersf @rao2z Check out the work by Helena Sarin @NeuralBricolage. She realizes some of her computer-generated art as ceramics. https://t.co/KSHFvOuNmd
2022-09-02 00:38:57 @GaryMarcus Question for everyone: What work trying to build dialogue systems on top of LLMs? (or is dialogue research pursuing other directions instead?). Not chatbots, but extended task-oriented dialogue
2022-09-02 00:34:19 @rao2z @kenarchersf Yesss!
2022-09-01 23:57:45 @kenarchersf @rao2z For example, inspired by Warhol, an artist could create a huge array of synthetic images as a commentary on the manipulation of modern society by bots or the "advances" of AI. Hmm, maybe I should give it a try?
2022-09-01 23:55:51 @kenarchersf @rao2z I think because the human artist controls and manipulates the AI image synthesis, it will be the human artist who uses the new medium to express their inner life and place their work within the context of all previous art
2022-09-01 21:04:21 @rao2z Yes, for formal portraits (although not completely
2022-09-01 20:52:13 @rao2z I wonder if AI-assisted art will follow the same path as photography. Rather than replacing painting, it became its own genre, and painting was indirectly changed as a result. The medium is (part of) the message
2022-09-01 18:03:15 RT @aihuborg: Get the highlights from @MihaelaVDS's talk at @IJCAIconf on her work in collaboration with clinicians to improve explainabili…
2022-09-01 15:44:47 @deliprao @ylecun ICLR was open from the start. It has been several years since NeurIPS and ICML became open. Of course, citation counts reflect the size of the field, and ML has become a big field.
2022-09-01 15:36:16 RT @tdietterich: @NickKristof I'd love to see the White House launch a "National Educational Catch Up" project. Let's create summer schools…
2022-09-01 15:36:12 @NickKristof I'd love to see the White House launch a "National Educational Catch Up" project. Let's create summer schools, after-school programs, etc. to get this generation of students back on track!
2022-09-01 15:34:04 @tasinger0001 @andreavhowe @NickKristof Yes. Teachers are essential workers, and just like the folks in health care and food production, some lives would have been lost in order that our children would thrive.
2022-09-01 15:29:26 @ChadNotChud Shouldn't there be two periods: the one you are quoting and the one ending your own sentence? As in: Nixon was lying when he said "I am not a crook.".
2022-09-01 15:18:54 @rudzinskimaciej @GaryMarcus I think the creation of "high art" will always be a human-centered process (with AI tools, of course). But I agree that the hotel rooms of the future will be decorated with machine-made art (if they aren't already)
2022-09-01 05:17:23 @GaryMarcus It is easy for a computer to create something novel. But hard to create something that is good art. The best art resonates with culture, history, and the human predicament.
2022-08-31 22:20:09 @adjiboussodieng Because my parents were educated, I didn't have to go through that mental revolution.
2022-08-31 22:19:28 @adjiboussodieng Oregon State has scholarships for first-generation-to-college students, but we need to do more on the mentoring side (and on many other aspects). It takes a huge mental revolution to think of yourself as an intellectual leader when your siblings are struggling every day.
2022-08-31 22:07:57 @adjiboussodieng I went to school on scholarships because my dad didn't make much money. But as a boy, I watched him struggle to finish his PhD, so I knew what it was like. (Of course, as a white male, no one was actively pushing me down either. Mentoring is not enough!)
2022-08-31 21:57:38 @adjiboussodieng I think it is also about the cultural insights that PhD-holding parents give to their PhD-seeking children. The children are "joining the family business". We need to mentor "outside the family" to reach everyone!
2022-08-31 17:53:16 @unsorsodicorda @rasbt That doesn't sound idiomatic to me. A company can be a "startup", but it is "founded" or "launched" or "spun up". Don't ask me why, I'm only a native speaker, so I have no idea why.
2022-08-31 17:42:40 RT @aihuborg: August has flown by! Here's everything we've covered: Dimitri Coelho Mollo is the latest in #NewVoicesInAI Awards galo…
2022-08-30 20:35:58 @zacharylipton @rasbt https://t.co/3BpC8LyiBc
2022-08-30 20:35:22 @zacharylipton @rasbt https://t.co/roB356nvzr
2022-08-30 19:39:38 @xfinity seems to have such a large outage that their status center web site isn't working. Anyone have any insight?
2022-08-30 19:00:26 @zeynep We listen to you!
2022-08-30 18:57:07 @rasbt @zacharylipton Just to clarify: The idea of an old-vs-new classifier is not mine. I don't know where I first heard about it, but similar ideas have been floating around for decades in the applied stats world
2022-08-30 18:25:38 @predict_addict @seanmylaw @rasbt @zacharylipton Thank you for this pointer!
2022-08-30 03:09:47 @tetraduzione @NeurIPSConf In the early days, there were tons of easy problems. Now we are making progress on the hard problems, but they are ... harder.
2022-08-30 02:37:53 @geomblog @boazbaraktcs @deliprao @ccanonne_ The answer should be "no". The law should ensure that the use of the data falls within the purposes for which data access was granted. It should impose a purpose-defined information bottleneck
2022-08-29 21:18:57 @ccanonne_ @boazbaraktcs I see your point, and I agree. The providers of LLMs must either provide a legal compliance guarantee or provide end users with meaningful tools for ensuring they are in compliance with the source data license terms.
2022-08-29 21:13:05 @ccanonne_ @boazbaraktcs The original motivation to allow Google to index your page was because it would help people find your page. Now Google has moved to synthesizing an answer to your query, but at least it gives you a link to the source. With LLMs, can it also link to the source?
2022-08-29 21:11:21 @ccanonne_ @boazbaraktcs Yet another reason that we need to solve the "attribution" problem: "Which training examples influenced this output?".
2022-08-29 19:53:42 @SeeBenjiRun @Noahpinion @PhilWMagness Time to invest in a second 220 outlet. I think you can get the IRA to pay part of the cost.
2022-08-29 19:47:27 @n_srishankar In other words, treat the outputs of the old classifier (preferably predicted probabilities rather than predicted class labels) as additional input features when training the new classifier. end/
2022-08-29 19:44:46 @n_srishankar If not, I think you would want to study methods for domain adaptation that do not require access to the training data. The first thing I would try is to "stack" a new classifier on top of the existing one. 2/
2022-08-29 19:43:02 @n_srishankar So your goal is to create a new (or updated) classifier that will perform well on future new data? First, I would run the old classifier on your new data and see if its accuracy is sufficiently high. If it is, you can just use it. 1/
2022-08-29 04:20:36 @krikamol @berkustun I came here to point to the same paper. https://t.co/gqamJEnsU7arxiv:1202.6389
2022-08-28 23:01:22 @n_srishankar Let me understand your problem better. Do you still have the trained classifier? Do you have new labeled data available or any old unlabeled data?
2022-08-28 22:25:04 @rasbt I like to use boosted trees with a logit link for this purpose
2022-08-28 17:48:41 @yudapearl @khademinori @ChrSzegedy Am I right that the key capability is to contrast two models: The original (opponent-independent) model and the revised model (updated to reflect the specific opponent and its weakness W)? 3/end
2022-08-28 17:46:12 @yudapearl @khademinori @ChrSzegedy But now suppose that Alphazero was trained against this particular opponent (e.g., by playing 100s of games against it). Now it has learned a model of how to play well against the opponent exploiting weakness W. IIUC @yudapearl, this doesn't qualify as counterfactual either 2/
2022-08-28 17:43:08 @yudapearl @khademinori @ChrSzegedy So my initial statement needs some refinement. If we are just looking at alphazero playing a game with learning "turned off", then it has no way of discovering weakness W in the opponent, and it can't answer counterfactual queries in @yudapearl's sense. 1/
2022-08-28 17:04:00 @VivaLaPanda_ @Noahpinion I think the issue is WHERE the water falls, not the total amount
2022-08-28 16:12:57 RT @zittrain: This @newsbeagle and @meharris piece describes a retinal implant used by hundreds of people to gain sight -- one that has bee…
2022-08-28 02:55:58 @ben_golub Agreed. So many important ideas, such as cross-validation, conformal prediction, and finite-sample concentration bounds, have been discovered in my lifetime.
2022-08-27 22:20:10 @yudapearl @khademinori @ChrSzegedy After the game ends, it can determine whether by making a different move, the game would have had a different outcome. @yudapearl, your examples usually involve latent randomness that is revealed by the factual experience. Does that extend to decisionmaking by the opponent?
2022-08-27 16:59:13 @wzuidema @chrmanning @soumithchintala @roydanroy @karpathy @ylecun @percyliang @RishiBommasani That suggests we are in for a rough ride.
2022-08-27 15:58:48 @khademinori @ChrSzegedy @yudapearl In my view, AlphaZero fits a causal model to its observations, so yes, it can answer counterfactual queries. It assumes a causal structure to the game and uses this structure when answering queries using MCTS
2022-08-27 15:22:06 @ChrSzegedy I used to subscribe to this view, but @yudapearl’s work suggests that we need to do more than compression to predict the effects of interventions and answer counterfactual queries
2022-08-27 03:39:10 @gideonmann Can it make and keep promises? Enter into contracts? Make new friends? Nah
2022-08-27 03:10:56 @deliprao I keep hearing good things about Rust. I plan to check it out
2022-08-27 03:10:10 @deliprao To be honest, it probably shortens my life span! But I suppose by reducing my productivity, it also reduces the amount of CPU time my programs consume :-)
2022-08-27 03:05:24 @deliprao I think they also need to consider the carbon footprint of the programmer. If it takes me 10x as long to write a program in C than in Lisp, I will have consumed 10x the calories and coffee :-).
2022-08-27 03:01:00 @rao2z @vardi @JmlrOrg @IEEE @Nature @ICAPSConference @demishassabis @jair @jmlr @nature isn't really a journal, it is a PR firm.
2022-08-26 21:24:26 @vardi @jmlrorg and @ieee PAMI are quite selective
2022-08-26 19:28:54 @gluonspring @IDoTheThinking I know retired scientists and engineers from industry who feel the same way.
2022-08-26 19:27:43 @gluonspring @IDoTheThinking There are even retired professors who would come back to teach, but the pay is so low that it is shocking
2022-08-26 19:23:14 @gluonspring @IDoTheThinking Yes!
2022-08-26 18:59:56 @gluonspring @IDoTheThinking For research faculty positions, this is true. But for instructor positions, the focus is on teaching experience. I think the main issue is low pay for lecturers.
2022-08-26 15:46:56 @jdcmedlock What we need is a "public option" for higher education. ... Oh wait
2022-08-26 05:15:19 @DavidAFrench This is an investment in a generation. It will result in folks starting new businesses and taking bigger risks. We all will benefit
2022-08-26 04:04:13 @BaldingsWorld What were once "state-funded public universities" have become "state-located universities" whose primary source of funding is tuition. I wouldn't say "extraordinarily", because most state universities have a huge deferred maintenance backlog. But our football stadiums look great!
2022-08-26 03:49:59 @matloff @BaldingsWorld Well, named professorships seem to be a thing at some schools. In the end, I'll be the X professor, in the Y hall, using the Z restroom in the W building in the V college.
2022-08-26 03:43:43 @matloff @BaldingsWorld Alumni and donors will pay for buildings but not salaries or even building maintenance
2022-08-25 20:57:15 Wonderful!!! https://t.co/hYfQGLmlvH
2022-08-25 20:55:14 @unsorsodicorda @dpkingma OCR and XCON are good examples. I don't think current spam filters match human performance yet. Targeted advertising is a nice example, but it also illustrates how AI go far beyond the original task to transform the entire approach.
2022-08-25 18:04:21 RT @tdietterich: @dpkingma Aside from game-like worlds (Go, Chess, etc.), are there ANY tasks that AI systems accomplish better and more ch…
2022-08-25 18:04:17 @dpkingma Aside from game-like worlds (Go, Chess, etc.), are there ANY tasks that AI systems accomplish better and more cheaply than human workers? To qualify, the human workers must no longer play any role the system. Please respond with pointers. Thanks
2022-08-25 00:20:57 With California banning new gasoline cars but not used ones, I wonder if a reasonable CO2 offset would be to buy used gasoline cars and destroy them. How would it compare to tree-based offsets?
2022-08-24 22:21:57 RT @FeiziSoheil: Adversarial examples, originally observed ~2014, were fascinating cases showing deep nets can be super sensitive to small…
2022-08-24 19:29:06 @wooldridgemike @terrible_archer @fchollet Another shortcoming of the "alignment" discourse is that it presumes that "alignment" is possible. Good engineering practice is to assume that we NEVER get the specs right, and we must consider how the overall social-technical system can compensate for this.
2022-08-24 19:26:25 @wooldridgemike @terrible_archer @fchollet Unfortunately, it isn't that simple. We have given AI systems control over social media, and this is raising serious planet-wide risks.
2022-08-24 19:23:27 @wooldridgemike @terrible_archer @fchollet I generally agree with Michael (as usual), but I think there could be x-risks if the AI system is given control over risky systems (e.g., nuclear weapons, biological weapons, etc.) But we should not give AI systems such access!
2022-08-24 19:07:13 @fchollet In particular, we seek a specification process that can tolerate/recover from specification errors. I am a big fan of the work that Stuart Russell has been doing on this. I would prefer to call it "agent-level specification analysis", but "alignment" is shorter. 2/2
2022-08-24 19:02:41 @fchollet When building some forms of AI software (e.g., RL and planning), we express the specifications in terms of goals and values. In such cases, the old problem of correctly specifying the function of a computer system becomes the problem of specifying goals and values. 1/
2022-08-24 14:30:57 @roydanroy @random_walker Nah, Geoff was overstating his case as he tends to do.
2022-08-24 04:50:58 @ScottTipple Well, every time you put on your mask, you have to check that it is sealing well. So it requires discipline and attention. What % of 5-year-olds can master it? As a former 5-year-old boy myself, I'm sure some could, but not many
2022-08-24 04:27:02 @taaltree Can you share a pointer to a good offset project?
2022-08-24 04:24:56 @kelley14419438 @HeyMmeG @Bob_Wachter @DrLeanaWen I am always wearing an N95 mask, and I haven't caught COVID yet. But most children I see are wearing cloth masks. And the studies I've read show that those are useless against Omicron. Of course I see a lot of noses and mask lifting/touching, too.
2022-08-20 01:15:05 @nerodenc @zeynep What monthly fee would you pay for such a service?
2022-08-19 22:35:16 @ryandcotterell @RishiBommasani @elgreco_winter Pardon my ignorance, but what is ARR?
2022-08-19 18:50:41 @ryandcotterell @elgreco_winter Leadership may be privileged or not, but either way, let’s be kind to each other
2022-08-19 15:56:45 @ShamezLadhani Might there be a dose-response phenomenon? If the long covid in parents led to greater doses in children, this might lead to more severe disease and sequelae. How much of the negative cohort were false negatives or failure to detect (because of test timing)?
2022-08-19 00:55:46 @adjiboussodieng A lobster roll is also not a sandwich https://t.co/itlFY8khOP
2022-08-19 00:27:59 @adjiboussodieng That does seem like a sandwich to me. As I said, I really have no idea.
2022-08-19 00:18:02 @adjiboussodieng As a “native eater”, I have no idea. But I agree that topology might be the reason. In a sandwich, the two pieces of bread are separated. In a taco and a hot dog, they are not. But I wonder if they arrived into English at different times so one meaning was established first
2022-08-18 05:03:54 @shmueli @ylecun @percyliang I think there have been some attempts to use "Ultra Large Scale Integration". But I don't think they have ever caught on. https://t.co/DobzU63ZnC
2022-08-18 04:27:05 @csabaveres @karpathy @stanfordnlp @soumithchintala @chrmanning @roydanroy @ylecun @percyliang Combining this idea with @apoorvkh's suggestion, we could call them "Large-Scale Corpus Models" (LSCMs) and drop the self-supervised method bit.
2022-08-18 04:23:46 @DavidZipper @NHTSAgov In my experience, high speed, reckless driving has increased during the pandemic. I recently drove up I5 in Oregon at 8am on a Sunday morning, and there were multiple crazy drivers on the road. Would dashcams help?
2022-08-18 02:36:41 RT @tdietterich: @NeelNanda5 @OpenAI Excellent analysis! It would be lovely to see more research of this quality
2022-08-18 02:36:37 @NeelNanda5 @OpenAI Excellent analysis! It would be lovely to see more research of this quality
2022-08-18 02:21:54 @cthorrez @ylecun @percyliang More than 2 hidden layers IMO
2022-08-18 01:52:57 @soumithchintala @chrmanning @roydanroy @karpathy @ylecun @percyliang @RishiBommasani Check out the etymology of jazz: https://t.co/RotA7McBSWOriginally meant "lively, energetic." Certainly less problematic than "foundation"
2022-08-18 01:10:38 @csabaveres @soumithchintala @roydanroy @karpathy @chrmanning @ylecun @percyliang Each new modality adds constraints. There are plenty of sentences that are ambiguous without an associated image.
2022-08-18 01:05:14 @markoff Fantastic to reread this!
2022-08-18 00:39:09 @soumithchintala @roydanroy @karpathy @chrmanning @ylecun @percyliang I didn’t want to engage without having a counter proposal
2022-08-18 00:38:05 @soumithchintala @roydanroy @karpathy @chrmanning @ylecun @percyliang That’s why I want to stay away from “foundation”. I worry that it is a foundation of sand. But I’m excited to see what we can build with these models. I’m especially interested in whether joint models of video, physical interaction, and language can get the causality right
2022-08-18 00:32:41 @soumithchintala @roydanroy @karpathy @chrmanning @ylecun @percyliang That’s not my view. I don’t have any flag to plant. My worry is that LSSMs are models of the shadows in Plato’s cave. While we can build lots of great stuff using them, they are likely to fail often because they don’t model the underlying causality
2022-08-16 21:28:56 @erikbryn I think Open Table might be using CF
2022-08-16 21:27:57 @zacharylipton Yes, and several other banned sounds. e.g., Barking dogs
2022-08-16 04:18:09 @BallouxFrancois @Jusrangers Prof. Balloux, I follow you to get a refreshingly different view on both the factual and policy questions surrounding the epidemic(s). Thanks!
2022-08-15 23:57:23 @BallouxFrancois @Jusrangers Twitter would be more effective if policy discussions focus on policy tweets and fact checking focuses on factual tweets. Personal attacks don’t persuade anyone
2022-08-15 23:52:14 @BallouxFrancois @Jusrangers I didn’t see any policy recommendation in that tweet. Isn’t ring vaccination the favored policy right now?
2022-08-15 23:41:18 @Jusrangers @BallouxFrancois Are saying that this statement is false? He may deserve your scorn, but not for this statement. Your tweet reveals more about you than about him
2022-08-15 23:03:16 @giffmana @francoisfleuret @ggdupont @percyliang My impression was that self-supervised is competitive with supervised in computer vision. Is this wrong? In particular, doesn't self-supervised permit training on much more data?
2022-08-15 21:13:31 @ISusmelj @TobyWalsh The fact that the model is making low-confidence, unstable predictions should enable a supervisory system to detect the problem
2022-08-15 17:52:38 @rao2z I am seeing some very cool art made by combining GANs or DALLE-type models with human creativity. For example, @neurobricolage (Helena Sarin). She designs digitally and then renders in ceramics. https://t.co/OLaI7fZrNu
2022-08-15 06:23:49 @unsorsodicorda @roydanroy @Sriraam_UTD @MaartenvSmeden I was thinking differently. Holding the network size fixed, the number of effective degrees of freedom increases with dataset size and complexity. In a sense, we overprovision network capacity and then only use as much as needed
2022-08-15 06:05:17 @Noahpinion I think Nuscale's first modular plant is supposed to cost only $3B. So that's 10x cheaper. But it suggests that your point stands.
2022-08-15 06:03:05 @unsorsodicorda @roydanroy @Sriraam_UTD @MaartenvSmeden How is this different? I think the key idea is that the number of degrees of freedom of a non-parametric model are chosen as part of the fitting process rather than determined beforehand.
2022-08-15 04:03:21 @LeeAdamWilshier @therealrukshan @asmithie90 Is there any evidence that this is happening? The cell is full of enzymes that degrade RNA, too.
2022-08-14 22:20:57 @unsorsodicorda @roydanroy @Sriraam_UTD @MaartenvSmeden I think this perspective ("NN as non-parametric models") is widely shared and very accurate.
2022-08-14 22:18:26 @sudhirPyadav @chrmanning @RishiBommasani @percyliang A language model is a probability distribution over the strings in a language. It is a well-defined technical concept. It explicitly about probabilities, so it is less likely to be mistaken for a deep knowledge base. "Foundation" invites that misunderstanding
2022-08-14 17:35:04 @sebjwallace @klausbrave Identifying to species level is hard unless you have lots of pixels on the bird. But video of the flight movements is probably also very informative
2022-08-14 17:30:41 @sebjwallace @klausbrave Yes. In fact Rich Caruana has worked on this, but I don’t know of anyone putting this into production.
2022-08-14 17:29:28 @chrmanning @roydanroy @ylecun @percyliang I didn’t know that “jazz” was marketing speak. Maybe it was?
2022-08-14 17:28:15 @chrmanning @ylecun @percyliang Isn’t “large” an adjective whose meaning shifts with time? Same with “big data” and “supercomputer”
2022-08-14 17:26:40 @rasbt @VictorButoi @percyliang I suppose it depends on the context. If you want to highlight the architecture, ViT is a great term.
2022-08-14 14:33:49 @csabaveres @percyliang or Large Corpus Model, LCM? That emphasizes the size of the corpus, which I think is the key
2022-08-14 05:20:26 @rishmishra I don't think that was a good thing for science and engineering. Only for lazy science writers.
2022-08-13 23:38:48 RT @sethlazar: 1. I think philosophers (and others) shouldn’t prioritise working on existential AGI risk. This is partly because I reject l…
2022-08-13 21:10:54 @ylecun @percyliang This is a revolutionary idea that overturns the standard practice of training only on data for a specific task, which has been the focus of ML and data analysis since the dawn of statistical modeling.
2022-08-13 21:09:35 @ylecun @percyliang The intent behind these system is to train them on a substantial fraction of the entire universe of available data so that they acquire knowledge of the full breadth of human activity. New software and hardware may empower academic labs to train them too.
2022-08-13 19:12:22 @zehavoc @percyliang This is another interesting proposal. An issue might be that in some systems, the core is actually a large database or corpus and the network is a summary and generalization of that corpus.
2022-08-13 19:10:30 @seth_stafford @percyliang I like "platform", another possibility is "substrate"
2022-08-13 06:19:34 @RishiBommasani @sudhirPyadav @percyliang The problem is hype, not comprehension
2022-08-13 05:55:33 @sudhirPyadav @RishiBommasani @percyliang Yes, that is exactly what I'm trying to do. The word "foundation" is too grandiose and smacks of the hype that infects too much of AI R&
2022-08-13 05:20:46 @percyliang Yes, but as you know, "Foundation" is too close to "Foundational", and many of us find that troubling. That is why I'm proposing a more neutral term. For use, maybe we could just call them "Upstream models".
2022-08-13 03:15:48 @joshua_saxe @percyliang Yes. But my guess is that self-supervision is the only way to scale to these immense data sets, so I think it is a safe term going forward.
2022-08-13 03:14:01 @RishiBommasani @percyliang In my view, "large" refers both to the model size and data size. But this is left to the interpretation of the user and reader. I'm sure you are familiar with the critiques of "foundation". I think we need a name that is descriptive and makes no claims about utility
2022-08-13 02:20:12 @frossi_t @percyliang Agreed. I wanted a term that could encompass models trained on multiple modalities: language + video + physical manipulation, etc.
2022-08-13 02:18:15 @rasbt @percyliang Don’t you think there will be improvements beyond transformers? I prefer a purely functional description
2022-08-13 01:17:35 @Noahpinion Do you know of any estimates of the current economic impact of lost time due to COVID illness? Are there statistics on hours worked per employee? How much of the worker shortage is due to COVID-reduced workforce?
2022-08-13 00:58:44 I propose that we adopt the term "Large Self-Supervised Models (LSSMs)" as a replacement for "Foundation Models" and "LLMs". "LLMs" don't capture non-linguistic data and "Foundation Models" is too grandiose. Thoughts? @percyliang
2022-08-13 00:16:53 RT @AdrianRaftery1: Our article, "Long‑term probabilistic temperature projections for all locations," w Xin Chen, David Battisti, @peiran_l…
2022-08-12 21:35:05 @narphorium @deliprao @github I was just trying to find some practical application for all of the academic papers on adversarial attacks on large models. :-)
2022-08-12 20:05:28 @deliprao @dsivakumar @iamtrask I remember feeling an immense sense of power when I realized I could add ANY two numbers of arbitrary length....my first glimpse of the power of algorithms.
2022-08-12 20:04:17 @daniela_witten @Lizstuartdc @raziehnabi @nataliexdean @BetsyOgburn @johnleibniz @analisereal @yayyyates Once you have flown with your own kids, other peoples' kids never bother you.
2022-08-12 16:59:54 @iamtrask @deliprao @github Yes, but the calculators are much more reliable than copilot as of today.
2022-08-12 03:49:29 @deliprao @github I see your point. I guess it depends on how reliable copilot-style systems are. Today, students need to learn how to check the correctness of the generated code.
2022-08-12 02:40:23 @deliprao @github Can we convince the copilot people to insert some backdoors so that copilot generates code with a specific flaw when given specific assignments? [Or, better yet, just rejects the request]
2022-08-12 00:18:21 RT @tommmitchell: Pleased to announce that my Machine Learning textbook is now a free download -- enjoy! https://t.co/F17h4YFLoo
2022-08-11 23:42:43 @Noahpinion My favorite line: "But why report any numbers at all if you don’t expect people to know what they mean?"
2022-08-11 22:13:27 @IDoTheThinking Rich people and business execs have a whole security team that manages their transport. It must be a strange and sad bubble to live in.
2022-08-11 20:51:09 @percyliang What is a chatbot and how is it different from a dialogue-based user interface? I'm interested in goal-oriented dialogue rather than natural-language entertainment.
2022-08-11 20:25:47 @KLdivergence @Dulles_Airport Do you find the map hard to interpret or find the airport layout hard to believe?
2022-08-10 21:37:37 RT @sethlazar: These days, there's a lot (too much) attention on philosophers highlighting 'longtermist' 'existential risks' supposedly pos…
2022-08-10 21:14:19 RT @paul_scharre: We're hiring! @CNASdc is growing our work on artificial intelligence safety and stability. We have 6 new positions open…
2022-08-10 17:48:12 Very interesting application of machine learning. It helps human video analysts focus their attention https://t.co/KiK5LyRucP
2022-08-10 00:14:24 RT @JAIR_Editor: Congrats to Nan Ye, Adhiraj Somani, David Hsu and Wee Sun Lee who have been awarded the 2022 IJCAI-JAIR best paper prize f…
2022-08-09 19:14:07 @rajiinio …which is not enough to cover the full PhD duration. It is frustrating.
2022-08-09 19:13:02 @BeltwayMap @Mathbarra @TheSotaSwede @stefista @_sadwaffle In my state, heavy trucks do the most damage followed by cars with studded tires. Weather causes all road surfaces to age. There are some paved bike paths around here that are quite hazardous as a result.
2022-08-09 17:34:53 @rajiinio Yes, that is an especially difficult situation for funding. As a professor, I’m always chasing funding. Federal grants are mostly not restricted to citizens or PRs, and they usually last 3 years
2022-08-09 16:12:08 @rajiinio There is no long term funding for research.
2022-08-08 22:23:27 @wilgieseler @TheSotaSwede @stefista @_sadwaffle What about Federal taxes? The Tax Foundation has a mixed record according to Politifact https://t.co/FOwfIJWxPV
2022-08-08 21:10:55 @TheSotaSwede @stefista @_sadwaffle Drivers do pay through gas taxes. And bicyclists need good streets too. (Ideally, safe, car-free streets!)
2022-08-08 20:50:23 @ylecun @JefferyAGoldst1 @GaryMarcus @miguelisolano Another factor to consider: dimensions of variability in the input. Learning is more successful when there is more variability along many dimensions. Playing with other people and interacting with the world gives more useful variability than passive observation.
2022-08-08 18:41:43 @RitaKonaev How much is due to insufficient supply of air traffic controllers?
2022-08-06 21:03:05 @BallouxFrancois It means schools open but either (a) unstaffed or (b) parents too afraid to send their kids to school. But you knew that. (And I agree with your position on school closings.)
2022-08-05 18:27:47 @balajis The US Ambassador to China R. Nicholas Burns doesn't speak Mandarin (according to his wikipedia page)
2022-08-05 16:22:51 @deliprao The success of automated translation is a dream come true! That UX is super cool and is a great learning tool, too.
2022-08-05 16:20:49 RT @markdalgleish: Being a self taught developer is pretty exciting because you're always being told that the code you wrote is actually us…
2022-08-05 03:24:59 @winequant @Noahpinion I think in this arrangement the sitter is still an independent contractor (https://t.co/npat13CBLB). If there are parts of the tax code that are unclear, Congress should clarify them.
2022-08-05 03:20:03 @mlittmancs @rao2z As @Bob_Wachter says, it has never been easier to catch COVID than it is right now.
2022-08-04 16:04:07 @jdcmedlock I stand with law enforcement!
2022-08-04 05:16:49 @francoisfleuret @j2bryson @blaiseaguera @marian_croak @dpetrou @AlisonLentz @bratton @jakesoll I agree with Joanna that much of our evolutionary ethical inheritance is in the "hardware".
2022-08-04 05:15:43 @francoisfleuret @j2bryson @blaiseaguera @marian_croak @dpetrou @AlisonLentz @bratton @jakesoll My hypothesis is that social interaction is necessary. Evolution has tuned us to make ethical decisions because they were successful in the long run. Maybe multi-agent game theory can help machines attain similar behavior?
2022-08-04 04:05:48 @blaiseaguera @marian_croak @dpetrou @AlisonLentz @j2bryson @bratton @jakesoll People are not able to articulate (linguistically) their values. Philosophers have been trying to do this for >
2022-07-27 03:23:06 @Grady_Booch Primrose managed to get and hold a security clearance for many years
2022-07-26 22:33:20 @johncarlosbaez @KyleCranmer @Twitter I believe you can do this by creating a List of the people you want to read.
2022-07-26 22:09:39 @ccanonne_ @roydanroy His papers are definitely important to read.
2022-07-26 20:04:17 @deliprao Reading the bios of people nominated for AAAI Fellow gives me “career envy”
2022-07-26 04:59:54 @trevor_strohman @cheer_wine @jdcmedlock Ouch!
2022-07-26 03:36:44 @Noahpinion I think it is a nonstationary, but memoryless, process.
2022-07-24 14:52:27 @negrosubversive There is another option: the free standing small town. Housing (and car use, alas) is like the burbs, but people know each other. Shopping for food can take a long time, because you stop to chat with everyone you know.
2022-07-24 14:37:23 @rao2z Yes! The real issues are about the correctness of the knowledge and the cost of maintaining correctness over time. Both approaches have weaknesses, but they are often complementary, and so benefit from hybridization
2022-07-24 14:12:55 @zittrain @KalieWertz Do we need a “right to disconnect”?
2022-07-23 15:32:17 @zicokolter Maybe causal modeling can give us a formal understanding of this. But it might work even in non-causal settings.
2022-07-23 15:30:58 @zicokolter It makes sense that if your original model was confident on the training data and the distribution shift is not too large, then you want to find the nearest model (in some magical sense) to your original model that is confident on the test data. "magical" = conjugate
2022-07-23 15:24:00 @zicokolter Very interesting. Did you do experiments with noise in the features of the test distribution? Could you force TENT to overfit in such cases? Maybe we can detect this case via other mechanisms?
2022-07-22 13:52:47 RT @aihuborg: Our Managing Editor Lucy Smith will be attending @IJCAIconf. Get in touch if you'd like to meet to share your research or fin…
2022-07-22 13:52:07 RT @Noahpinion: An intrepid team of researchers has been studying why the U.S. has such trouble building trains.@ericgoldwyn, one of the…
2022-07-22 13:46:07 @rajiinio @AndrewYNg 20+ years ago, according to Piero Bonissone, GE was already using neural networks for loan risk assessment. The training data was controlled by a committee that voted whether to include each data point to ensure legal compliance.
2022-07-22 13:43:45 @rajiinio Is this what Andrew Ng @AndrewYNg means by data-centric AI? The data in ML provide most of the problem specification, so careful managemt of the data is required to ensure the correct specification.
2022-07-21 14:03:55 RT @trvrb: There seems to be a worry that telling people we've exited the "pandemic phase" will lead to further reduced precautions. As alw…
2022-07-20 23:34:48 @MazMHussain I suspect that the collapse of the twin towers was more astonishing than the death toll. Two iconic and very solid things disappeared.
2022-07-20 15:39:06 @siminyu_kat Fascinating!
2022-07-20 13:07:19 RT @bschoelkopf: Our 2012 paper ‘On causal and anticausal learning’ just received a Test of Time Honorable Mention at @icmlconf #ICML2022:…
2022-07-19 19:28:30 @Eli_Schwartz @agrover112 @thegautamkamath They didn't compare against max logit or Open Hybrid. Our group has never gotten good results with knn in the open category setting. Reviewers need to insist on better experiments.
2022-07-18 19:31:43 @kphextwin @Noahpinion And Oregon, thank goodness!
2022-07-18 15:27:20 Is it possible that @RichardSSutton and Jack Dongarra are the same person? I've never seen them in the same room together. Hmmm... https://t.co/gfSxhWj2AI
2022-07-18 15:25:07 @zendaimyo @MazMHussain You also have media companies whose business model requires distortion of the news. Even the best companies engage in click bait advertising, which erodes trust.
2022-07-18 15:20:08 @rryakitimbo1 @dsa_org @mozilla Wish I could be there!
2022-07-18 13:23:59 RT @aihuborg: We'll be (virtually) at #ICML2022 next week. If you are attending the conference and would like to write a blog post about yo…
2022-07-17 23:00:42 @JGatsby1923 @ryxcommar If they observed 0 bots in a random sample of 100, a 95% confidence interval on the true proportion of bots would be [0,0.036] (Clopper-Pearson "exact" test). As others point out, they will have a sample of more than 2000 within a month.
2022-07-17 16:29:00 @joao_semnome9 @random_walker It doesn't tell you anything about the labels, but the test set will no longer give you an accurate estimate of performance. If you use the test set for PCA, binning, or normalizing, your "test set" performance will be biased compared to fresh data
2022-07-17 14:36:52 @rasbt Does anyone apply differential privacy methods to obtain re-usable validation sets as in https://t.co/t9qTitAw1l
2022-07-17 14:34:30 @rasbt Agreed. On problems with scarce data, it is best to build a data simulator and make design choices on simulated data.
2022-07-17 14:31:33 @joao_semnome9 @random_walker I strongly disagree. Every ML/statistical procedure (whether supervised or unsupervised) is vulnerable to underestimating variability in the inputs. If you keep the test set separate, you obtain an independent estimate of the effect of variability.
2022-07-17 03:38:51 @davidminpdx Time to brand them for what they are: Extremists without compassion.
2022-07-17 00:31:50 @rasbt @AlexGDimakis I tell my students that the goal of research experiments is to minimize the time required to get an answer we can trust. This is not usu the goal of software engineering, but it involves many of the same practices
2022-07-16 21:14:49 @Noahpinion Love the joke, but there is also an amazingly vital literature in Spanish, only a small fraction of which is translated into English. (Oh, and Brazil would definitely be a member of the Telenovela Empire.)
2022-07-15 23:12:32 @RMac18 @yangyang_cheng What’s weird is that the Apple person looks like they are photographing us. They may be hidden, but we are captured
2022-07-15 07:51:33 After experiencing long security delays at AMS, I think we Americans should pause for a moment and congratulate @TSA on making it through the pandemic without huge security problems.
2022-07-15 07:24:33 @BallouxFrancois @zeynep @RidleyDM This is anecdote, not evidence. In any case, if we are going to be fighting these airborne infections for a long time, we need to consider how we can improve our infrastructure and our social norms. Repeated infection is a drag on the economy AND the society
2022-07-15 07:21:27 RT @EricTopol: The BA.5 variant and what we should do about it. My @latimes oped today https://t.co/JjVsFmTJtO https://t.co/jIyjPTK5m8
2022-07-15 07:10:29 @boulderfish @SonnyCrockett04 @BallouxFrancois @zeynep My experience is merely an anecdote and proves nothing. We know from health care workers that even careful masking with high quality masks is not a perfect defense. But it is not useless either. We must live with courage.
2022-07-14 21:27:06 @SonnyCrockett04 @BallouxFrancois @zeynep And if you aren't going to enforce a mandate, you are damaging your credibility for future cases where public health mandates might be essential.
2022-07-14 21:25:28 @SonnyCrockett04 @BallouxFrancois @zeynep On the other hand, I just got off a Lufthansa flight that MANDATES N95 level masks
2022-07-14 21:23:51 @SonnyCrockett04 @BallouxFrancois @zeynep I'm just wrapping up a 2 week trip to Europe from the US. I've been wearing N95 masks. I have been in the presence of two people who were likely contagious at the time. Both tested positive within 48 hours after our meeting. I have not gotten it (yet). Survivor bias? Maybe
2022-07-14 15:13:49 @zeynep It would be great if some neutral organization could audit such claims and provide a seal of approval.
2022-07-14 15:12:50 @zeynep I think we will be battling this virus family for a long time. I will be watching for restaurants and theaters that provide a ventilation guarantee before I patronize them.
2022-07-14 15:08:56 RT @magi_jay: My initial reaction to this piece was that it was "ok" to fact-check the story, but that the article was poorly executed, par…
2022-07-12 21:18:21 RT @etzioni: Please RT: Hundreds of tech, business and nonprofit leaders urge states to boost computer science education https://t.co/UIiX1…
2022-07-12 20:41:24 @KLdivergence @matloff I was replying to Norm, but messed up the threading, sorry
2022-07-12 20:40:33 @KLdivergence @matloff
2022-07-12 20:39:20 @KLdivergence I think this applies to all applications. If you see a shift in subgroup accuracy, you will want to understand what’s happening
2022-07-12 20:38:22 @KLdivergence Checking accuracy on subgroups is a basic diagnostic, just like plotting your regression residuals and checking for patterns. Very important to automate this and recheck after every model update (for models in production)
2022-07-11 17:58:41 RT @fchollet: This is a joke, but it's also not. You know there *will* be benchmarks for joke understanding, and LLMs *will* get superhuman…
2022-07-11 09:46:33 @realeigenvalues @InCogni87402288 @IDoTheThinking Certainly intuition can come first, but I have lots of "insights" that turn out to be bogus when I check the math.
2022-07-11 09:45:16 @BallouxFrancois I must have chosen a bad reference. I only mean to say that public health messaging is complex
2022-07-11 08:46:43 @wellingmax Exactly! This kind of scare talk is absurd. Without physical constraints, ML will smooth and extrapolate badly.
2022-07-11 08:44:50 @BallouxFrancois At least 2 M women contracted AIDS.
2022-07-11 08:12:03 @InCogni87402288 @IDoTheThinking You need both inductive and deductive reasoning in science.
2022-07-11 08:09:16 @JosephKahn @Noahpinion This is wonderful. However, with so much absenteeism due to COVID, what is the effective reduction in the size of the workforce? Is it an important contributor to inflation?
2022-07-11 08:06:39 RT @rao2z: There seems to be an almost willful confusion about the need and role for explainability of #AI systems on #AI twitter. Contr…
2022-07-10 20:26:06 @DrAEHernandez @fernandaedi @EricTopol @phylogenomics @SarahNev @jburnmurdoch @MetroCDMX I'm in Spain right now. Masking is required on public transport, taxis, airports. However, I would estimate no more than 20% are wearing good N95/KN95 masks. Lots of noses visible among the surgical mask/cloth mask wearers.
2022-07-10 18:30:29 @NickKristof Of course, life began when life began (about 3.7 billion years ago). Every one of us traces our ancestry back to that beginning. Reproduction and development are complex
2022-07-10 18:11:27 @vardi COBOL was my third language, after Fortran and Sigma 9 Assembly. "picture" statements and "move corresponding" are some of my favorites. I had fun trying to make the code read like English. Similar to the challenge of making prolog read like logical inference!
2022-07-10 15:23:42 @Ar_Douillard That is bad editor behavior. At some point, the editor needs to read the paper and make a decision. When I was a young editor, a senior reviewer (the wise John Holland) gave me some good advice about this.
2022-07-10 10:11:45 @roydanroy @Meaningness @Ar_Douillard I think we want to create agents that engage in a dialogue to achieve goals, not just prompt-answering machines. Search engines can answer prompts, and they are more reliable than LLMs, because they don't make stuff up.
2022-07-10 09:21:59 @Meaningness @Ar_Douillard Agreed. But what is missing from both accounts is a description of the underlying communicative goals and how those connect to the assembly of fluent utterances. Today, we have the fluency, but it is goal free.
2022-07-10 09:01:12 Question of the day: Which will arrive first: AGI or @microsoft Outlook correctly remembering the user's customizations of the ribbon? @Office
2022-07-10 08:40:12 @TobyWalsh Basically, the article confirms that the iPhone hardware is much more capable that the Apollo Guidance Computer. The software, on the other hand, is bloated and unreliable
2022-07-09 16:53:18 @lazowska Yes, spin off college sports as commercial companies! They warp the academic enterprise and, from time-to-time, require big injections of cash from the state.
2022-07-09 16:49:37 @gdb *it. I'm using a different phone today and it has terrible autocorrect.
2022-07-09 15:35:47 @matloff @daniela_witten I guess that's why I'm always getting tripped up by the differences in indexing and accessors. I keep trying to use "." in my python variable names too :-)
2022-07-09 15:35:01 @matloff @daniela_witten I agree. I train my deep nets in python, then export the data and analyze it in R.
2022-07-09 15:34:06 RT @tdietterich: @MaxCRoser Regarding limits, the big open question in my mind is to what extent does understanding culture, history, langu…
2022-07-09 15:33:51 @MaxCRoser I prefer not to think in terms of specific "tasks". Indeed, I think the AI research community has over-valued "matching human performance on tasks" rather than understanding the basis of intelligent behavior. end/
2022-07-09 15:32:41 @MaxCRoser But if such understanding through self-projection is not required, then it should be possible to program computers to read/watch/listen to all of the products of human culture and understand them, to observe social interactions and participate in them, etc. 3/
2022-07-09 15:30:12 @MaxCRoser If this relies critically on our internal experiences and our projection of those to reason about other people, then digital computers may be at a disadvantage. We will need to program good simulations of these experiences, which is difficult, as they are unobservable. 2/
2022-07-09 15:29:04 @MaxCRoser Regarding limits, the big open question in my mind is to what extent does understanding culture, history, language, emotion, and social interaction rely on the physical and psychological experience of being human?
2022-07-09 08:11:28 @dustinvtran @zacharylipton Very common: UCSD, UCSC, UCI, UCB, UCLA, ...
2022-07-09 08:10:12 @matloff @daniela_witten I was thinking more of notations like [,1] and [:,1] not to mention the variety of accessors in R: $, @, [], [[]]. I guess each of them made sense at the time...
2022-07-09 06:33:12 @matloff @daniela_witten And you will curse the designers of both for their indexing design choices
2022-07-08 16:05:49 @Usman_skhan @jmhernandez233 @roydanroy @RealAAAI No journal that I've been involved with (Machine Learning, JMLR) has had this problem, to my knowledge.
2022-07-08 10:30:13 @gdb What program will is want to write? Programming, like writing, is in part a process of figuring out what we want the program to do.
2022-07-08 10:27:08 Data point for travelers: Wait today at Schipol terminal 1 was 80 minutes. My guess is 2/3 of the scanning stations were operating
2022-07-08 10:22:48 @jmhernandez233 @roydanroy @RealAAAI locked down authorship several years ago to address this.
2022-07-08 05:56:13 @LeonDerczynski I thought author changes are not permitted after submission
2022-07-07 16:53:13 @LukasGalke @ICBINBWorkshop Yes, the recording will be posted to Youtube
2022-07-07 06:16:59 @DavidZipper Does lane-keeping assistance reduce accidents? I have always found it very annoying and disabled it. But I'm curious about any careful studies of effectiveness.
2022-07-06 21:20:46 I’m looking forward to presenting our work on deep anomaly detection. We recently posted a substantial revision of the paper on arxiv https://t.co/0KUqdPh6dH https://t.co/TkQFL0jtYJ
2022-07-05 19:00:32 @SergeBelongie It's so wonderful when you are in a new place and it starts listing off species you've never heard before. It makes learning the birdsongs easy
2022-07-05 15:51:43 @Abebab @emilymbender Maybe we could form a team of people and take turns engaging with reporters? I agree it is exhausting
2022-07-04 21:22:50 @roydanroy @natefoster Cdr and car?
2022-07-04 21:22:11 @mraginsky The big theater is where the lamp post is located
2022-07-04 21:20:28 @emjotde It’s true. Many of us do
2022-07-04 11:35:17 @roydanroy I flew SFO to AMS yesterday on United without problems.
2022-07-04 10:11:06 RT @bigmlcom: Happening now at the #DutchMLSchool: #LessonsLearned Applying #AnomalyDetection at Scale, by @AlvaroClementeV, #MLengineer at…
2022-07-04 08:38:33 Beautiful work from @MercedesBenz in collaboration with @bigmlcom to empower production engineers to develop their own ML applications. Huge application of anomaly detection to detect and correct welding problems. Led by Abdelghani Khaldi https://t.co/QgarC1Kq7d
2022-07-04 04:20:40 @sarahookr When we were building the CALO "AI Assistant" back in 2006, we were saving all interactions (email, speech, desktop activity). Our system would exhaust memory in 4 days. Knowing what to ignore is the key to "sustainability" of AI systems
2022-07-03 10:55:32 @Sobheomiid @fchollet Never. Use a validation set instead.
2022-07-03 10:53:17 @unixpickle I agree that RL is smoothed out search. And search is a form of reasoning when it uses a model. What else could it be? Obviously the problem space matters, as any AI textbook will tell you
2022-07-02 22:50:10 @rao2z Maybe
2022-07-02 20:34:44 I think it is time to rethink in person conferences https://t.co/X62rW6KCpF
2022-07-02 20:14:54 @DavidSKrueger Maybe it is inductive skepticism. During my whole time as an AI researcher, there have been advances described breathlessly as breakthroughs that turned out to be only incremental advances. The little AI Hype boy has cried wolf too many times.
2022-07-01 23:33:11 @erikbryn @percyliang @ilyasut @GaryMarcus @fchollet @JeffDean @DigEconLab This reminds me of a statement Stuart Russell once made. "A computer that simulates playing chess is actually playing chess". That applies here as well. But this is true only of formal, symbol-manipulation tasks, ironically.
2022-07-01 18:47:13 @IDoTheThinking Do you know how many people need to be within walking distance of a small grocery to make it financially viable?
2022-07-01 18:24:58 @MFordFuture @vardi Complete nonsense.
2022-07-01 17:00:14 @morris_frank_ @deliprao I still need to be convinced that these models are not just bad search engines that produce fluent sentences. Google can retrieve the wikipedia page for "overfitting", but that doesn't mean it understands the concept.
2022-07-01 16:09:24 @morris_frank_ @deliprao It can produce the right words. But how could it turn those words into actions? Let’s ask it what parts of its knowledge are the result of overfitting
2022-07-01 16:06:26 @IJCAIconf @rao2z But I agree with Rao here. Her intellectual contributions are not in deep learning. Aren’t they important? (And to whom am I speaking?)
2022-07-01 15:59:06 @replyallguy @Noahpinion Because traffic is a flow in a network , adding lanes can change which roads people choose to drive on. And traffic on some roads can increase even if no cars are added to the system. https://t.co/A4T6pmFzvS
2022-07-01 15:55:01 @deliprao More like just quoting the Bible. Does the system even know what overfitting is? No, it just knows the textual contexts in which the word is used
2022-07-01 15:53:14 @giffmana This is the same reason not to trust the original code either. Two independent implementations debugged against each other ftw.
2022-07-01 04:10:28 @3rdonthematch @ClaudeMotley @michaelharriot I agree she hasn't redeemed herself yet. But I hope she is on a path that leads there.
2022-07-01 00:32:00 @olcan @percyliang @RishiBommasani Isn't it surprising how confusing those two words are? Almost like it was a poor choice for that very reason...
2022-07-01 00:01:11 @percyliang @RishiBommasani Many of us think the reason it is still controversial is that you are still calling them "foundation" models. :-)
2022-07-01 00:00:10 @AiSimonThompson @deliprao That is obviously circular reasoning, but I wanted to move a step beyond Blum &
2022-06-30 23:58:48 @AiSimonThompson @deliprao Suppose the meta-level is a system for causal discovery and causal reasoning. It would naturally try to perform interventions on itself. Of course the conclusion (that it has free will) is presupposed by its ability to treat its experiments as interventions. 1/
2022-06-30 23:57:17 RT @brandonsilverm: I've been offline for most of the last week but thought I'd jump in with a few thoughts about the article below. belo…
2022-06-30 19:18:22 RT @maosbot: What is Probabilistic Numerics (PN)? To illustrate, take one core use case of PN— computing integrals. Most integrals are intr…
2022-06-30 19:12:59 @predict_addict @unsorsodicorda @Nils_Rethmeier @Aaroth @rohitrango @uai Thanks
2022-06-30 05:32:01 @michaelharriot Maybe not "hero". But it took and will continue to take a lot of courage to face down the journalists and the MAGA mob. Don't any of you believe in redemption? Or do you just want to pile on too?
2022-06-30 05:25:20 @adad8m @thomasahle @thegautamkamath Interestingness of the research, easy of entry to the field, ease of doing publishable research, job market after getting your degree, salaries of those jobs
2022-06-30 03:39:37 RT @tdietterich: @deliprao Very nice paper. I would add that a CTM that also observes and reasons about itself can conduct experiments to t…
2022-06-28 02:43:58 @Aaroth @rohitrango Exactly. Calibrated probabilities, prediction intervals, OOD detection, abstention, etc.
2022-06-27 22:09:39 @EthanJPerez Even curves like the one of the left should be ok IF the amount of data needs to scale exponentially. Perhaps limit the data scaling somehow?
2022-06-27 19:29:43 @Noahpinion What is the carbon footprint of building the runway capable of supporting takeoff and landing?
2022-06-27 19:08:43 Interesting @bing behavior. Is this wise, @microsoft?https://t.co/soI0qzL6JP
2022-06-27 18:59:15 @POETSorg "eyes closed when they should have been openeyes open when they should have been closed"
2022-06-27 05:41:38 @butts_4_jesus I’d love to have this kind of info for flu and cold viruses too. Maybe we’ll have it in a few years
2022-06-27 05:39:29 @butts_4_jesus Well, suppose I m traveling internationally and want to estimate my risk of getting sick and stuck. Knowing the prevalence of different strains in different places can help.
2022-06-26 16:24:23 @florian_tramer @thegautamkamath @rasbt AI2 @allen_ai is an endowed research organization isn’t it?
2022-06-26 16:14:43 @rasbt @thegautamkamath I don’t think so. With a profile, you can clean up all of the mistakes: merge publications, delete ones that you didn’t write, etc.
2022-06-26 16:11:09 @Ar_Douillard Yes!
2022-06-26 16:02:13 @florian_tramer @thegautamkamath @rasbt @SemanticScholar
2022-06-26 16:00:05 @florian_tramer @thegautamkamath @rasbt “Pet project”?
2022-06-26 15:59:37 @florian_tramer @thegautamkamath @rasbt When I chaired the AAAI Fellows committee, we carefully normalized. 1000 citations in computer vision is roughly 500 in machine learning and 30 in AI planning, IIRC
2022-06-26 15:56:43 @florian_tramer @thegautamkamath @rasbt I like the metrics. Like all metrics, they must be interpreted with care. It would be nice to normalize by the size of the scientific community in which each paper appears.
2022-06-26 15:54:07 @jacyanthis @russelljkaplan The LLMs show much more evidence of memorized associations than of causal understanding. I am eager to see more analysis of the learned representations
2022-06-26 14:42:31 @thegautamkamath @rasbt Dblp + semanticscholar ftw
2022-06-26 14:27:32 @russelljkaplan I agree that more data would help, but we need better representations. Representations of the causal structure that produces the data. Transformers are not discovering them.
2022-06-26 03:32:22 @andyroyle_pwrc As someone who works at another OSU, I think Ohio State should be required to change their domain name to https://t.co/LB2L84UbkY and relinquish https://t.co/ZVCttwcDAw
2022-06-25 20:30:12 @brewster_kahle Because babies smell wonderful
2022-06-25 18:28:12 @kchonyc and I would like to live in a country governed by the principle of one person-one vote
2022-06-25 01:29:16 @rajiinio @arxiv uses iThenticate too, I believe. So does AAAI, IIRC
2022-06-24 23:57:04 @TobyWalsh @BBC_Future Aha! That must be the secret to your productivity!
2022-06-24 20:01:46 From @zeynep. "I am not a technophobe. [...] I’m often an early adopter of tech, and get enthusiastic for its many potential uses. But I’m also a sociologist studying authoritarianism, and our digital infrastructure has become the infrastructure of authoritarianism." https://t.co/pQG0V78JHq
2022-06-24 05:38:27 @thegautamkamath @CVPR I don't know the answer, but you can certainly report it to arXiv
2022-06-24 00:17:56 RT @sethmoulton: Just so everyone knows, these are the states - including Massachusetts - that had one of their gun laws overturned today.…
2022-06-23 21:50:09 @richardtomsett @histoftech DARPA has funded several relevant programs (Assured Autonomy, SAIL-ON, CAML, etc.). Several million dollars have been invested. Maybe the issue is one of tech transfer, of getting this stuff into production? @bigml is working in this space.
2022-06-23 16:15:22 @histoftech Workshop at last year's ICML: https://t.co/L7fiCBbVhx
2022-06-23 16:14:46 @histoftech Upcoming KDD workshop: https://t.co/toi1Yc2Rxc
2022-06-23 16:14:03 @histoftech There is a very active research community on uncertainty quantification and out-of-distribution detection in machine learning. ICML workshop: https://t.co/MU9DQObXBJ
2022-06-23 16:08:51 @rao2z @roydanroy @_onionesque My advisor told me to only put into the future work section the work that aspects I did not intend to work on. Certainly don't make any promises!
2022-06-23 15:48:38 RT @roydanroy: @_onionesque Future work statements are best avoided in my opinion.
2022-06-23 03:28:57 @ChillAnonDragon @JesseDodge This is entirely irrelevant. The deep learning frameworks compile to cuda.
2022-06-22 16:37:04 @roydanroy @thegautamkamath @dustinvtran @ledell That's why I focused on how the information is used rather than permission to read it. But the ability to do probabilistic inference at scale may change society's view of this and lead to disallowing certain inferences (e.g., about diseases, pregnancies, etc.).
2022-06-22 16:33:03 @roydanroy @thegautamkamath @dustinvtran @ledell If you read a web site and then uses the information so-acquired to perform an illegal action, that is bad, right?
2022-06-22 16:24:21 The only thing more annoying that constant software updates is constant updates to the legal terms and conditions. @paypal just updated for, what, the third time this year? Frequent updates = dark pattern to discourage careful checking of the terms.
2022-06-22 16:22:17 @roydanroy @thegautamkamath @dustinvtran @ledell I think we, as a society, are still trying to understand what the terms of use should be for web sites. The acceptability of scraping to fit a model must depend on how that model will be used. Some uses are already illegal
2022-06-22 16:04:50 @ecsquendor @dustinvtran @ledell Society has long recognized the need for laws and norms, and the right for all members of society to participate in creating and enforcing them.
2022-06-22 05:21:55 @dustinvtran @ledell This letter is aimed at the entire R&
2022-06-22 04:18:04 @AlexGDimakis @tallinzen Yes, it means you were lucky and had a positive outcome. Also implies that you were not expecting it.
2022-06-22 04:16:53 @AlexGDimakis @tallinzen I still remember when I moved to New Jersey as a 12 year old and heard people using “lucked out”. I had no idea what it meant. So much for compositional semantics!
2022-06-22 02:47:23 @tejuafonja I looked at the Federal Register for this, since I had never heard of it. It seems that the intent is to reduce the brain drain from developing countries toward the US. I can see why some people would regard this as an anti-colonial rule. I don't know what to think...
2022-06-21 18:00:45 RT @percyliang: There are legitimate and scientifically valuable reasons to train a language model on toxic text, but the deployment of GPT…
2022-06-21 03:02:19 @roydanroy If Deep Learning didn't work so well, we would definitely not use it.
2022-06-21 03:01:46 @roydanroy Did we really say non-convex was wrong? It was just awkward and inconvenient. It still is.
2022-06-20 22:22:48 @AnikaFreeindeed Yes! Our challenge is to find a way to govern ourselves with kindness and respect
2022-06-20 21:03:32 @AmandaAskell I put a lot of weight on their learning ability rather than their knowledge or skill. They are already learning in utero.
2022-06-20 19:32:10 @Noahpinion How many times have you had it, @Noahpinion? It is easy to catch in SFO these days.
2022-06-20 19:04:29 RT @benbendc: Guidelines for journalists and editors about reporting on robots, AI, and computers https://t.co/gorYo1VZib @hcil_umd @umdcs…
2022-06-19 21:47:00 @ducha_aiki @arxiv @CVPR I just sent an email to the arXiv folks. Monday is a holiday, so I think their help desk is not operating tomorrow. Hopefully someone will see my emails.
2022-06-19 20:51:22 @dwf Yes, but I'll bet she isn't trying to show a powerpoint presentation with presenter mode on her screen and regular projection on the zoom screen and an in-room projector.
2022-06-19 18:58:23 It's interesting to watch industry folks (who use their own corporate telecon systems) struggle with Zoom. I'm sure I'd struggle when learning to use their systems too. It's just evidence of subcultures in tech.
2022-06-19 18:07:46 @IDoTheThinking This isn't strictly true. There is mandatory retesting in several states once you get old. But I agree that you should have to recertify just like pilots do. It would cost a lot to scale up testing capacity. Would the insurance companies pay for it?
2022-06-19 05:27:57 @Noahpinion Google is fighting a continual battle against spammers and vice versa. What is the equilibrium? Does web search become unusable or does spam have to start producing useful information?
2022-06-18 16:56:26 @rgblong Excellent thread
2022-06-18 05:52:05 @rao2z @GaryMarcus The LLM should be able to score the probability of the prompt. So maybe it will be easy to check for weird prompts?
2022-06-17 04:17:32 @GaryMarcus This incoherent prompt is presumably "out of distribution" for the training data, and in such cases, statistical learning shows its weaknesses! I wonder if anyone is working on OOD detection for LLMs.
2022-06-17 03:39:26 @DuncanARiach @TheSeaMouse @yaroslavvb I agree that we should strive to fix bugs. But these large models run in parallel on multiple processors, so they are stochastic and it can be very hard to replicate bugs
2022-06-16 23:31:28 Hey @nytimes. As a subscriber, I click on Twitter links to get to your stories and then click on the “read in app” link. Your new popup covers that link, which makes it useless. Please fix!
2022-06-16 20:02:54 RT @themarkup: NEW: Some of the country’s top hospitals have been sending sensitive patient information to Facebook.These hospitals may h…
2022-06-16 17:05:18 @roydanroy @otis_reid It has an asymptote at 1, so it can’t be sqrt
2022-06-16 15:20:51 @axelsoiv @jdcmedlock @mattyglesias And flat vs hilly (looking at biking in the Netherlands and Denmark)
2022-06-16 04:57:31 RT @seanmcgregor: "The Cavalcade of Unimagined AI Consequences"I am slotted to present 20 slides where each slide advances automatically…
2022-06-16 04:04:56 @chipro Congratulations!
2022-06-16 03:35:14 @demishassabis @geoffreyhinton @ylecun @DeepMind @fpa Congratulations!
2022-06-16 03:33:27 @GaryMarcus @AndrewLampinen @dileeplearning @peabody124 @spiantado @FelixHill84 @emilymbender Or better example: That's why we teach the heat equation.
2022-06-16 03:33:10 @GaryMarcus @AndrewLampinen @dileeplearning @peabody124 @spiantado @FelixHill84 @emilymbender That is why we teach Newton's physics in high school rather than having students memorize tables of the return times of comets.
2022-06-16 03:32:23 @GaryMarcus @AndrewLampinen @dileeplearning @peabody124 @spiantado @FelixHill84 @emilymbender But if there are systematic regularities concerning a concept such as "heat", and if I can reason with those, I can make actionable predictions in a much wider range of situations.
2022-06-16 03:31:11 @GaryMarcus @AndrewLampinen @dileeplearning @peabody124 @spiantado @FelixHill84 @emilymbender But I understand hot surfaces in many other ways if I can predict how long they take to heat up and cool off
2022-06-16 03:29:08 @GaryMarcus @AndrewLampinen @dileeplearning @peabody124 @spiantado @FelixHill84 @emilymbender As I've argued elsewhere, "understanding" is a matter of degree (or variety). Some understanding can just be a reflex and doesn't require simulation. My body understands extremely hot surfaces and retracts automatically when I touch one.
2022-06-16 02:48:14 @GaryMarcus @AndrewLampinen @dileeplearning @peabody124 @spiantado @FelixHill84 @emilymbender If you compute 5*7 by table lookup, have you done multiplication?
2022-06-16 01:42:38 @BallouxFrancois @seventiessally This will come in handy during fire season this summer…
2022-06-15 19:57:06 @fchollet The rest of us are the takers of posts
2022-06-15 19:45:01 @etzioni @geekwire Oren @etzioni, you have done such a fantastic job at AI2! Congratulations
2022-06-15 17:58:24 @peabody124 @dileeplearning @AndrewLampinen @spiantado @GaryMarcus @FelixHill84 @emilymbender I didn't say there is no meaning captured by LLMs. But it only relates one linguistic utterance to another without any grounding in non-linguistic action. That may be sufficient for some applications, but most applications will require connection to other effectors.
2022-06-15 16:49:44 @AndrewLampinen @dileeplearning @peabody124 @spiantado @GaryMarcus @FelixHill84 @emilymbender Yes, because in the robot case, the robot is doing the mapping. In the language case, WE are doing the mapping, not the LLM. Turning linguistic output into action requires language understanding
2022-06-15 16:19:58 @GaryMarcus @Tesla @elonmusk Everyone I know has purchased a Tesla because it is a great car and fun to drive. I don't see why FSD is important for Tesla's future as long as they continue to produce great cars.
2022-06-15 16:02:19 @willknight If glowing eyes are a requirement for sentience, we humans are in trouble...
2022-06-15 01:51:46 @Bobpi11Bobpi @capnsnarky For example, I could program my iphone to monitor the accelerometer signals and if I dropped it, the phone could play an audio clip of a scream. But I don't think it is feeling any pain.
2022-06-15 01:50:19 @Bobpi11Bobpi @capnsnarky The only known sentient systems (in the sense of having feelings) are biological. As a materialist, I don't rule out other possibilities. My only point is that I could create a digital mimic that was clearly not experiencing pain. 1/
2022-06-14 19:40:45 @emilymbender Excellent essay. I know these take a long time to write well
2022-06-14 19:40:03 RT @emilymbender: I think the main lesson from this week's AI sentience debate is the urgent need for transparency around so-called AI syst…
2022-06-14 16:00:41 @LucaAmb I thought this was a joke. Is there actually a paper?
2022-06-14 15:58:18 @minsukkahng @GoogleAI @OregonState We will miss you @OregonState Minsuk! Best wishes in your new position
2022-06-14 05:31:15 @ledell @Bob_Wachter It occurs to me that the age profile of the car accident deaths might be very different (younger) from the COVID deaths (older). So these may not be meaningful things to compare.
2022-06-13 23:11:29 RT @jeffrsebo: Today at the UK border:Border Guard: "Business or pleasure?"Me: "Business."BG: "What kind of business?"Me: "Academic ph…
2022-06-13 22:47:55 @bertmorrien about whether it is achieving the task, and if it is failing, it takes an additional action--alerting the home owner (or the pilot in the case of an autopilot).
2022-06-13 22:47:06 @bertmorrien I chose this example as an instance of the simplest possible form of self-awareness. In addition to performing the basic task (turning heat on/off based on the difference between target temperature and set point), the thermostat reasons at a "meta level" 1/
2022-06-13 22:03:47 @bertmorrien A standard thermostat doesn’t know when it is failing. If you leave the house windows open, it will waste huge amounts of energy trying in vain to heat the house
2022-06-13 22:01:02 @hobbesdream @LibertyDevin Indeed, there is a viral thread you may have seen today from a patient in extreme pain who was dismissed by hospital staff at NYU who didn’t believe him.
2022-06-13 21:53:11 @kareem_carr Sure would be nice to know
2022-06-13 19:15:42 @LibertyDevin I don't think we can let ourselves off the hook so easily. Our attitude toward, say, torture would be very different if we thought prisoners were just faking pain convincingly versus actually experiencing it.
2022-06-13 19:11:23 @unsorsodicorda @ml_angelopoulos @stats_stephen Sure
2022-06-13 19:03:38 @LibertyDevin The Turing Test is a test of convincing fakery. So while we are close to passing the test, it doesn't give us very useful capabilities (except perhaps for phishing attacks). It is not an important milestone.
2022-06-13 16:14:40 @unsorsodicorda @ml_angelopoulos @stats_stephen Yes, I’m cleaning up the code for release. Yes, this all presumes that the policy (and the world) is not changing its behavior
2022-06-13 03:48:13 @Noahpinion Caution: This is about "Large Scale AI Results". By definition, these require the large scale compute of the big internet companies. But this is a tiny fraction of "important AI research results".
2022-06-13 03:37:13 @__picaro8 I will be speaking about this at TU Delft on July 7, but I don't know whether that presentation will be recorded.
2022-06-13 01:06:38 Here are three prediction intervals (black lines) for three different starting states in an invasive species management problem. 18 was an easy starting state
2022-06-13 01:06:37 We build on beautiful work by Lei, et al (JASA 2018), Romano et al (https://t.co/F2E9ouUB9J), and Diquigiovanni, et al (https://t.co/r7PCsZz7zJ).
2022-06-13 01:06:36 We provide lower and upper prediction intervals lo(1),...,lo(H) and hi(1),...,hi(H) such that with probability 1 - delta, lo(t) <
2022-06-13 00:48:23 Before you press GO, you would like to have confidence that your robot will succeed at this task. Let b(1), ..., b(H) be real values that quantify the performance of the robot along the H-step trajectory of states, actions, and rewards. 4/
2022-06-12 04:30:14 @GaryMarcus Evergreen tweet! @MFordFuture please please read the critiques of the Turing Test.
2022-06-12 02:56:54 @sdbaral Thank you
2022-06-11 23:45:26 @sdbaral I can imagine some cases where testing *might* make sense depending on the accuracy of the test and prevalence of the disease: visiting people in hospital or nursing homes
2022-06-11 23:25:55 @sdbaral Those of us who like collecting data on ourselves (fitbit and Apple Watch owners) will provide a market for more and better testing. So the industry will survive and improve over time.
2022-06-11 03:54:49 @jaschasd @ethansdyer Cool! I was just submitting a paper to JMLR, and you have to click to add each new author.
2022-06-11 02:15:33 @mark_riedl @isbellHFh Kind of like the Canadian system
2022-06-10 19:05:51 @emilymbender @rajiinio @timnitGebru Yes. I should have said "additional" rather than "alternative". I agree they are all operating together
2022-06-10 18:54:39 @emilymbender @rajiinio @timnitGebru Two other hypotheses: (i) western civilization has a long and deep obsession with creating a human-like being, and AI is a continuation of this fascination. (ii) human-assistive systems are much harder to build and evaluate.
2022-06-10 18:26:15 @jaschasd I can imagine it took a long time to fill out the arxiv submission form! Scaling breaks a lot of things :-)
2022-06-10 17:41:35 @beenwrekt @ajlamesa @Bob_Wachter I don't find this argument convincing at all. Everyone will die eventually, too.
2022-06-10 17:41:10 @ajlamesa @beenwrekt @Bob_Wachter OK, what is the relative risk of long flu vs long COVID?
2022-06-10 17:39:16 @beenwrekt @ajlamesa @Bob_Wachter AFAIK, the seasonal flu doesn't give you long COVID.
2022-06-10 15:45:12 @beenwrekt @ajlamesa What would be the best metrics to publish for those of us trying to avoid catching COVID? I like @Bob_Wachter’s test positivity of non-COVID hospital admits at UCSF. I wish it were available everywhere
2022-06-10 14:40:41 RT @MarkHertling: Normally I have no problem sleeping. But after watching the first Congressional 1/6 report, I needed to write some thing…
2022-06-10 06:07:27 RT @togelius: To people who complain about moving goalposts for AI: this is how it has always been, since the inception of the field. This…
2022-06-10 02:25:07 @Noahpinion Nice line: "for a great-power conflict to be prosecuted so cheaply is unprecedented in our history". I keep thinking the Russians forgot that proxy wars should be fought by proxies.
2022-06-10 01:59:02 @BallouxFrancois So I think it is primarily regional variations.
2022-06-10 01:58:30 @BallouxFrancois I took a 2-hour flight yesterday from Los Angeles to Eugene Oregon. I estimate 80% of passengers were masked on board. However, multiple pilots and flight attendants in LAX were unmasked
2022-06-10 00:14:59 @gideonmann To be more precise, I'd like to see which training sentences are responsible for this generalization. This is a form of "explanation by attribution". Some classifiers, such as k-nearest neighbors, can easily provide this. Is anyone looking at this for transformer networks?
2022-06-09 23:31:15 @gideonmann I’m really interested in questions like this—where we try to understand the statistical basis for an observed generalization. Who is studying this?
2022-06-09 21:26:09 @PartnershipAI @cao_xuenan @1optimizer @mtlaiethics Yes
2022-06-09 18:44:05 (Joking / Not Joking)
2022-06-09 18:42:57 I propose that we call large pretrained language models Sorry Models. They are trained on a sorry excuse for language use, and when they are applied to subsequent tasks, we keep having to say "I'm sorry" when they make stupid and/or dangerous mistakes.
2022-06-09 04:43:07 @roydanroy @AlexGDimakis @pfau @raphaelmilliere @ylecun Also linear predictive coding
2022-06-09 04:14:30 @compthink @fchollet Another version of this was "if an AI can perform Y using mechanism X, then X must be a lower bound on the (complexity of the) mechanism that humans use." As a grad student, I believed this until my psych professor showed me how misframed the whole line of reasoning is.
2022-06-09 04:11:56 @yudapearl Congratulations!
2022-06-09 00:01:31 @MelMitchell1 by @hannawallach @az_jacobs
2022-06-08 23:58:59 @MelMitchell1 But it is worth asking, "Suppose we took these constructions seriously as models, how would we evaluate them?" What phenomena are they modeling? How should we measure the correctness of the models. That leads us measurement theory (e.g., https://t.co/c9BbTD0We3)
2022-06-08 23:54:10 @MelMitchell1 We could decide to banish the word "model" from our vocabulary and just call these things "systems" or "constructions". 3/
2022-06-08 23:52:28 @MelMitchell1 Statisticians call their fitted equations "models" meaning statistical models. ML people borrowed the term from both of these traditions. 2/
2022-06-08 23:51:34 @MelMitchell1 In the early days of AI, "models" usually meant models of human performance (cognitive models), although they rarely deserved the name. Only a few of them were actually compared against observational data. 1/
2022-06-08 23:02:01 @cgarciae88 It seems to me that we need better infrastructure for key management. Students make mistakes like this all the time
2022-06-08 22:02:48 @avehtari Cool ideas
2022-06-08 22:01:55 @dustinvtran @avehtari It is fine to update an arxiv submission to fix typos. But your first version should be finished and ready for people to read (ie not just a first draft). Please add a comment summarizing the changes
2022-06-08 21:20:37 @igrgavilan @GaryMarcus I've also seen product-oriented corporate groups that publish their results on public benchmarks because their in-house data is proprietary. This can appear to outsiders as benchmark chasing.
2022-06-08 21:08:44 @igrgavilan @GaryMarcus Surely companies pursue products rather than either benchmarks or foundational ideas. But I agree that many research groups design and pursue benchmarks. Gary himself has proposed some benchmarks as part of his bet with Elon Musk
2022-06-08 20:54:35 @boazbaraktcs Centuries of technology and experiments.
2022-06-08 18:16:04 @TheSeaMouse @yaroslavvb We can do better
2022-06-07 23:28:56 @JHWeissmann There is also the question of how many kids to have. I chose to stop at 2 for ecological reasons. That was 35 years ago, when climate was just one of many ecological considerations
2022-06-07 21:49:52 @blairasaservice @SkotBotCambo Obviously both the technology and the organizations need to change. As a technologist, I will focus on the tech and collaborate with people who are experts in organizations. Please read beyond the headlines.
2022-06-07 04:23:17 @fchollet The agent perspective is fine as long as the operations are at a sufficiently high level. Looping over the elements of an array is difficult for students to wrap their minds around and teaches them bad habits. I love vector and matrix commands as single instructions!
2022-06-07 04:20:37 @fchollet Programs in prolog have both a declarative and an imperative reading. To know how to use "cut", you need to understand the imperative reading. The really cool idea is parameter passing by unification (bi-directional matching).
2022-06-07 04:11:45 @roydanroy @edwardhkennedy Twitter need a *groan* button
2022-06-06 05:59:35 RT @mariashriver: On this day in 1968, before many of you might even have been born, my uncle, Robert F. Kennedy was gunned down while runn…
2022-06-06 05:24:28 @AlexGDimakis @MikePFrank This is an area where I hope the cognitive scientists can help us. Maybe they can tell us how the brain decides what to save and what to discard. Maybe this can also be formulated as an RL problem?
2022-06-05 21:31:03 @MikePFrank @AlexGDimakis Many of us think this is the purpose of episodic memory. Our brains learn what is worth remembering and index it for later retrieval. This scales better than a naive implementation of attention.
2022-06-05 19:38:46 @fchollet I suspect that on the academic side, this reflects the fact that the research is being done by graduate students who are just learning how to do research. The PhD is primarily a training process. The 3-year scope of grant funding may play a big role too
2022-06-05 05:42:55 @TheSeaMouse @yaroslavvb But we should not be satisfied with such solutions. Our goal should be fully automated ML.
2022-06-05 05:12:34 @tejuafonja @black_in_ai @CarnegieMellon @datascifellows How exciting! Looking forward to learning about your project
2022-06-05 01:00:01 @matloff @banburismus_ Yes, the RKHS Representer Theorem is definitely a thing of beauty!
2022-06-04 22:03:58 @matloff @banburismus_ For me, elegance means we understand why it works and when it fails, and the explanations are simple.
2022-06-04 22:03:12 @matloff @banburismus_ Not closed form, but it IS convex. I mentioned it because theory more or less preceded experiment (or emerged simultaneously). And the main design decisions were driven by theoretical understanding. But I guess we have different definitions of elegance
2022-06-04 19:05:20 @matloff @banburismus_ I think the thread listed lots of elegance. But elegance usually lags behind experimental engineering. Two exceptions: boosting and SVMs
2022-06-04 16:06:29 @jason_pontin Yes, all of these "are indicative of" academic writing disease
2022-06-03 22:52:57 @ShannonSkalos Agreed! R doesn’t wiggle and actively oppose me
2022-06-03 17:50:21 @GaryMarcus If we focus on the right goal, then implementation issues (such as vigilance) will get the right attention that they need :-)
2022-06-03 17:23:12 @GaryMarcus As many others have pointed out, the goal (and metric) should be the safety and efficiency of the combined human+car system. The focus on 100% automation distracts from this goal
2022-06-03 05:40:44 @sethlazar @FAccTConference It is so exciting to see this field grow and mature! Thank you to everyone
2022-06-02 19:32:46 @Miles_Brundage Such fun!
2022-06-02 03:44:03 @plevy But of course the learning agent has given up the ability to learn or to detect that the environment has changed. So under the definition of intelligence as "flexible ability to solve new problems", it has ceased to be intelligent @GaryMarcus
2022-06-02 03:41:36 @plevy You can argue that in the closed world of ALE, this is optimal intelligent behavior. It achieves the task with minimum energy expenditure.
2022-06-02 03:40:31 @plevy Correction: This was for Pong and a couple of other Atari Learning Environment games. Read more here: https://t.co/zi6WjfUD7s
2022-06-02 03:33:07 @plevy Alan Fern has an interesting finding about DQN and Breakout. It turns out that in the ALE setup, Breakout is completely deterministic. Once DQN learns a policy, it then learns to ignore its percepts and play open loop. It is basically a sea squirt. 1/
2022-06-01 21:44:15 I recall seeing some work (from Berkeley)? Where the supervision at the image level just listed the set of object classes present, they were able to learn to recognize the objects (in new images).
2022-06-01 21:26:40 Computer vision twitter: I'm trying to find the earliest work where researchers matched captions to detections to learn to classify people. I'm also interested in the earliest work for learning to classify objects from image-level weak supervision. References?
2022-06-01 21:05:34 @GaryMarcus It is certainly a measure of people. Each paper has 3-4 authors. It is also a measure of the mind share of the research community as distinct from the twitter community
2022-06-01 19:54:43 @GaryMarcus Every day, cs.LG on arXiv releases around 100 papers in machine learning. I would guess only 2-3 per day relate to LLMs. The most popular category is federated learning. Physics-informed ML and ML-informed physics is also bigger than LLMs on ML arXiv.
2022-06-01 19:24:03 @GaryMarcus Is that true? Even inside Google, what fraction of the budget goes to LLMs? Prominence on Twitter can be misleading
2022-06-01 18:53:04 @GaryMarcus The first four apply to all AI research, don't they Gary? CYC was never published
2022-06-01 04:55:45 @Plinz @rao2z @giannis_daras So if asked to give subtitles, it will potentially generate different letters each time? Will they also project to the same place? Maybe not...
2022-06-01 04:28:53 @rao2z @Plinz @giannis_daras Is the text idempotent? If not, can we iterate to find the attractors?
2022-06-01 00:04:00 @mark_riedl Your shibboleth is an absence
2022-05-31 19:29:26 @DrTrapezio @FryRsquared She is such an amazing human being and so good at explaining things--even such personal things.
2022-05-31 19:25:14 @mark_riedl We've certainly observed before that cluster centers (found by clustering algorithms) may not correspond to realizable data. Is this a technique for finding those centers? Can this work for non-text too?
2022-05-31 17:50:38 RT @giannis_daras: DALLE-2 has a secret language. "Apoploe vesrreaitais" means birds. "Contarra ccetnxniams luryca tanniounons" means bug…
2022-05-31 17:50:35 @giannis_daras This is wild!
2022-05-31 16:05:26 @beenwrekt @ducha_aiki @drlucymcbride @ShiraDoronMD @TuftsMedicalCtr @DLeonhardt I trust Bob Wachter, who they quote.
2022-05-31 15:33:21 @beenwrekt @drlucymcbride @ShiraDoronMD @TuftsMedicalCtr @DLeonhardt Well, I'll keep reading your excellent papers despite our disagreement about masks :-)
2022-05-31 15:32:23 @beenwrekt @drlucymcbride @ShiraDoronMD @TuftsMedicalCtr @DLeonhardt The Cochrane article mostly concludes that existing studies are too flawed to provide much evidence. That's disappointing
2022-05-31 15:26:49 @beenwrekt @drlucymcbride @ShiraDoronMD @TuftsMedicalCtr @DLeonhardt I'll take a look. I thought the @nytimes had a good article today: https://t.co/yUWG3Ukc64
2022-05-31 15:26:18 @beenwrekt @drlucymcbride @ShiraDoronMD @TuftsMedicalCtr @DLeonhardt Uh, this is not a valid argument
2022-05-31 15:19:32 @beenwrekt @drlucymcbride @ShiraDoronMD @TuftsMedicalCtr @DLeonhardt This needs to be qualified, yes? Are you specifically talking about masks in lower grades in schools? There seem to be many studies showing effectiveness in general. https://t.co/xJTKPiXuZU
2022-05-31 15:13:07 @deliprao I don't think you have been imprisoned. That would limit many other freedoms in addition to free speech. There is nothing stopping you from publishing a newspaper, starting a blog, etc. You just can't demand that, say, the New York Times publishes your writing.
2022-05-31 15:10:55 @deliprao It doesn't limit your reach to your own followers.
2022-05-31 14:15:33 @deliprao Free speech doesn’t mean you are guaranteed an audience.
2022-05-31 14:14:24 @beenwrekt @drlucymcbride @ShiraDoronMD @TuftsMedicalCtr @DLeonhardt All our lives have been disrupted. But by the virus, not the masks. For me, the question is whether masks can help us return to normal
2022-05-30 21:40:59 @taaltree Principal component regression combines both the rotation of the coordinate axes and the selection of variables. Maybe worth considering
2022-05-30 16:56:27 @_onionesque This is an all-star team. I love their work and their writing
2022-05-30 03:19:28 @hangingnoodles @GaryMarcus It is a safe bet that current transformer models can't adequately represent this knowledge. But they capture SOME of it through its connection to prior and subsequent utterances in the training data.
2022-05-30 03:17:51 @hangingnoodles @GaryMarcus Alas, we can't easily access "what we know about their referents". It manifests in hundreds of (often small) changes in behavior (including subsequent utterances). It is not at all obvious how to represent this knowledge.
2022-05-30 01:42:35 @flipphillips @seeingwithsound @rasbt @LucaAmb But of course that wasn’t the point of the experiment. Sheesh
2022-05-29 20:42:48 @nelslindahl @GaryMarcus I share your concern. At the undergraduate level, I advise most research-oriented students to major in math or physics rather than CS. Does anyone have pointers to an undergraduate ML/data science program that matches the rigor of the best EE or MechE programs?
2022-05-29 20:31:13 @nelslindahl @GaryMarcus Meanwhile, there is excellent research in applying ML in engineering problems in materials science, mechanical design, VLSI layout, and so on.
2022-05-29 20:30:37 @nelslindahl @GaryMarcus I don't know about schools in India, but in my experience in the US, there are some engineering faculty who just surf one fad after another. "Crowding out" doesn't seem to be an apt description of that behavior.
2022-05-29 20:13:27 @nelslindahl @GaryMarcus I see a lot of these papers submitted to @arxiv. We try to reject simple data analysis exercises that lack a research contribution either to ML or to the application field.
2022-05-29 20:12:05 @nelslindahl @GaryMarcus Every field that builds easy-to-use tools risks creating a broad user community that doesn't understand the underlying techniques and assumptions. But this posting isn't about that. It is about doing superficial applied ML to exploit what we could call the "network fad effect"
2022-05-29 17:27:43 @GaryMarcus @rogerkmoore I remember when I first studied formal language theory that viewing a language as a set of strings seems like such an impoverished view. But over time we become accustomed to our definitions and forget their unnaturalness.
2022-05-29 17:03:21 @GaryMarcus @rogerkmoore The Chomsky hierarchy describes formal languages. Perhaps this is where the error began? Or was the error to confuse formal languages with natural languages?
2022-05-29 16:48:56 @TaliaRinger Interesting! Can you point us to something that explains non-discrete logics?
2022-05-29 16:02:28 @rasbt @professorwcohen I’m not going to read either one on my phone. Just bookmark for later.
2022-05-29 15:53:36 @MatthewJBar Separately, are some non consequentialist rules stable equilibria for multiparty decision making?
2022-05-29 15:51:42 @MatthewJBar Aren’t there computational limits to computing the consequences of actions? Having simpler rules for decision making avoids this problem.
2022-05-29 02:13:46 @sarahookr @_joaogui1 I think we are moving into a period of hardware/software co-design. We can’t design them separately anymore and expect good results
2022-05-28 19:29:16 @RichardSSutton This definition doesn't appear to encompass learning. It could describe a system that is just applying an innate skill.
2022-05-28 02:00:24 @kchonyc An attempt by non authors to muscle their way onto a paper?
2022-05-27 23:37:00 @Noahpinion The Left has been out of power so long they have forgotten what winning looks like.
2022-05-27 21:02:08 RT @tdietterich: @rao2z Drew McDermott was a giant of AI. He inspired and challenged the whole AI research community
2022-05-27 21:01:59 @rao2z Drew McDermott was a giant of AI. He inspired and challenged the whole AI research community
2022-05-27 04:32:04 @chris_j_paxton Maybe legged instead of wheeled?
2022-05-26 19:37:44 RT @AaronHertzmann: Here are some impressions from using DALL-E 2 to make art, and what it tells us about computational creativity and the…
2022-05-26 18:32:02 @NandoDF @roydanroy @JayAlammar Thank you for correcting me. I read the paper last week and I'm already confusing some of the details. Sorry
2022-05-26 17:52:41 @MWCvitkovic One can argue that as a scientist you shouldn't run a safety trial on human volunteers unless you have first run it on yourself. I know of at least one pharma researcher with this attitude
2022-05-26 15:50:46 RT @arxiv: We're hiring! The arXiv Technical Director will lead the effort to update arXiv’s technical design and implementation, in order…
2022-05-26 15:18:55 @roydanroy @JayAlammar @NandoDF
2022-05-26 15:17:44 @roydanroy @JayAlammar AI researchers have long wondered about the degree of overlap between different tasks or bodies of knowledge. Perhaps analysis of GATO can tell us something about that
2022-05-26 15:15:58 @roydanroy @JayAlammar Yes, this is joint distillation of multiple previously-learned teacher models into a single student model. So the interesting question to me is how much structure it is able to share across the tasks
2022-05-26 15:05:09 @emilymbender Wow. Just wow.
2022-05-26 04:52:45 @emilymbender A good option is to reply to the thread (by replying to the first tweet) and then retweet your reply if you want your followers to contribute to the original conversation. A quote tweet creates a separate conversation and invites chaos
2022-05-26 00:26:53 @_julesh_ Leo Breiman used to call these “comfort theorems”. Any model family that does not have such a theorem is limited.
2022-05-25 23:32:39 @NandoDF @JayAlammar [Obviously this requires more thought than fits into a tweet. I don't think I have any particularly insightful experiments to propose.]
2022-05-25 23:31:05 @NandoDF @JayAlammar I read that part, and it is interesting, but I was wondering if there is a way to analyze shared activations or other forms of explanation across the training tasks. Maybe ablating some network regions would leave some tasks intact but destroy others?
2022-05-25 22:59:09 @JayAlammar Is there any evidence that GATO is anything more than a big switch? Are there interesting nuggets of shareable knowledge identified across tasks? It certainly seems like GATO is using a lot fewer parameters than the teacher networks from which it was trained
2022-05-25 22:54:53 @Noahpinion In hard times, every incumbent is unpopular. We have a huge "hope deficit" about the economy, COVID, crime, climate change, etc. Time for a new hope candidate!
2022-05-25 18:03:56 @haldaume3 We were fortunate to have Priscilla's help in several ICML meetings as well. Thank you!
2022-05-25 18:03:07 @jbensnyder My daughter does field work in bear country, and she can use a revolver or shotgun for bear defense.
2022-05-25 02:48:07 @joeyliaw Both systems create amazing images. Very nice!
2022-05-25 02:12:36 @rogtron Exactly! The murderers are people who own guns or who have stolen guns from their owners. Can we think creatively about how to achieve a huge reduction in gun crime?
2022-05-25 00:15:21 The onus is on gun owners to develop an effective proposal for ending gun violence. If they can't, then gun ownership must end.
2022-05-24 23:57:45 @gluonspring @Noahpinion Yes, and the most innovative engineers operate at the frontier of what is possible. They confront immense uncertainty AND risks to human lives and well being.
2022-05-24 23:18:35 @gluonspring @Noahpinion I don't know, but engineers generally don't need to grapple with the same levels of extreme uncertainty as scientists do.
2022-05-24 23:01:49 @Noahpinion Engineers as a group are politically more conservative than scientists in my experience. (Although this may be changing with more women engineering faculty members.)
2022-05-24 16:31:01 @gautamcgoel It depends on the system under control. RL for Go or theorem proving doesn’t raise any of these issues. But real-time physical interaction is a “different game”
2022-05-24 15:51:19 @AlexGDimakis Is there an LLM that solves reasoning puzzles? I.e., as opposed to replaying a solution discovered by some other mechanism like Gato?
2022-05-24 06:03:52 @roydanroy @Uber Uber offers full recording via the app in some countries, why not all?
2022-05-23 21:57:35 @MaximZiatdinov @GaryMarcus @kristintynski @ylecun @JefferyAGoldst1 @miguelisolano @peterfoldiak Are there any papers where people have applied GAN or VAE-type models to learn novel augmentations?
2022-05-22 23:46:13 @miguelisolano @GaryMarcus What is the best statement of the scaling hypothesis?
2022-05-22 15:55:04 @TDeryugina @RicciMilstein I want the right to be able to use my laptop. If the seats are spaced sufficiently far apart, they can recline and I can still use my laptop.
2022-05-21 20:56:16 @laurajfagan Yes
2022-05-21 20:51:43 Very disappointed in this @united https://t.co/V6ZpMARxVP
2022-05-21 16:24:29 RT @dribnet: Homage to the Pixel: a generative study of computational color perception. Often I manipulate shapes and colors using machine…
2022-05-21 15:21:06 @mlittmancs How wonderful! Thank you for taking on this service to the country and the research community
2022-05-21 14:34:34 @sarameghanbeery Congratulations!
2022-05-20 22:25:33 RT @TobyWalsh: https://t.co/tF93oyBx3g
2022-05-20 14:50:32 @wooldridgemike Congratulations!
2022-05-20 14:41:18 @pfau Two data points should be enough to define a trend
2022-05-20 08:11:00 CAFIAC FIX
2022-10-29 05:45:04 @matloff Yeah
2022-10-29 05:37:23 @ylecun Is there any evidence that this is more than a preprogrammed procedure?
2022-10-29 05:34:08 @boazbaraktcs Academic Twitter includes Academic #metoo (as it should). Then come the trolls. But presumably the trolls will come to mastodon too.
2022-10-29 05:30:09 @DKThomp @Noahpinion We always want to find a scapegoat. That’s why my prior is for a wild origin. The evidence continues to be ambiguous
2022-10-29 05:23:11 @deliprao This is why I liked teaching intro algorithms. You can teach TRUTH (eg sorting).
2022-10-29 05:05:04 Great thread by my colleague @taaltree on why being a professor is a great job https://t.co/sh1idUdZP7
2022-10-28 04:56:24 I am now @tdietterich@mastodon.social. Will cross-tweet for now
2022-10-27 23:06:50 @aaronharris @Noahpinion I guess the courts will decide that question: https://t.co/k1ROaMBVs9
2022-10-27 21:34:36 @aaronharris @Noahpinion Monopoly over the Apple App store and all financial transactions that go through it.
2022-10-27 16:09:35 RT @predict_addict: The long (17 years!) wait is over.The second edition of the ALRW book (Conformal Prediction Bible) written by the c…
2022-10-27 05:40:36 @roydanroy @KevinKaichuang Me too. Entrapment effort?
2022-10-27 04:59:32 @Noahpinion The harms caused by Apple are the standard harms of a monopoly. The harms of Facebook (and social media in general) threaten societies around the world through the toxic behavior that they enable. I agree that anti-trust is more relevant to the former.
2022-10-27 02:13:32 @Abebab Just being true to its history…
2022-10-26 03:05:48 @mhar4 @yangyang_cheng I think the two generations are very different. The first was triumphal, but today we are all extremely mindful of the fragility of freedom, health, and peace. end/
2022-10-26 03:03:52 @mhar4 @yangyang_cheng She writes "The new cold warriors are not so different from their counterparts a generation ago, who preached that marketisation and free trade would usher in political liberalisation. Both narratives stem from an unabashed belief in the supremacy of their own system..." 2/
2022-10-26 03:03:27 @mhar4 @yangyang_cheng I am always moved by @yangyang_cheng's writing, and this piece is no exception. For a whole host of reasons, we have all entered a state of perpetual mourning. But I disagree with one statement. 1/
2022-10-25 15:27:22 @Raza_Habib496 @yoavgo Interesting, thanks.
2022-10-25 00:11:52 @yoavgo I'm not sure what it would mean to be "about language", but existing models are certainly not "about communication" because an LLM is not an agent and has no communicative goals
2022-10-25 00:08:41 @schwellenbach But it isn't exactly a surprise is it? Anyone paying attention knows that Eric Schmidt has been devoting a lot of his time to AI. (I have no special insight
2022-10-24 17:33:48 RT @GaelVaroquaux: #NeurIPS2022 paper: https://t.co/qft4LgZTQ7An easily reusable benchmark that shows that deep learning underperforms…
2022-10-24 15:52:43 @MelMitchell1 @arxiv You might consider withdrawing and resubmitting the paper with the correct categories. Long delays are usu caused by discussions among moderators about the best category. These can get stalled because the discussion UI is clunky.
2022-10-24 15:49:40 @tinaeliassi @davecliff @MelMitchell1 @arxiv We clear the cs.LG category every day. (If we didn't, we would drown in submissions and never recover.) That should never be the category that is holding up a release.
2022-10-24 04:17:28 Interesting housing story in the @latimes. I knew the racist history of Los Angeles was bad, but it was much worse than I had realized. https://t.co/uHRFNnjro5
2022-10-24 01:27:17 @cjmaddison I would rephrase this as" It is by the grace of God AND the measures you are taking (vaccination, masking)". Those measures are not useless, but they provide no guarantees
2022-10-24 01:23:40 @rasbt I'd like to see metrics for ease of tuning or insensitivity to hyperparameters. More radical: It is not an algorithm until all tuning is automated. I guess that would be required for third-party evaluation.
2022-10-23 22:02:41 @cloudquistador @emilymbender Excellent point! Thanks for clarifying
2022-10-23 21:25:39 @AnthroPunk @emilymbender Yes, and I take very seriously the idea that human intelligence may have evolved primarily to support social interaction. I think that is why so many of our AI benchmarks seem to be mis-targeted.
2022-10-23 21:23:10 @cloudquistador @emilymbender Yes. I'm not sure what you mean by "aiming in error". Are we focusing on the wrong target behaviors?
2022-10-23 20:17:48 @cloudquistador @emilymbender The achievement of aerodynamics is that we (mostly) understand what material properties are required to mimic birds. It would be wonderful to have similar insights into the computational/biological substrates required to mimic human cognition in detail.
2022-10-23 19:50:25 @emilymbender At some level of detail, they fly for the same reasons.
2022-10-23 19:49:38 @emilymbender They fly.
2022-10-23 19:44:17 @emilymbender But I agree it is difficult to define the "functional behavior" of cognition, and that seems to be a major stumbling block in these discussions
2022-10-23 19:43:22 @emilymbender Hmm. If airplanes are not models of avian flight behavior, comparisons to avian flight behavior are only marketing/hype? I think there can be global functional comparisons without any claim that the underlying detailed structure is a "model"
2022-10-23 04:19:29 @deliprao I do love the feeling that the computer is going to work all night while I get to sleep.
2022-10-23 03:08:42 @deliprao This is an opportunity to get in touch with the early years of computing. In every generation, our ambitions outstrip the hardware and we must wait for computations to finish.
2022-10-22 22:47:02 @earnmyturns @Noahpinion And Spain was having enough trouble holding on to Catalonia, which continues to this day.
2022-10-22 21:21:07 @GaryMarcus I don't know how to quantify progress, so I don't know if it is exponential. But we have been making progress on dealing with change/novelty. e.g., Bryan Loyall at CRA leads an exciting project: https://t.co/dggKFIm3Cf
2022-10-22 21:15:51 @tejuafonja Instead, it is the one that is most clearly written and that clarifies the fundamentals. The paper that explains what is happening and why.
2022-10-22 21:14:53 @tejuafonja A reason not to wait is if you are in a race with some other research groups. But it is easy to worry too much about this. Good, mature work is a more valuable contribution than rushed, immature work. Usually the most cited/appreciated work is not the very first paper on a topic
2022-10-22 21:13:12 @tejuafonja One reason to wait is that if you publicize the unimproved version of your paper, that is the version people will read. And they may dismiss it or misunderstand it. After improvement via peer review, they will hopefully understand and appreciate it much more
2022-10-22 21:11:47 @tejuafonja It depends on how confident they are that the work is ready for public examination. I usually prefer to have a round of peer review before going public with my work. That can be internal peer review if you are part of a large enough organization.
2022-10-22 19:29:36 RT @tdietterich: @denny_zhou Nonsense. Each deployed app is a choice. I hope this attitude is not representative of Google brain as an orga…
2022-10-22 19:29:24 @denny_zhou Nonsense. Each deployed app is a choice. I hope this attitude is not representative of Google brain as an organization
2022-10-21 23:23:23 @RRamirezMorales @mbeisen *Most of the areas within these fields
2022-10-21 23:21:26 @RRamirezMorales @mbeisen But most of these fields don't need gate keeping.
2022-10-21 23:20:42 @RRamirezMorales @mbeisen The physicists have to deal with variations on perpetual motion and theories of everything.
2022-10-21 23:20:17 @RRamirezMorales @mbeisen In computer science, the sensitive areas are (a) proofs that P=NP, (b) papers claiming to diagnose COVID using chest x-rays and deep learning, (c) papers claiming to forecast COVID using fancy time series analysis.
2022-10-21 19:47:20 @RRamirezMorales @mbeisen Most areas of science and engineering don’t draw the attention of the media or the trolls. But I agree that there are some areas that require careful filtering
2022-10-21 00:15:06 @scottjshapiro I assume this particular tweet was aimed at philosophers, but it certainly applies generally. Technologists need to have a basic understanding of political theory, sociology, psychology, philosophy, and economics to analyze the broader context of their work.
2022-10-21 00:13:00 @scottjshapiro Whenever you cross disciplinary borders, it is best to have someone from that discipline as an escort. Interdisciplinary work requires a big investment to understand the concerns and conceptual structures of each discipline.
2022-10-20 20:33:05 @randall_balestr @sindero Tree ensembles create empty partitions, too, by intersecting no empty partitions. It is not obvious that this is a good thing
2022-10-20 16:59:00 @rasbt There is usually one spiking net paper per day on ArXiv
2022-10-20 14:53:59 @MazMHussain I always figured South Asian was adopted to remind people in the West that there are lots more Asians than just Chinese, Japanese, and Koreans.
2022-10-20 00:39:13 @_joaogui1 @FelixHill84 @erikbryn @gdb @OpenAI @scaleai The DL frameworks make neural net assignments more reliable, but tuning is still a black hole for students. end/
2022-10-20 00:37:22 @_joaogui1 @FelixHill84 @erikbryn @gdb @OpenAI @scaleai Back in the 90s, a colleague of mine said "If I give a neural network assignment, only 10% of the students can get it to work. If I assign an SVM, 90% get it to work, and they all get the same answer." 2/
2022-10-20 00:35:48 @_joaogui1 @FelixHill84 @erikbryn @gdb @OpenAI @scaleai Even in the 1990s, shallow neural networks could beat SVMs in the right hands. But they were much more difficult to tune and slower to train. The big advantage of SVMs was that the optimization problem was convex and they are very insensitive to the margin parameter C. 1/
2022-10-19 22:43:20 @davidthewid And one junior faculty member stepped up and did a lot of informal advising too
2022-10-19 22:42:19 @davidthewid In countries such as Canada, where funding goes directly to students, the power issues are reduced. But that still doesn’t fix the single advisor problem. My fellow students in grad school ended up advising each other during the 1980s AI boom
2022-10-19 19:10:33 @erikbryn @gdb @OpenAI @scaleai What has increased monotonically is the amount of data available for training
2022-10-19 19:10:02 @erikbryn @gdb @OpenAI @scaleai This is a simplification. There were intermediate points along the way where compute loads decreased. For example, when we switched to SVMs from 1980s neural networks.
2022-10-19 16:54:03 @davidmanheim @Noahpinion As long as the US continues to send $3B per year in military aid to Israel, can't it expect cooperation in military affairs?
2022-10-30 05:27:59 @Miles_Brundage @Lee_Morgan7 In a previous Twitter discussion, someone suggested "Large Corpus Models"
2022-10-29 22:32:04 @adawan919 @ylecun Evolution has no goal. It is stochastic forward search followed by natural selection (eg reproductive success or failure).
2022-10-29 22:30:42 @adawan919 @ylecun “Planning”, as we use it in AI, involves assembling a sequence of actions to achieve a goal. Efficient planning algorithms incorporate the goal into the search.
2022-10-29 16:15:21 @adawan919 @ylecun Natural selection does not plan
2022-10-29 15:48:28 @rao2z @SheilaMcIlraith That paper shows how to automatically extend the state space to convert non-Markovian reward specifications into Markovian reward functions
2022-10-29 15:47:38 @rao2z Yes, when teaching MDPs, I tend to focus too much on the Markovian transitions and not enough on the Markovian rewards. I love @SheilaMcIlraith's work on more expressive languages for expressing reward functions: https://t.co/7jIG8mvc6Y
2022-10-29 05:45:04 @matloff Yeah
2022-10-29 05:37:23 @ylecun Is there any evidence that this is more than a preprogrammed procedure?
2022-10-29 05:34:08 @boazbaraktcs Academic Twitter includes Academic #metoo (as it should). Then come the trolls. But presumably the trolls will come to mastodon too.
2022-10-29 05:30:09 @DKThomp @Noahpinion We always want to find a scapegoat. That’s why my prior is for a wild origin. The evidence continues to be ambiguous
2022-10-29 05:23:11 @deliprao This is why I liked teaching intro algorithms. You can teach TRUTH (eg sorting).
2022-10-29 05:05:04 Great thread by my colleague @taaltree on why being a professor is a great job https://t.co/sh1idUdZP7
2022-10-28 04:56:24 I am now @tdietterich@mastodon.social. Will cross-tweet for now
2022-10-27 23:06:50 @aaronharris @Noahpinion I guess the courts will decide that question: https://t.co/k1ROaMBVs9
2022-10-27 21:34:36 @aaronharris @Noahpinion Monopoly over the Apple App store and all financial transactions that go through it.
2022-10-27 16:09:35 RT @predict_addict: The long (17 years!) wait is over.The second edition of the ALRW book (Conformal Prediction Bible) written by the c…
2022-10-27 05:40:36 @roydanroy @KevinKaichuang Me too. Entrapment effort?
2022-10-27 04:59:32 @Noahpinion The harms caused by Apple are the standard harms of a monopoly. The harms of Facebook (and social media in general) threaten societies around the world through the toxic behavior that they enable. I agree that anti-trust is more relevant to the former.
2022-10-27 02:13:32 @Abebab Just being true to its history…
2022-10-26 03:05:48 @mhar4 @yangyang_cheng I think the two generations are very different. The first was triumphal, but today we are all extremely mindful of the fragility of freedom, health, and peace. end/
2022-10-26 03:03:52 @mhar4 @yangyang_cheng She writes "The new cold warriors are not so different from their counterparts a generation ago, who preached that marketisation and free trade would usher in political liberalisation. Both narratives stem from an unabashed belief in the supremacy of their own system..." 2/
2022-10-26 03:03:27 @mhar4 @yangyang_cheng I am always moved by @yangyang_cheng's writing, and this piece is no exception. For a whole host of reasons, we have all entered a state of perpetual mourning. But I disagree with one statement. 1/
2022-10-25 15:27:22 @Raza_Habib496 @yoavgo Interesting, thanks.
2022-10-25 00:11:52 @yoavgo I'm not sure what it would mean to be "about language", but existing models are certainly not "about communication" because an LLM is not an agent and has no communicative goals
2022-10-25 00:08:41 @schwellenbach But it isn't exactly a surprise is it? Anyone paying attention knows that Eric Schmidt has been devoting a lot of his time to AI. (I have no special insight
2022-10-24 17:33:48 RT @GaelVaroquaux: #NeurIPS2022 paper: https://t.co/qft4LgZTQ7An easily reusable benchmark that shows that deep learning underperforms…
2022-10-24 15:52:43 @MelMitchell1 @arxiv You might consider withdrawing and resubmitting the paper with the correct categories. Long delays are usu caused by discussions among moderators about the best category. These can get stalled because the discussion UI is clunky.
2022-10-24 15:49:40 @tinaeliassi @davecliff @MelMitchell1 @arxiv We clear the cs.LG category every day. (If we didn't, we would drown in submissions and never recover.) That should never be the category that is holding up a release.
2022-10-24 04:17:28 Interesting housing story in the @latimes. I knew the racist history of Los Angeles was bad, but it was much worse than I had realized. https://t.co/uHRFNnjro5
2022-10-24 01:27:17 @cjmaddison I would rephrase this as" It is by the grace of God AND the measures you are taking (vaccination, masking)". Those measures are not useless, but they provide no guarantees
2022-10-24 01:23:40 @rasbt I'd like to see metrics for ease of tuning or insensitivity to hyperparameters. More radical: It is not an algorithm until all tuning is automated. I guess that would be required for third-party evaluation.
2022-10-23 22:02:41 @cloudquistador @emilymbender Excellent point! Thanks for clarifying
2022-10-23 21:25:39 @AnthroPunk @emilymbender Yes, and I take very seriously the idea that human intelligence may have evolved primarily to support social interaction. I think that is why so many of our AI benchmarks seem to be mis-targeted.
2022-10-23 21:23:10 @cloudquistador @emilymbender Yes. I'm not sure what you mean by "aiming in error". Are we focusing on the wrong target behaviors?
2022-10-23 20:17:48 @cloudquistador @emilymbender The achievement of aerodynamics is that we (mostly) understand what material properties are required to mimic birds. It would be wonderful to have similar insights into the computational/biological substrates required to mimic human cognition in detail.
2022-10-23 19:50:25 @emilymbender At some level of detail, they fly for the same reasons.
2022-10-23 19:49:38 @emilymbender They fly.
2022-10-23 19:44:17 @emilymbender But I agree it is difficult to define the "functional behavior" of cognition, and that seems to be a major stumbling block in these discussions
2022-10-23 19:43:22 @emilymbender Hmm. If airplanes are not models of avian flight behavior, comparisons to avian flight behavior are only marketing/hype? I think there can be global functional comparisons without any claim that the underlying detailed structure is a "model"
2022-10-23 04:19:29 @deliprao I do love the feeling that the computer is going to work all night while I get to sleep.
2022-10-23 03:08:42 @deliprao This is an opportunity to get in touch with the early years of computing. In every generation, our ambitions outstrip the hardware and we must wait for computations to finish.
2022-10-22 22:47:02 @earnmyturns @Noahpinion And Spain was having enough trouble holding on to Catalonia, which continues to this day.
2022-10-22 21:21:07 @GaryMarcus I don't know how to quantify progress, so I don't know if it is exponential. But we have been making progress on dealing with change/novelty. e.g., Bryan Loyall at CRA leads an exciting project: https://t.co/dggKFIm3Cf
2022-10-22 21:15:51 @tejuafonja Instead, it is the one that is most clearly written and that clarifies the fundamentals. The paper that explains what is happening and why.
2022-10-22 21:14:53 @tejuafonja A reason not to wait is if you are in a race with some other research groups. But it is easy to worry too much about this. Good, mature work is a more valuable contribution than rushed, immature work. Usually the most cited/appreciated work is not the very first paper on a topic
2022-10-22 21:13:12 @tejuafonja One reason to wait is that if you publicize the unimproved version of your paper, that is the version people will read. And they may dismiss it or misunderstand it. After improvement via peer review, they will hopefully understand and appreciate it much more
2022-10-22 21:11:47 @tejuafonja It depends on how confident they are that the work is ready for public examination. I usually prefer to have a round of peer review before going public with my work. That can be internal peer review if you are part of a large enough organization.
2022-10-22 19:29:36 RT @tdietterich: @denny_zhou Nonsense. Each deployed app is a choice. I hope this attitude is not representative of Google brain as an orga…
2022-10-22 19:29:24 @denny_zhou Nonsense. Each deployed app is a choice. I hope this attitude is not representative of Google brain as an organization
2022-10-21 23:23:23 @RRamirezMorales @mbeisen *Most of the areas within these fields
2022-10-21 23:21:26 @RRamirezMorales @mbeisen But most of these fields don't need gate keeping.
2022-10-21 23:20:42 @RRamirezMorales @mbeisen The physicists have to deal with variations on perpetual motion and theories of everything.
2022-10-21 23:20:17 @RRamirezMorales @mbeisen In computer science, the sensitive areas are (a) proofs that P=NP, (b) papers claiming to diagnose COVID using chest x-rays and deep learning, (c) papers claiming to forecast COVID using fancy time series analysis.
2022-10-21 19:47:20 @RRamirezMorales @mbeisen Most areas of science and engineering don’t draw the attention of the media or the trolls. But I agree that there are some areas that require careful filtering
2022-10-21 00:15:06 @scottjshapiro I assume this particular tweet was aimed at philosophers, but it certainly applies generally. Technologists need to have a basic understanding of political theory, sociology, psychology, philosophy, and economics to analyze the broader context of their work.
2022-10-21 00:13:00 @scottjshapiro Whenever you cross disciplinary borders, it is best to have someone from that discipline as an escort. Interdisciplinary work requires a big investment to understand the concerns and conceptual structures of each discipline.
2022-10-20 20:33:05 @randall_balestr @sindero Tree ensembles create empty partitions, too, by intersecting no empty partitions. It is not obvious that this is a good thing
2022-10-20 16:59:00 @rasbt There is usually one spiking net paper per day on ArXiv
2022-10-20 14:53:59 @MazMHussain I always figured South Asian was adopted to remind people in the West that there are lots more Asians than just Chinese, Japanese, and Koreans.
2022-10-20 00:39:13 @_joaogui1 @FelixHill84 @erikbryn @gdb @OpenAI @scaleai The DL frameworks make neural net assignments more reliable, but tuning is still a black hole for students. end/
2022-10-20 00:37:22 @_joaogui1 @FelixHill84 @erikbryn @gdb @OpenAI @scaleai Back in the 90s, a colleague of mine said "If I give a neural network assignment, only 10% of the students can get it to work. If I assign an SVM, 90% get it to work, and they all get the same answer." 2/
2022-10-20 00:35:48 @_joaogui1 @FelixHill84 @erikbryn @gdb @OpenAI @scaleai Even in the 1990s, shallow neural networks could beat SVMs in the right hands. But they were much more difficult to tune and slower to train. The big advantage of SVMs was that the optimization problem was convex and they are very insensitive to the margin parameter C. 1/
2022-10-19 22:43:20 @davidthewid And one junior faculty member stepped up and did a lot of informal advising too
2022-10-19 22:42:19 @davidthewid In countries such as Canada, where funding goes directly to students, the power issues are reduced. But that still doesn’t fix the single advisor problem. My fellow students in grad school ended up advising each other during the 1980s AI boom
2022-10-19 19:10:33 @erikbryn @gdb @OpenAI @scaleai What has increased monotonically is the amount of data available for training
2022-10-19 19:10:02 @erikbryn @gdb @OpenAI @scaleai This is a simplification. There were intermediate points along the way where compute loads decreased. For example, when we switched to SVMs from 1980s neural networks.
2022-10-19 16:54:03 @davidmanheim @Noahpinion As long as the US continues to send $3B per year in military aid to Israel, can't it expect cooperation in military affairs?
2022-11-17 16:59:52 @lilkuo "surrealism is the rational response to dictatorship" -- Carles Sierra
2022-11-17 04:17:32 @CT_Bergstrom Is this because it is trained on a lot of papers written by white men?
2022-11-17 04:17:03 @dcbaok @CT_Bergstrom Same with papers in Science and Nature (with some exceptions such as the DeepMind papers)
2022-11-16 22:42:42 @GaryMarcus On the other hand, after playing with it, I think most of its hypotheses are false
2022-11-16 22:40:27 @MattNiessner Just like startups, a lot of cutting edge tech fails to meet a real need. So I think people rationally adopt a negatively-biased prior
2022-11-16 16:58:53 @deliprao More than 11 people have died because of rumors and misinformation spread on social networks
2022-11-16 05:55:24 https://t.co/YaH8PuHnIU
2022-11-16 05:54:57 Giving Galactica a try. This mainly tells me that the face recognition community use this term differently than I do. For me, open category == open set. https://t.co/jGYx8CbxJq
2022-11-16 03:45:41 @GaryMarcus But the great thing about proofs is that they can be checked. Think of it has an hypothesis generator, not a source of truth. Same applies to generating code.
2022-11-15 22:12:55 RT @MattBetts11: Excited for this week’s EECB talk from @karst_justine !
2022-11-15 21:43:17 @noop_noob @rao2z Yes, some theories of consciousness would consider that a consciousness mechanism.
2022-11-15 20:43:53 @AmanBitz @ylecun @DebasmitDas1 @roydanroy Current SOTA according to https://t.co/aLwWum6M4Z is 96.81, and it appears to be asymptoting.
2022-11-15 20:42:44 @AmanBitz @ylecun @DebasmitDas1 @roydanroy But suppose you could raise CIFAR100 accuracy to 97%, I wouldn't be interested unless the same techniques worked on ImageNet and COCO and were practical at scale. Working well on CIFAR100 alone doesn't tell us much.
2022-11-15 18:37:29 @ylecun @DebasmitDas1 @roydanroy Well, it depends on what the paper is about. A accuracy-oriented computer vision paper evaluated on MNIST, CIFAR10, and CIFAR100 is no longer interesting. But, a paper studying the open category problem using these data sets is interesting (IMO: https://t.co/0KUqdPhE3f)
2022-11-15 18:15:14 @INSMI_CNRS @BachFrancis @giry_claire @sup_recherche @ENS_ULM @INS2I_CNRS Congratulations!
2022-11-15 16:40:40 @falsalem76 @_jasonwei Not unless it is something that can be learned from data, which I doubt. It is an architectural property.
2022-11-15 15:54:24 @mhudack @AutismCapital No, it’s anarchist
2022-11-15 04:42:08 @DebasmitDas1 @roydanroy That’ll be obsolete tomorrow
2022-11-15 04:39:36 @andyroyle_pwrc Much cheaper in Spain. Maybe subsidized there?
2022-11-15 04:39:00 @MarkSorel8 @andyroyle_pwrc My pharmacist did the paperwork for me. Thanks @FredMeyerInc
2022-11-15 00:33:11 @rao2z I agree that an agent that can detect that the anticipated effects of actions were not observed would be more robust (and could detect novelty). But is this capability required for a system to be treated as an "agent"?
2022-11-14 23:59:23 @rao2z Similar models can be trained to predict the next action in a policy. So they could *execute* a policy, but not create one. If we take the "agent stance", we can predict their behavior by attributing goals to them. Are you wanting an agent to do more?
2022-11-14 16:20:50 @jacobpattywagon The later shots are indoors. Outdoors there would be hats (but different ones)
2022-11-14 06:32:09 @Raza_Habib496 If we can predict the actions of a system by attributing to it goals and beliefs, then it is useful to treat it as an agent. What are the goals, beliefs, and actions of an LLM?
2022-11-17 16:59:52 @lilkuo "surrealism is the rational response to dictatorship" -- Carles Sierra
2022-11-17 04:17:32 @CT_Bergstrom Is this because it is trained on a lot of papers written by white men?
2022-11-17 04:17:03 @dcbaok @CT_Bergstrom Same with papers in Science and Nature (with some exceptions such as the DeepMind papers)
2022-11-16 22:42:42 @GaryMarcus On the other hand, after playing with it, I think most of its hypotheses are false
2022-11-16 22:40:27 @MattNiessner Just like startups, a lot of cutting edge tech fails to meet a real need. So I think people rationally adopt a negatively-biased prior
2022-11-16 16:58:53 @deliprao More than 11 people have died because of rumors and misinformation spread on social networks
2022-11-16 05:55:24 https://t.co/YaH8PuHnIU
2022-11-16 05:54:57 Giving Galactica a try. This mainly tells me that the face recognition community use this term differently than I do. For me, open category == open set. https://t.co/jGYx8CbxJq
2022-11-16 03:45:41 @GaryMarcus But the great thing about proofs is that they can be checked. Think of it has an hypothesis generator, not a source of truth. Same applies to generating code.
2022-11-15 22:12:55 RT @MattBetts11: Excited for this week’s EECB talk from @karst_justine !
2022-11-15 21:43:17 @noop_noob @rao2z Yes, some theories of consciousness would consider that a consciousness mechanism.
2022-11-15 20:43:53 @AmanBitz @ylecun @DebasmitDas1 @roydanroy Current SOTA according to https://t.co/aLwWum6M4Z is 96.81, and it appears to be asymptoting.
2022-11-15 20:42:44 @AmanBitz @ylecun @DebasmitDas1 @roydanroy But suppose you could raise CIFAR100 accuracy to 97%, I wouldn't be interested unless the same techniques worked on ImageNet and COCO and were practical at scale. Working well on CIFAR100 alone doesn't tell us much.
2022-11-15 18:37:29 @ylecun @DebasmitDas1 @roydanroy Well, it depends on what the paper is about. A accuracy-oriented computer vision paper evaluated on MNIST, CIFAR10, and CIFAR100 is no longer interesting. But, a paper studying the open category problem using these data sets is interesting (IMO: https://t.co/0KUqdPhE3f)
2022-11-15 18:15:14 @INSMI_CNRS @BachFrancis @giry_claire @sup_recherche @ENS_ULM @INS2I_CNRS Congratulations!
2022-11-15 16:40:40 @falsalem76 @_jasonwei Not unless it is something that can be learned from data, which I doubt. It is an architectural property.
2022-11-15 15:54:24 @mhudack @AutismCapital No, it’s anarchist
2022-11-15 04:42:08 @DebasmitDas1 @roydanroy That’ll be obsolete tomorrow
2022-11-15 04:39:36 @andyroyle_pwrc Much cheaper in Spain. Maybe subsidized there?
2022-11-15 04:39:00 @MarkSorel8 @andyroyle_pwrc My pharmacist did the paperwork for me. Thanks @FredMeyerInc
2022-11-15 00:33:11 @rao2z I agree that an agent that can detect that the anticipated effects of actions were not observed would be more robust (and could detect novelty). But is this capability required for a system to be treated as an "agent"?
2022-11-14 23:59:23 @rao2z Similar models can be trained to predict the next action in a policy. So they could *execute* a policy, but not create one. If we take the "agent stance", we can predict their behavior by attributing goals to them. Are you wanting an agent to do more?
2022-11-14 16:20:50 @jacobpattywagon The later shots are indoors. Outdoors there would be hats (but different ones)
2022-11-14 06:32:09 @Raza_Habib496 If we can predict the actions of a system by attributing to it goals and beliefs, then it is useful to treat it as an agent. What are the goals, beliefs, and actions of an LLM?
2022-11-18 22:29:56 @taaltree But lots of former Twitter engineers are posting on this and other sites, and I believe them.
2022-11-17 16:59:52 @lilkuo "surrealism is the rational response to dictatorship" -- Carles Sierra
2022-11-17 04:17:32 @CT_Bergstrom Is this because it is trained on a lot of papers written by white men?
2022-11-17 04:17:03 @dcbaok @CT_Bergstrom Same with papers in Science and Nature (with some exceptions such as the DeepMind papers)
2022-11-16 22:42:42 @GaryMarcus On the other hand, after playing with it, I think most of its hypotheses are false
2022-11-16 22:40:27 @MattNiessner Just like startups, a lot of cutting edge tech fails to meet a real need. So I think people rationally adopt a negatively-biased prior
2022-11-16 16:58:53 @deliprao More than 11 people have died because of rumors and misinformation spread on social networks
2022-11-16 05:55:24 https://t.co/YaH8PuHnIU
2022-11-16 05:54:57 Giving Galactica a try. This mainly tells me that the face recognition community use this term differently than I do. For me, open category == open set. https://t.co/jGYx8CbxJq
2022-11-16 03:45:41 @GaryMarcus But the great thing about proofs is that they can be checked. Think of it has an hypothesis generator, not a source of truth. Same applies to generating code.
2022-11-15 22:12:55 RT @MattBetts11: Excited for this week’s EECB talk from @karst_justine !
2022-11-15 21:43:17 @noop_noob @rao2z Yes, some theories of consciousness would consider that a consciousness mechanism.
2022-11-15 20:43:53 @AmanBitz @ylecun @DebasmitDas1 @roydanroy Current SOTA according to https://t.co/aLwWum6M4Z is 96.81, and it appears to be asymptoting.
2022-11-15 20:42:44 @AmanBitz @ylecun @DebasmitDas1 @roydanroy But suppose you could raise CIFAR100 accuracy to 97%, I wouldn't be interested unless the same techniques worked on ImageNet and COCO and were practical at scale. Working well on CIFAR100 alone doesn't tell us much.
2022-11-15 18:37:29 @ylecun @DebasmitDas1 @roydanroy Well, it depends on what the paper is about. A accuracy-oriented computer vision paper evaluated on MNIST, CIFAR10, and CIFAR100 is no longer interesting. But, a paper studying the open category problem using these data sets is interesting (IMO: https://t.co/0KUqdPhE3f)
2022-11-15 18:15:14 @INSMI_CNRS @BachFrancis @giry_claire @sup_recherche @ENS_ULM @INS2I_CNRS Congratulations!
2022-11-15 16:40:40 @falsalem76 @_jasonwei Not unless it is something that can be learned from data, which I doubt. It is an architectural property.
2022-11-15 15:54:24 @mhudack @AutismCapital No, it’s anarchist
2022-11-15 04:42:08 @DebasmitDas1 @roydanroy That’ll be obsolete tomorrow
2022-11-15 04:39:36 @andyroyle_pwrc Much cheaper in Spain. Maybe subsidized there?
2022-11-15 04:39:00 @MarkSorel8 @andyroyle_pwrc My pharmacist did the paperwork for me. Thanks @FredMeyerInc
2022-11-15 00:33:11 @rao2z I agree that an agent that can detect that the anticipated effects of actions were not observed would be more robust (and could detect novelty). But is this capability required for a system to be treated as an "agent"?
2022-11-14 23:59:23 @rao2z Similar models can be trained to predict the next action in a policy. So they could *execute* a policy, but not create one. If we take the "agent stance", we can predict their behavior by attributing goals to them. Are you wanting an agent to do more?
2022-11-14 16:20:50 @jacobpattywagon The later shots are indoors. Outdoors there would be hats (but different ones)
2022-11-14 06:32:09 @Raza_Habib496 If we can predict the actions of a system by attributing to it goals and beliefs, then it is useful to treat it as an agent. What are the goals, beliefs, and actions of an LLM?
2022-11-20 03:59:54 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 I agree it would be great for the research community to have access to these models. I'm not sure that full open source would be wise. The trolls and nation state actors would then also have access to refine their methods.
2022-11-20 03:49:16 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 This will still make it easy to generate fluent misinformation: Just build a collection of documents full of conspiracy theories, racism, etc.
2022-11-20 03:46:47 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 LLMs can generate lots of plausible hypotheses, but these need to be checked against reliable knowledge sources.
2022-11-20 03:45:57 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 I think LLMs will be replaced with models that make heavy use of external resources (document stores, databases, theorem provers, simulators, etc.). LLMs as we know them today are best viewed as pattern matching heuristics, not knowledge bases
2022-11-20 03:28:10 @deliprao @ben_golub If a grammar rule is consistent, it is quite easy for the net to learn to follow the rule without ever learning an explicit rule.
2022-11-20 03:21:05 @CharlestonCWKC @blekhman Yes, they all do.
2022-11-20 03:19:25 @aaronsibarium Rankings are a racket to sell magazines. The lower-ranked schools don’t dare challenge the rankings, so it is wonderful to see the top schools demolishing the BS-driven enterprise
2022-11-20 00:24:13 @ylecun @GaryMarcus @mrgreene1977 The last chapter of deep fakes hasn’t been written yet.
2022-11-19 21:05:30 @vgcerf He gave an inspiring series of lectures at UIUC when I was a grad student there. I particularly liked his advice about moving back to academia: He said if that was your plan, you should always live as if you were being paid an academic salary.
2022-11-19 20:58:00 @PalliThordarson These are also more likely to be the people who will be recruited to move to a top-ranked institution. And sure enough, with access to more resources, they will become more productive. end/
2022-11-19 20:56:25 @PalliThordarson If you are a student or a faculty member at a lower-ranked institution, you need to remind yourself to "work as if I was at Stanford"--that is, work on the most important problems you can. People who do world class research at such institutions have to work harder. 4/
2022-11-19 20:52:20 @PalliThordarson I think faculty and students lower-ranked institutions can suffer from the "tyranny of low expectations". This combines with less access to strong students and facilities can easily lead to less significant research outcomes. 3/
2022-11-19 20:50:13 @PalliThordarson They think "I'm at Harvard, I need to do `Harvard-level' work." Of course, if such work requires more resources and stronger students and postdocs, they have those too. So there is potential confounding. At lower-ranked institutions, do faculty feel less of this pressure? 2/
2022-11-19 20:48:13 @PalliThordarson Note that they did not analyze whether people at elite institutions studied more interesting/fundamental research questions. It's a difficult question to study, but my pet hypothesis is that faculty at elite institutions feel pressure to work on those fundamental questions 1/
2022-11-19 20:22:19 @B1ar2n3a Crazy. People should have the freedom to have kids whenever they think it best. I’ve suggested the PostDoc years as the time that is least stressful and hours are most flexible. But I didn’t do a PostDoc myself, so we had our kids while we were assistant profs
2022-11-19 04:17:49 @rlucas @netcapgirl TCP/IP routes around problems
2022-11-19 04:09:53 @paulnovosad Would you like to make a wager?
2022-11-19 04:08:53 @deliprao Same here. I asked a question for which wanted the answer. The result was partly useful and partly nonsense. Is it a symptom of bad benchmarks that Meta thought this was ready to release? Or maybe an excess of transparency?
2022-11-19 04:04:33 @RiverTamYDN Everyone wants to be treated with dignity
2022-11-19 03:09:01 @moyix Is there a betting market? I’d be curious to see the odds
2022-11-18 22:29:56 @taaltree But lots of former Twitter engineers are posting on this and other sites, and I believe them.
2022-11-17 16:59:52 @lilkuo "surrealism is the rational response to dictatorship" -- Carles Sierra
2022-11-17 04:17:32 @CT_Bergstrom Is this because it is trained on a lot of papers written by white men?
2022-11-17 04:17:03 @dcbaok @CT_Bergstrom Same with papers in Science and Nature (with some exceptions such as the DeepMind papers)
2022-11-16 22:42:42 @GaryMarcus On the other hand, after playing with it, I think most of its hypotheses are false
2022-11-16 22:40:27 @MattNiessner Just like startups, a lot of cutting edge tech fails to meet a real need. So I think people rationally adopt a negatively-biased prior
2022-11-16 16:58:53 @deliprao More than 11 people have died because of rumors and misinformation spread on social networks
2022-11-16 05:55:24 https://t.co/YaH8PuHnIU
2022-11-16 05:54:57 Giving Galactica a try. This mainly tells me that the face recognition community use this term differently than I do. For me, open category == open set. https://t.co/jGYx8CbxJq
2022-11-16 03:45:41 @GaryMarcus But the great thing about proofs is that they can be checked. Think of it has an hypothesis generator, not a source of truth. Same applies to generating code.
2022-11-15 22:12:55 RT @MattBetts11: Excited for this week’s EECB talk from @karst_justine !
2022-11-15 21:43:17 @noop_noob @rao2z Yes, some theories of consciousness would consider that a consciousness mechanism.
2022-11-15 20:43:53 @AmanBitz @ylecun @DebasmitDas1 @roydanroy Current SOTA according to https://t.co/aLwWum6M4Z is 96.81, and it appears to be asymptoting.
2022-11-15 20:42:44 @AmanBitz @ylecun @DebasmitDas1 @roydanroy But suppose you could raise CIFAR100 accuracy to 97%, I wouldn't be interested unless the same techniques worked on ImageNet and COCO and were practical at scale. Working well on CIFAR100 alone doesn't tell us much.
2022-11-15 18:37:29 @ylecun @DebasmitDas1 @roydanroy Well, it depends on what the paper is about. A accuracy-oriented computer vision paper evaluated on MNIST, CIFAR10, and CIFAR100 is no longer interesting. But, a paper studying the open category problem using these data sets is interesting (IMO: https://t.co/0KUqdPhE3f)
2022-11-15 18:15:14 @INSMI_CNRS @BachFrancis @giry_claire @sup_recherche @ENS_ULM @INS2I_CNRS Congratulations!
2022-11-15 16:40:40 @falsalem76 @_jasonwei Not unless it is something that can be learned from data, which I doubt. It is an architectural property.
2022-11-15 15:54:24 @mhudack @AutismCapital No, it’s anarchist
2022-11-15 04:42:08 @DebasmitDas1 @roydanroy That’ll be obsolete tomorrow
2022-11-15 04:39:36 @andyroyle_pwrc Much cheaper in Spain. Maybe subsidized there?
2022-11-15 04:39:00 @MarkSorel8 @andyroyle_pwrc My pharmacist did the paperwork for me. Thanks @FredMeyerInc
2022-11-15 00:33:11 @rao2z I agree that an agent that can detect that the anticipated effects of actions were not observed would be more robust (and could detect novelty). But is this capability required for a system to be treated as an "agent"?
2022-11-14 23:59:23 @rao2z Similar models can be trained to predict the next action in a policy. So they could *execute* a policy, but not create one. If we take the "agent stance", we can predict their behavior by attributing goals to them. Are you wanting an agent to do more?
2022-11-14 16:20:50 @jacobpattywagon The later shots are indoors. Outdoors there would be hats (but different ones)
2022-11-14 06:32:09 @Raza_Habib496 If we can predict the actions of a system by attributing to it goals and beliefs, then it is useful to treat it as an agent. What are the goals, beliefs, and actions of an LLM?
2022-11-20 20:22:48 RT @gradydoctor: 1/ I reached into the console and fished around for my badge. It wasn’t there. My brow furrowed. Me: “Badge, we are not d…
2022-11-20 20:19:12 RT @hholdenthorp: People are rightly focused on how the grad students at UC will get what they deserve for carrying most of the load of the…
2022-11-20 17:04:59 @McknightLaura Yes. Funding agencies need to increase award amounts to compensate
2022-11-20 17:02:55 @deliprao The student is “getting” 80K that they would otherwise have to raise from other sources. That said, I’m all in favor of increasing student salaries
2022-11-20 16:42:10 @deliprao “Stipend” is a misnomer. We pay a salary for the work the student does teaching or doing research. “Tuition” is what the student pays for getting an education (classroom and mentoring). Grants cover both of these (in US)
2022-11-20 03:59:54 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 I agree it would be great for the research community to have access to these models. I'm not sure that full open source would be wise. The trolls and nation state actors would then also have access to refine their methods.
2022-11-20 03:49:16 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 This will still make it easy to generate fluent misinformation: Just build a collection of documents full of conspiracy theories, racism, etc.
2022-11-20 03:46:47 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 LLMs can generate lots of plausible hypotheses, but these need to be checked against reliable knowledge sources.
2022-11-20 03:45:57 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 I think LLMs will be replaced with models that make heavy use of external resources (document stores, databases, theorem provers, simulators, etc.). LLMs as we know them today are best viewed as pattern matching heuristics, not knowledge bases
2022-11-20 03:28:10 @deliprao @ben_golub If a grammar rule is consistent, it is quite easy for the net to learn to follow the rule without ever learning an explicit rule.
2022-11-20 03:21:05 @CharlestonCWKC @blekhman Yes, they all do.
2022-11-20 03:19:25 @aaronsibarium Rankings are a racket to sell magazines. The lower-ranked schools don’t dare challenge the rankings, so it is wonderful to see the top schools demolishing the BS-driven enterprise
2022-11-20 00:24:13 @ylecun @GaryMarcus @mrgreene1977 The last chapter of deep fakes hasn’t been written yet.
2022-11-19 21:05:30 @vgcerf He gave an inspiring series of lectures at UIUC when I was a grad student there. I particularly liked his advice about moving back to academia: He said if that was your plan, you should always live as if you were being paid an academic salary.
2022-11-19 20:58:00 @PalliThordarson These are also more likely to be the people who will be recruited to move to a top-ranked institution. And sure enough, with access to more resources, they will become more productive. end/
2022-11-19 20:56:25 @PalliThordarson If you are a student or a faculty member at a lower-ranked institution, you need to remind yourself to "work as if I was at Stanford"--that is, work on the most important problems you can. People who do world class research at such institutions have to work harder. 4/
2022-11-19 20:52:20 @PalliThordarson I think faculty and students lower-ranked institutions can suffer from the "tyranny of low expectations". This combines with less access to strong students and facilities can easily lead to less significant research outcomes. 3/
2022-11-19 20:50:13 @PalliThordarson They think "I'm at Harvard, I need to do `Harvard-level' work." Of course, if such work requires more resources and stronger students and postdocs, they have those too. So there is potential confounding. At lower-ranked institutions, do faculty feel less of this pressure? 2/
2022-11-19 20:48:13 @PalliThordarson Note that they did not analyze whether people at elite institutions studied more interesting/fundamental research questions. It's a difficult question to study, but my pet hypothesis is that faculty at elite institutions feel pressure to work on those fundamental questions 1/
2022-11-19 20:22:19 @B1ar2n3a Crazy. People should have the freedom to have kids whenever they think it best. I’ve suggested the PostDoc years as the time that is least stressful and hours are most flexible. But I didn’t do a PostDoc myself, so we had our kids while we were assistant profs
2022-11-19 04:17:49 @rlucas @netcapgirl TCP/IP routes around problems
2022-11-19 04:09:53 @paulnovosad Would you like to make a wager?
2022-11-19 04:08:53 @deliprao Same here. I asked a question for which wanted the answer. The result was partly useful and partly nonsense. Is it a symptom of bad benchmarks that Meta thought this was ready to release? Or maybe an excess of transparency?
2022-11-19 04:04:33 @RiverTamYDN Everyone wants to be treated with dignity
2022-11-19 03:09:01 @moyix Is there a betting market? I’d be curious to see the odds
2022-11-18 22:29:56 @taaltree But lots of former Twitter engineers are posting on this and other sites, and I believe them.
2022-11-17 16:59:52 @lilkuo "surrealism is the rational response to dictatorship" -- Carles Sierra
2022-11-17 04:17:32 @CT_Bergstrom Is this because it is trained on a lot of papers written by white men?
2022-11-17 04:17:03 @dcbaok @CT_Bergstrom Same with papers in Science and Nature (with some exceptions such as the DeepMind papers)
2022-11-16 22:42:42 @GaryMarcus On the other hand, after playing with it, I think most of its hypotheses are false
2022-11-16 22:40:27 @MattNiessner Just like startups, a lot of cutting edge tech fails to meet a real need. So I think people rationally adopt a negatively-biased prior
2022-11-16 16:58:53 @deliprao More than 11 people have died because of rumors and misinformation spread on social networks
2022-11-16 05:55:24 https://t.co/YaH8PuHnIU
2022-11-16 05:54:57 Giving Galactica a try. This mainly tells me that the face recognition community use this term differently than I do. For me, open category == open set. https://t.co/jGYx8CbxJq
2022-11-16 03:45:41 @GaryMarcus But the great thing about proofs is that they can be checked. Think of it has an hypothesis generator, not a source of truth. Same applies to generating code.
2022-11-15 22:12:55 RT @MattBetts11: Excited for this week’s EECB talk from @karst_justine !
2022-11-15 21:43:17 @noop_noob @rao2z Yes, some theories of consciousness would consider that a consciousness mechanism.
2022-11-15 20:43:53 @AmanBitz @ylecun @DebasmitDas1 @roydanroy Current SOTA according to https://t.co/aLwWum6M4Z is 96.81, and it appears to be asymptoting.
2022-11-15 20:42:44 @AmanBitz @ylecun @DebasmitDas1 @roydanroy But suppose you could raise CIFAR100 accuracy to 97%, I wouldn't be interested unless the same techniques worked on ImageNet and COCO and were practical at scale. Working well on CIFAR100 alone doesn't tell us much.
2022-11-15 18:37:29 @ylecun @DebasmitDas1 @roydanroy Well, it depends on what the paper is about. A accuracy-oriented computer vision paper evaluated on MNIST, CIFAR10, and CIFAR100 is no longer interesting. But, a paper studying the open category problem using these data sets is interesting (IMO: https://t.co/0KUqdPhE3f)
2022-11-15 18:15:14 @INSMI_CNRS @BachFrancis @giry_claire @sup_recherche @ENS_ULM @INS2I_CNRS Congratulations!
2022-11-15 16:40:40 @falsalem76 @_jasonwei Not unless it is something that can be learned from data, which I doubt. It is an architectural property.
2022-11-15 15:54:24 @mhudack @AutismCapital No, it’s anarchist
2022-11-15 04:42:08 @DebasmitDas1 @roydanroy That’ll be obsolete tomorrow
2022-11-15 04:39:36 @andyroyle_pwrc Much cheaper in Spain. Maybe subsidized there?
2022-11-15 04:39:00 @MarkSorel8 @andyroyle_pwrc My pharmacist did the paperwork for me. Thanks @FredMeyerInc
2022-11-15 00:33:11 @rao2z I agree that an agent that can detect that the anticipated effects of actions were not observed would be more robust (and could detect novelty). But is this capability required for a system to be treated as an "agent"?
2022-11-14 23:59:23 @rao2z Similar models can be trained to predict the next action in a policy. So they could *execute* a policy, but not create one. If we take the "agent stance", we can predict their behavior by attributing goals to them. Are you wanting an agent to do more?
2022-11-14 16:20:50 @jacobpattywagon The later shots are indoors. Outdoors there would be hats (but different ones)
2022-11-14 06:32:09 @Raza_Habib496 If we can predict the actions of a system by attributing to it goals and beliefs, then it is useful to treat it as an agent. What are the goals, beliefs, and actions of an LLM?
2022-11-22 03:02:20 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI @ylecun What was the first good LLM whose weights were released?
2022-11-22 02:46:31 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI But I should have included entrepreneurs who want to build on top of these models, not just researchers
2022-11-22 02:45:16 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI Same rules for them, IMO.
2022-11-22 00:23:57 @wightmanr @EMostaque @paperswithcode @MetaAI Don't you think there are risks that the model could be deployed for disinformation campaigns, phishing, and so on? How do you enforce the license without some lock and key?
2022-11-22 00:01:43 @TheRandomMtrix @ChristophMolnar I think this is incorrect. There are many ML models that do not support statistical inferences, and some statistical models that are poor predictive models but do support inference.
2022-11-21 05:23:55 @EMostaque @paperswithcode @MetaAI I think you should put it behind a security wall but give access to authenticated researchers with tight license conditions.
2022-11-20 20:22:48 RT @gradydoctor: 1/ I reached into the console and fished around for my badge. It wasn’t there. My brow furrowed. Me: “Badge, we are not d…
2022-11-20 20:19:12 RT @hholdenthorp: People are rightly focused on how the grad students at UC will get what they deserve for carrying most of the load of the…
2022-11-20 17:04:59 @McknightLaura Yes. Funding agencies need to increase award amounts to compensate
2022-11-20 17:02:55 @deliprao The student is “getting” 80K that they would otherwise have to raise from other sources. That said, I’m all in favor of increasing student salaries
2022-11-20 16:42:10 @deliprao “Stipend” is a misnomer. We pay a salary for the work the student does teaching or doing research. “Tuition” is what the student pays for getting an education (classroom and mentoring). Grants cover both of these (in US)
2022-11-20 03:59:54 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 I agree it would be great for the research community to have access to these models. I'm not sure that full open source would be wise. The trolls and nation state actors would then also have access to refine their methods.
2022-11-20 03:49:16 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 This will still make it easy to generate fluent misinformation: Just build a collection of documents full of conspiracy theories, racism, etc.
2022-11-20 03:46:47 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 LLMs can generate lots of plausible hypotheses, but these need to be checked against reliable knowledge sources.
2022-11-20 03:45:57 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 I think LLMs will be replaced with models that make heavy use of external resources (document stores, databases, theorem provers, simulators, etc.). LLMs as we know them today are best viewed as pattern matching heuristics, not knowledge bases
2022-11-20 03:28:10 @deliprao @ben_golub If a grammar rule is consistent, it is quite easy for the net to learn to follow the rule without ever learning an explicit rule.
2022-11-20 03:21:05 @CharlestonCWKC @blekhman Yes, they all do.
2022-11-20 03:19:25 @aaronsibarium Rankings are a racket to sell magazines. The lower-ranked schools don’t dare challenge the rankings, so it is wonderful to see the top schools demolishing the BS-driven enterprise
2022-11-20 00:24:13 @ylecun @GaryMarcus @mrgreene1977 The last chapter of deep fakes hasn’t been written yet.
2022-11-19 21:05:30 @vgcerf He gave an inspiring series of lectures at UIUC when I was a grad student there. I particularly liked his advice about moving back to academia: He said if that was your plan, you should always live as if you were being paid an academic salary.
2022-11-19 20:58:00 @PalliThordarson These are also more likely to be the people who will be recruited to move to a top-ranked institution. And sure enough, with access to more resources, they will become more productive. end/
2022-11-19 20:56:25 @PalliThordarson If you are a student or a faculty member at a lower-ranked institution, you need to remind yourself to "work as if I was at Stanford"--that is, work on the most important problems you can. People who do world class research at such institutions have to work harder. 4/
2022-11-19 20:52:20 @PalliThordarson I think faculty and students lower-ranked institutions can suffer from the "tyranny of low expectations". This combines with less access to strong students and facilities can easily lead to less significant research outcomes. 3/
2022-11-19 20:50:13 @PalliThordarson They think "I'm at Harvard, I need to do `Harvard-level' work." Of course, if such work requires more resources and stronger students and postdocs, they have those too. So there is potential confounding. At lower-ranked institutions, do faculty feel less of this pressure? 2/
2022-11-19 20:48:13 @PalliThordarson Note that they did not analyze whether people at elite institutions studied more interesting/fundamental research questions. It's a difficult question to study, but my pet hypothesis is that faculty at elite institutions feel pressure to work on those fundamental questions 1/
2022-11-19 20:22:19 @B1ar2n3a Crazy. People should have the freedom to have kids whenever they think it best. I’ve suggested the PostDoc years as the time that is least stressful and hours are most flexible. But I didn’t do a PostDoc myself, so we had our kids while we were assistant profs
2022-11-19 04:17:49 @rlucas @netcapgirl TCP/IP routes around problems
2022-11-19 04:09:53 @paulnovosad Would you like to make a wager?
2022-11-19 04:08:53 @deliprao Same here. I asked a question for which wanted the answer. The result was partly useful and partly nonsense. Is it a symptom of bad benchmarks that Meta thought this was ready to release? Or maybe an excess of transparency?
2022-11-19 04:04:33 @RiverTamYDN Everyone wants to be treated with dignity
2022-11-19 03:09:01 @moyix Is there a betting market? I’d be curious to see the odds
2022-11-18 22:29:56 @taaltree But lots of former Twitter engineers are posting on this and other sites, and I believe them.
2022-11-17 16:59:52 @lilkuo "surrealism is the rational response to dictatorship" -- Carles Sierra
2022-11-17 04:17:32 @CT_Bergstrom Is this because it is trained on a lot of papers written by white men?
2022-11-17 04:17:03 @dcbaok @CT_Bergstrom Same with papers in Science and Nature (with some exceptions such as the DeepMind papers)
2022-11-16 22:42:42 @GaryMarcus On the other hand, after playing with it, I think most of its hypotheses are false
2022-11-16 22:40:27 @MattNiessner Just like startups, a lot of cutting edge tech fails to meet a real need. So I think people rationally adopt a negatively-biased prior
2022-11-16 16:58:53 @deliprao More than 11 people have died because of rumors and misinformation spread on social networks
2022-11-16 05:55:24 https://t.co/YaH8PuHnIU
2022-11-16 05:54:57 Giving Galactica a try. This mainly tells me that the face recognition community use this term differently than I do. For me, open category == open set. https://t.co/jGYx8CbxJq
2022-11-16 03:45:41 @GaryMarcus But the great thing about proofs is that they can be checked. Think of it has an hypothesis generator, not a source of truth. Same applies to generating code.
2022-11-15 22:12:55 RT @MattBetts11: Excited for this week’s EECB talk from @karst_justine !
2022-11-15 21:43:17 @noop_noob @rao2z Yes, some theories of consciousness would consider that a consciousness mechanism.
2022-11-15 20:43:53 @AmanBitz @ylecun @DebasmitDas1 @roydanroy Current SOTA according to https://t.co/aLwWum6M4Z is 96.81, and it appears to be asymptoting.
2022-11-15 20:42:44 @AmanBitz @ylecun @DebasmitDas1 @roydanroy But suppose you could raise CIFAR100 accuracy to 97%, I wouldn't be interested unless the same techniques worked on ImageNet and COCO and were practical at scale. Working well on CIFAR100 alone doesn't tell us much.
2022-11-15 18:37:29 @ylecun @DebasmitDas1 @roydanroy Well, it depends on what the paper is about. A accuracy-oriented computer vision paper evaluated on MNIST, CIFAR10, and CIFAR100 is no longer interesting. But, a paper studying the open category problem using these data sets is interesting (IMO: https://t.co/0KUqdPhE3f)
2022-11-15 18:15:14 @INSMI_CNRS @BachFrancis @giry_claire @sup_recherche @ENS_ULM @INS2I_CNRS Congratulations!
2022-11-15 16:40:40 @falsalem76 @_jasonwei Not unless it is something that can be learned from data, which I doubt. It is an architectural property.
2022-11-15 15:54:24 @mhudack @AutismCapital No, it’s anarchist
2022-11-15 04:42:08 @DebasmitDas1 @roydanroy That’ll be obsolete tomorrow
2022-11-15 04:39:36 @andyroyle_pwrc Much cheaper in Spain. Maybe subsidized there?
2022-11-15 04:39:00 @MarkSorel8 @andyroyle_pwrc My pharmacist did the paperwork for me. Thanks @FredMeyerInc
2022-11-15 00:33:11 @rao2z I agree that an agent that can detect that the anticipated effects of actions were not observed would be more robust (and could detect novelty). But is this capability required for a system to be treated as an "agent"?
2022-11-14 23:59:23 @rao2z Similar models can be trained to predict the next action in a policy. So they could *execute* a policy, but not create one. If we take the "agent stance", we can predict their behavior by attributing goals to them. Are you wanting an agent to do more?
2022-11-14 16:20:50 @jacobpattywagon The later shots are indoors. Outdoors there would be hats (but different ones)
2022-11-14 06:32:09 @Raza_Habib496 If we can predict the actions of a system by attributing to it goals and beliefs, then it is useful to treat it as an agent. What are the goals, beliefs, and actions of an LLM?
2022-11-22 23:35:17 @ylecun @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI But as @Michael_J_Black points out, Galactica may open the door to a new set of fake science "attacks" on the publication system. I think it might also support a great semantic auto-correct for authors (esp ones with less skill in English)
2022-11-22 22:44:16 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI @ylecun Oh, and GPT-NeoX was released in April of 2022, I believe (20B parameters). It isn't obvious that Galactica is dangerously more powerful than these existing open models.
2022-11-22 22:42:12 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI @ylecun I did a bit of searching. XLNet was released in 2019, I believe. The weights for RoBERTa, DeBERTa, and XLM-RoBERTa are available I think. I'm not sure about OPT-175B. The fact that these are all available changes my position regarding Galactica. 1/
2022-11-22 22:00:18 @GaryMarcus @___merc___ Another hypothesis is that Amazon never figured out how to make a business out of it.
2022-11-22 20:48:37 @eugediana @MetaAI I think we can learn a lot by studying network internals and activations and their relationship to the training data. For this, we ideally have access to the network weights and training data.
2022-11-22 20:11:39 @MetaAI I'm looking forward to studying this system in detail.
2022-11-22 17:46:53 "They would ache with misplaced vigilance, they would lash out in anger, never be able to scratch the itch of fear, be torn by the longing to forget and the urge to always remember." Amazing writing by Dave Philips https://t.co/ia2bdDL0cY
2022-11-22 03:02:20 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI @ylecun What was the first good LLM whose weights were released?
2022-11-22 02:46:31 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI But I should have included entrepreneurs who want to build on top of these models, not just researchers
2022-11-22 02:45:16 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI Same rules for them, IMO.
2022-11-22 00:23:57 @wightmanr @EMostaque @paperswithcode @MetaAI Don't you think there are risks that the model could be deployed for disinformation campaigns, phishing, and so on? How do you enforce the license without some lock and key?
2022-11-22 00:01:43 @TheRandomMtrix @ChristophMolnar I think this is incorrect. There are many ML models that do not support statistical inferences, and some statistical models that are poor predictive models but do support inference.
2022-11-21 05:23:55 @EMostaque @paperswithcode @MetaAI I think you should put it behind a security wall but give access to authenticated researchers with tight license conditions.
2022-11-20 20:22:48 RT @gradydoctor: 1/ I reached into the console and fished around for my badge. It wasn’t there. My brow furrowed. Me: “Badge, we are not d…
2022-11-20 20:19:12 RT @hholdenthorp: People are rightly focused on how the grad students at UC will get what they deserve for carrying most of the load of the…
2022-11-20 17:04:59 @McknightLaura Yes. Funding agencies need to increase award amounts to compensate
2022-11-20 17:02:55 @deliprao The student is “getting” 80K that they would otherwise have to raise from other sources. That said, I’m all in favor of increasing student salaries
2022-11-20 16:42:10 @deliprao “Stipend” is a misnomer. We pay a salary for the work the student does teaching or doing research. “Tuition” is what the student pays for getting an education (classroom and mentoring). Grants cover both of these (in US)
2022-11-20 03:59:54 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 I agree it would be great for the research community to have access to these models. I'm not sure that full open source would be wise. The trolls and nation state actors would then also have access to refine their methods.
2022-11-20 03:49:16 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 This will still make it easy to generate fluent misinformation: Just build a collection of documents full of conspiracy theories, racism, etc.
2022-11-20 03:46:47 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 LLMs can generate lots of plausible hypotheses, but these need to be checked against reliable knowledge sources.
2022-11-20 03:45:57 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 I think LLMs will be replaced with models that make heavy use of external resources (document stores, databases, theorem provers, simulators, etc.). LLMs as we know them today are best viewed as pattern matching heuristics, not knowledge bases
2022-11-20 03:28:10 @deliprao @ben_golub If a grammar rule is consistent, it is quite easy for the net to learn to follow the rule without ever learning an explicit rule.
2022-11-20 03:21:05 @CharlestonCWKC @blekhman Yes, they all do.
2022-11-20 03:19:25 @aaronsibarium Rankings are a racket to sell magazines. The lower-ranked schools don’t dare challenge the rankings, so it is wonderful to see the top schools demolishing the BS-driven enterprise
2022-11-20 00:24:13 @ylecun @GaryMarcus @mrgreene1977 The last chapter of deep fakes hasn’t been written yet.
2022-11-19 21:05:30 @vgcerf He gave an inspiring series of lectures at UIUC when I was a grad student there. I particularly liked his advice about moving back to academia: He said if that was your plan, you should always live as if you were being paid an academic salary.
2022-11-19 20:58:00 @PalliThordarson These are also more likely to be the people who will be recruited to move to a top-ranked institution. And sure enough, with access to more resources, they will become more productive. end/
2022-11-19 20:56:25 @PalliThordarson If you are a student or a faculty member at a lower-ranked institution, you need to remind yourself to "work as if I was at Stanford"--that is, work on the most important problems you can. People who do world class research at such institutions have to work harder. 4/
2022-11-19 20:52:20 @PalliThordarson I think faculty and students lower-ranked institutions can suffer from the "tyranny of low expectations". This combines with less access to strong students and facilities can easily lead to less significant research outcomes. 3/
2022-11-19 20:50:13 @PalliThordarson They think "I'm at Harvard, I need to do `Harvard-level' work." Of course, if such work requires more resources and stronger students and postdocs, they have those too. So there is potential confounding. At lower-ranked institutions, do faculty feel less of this pressure? 2/
2022-11-19 20:48:13 @PalliThordarson Note that they did not analyze whether people at elite institutions studied more interesting/fundamental research questions. It's a difficult question to study, but my pet hypothesis is that faculty at elite institutions feel pressure to work on those fundamental questions 1/
2022-11-19 20:22:19 @B1ar2n3a Crazy. People should have the freedom to have kids whenever they think it best. I’ve suggested the PostDoc years as the time that is least stressful and hours are most flexible. But I didn’t do a PostDoc myself, so we had our kids while we were assistant profs
2022-11-19 04:17:49 @rlucas @netcapgirl TCP/IP routes around problems
2022-11-19 04:09:53 @paulnovosad Would you like to make a wager?
2022-11-19 04:08:53 @deliprao Same here. I asked a question for which wanted the answer. The result was partly useful and partly nonsense. Is it a symptom of bad benchmarks that Meta thought this was ready to release? Or maybe an excess of transparency?
2022-11-19 04:04:33 @RiverTamYDN Everyone wants to be treated with dignity
2022-11-19 03:09:01 @moyix Is there a betting market? I’d be curious to see the odds
2022-11-18 22:29:56 @taaltree But lots of former Twitter engineers are posting on this and other sites, and I believe them.
2022-11-17 16:59:52 @lilkuo "surrealism is the rational response to dictatorship" -- Carles Sierra
2022-11-17 04:17:32 @CT_Bergstrom Is this because it is trained on a lot of papers written by white men?
2022-11-17 04:17:03 @dcbaok @CT_Bergstrom Same with papers in Science and Nature (with some exceptions such as the DeepMind papers)
2022-11-16 22:42:42 @GaryMarcus On the other hand, after playing with it, I think most of its hypotheses are false
2022-11-16 22:40:27 @MattNiessner Just like startups, a lot of cutting edge tech fails to meet a real need. So I think people rationally adopt a negatively-biased prior
2022-11-16 16:58:53 @deliprao More than 11 people have died because of rumors and misinformation spread on social networks
2022-11-16 05:55:24 https://t.co/YaH8PuHnIU
2022-11-16 05:54:57 Giving Galactica a try. This mainly tells me that the face recognition community use this term differently than I do. For me, open category == open set. https://t.co/jGYx8CbxJq
2022-11-16 03:45:41 @GaryMarcus But the great thing about proofs is that they can be checked. Think of it has an hypothesis generator, not a source of truth. Same applies to generating code.
2022-11-15 22:12:55 RT @MattBetts11: Excited for this week’s EECB talk from @karst_justine !
2022-11-15 21:43:17 @noop_noob @rao2z Yes, some theories of consciousness would consider that a consciousness mechanism.
2022-11-15 20:43:53 @AmanBitz @ylecun @DebasmitDas1 @roydanroy Current SOTA according to https://t.co/aLwWum6M4Z is 96.81, and it appears to be asymptoting.
2022-11-15 20:42:44 @AmanBitz @ylecun @DebasmitDas1 @roydanroy But suppose you could raise CIFAR100 accuracy to 97%, I wouldn't be interested unless the same techniques worked on ImageNet and COCO and were practical at scale. Working well on CIFAR100 alone doesn't tell us much.
2022-11-15 18:37:29 @ylecun @DebasmitDas1 @roydanroy Well, it depends on what the paper is about. A accuracy-oriented computer vision paper evaluated on MNIST, CIFAR10, and CIFAR100 is no longer interesting. But, a paper studying the open category problem using these data sets is interesting (IMO: https://t.co/0KUqdPhE3f)
2022-11-15 18:15:14 @INSMI_CNRS @BachFrancis @giry_claire @sup_recherche @ENS_ULM @INS2I_CNRS Congratulations!
2022-11-15 16:40:40 @falsalem76 @_jasonwei Not unless it is something that can be learned from data, which I doubt. It is an architectural property.
2022-11-15 15:54:24 @mhudack @AutismCapital No, it’s anarchist
2022-11-15 04:42:08 @DebasmitDas1 @roydanroy That’ll be obsolete tomorrow
2022-11-15 04:39:36 @andyroyle_pwrc Much cheaper in Spain. Maybe subsidized there?
2022-11-15 04:39:00 @MarkSorel8 @andyroyle_pwrc My pharmacist did the paperwork for me. Thanks @FredMeyerInc
2022-11-15 00:33:11 @rao2z I agree that an agent that can detect that the anticipated effects of actions were not observed would be more robust (and could detect novelty). But is this capability required for a system to be treated as an "agent"?
2022-11-14 23:59:23 @rao2z Similar models can be trained to predict the next action in a policy. So they could *execute* a policy, but not create one. If we take the "agent stance", we can predict their behavior by attributing goals to them. Are you wanting an agent to do more?
2022-11-14 16:20:50 @jacobpattywagon The later shots are indoors. Outdoors there would be hats (but different ones)
2022-11-14 06:32:09 @Raza_Habib496 If we can predict the actions of a system by attributing to it goals and beliefs, then it is useful to treat it as an agent. What are the goals, beliefs, and actions of an LLM?
2022-11-25 02:17:39 @KolotaTyler @Noahpinion I think education is a practice in many cultures. Jews have learned to read in order to study Torah and Talmud. Chinese studied for the national exam.
2022-11-24 08:04:09 @houmanhemmati The rules you list absolutely do apply to clinical trials but not absolutely to public health.
2022-11-24 08:02:52 @houmanhemmati These rules are not absolute. Extreme case: If a person is infectious and would infect 100s of other people with a lethal disease if not quarantined, the autonomy of those 100 people outweighs the individual. It would be unethical not to confine him.
2022-11-23 22:03:09 @EMostaque @rasbt @ylecun @leonpalafox @wightmanr @paperswithcode @MetaAI @Michael_J_Black Suppose 10% of future papers are synthesized fakes. All of us will need to review 10% more papers. We are already straining under the current reviewing load. "Things like peer review and curation" are not free. To be continued after I process today's 201 cs.LG submissions
2022-11-23 19:48:49 @rao2z @AIESConf No, I would definitely consider that deception. But, at least according to Dennett, "wrong belief" would also be attributed by an observer (along with "cause"). I haven't read Bratman
2022-11-23 19:28:37 @rao2z @AIESConf I'm with Dennett here: Intention is attributed, not necessarily captured in an explicit knowledge structure
2022-11-23 19:20:08 @rao2z @AIESConf Even without an explicit theory-of-mind, game-playing agents can deceive. We've had this capability since at least the invention of the AO* algorithm (Martelli &
2022-11-23 19:15:41 @IDoTheThinking I think you are mistaken. There are many couples choosing not to have children because of concerns about the impact of large human populations.
2022-11-22 23:35:17 @ylecun @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI But as @Michael_J_Black points out, Galactica may open the door to a new set of fake science "attacks" on the publication system. I think it might also support a great semantic auto-correct for authors (esp ones with less skill in English)
2022-11-22 22:44:16 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI @ylecun Oh, and GPT-NeoX was released in April of 2022, I believe (20B parameters). It isn't obvious that Galactica is dangerously more powerful than these existing open models.
2022-11-22 22:42:12 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI @ylecun I did a bit of searching. XLNet was released in 2019, I believe. The weights for RoBERTa, DeBERTa, and XLM-RoBERTa are available I think. I'm not sure about OPT-175B. The fact that these are all available changes my position regarding Galactica. 1/
2022-11-22 22:00:18 @GaryMarcus @___merc___ Another hypothesis is that Amazon never figured out how to make a business out of it.
2022-11-22 20:48:37 @eugediana @MetaAI I think we can learn a lot by studying network internals and activations and their relationship to the training data. For this, we ideally have access to the network weights and training data.
2022-11-22 20:11:39 @MetaAI I'm looking forward to studying this system in detail.
2022-11-22 17:46:53 "They would ache with misplaced vigilance, they would lash out in anger, never be able to scratch the itch of fear, be torn by the longing to forget and the urge to always remember." Amazing writing by Dave Philips https://t.co/ia2bdDL0cY
2022-11-22 03:02:20 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI @ylecun What was the first good LLM whose weights were released?
2022-11-22 02:46:31 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI But I should have included entrepreneurs who want to build on top of these models, not just researchers
2022-11-22 02:45:16 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI Same rules for them, IMO.
2022-11-22 00:23:57 @wightmanr @EMostaque @paperswithcode @MetaAI Don't you think there are risks that the model could be deployed for disinformation campaigns, phishing, and so on? How do you enforce the license without some lock and key?
2022-11-22 00:01:43 @TheRandomMtrix @ChristophMolnar I think this is incorrect. There are many ML models that do not support statistical inferences, and some statistical models that are poor predictive models but do support inference.
2022-11-21 05:23:55 @EMostaque @paperswithcode @MetaAI I think you should put it behind a security wall but give access to authenticated researchers with tight license conditions.
2022-11-20 20:22:48 RT @gradydoctor: 1/ I reached into the console and fished around for my badge. It wasn’t there. My brow furrowed. Me: “Badge, we are not d…
2022-11-20 20:19:12 RT @hholdenthorp: People are rightly focused on how the grad students at UC will get what they deserve for carrying most of the load of the…
2022-11-20 17:04:59 @McknightLaura Yes. Funding agencies need to increase award amounts to compensate
2022-11-20 17:02:55 @deliprao The student is “getting” 80K that they would otherwise have to raise from other sources. That said, I’m all in favor of increasing student salaries
2022-11-20 16:42:10 @deliprao “Stipend” is a misnomer. We pay a salary for the work the student does teaching or doing research. “Tuition” is what the student pays for getting an education (classroom and mentoring). Grants cover both of these (in US)
2022-11-20 03:59:54 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 I agree it would be great for the research community to have access to these models. I'm not sure that full open source would be wise. The trolls and nation state actors would then also have access to refine their methods.
2022-11-20 03:49:16 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 This will still make it easy to generate fluent misinformation: Just build a collection of documents full of conspiracy theories, racism, etc.
2022-11-20 03:46:47 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 LLMs can generate lots of plausible hypotheses, but these need to be checked against reliable knowledge sources.
2022-11-20 03:45:57 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 I think LLMs will be replaced with models that make heavy use of external resources (document stores, databases, theorem provers, simulators, etc.). LLMs as we know them today are best viewed as pattern matching heuristics, not knowledge bases
2022-11-20 03:28:10 @deliprao @ben_golub If a grammar rule is consistent, it is quite easy for the net to learn to follow the rule without ever learning an explicit rule.
2022-11-20 03:21:05 @CharlestonCWKC @blekhman Yes, they all do.
2022-11-20 03:19:25 @aaronsibarium Rankings are a racket to sell magazines. The lower-ranked schools don’t dare challenge the rankings, so it is wonderful to see the top schools demolishing the BS-driven enterprise
2022-11-20 00:24:13 @ylecun @GaryMarcus @mrgreene1977 The last chapter of deep fakes hasn’t been written yet.
2022-11-19 21:05:30 @vgcerf He gave an inspiring series of lectures at UIUC when I was a grad student there. I particularly liked his advice about moving back to academia: He said if that was your plan, you should always live as if you were being paid an academic salary.
2022-11-19 20:58:00 @PalliThordarson These are also more likely to be the people who will be recruited to move to a top-ranked institution. And sure enough, with access to more resources, they will become more productive. end/
2022-11-19 20:56:25 @PalliThordarson If you are a student or a faculty member at a lower-ranked institution, you need to remind yourself to "work as if I was at Stanford"--that is, work on the most important problems you can. People who do world class research at such institutions have to work harder. 4/
2022-11-19 20:52:20 @PalliThordarson I think faculty and students lower-ranked institutions can suffer from the "tyranny of low expectations". This combines with less access to strong students and facilities can easily lead to less significant research outcomes. 3/
2022-11-19 20:50:13 @PalliThordarson They think "I'm at Harvard, I need to do `Harvard-level' work." Of course, if such work requires more resources and stronger students and postdocs, they have those too. So there is potential confounding. At lower-ranked institutions, do faculty feel less of this pressure? 2/
2022-11-19 20:48:13 @PalliThordarson Note that they did not analyze whether people at elite institutions studied more interesting/fundamental research questions. It's a difficult question to study, but my pet hypothesis is that faculty at elite institutions feel pressure to work on those fundamental questions 1/
2022-11-19 20:22:19 @B1ar2n3a Crazy. People should have the freedom to have kids whenever they think it best. I’ve suggested the PostDoc years as the time that is least stressful and hours are most flexible. But I didn’t do a PostDoc myself, so we had our kids while we were assistant profs
2022-11-19 04:17:49 @rlucas @netcapgirl TCP/IP routes around problems
2022-11-19 04:09:53 @paulnovosad Would you like to make a wager?
2022-11-19 04:08:53 @deliprao Same here. I asked a question for which wanted the answer. The result was partly useful and partly nonsense. Is it a symptom of bad benchmarks that Meta thought this was ready to release? Or maybe an excess of transparency?
2022-11-19 04:04:33 @RiverTamYDN Everyone wants to be treated with dignity
2022-11-19 03:09:01 @moyix Is there a betting market? I’d be curious to see the odds
2022-11-18 22:29:56 @taaltree But lots of former Twitter engineers are posting on this and other sites, and I believe them.
2022-11-17 16:59:52 @lilkuo "surrealism is the rational response to dictatorship" -- Carles Sierra
2022-11-17 04:17:32 @CT_Bergstrom Is this because it is trained on a lot of papers written by white men?
2022-11-17 04:17:03 @dcbaok @CT_Bergstrom Same with papers in Science and Nature (with some exceptions such as the DeepMind papers)
2022-11-16 22:42:42 @GaryMarcus On the other hand, after playing with it, I think most of its hypotheses are false
2022-11-16 22:40:27 @MattNiessner Just like startups, a lot of cutting edge tech fails to meet a real need. So I think people rationally adopt a negatively-biased prior
2022-11-16 16:58:53 @deliprao More than 11 people have died because of rumors and misinformation spread on social networks
2022-11-16 05:55:24 https://t.co/YaH8PuHnIU
2022-11-16 05:54:57 Giving Galactica a try. This mainly tells me that the face recognition community use this term differently than I do. For me, open category == open set. https://t.co/jGYx8CbxJq
2022-11-16 03:45:41 @GaryMarcus But the great thing about proofs is that they can be checked. Think of it has an hypothesis generator, not a source of truth. Same applies to generating code.
2022-11-15 22:12:55 RT @MattBetts11: Excited for this week’s EECB talk from @karst_justine !
2022-11-15 21:43:17 @noop_noob @rao2z Yes, some theories of consciousness would consider that a consciousness mechanism.
2022-11-15 20:43:53 @AmanBitz @ylecun @DebasmitDas1 @roydanroy Current SOTA according to https://t.co/aLwWum6M4Z is 96.81, and it appears to be asymptoting.
2022-11-15 20:42:44 @AmanBitz @ylecun @DebasmitDas1 @roydanroy But suppose you could raise CIFAR100 accuracy to 97%, I wouldn't be interested unless the same techniques worked on ImageNet and COCO and were practical at scale. Working well on CIFAR100 alone doesn't tell us much.
2022-11-15 18:37:29 @ylecun @DebasmitDas1 @roydanroy Well, it depends on what the paper is about. A accuracy-oriented computer vision paper evaluated on MNIST, CIFAR10, and CIFAR100 is no longer interesting. But, a paper studying the open category problem using these data sets is interesting (IMO: https://t.co/0KUqdPhE3f)
2022-11-15 18:15:14 @INSMI_CNRS @BachFrancis @giry_claire @sup_recherche @ENS_ULM @INS2I_CNRS Congratulations!
2022-11-15 16:40:40 @falsalem76 @_jasonwei Not unless it is something that can be learned from data, which I doubt. It is an architectural property.
2022-11-15 15:54:24 @mhudack @AutismCapital No, it’s anarchist
2022-11-15 04:42:08 @DebasmitDas1 @roydanroy That’ll be obsolete tomorrow
2022-11-15 04:39:36 @andyroyle_pwrc Much cheaper in Spain. Maybe subsidized there?
2022-11-15 04:39:00 @MarkSorel8 @andyroyle_pwrc My pharmacist did the paperwork for me. Thanks @FredMeyerInc
2022-11-15 00:33:11 @rao2z I agree that an agent that can detect that the anticipated effects of actions were not observed would be more robust (and could detect novelty). But is this capability required for a system to be treated as an "agent"?
2022-11-14 23:59:23 @rao2z Similar models can be trained to predict the next action in a policy. So they could *execute* a policy, but not create one. If we take the "agent stance", we can predict their behavior by attributing goals to them. Are you wanting an agent to do more?
2022-11-14 16:20:50 @jacobpattywagon The later shots are indoors. Outdoors there would be hats (but different ones)
2022-11-14 06:32:09 @Raza_Habib496 If we can predict the actions of a system by attributing to it goals and beliefs, then it is useful to treat it as an agent. What are the goals, beliefs, and actions of an LLM?
2022-11-29 04:24:55 @pbreit @Plinz Maybe it is the result of an interaction between my network latency and various microservice timeouts? I did not experience these problems prior to the ownership change. I see them in both the iPhone app and in Firefox
2022-11-29 03:11:57 @Plinz I am experiencing lots of problems with Twitter. Expanding a tweet to see replies is often slow or fails with “something went wrong, try again”. Tweets fail to send or send but still give errors. I hope it keeps running!
2022-11-29 02:38:49 @gchrupala Ah, I misread the poll question. Thanks!
2022-11-29 01:45:42 @gchrupala Can anyone confirm that those predictive typing systems are based on LLMs?
2022-11-29 00:55:11 @gchrupala What models have actually been deployed?
2022-11-28 23:31:51 @emilymbender @fchollet I agree with @fchollet on many points, and I like his 3-way categorization. However, all three forms of AI have a shared set of requirements of which a large knowledge base is one. And I don't think it is correct to view emerging LLM capabilities as betters clocks and ladders.
2022-11-28 19:59:32 @emilymbender @fchollet The refrain in the 1980s was "In the knowledge lies the power", i.e., AI performance requires large amounts of world knowledge. LLMs fit to curated data sets and coupled with curated external resources bring us closer to meeting this requirement. They are not just ladder rungs
2022-11-28 16:13:53 @SeanMcCarthyAZ @mfl_normal @MsMelChen I take the learning loss very seriously. I’ve been advocating for a nationwide summer school to help kids catch up. I believe the lost ground can be regained, but we need federal funds to make it happen
2022-11-28 05:07:18 @mfl_normal @MsMelChen Your children missed school. But they were not killed by the police. You were not locked into your house or threatened with prison. Try to get some sense of proportion. We have all suffered in one way or another
2022-11-28 05:05:07 @mfl_normal @MsMelChen Some of these people protesting will get Covid
2022-11-27 23:29:43 @Grady_Booch @_ashawndabney me, so far
2022-11-27 21:08:35 @andrewgwils At least YOU have a stopping rule. I'm still trying to find one for myself
2022-11-27 06:56:34 @mattyglesias The real test will come when they start pushing changes to the system. But they might just succeed. I wish them well!
2022-11-27 06:55:02 @mattyglesias Opening a tweet to see the replies has been consistently failing for me on both iPhone and web browser.
2022-11-27 00:16:25 @jerrellahaynes @TheRealAdamG @JordanJacobs10 @geoffreyhinton The central challenge of software engineering is to specify the desired behavior of the system. AI can't do it alone unless it can infer the desired behavior somehow. But I agree in the near future LLMs will allow us to raise the level of abstraction of programming
2022-11-26 15:59:30 @martinmbauer For me personally, I use OneNote for note taking, so I prefer to have my laptop in class and seminars.
2022-11-26 04:22:41 @martinmbauer During tenure mentoring, I sat in on a colleague’s class. Some students were watching videos instead of the lecture. We ban electronics in our seminars and many of our classes
2022-11-26 04:17:24 @ljbuturovic @wightmanr @ylecun @EMostaque @rasbt @leonpalafox @paperswithcode @MetaAI @Michael_J_Black One word: Hydroxychloroquine
2022-11-26 02:35:09 @andrewgwils Beethoven died age 57. And of course Mozart and Schubert both died very young. Short summary of research: https://t.co/OSd5raWtMh
2022-11-26 00:51:26 @tsungxu I see. So is the lesson that future big wins will come from combining software with some other enabling technology (e.g., advanced materials, synthetic biology, etc.)?
2022-11-25 23:49:46 @tsungxu Most companies rely on multiple technologies, including software, and this has been true for some time. Examples: jet engines, aircraft, automobile manufacturing, etc. What is different?
2022-11-25 21:46:31 What is the "Asia-pacific Artificial Intelligence Association" and why are so many non-Asians on its Executive Committee?
2022-11-25 18:17:33 @wightmanr @ylecun @EMostaque @rasbt @leonpalafox @paperswithcode @MetaAI @Michael_J_Black Fake science is dangerous and wasteful. Danger: wrong therapies kill people
2022-11-25 06:14:29 @ylecun @EMostaque @rasbt @leonpalafox @wightmanr @paperswithcode @MetaAI @Michael_J_Black This clarifies for me that the main risk is undetected fake papers (I.e. faked results).
2022-11-25 05:48:08 @roydanroy @xwang_lk @RealAAAI Lots of good work is appearing at AAAI, the AAAI deadline comes soon after NEURIPS decisions. Many strong papers from Chinese groups
2022-11-25 02:17:39 @KolotaTyler @Noahpinion I think education is a practice in many cultures. Jews have learned to read in order to study Torah and Talmud. Chinese studied for the national exam.
2022-11-24 08:04:09 @houmanhemmati The rules you list absolutely do apply to clinical trials but not absolutely to public health.
2022-11-24 08:02:52 @houmanhemmati These rules are not absolute. Extreme case: If a person is infectious and would infect 100s of other people with a lethal disease if not quarantined, the autonomy of those 100 people outweighs the individual. It would be unethical not to confine him.
2022-11-23 22:03:09 @EMostaque @rasbt @ylecun @leonpalafox @wightmanr @paperswithcode @MetaAI @Michael_J_Black Suppose 10% of future papers are synthesized fakes. All of us will need to review 10% more papers. We are already straining under the current reviewing load. "Things like peer review and curation" are not free. To be continued after I process today's 201 cs.LG submissions
2022-11-23 19:48:49 @rao2z @AIESConf No, I would definitely consider that deception. But, at least according to Dennett, "wrong belief" would also be attributed by an observer (along with "cause"). I haven't read Bratman
2022-11-23 19:28:37 @rao2z @AIESConf I'm with Dennett here: Intention is attributed, not necessarily captured in an explicit knowledge structure
2022-11-23 19:20:08 @rao2z @AIESConf Even without an explicit theory-of-mind, game-playing agents can deceive. We've had this capability since at least the invention of the AO* algorithm (Martelli &
2022-11-23 19:15:41 @IDoTheThinking I think you are mistaken. There are many couples choosing not to have children because of concerns about the impact of large human populations.
2022-11-22 23:35:17 @ylecun @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI But as @Michael_J_Black points out, Galactica may open the door to a new set of fake science "attacks" on the publication system. I think it might also support a great semantic auto-correct for authors (esp ones with less skill in English)
2022-11-22 22:44:16 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI @ylecun Oh, and GPT-NeoX was released in April of 2022, I believe (20B parameters). It isn't obvious that Galactica is dangerously more powerful than these existing open models.
2022-11-22 22:42:12 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI @ylecun I did a bit of searching. XLNet was released in 2019, I believe. The weights for RoBERTa, DeBERTa, and XLM-RoBERTa are available I think. I'm not sure about OPT-175B. The fact that these are all available changes my position regarding Galactica. 1/
2022-11-22 22:00:18 @GaryMarcus @___merc___ Another hypothesis is that Amazon never figured out how to make a business out of it.
2022-11-22 20:48:37 @eugediana @MetaAI I think we can learn a lot by studying network internals and activations and their relationship to the training data. For this, we ideally have access to the network weights and training data.
2022-11-22 20:11:39 @MetaAI I'm looking forward to studying this system in detail.
2022-11-22 17:46:53 "They would ache with misplaced vigilance, they would lash out in anger, never be able to scratch the itch of fear, be torn by the longing to forget and the urge to always remember." Amazing writing by Dave Philips https://t.co/ia2bdDL0cY
2022-11-22 03:02:20 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI @ylecun What was the first good LLM whose weights were released?
2022-11-22 02:46:31 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI But I should have included entrepreneurs who want to build on top of these models, not just researchers
2022-11-22 02:45:16 @leonpalafox @wightmanr @EMostaque @paperswithcode @MetaAI Same rules for them, IMO.
2022-11-22 00:23:57 @wightmanr @EMostaque @paperswithcode @MetaAI Don't you think there are risks that the model could be deployed for disinformation campaigns, phishing, and so on? How do you enforce the license without some lock and key?
2022-11-22 00:01:43 @TheRandomMtrix @ChristophMolnar I think this is incorrect. There are many ML models that do not support statistical inferences, and some statistical models that are poor predictive models but do support inference.
2022-11-21 05:23:55 @EMostaque @paperswithcode @MetaAI I think you should put it behind a security wall but give access to authenticated researchers with tight license conditions.
2022-11-20 20:22:48 RT @gradydoctor: 1/ I reached into the console and fished around for my badge. It wasn’t there. My brow furrowed. Me: “Badge, we are not d…
2022-11-20 20:19:12 RT @hholdenthorp: People are rightly focused on how the grad students at UC will get what they deserve for carrying most of the load of the…
2022-11-20 17:04:59 @McknightLaura Yes. Funding agencies need to increase award amounts to compensate
2022-11-20 17:02:55 @deliprao The student is “getting” 80K that they would otherwise have to raise from other sources. That said, I’m all in favor of increasing student salaries
2022-11-20 16:42:10 @deliprao “Stipend” is a misnomer. We pay a salary for the work the student does teaching or doing research. “Tuition” is what the student pays for getting an education (classroom and mentoring). Grants cover both of these (in US)
2022-11-20 03:59:54 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 I agree it would be great for the research community to have access to these models. I'm not sure that full open source would be wise. The trolls and nation state actors would then also have access to refine their methods.
2022-11-20 03:49:16 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 This will still make it easy to generate fluent misinformation: Just build a collection of documents full of conspiracy theories, racism, etc.
2022-11-20 03:46:47 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 LLMs can generate lots of plausible hypotheses, but these need to be checked against reliable knowledge sources.
2022-11-20 03:45:57 @joshua_saxe @GaryMarcus @ylecun @mrgreene1977 I think LLMs will be replaced with models that make heavy use of external resources (document stores, databases, theorem provers, simulators, etc.). LLMs as we know them today are best viewed as pattern matching heuristics, not knowledge bases
2022-11-20 03:28:10 @deliprao @ben_golub If a grammar rule is consistent, it is quite easy for the net to learn to follow the rule without ever learning an explicit rule.
2022-11-20 03:21:05 @CharlestonCWKC @blekhman Yes, they all do.
2022-11-20 03:19:25 @aaronsibarium Rankings are a racket to sell magazines. The lower-ranked schools don’t dare challenge the rankings, so it is wonderful to see the top schools demolishing the BS-driven enterprise
2022-11-20 00:24:13 @ylecun @GaryMarcus @mrgreene1977 The last chapter of deep fakes hasn’t been written yet.
2022-11-19 21:05:30 @vgcerf He gave an inspiring series of lectures at UIUC when I was a grad student there. I particularly liked his advice about moving back to academia: He said if that was your plan, you should always live as if you were being paid an academic salary.
2022-11-19 20:58:00 @PalliThordarson These are also more likely to be the people who will be recruited to move to a top-ranked institution. And sure enough, with access to more resources, they will become more productive. end/
2022-11-19 20:56:25 @PalliThordarson If you are a student or a faculty member at a lower-ranked institution, you need to remind yourself to "work as if I was at Stanford"--that is, work on the most important problems you can. People who do world class research at such institutions have to work harder. 4/
2022-11-19 20:52:20 @PalliThordarson I think faculty and students lower-ranked institutions can suffer from the "tyranny of low expectations". This combines with less access to strong students and facilities can easily lead to less significant research outcomes. 3/
2022-11-19 20:50:13 @PalliThordarson They think "I'm at Harvard, I need to do `Harvard-level' work." Of course, if such work requires more resources and stronger students and postdocs, they have those too. So there is potential confounding. At lower-ranked institutions, do faculty feel less of this pressure? 2/
2022-11-19 20:48:13 @PalliThordarson Note that they did not analyze whether people at elite institutions studied more interesting/fundamental research questions. It's a difficult question to study, but my pet hypothesis is that faculty at elite institutions feel pressure to work on those fundamental questions 1/
2022-11-19 20:22:19 @B1ar2n3a Crazy. People should have the freedom to have kids whenever they think it best. I’ve suggested the PostDoc years as the time that is least stressful and hours are most flexible. But I didn’t do a PostDoc myself, so we had our kids while we were assistant profs
2022-11-19 04:17:49 @rlucas @netcapgirl TCP/IP routes around problems
2022-11-19 04:09:53 @paulnovosad Would you like to make a wager?
2022-11-19 04:08:53 @deliprao Same here. I asked a question for which wanted the answer. The result was partly useful and partly nonsense. Is it a symptom of bad benchmarks that Meta thought this was ready to release? Or maybe an excess of transparency?
2022-11-19 04:04:33 @RiverTamYDN Everyone wants to be treated with dignity
2022-11-19 03:09:01 @moyix Is there a betting market? I’d be curious to see the odds
2022-11-18 22:29:56 @taaltree But lots of former Twitter engineers are posting on this and other sites, and I believe them.
2022-11-17 16:59:52 @lilkuo "surrealism is the rational response to dictatorship" -- Carles Sierra
2022-11-17 04:17:32 @CT_Bergstrom Is this because it is trained on a lot of papers written by white men?
2022-11-17 04:17:03 @dcbaok @CT_Bergstrom Same with papers in Science and Nature (with some exceptions such as the DeepMind papers)
2022-11-16 22:42:42 @GaryMarcus On the other hand, after playing with it, I think most of its hypotheses are false
2022-11-16 22:40:27 @MattNiessner Just like startups, a lot of cutting edge tech fails to meet a real need. So I think people rationally adopt a negatively-biased prior
2022-11-16 16:58:53 @deliprao More than 11 people have died because of rumors and misinformation spread on social networks
2022-11-16 05:55:24 https://t.co/YaH8PuHnIU
2022-11-16 05:54:57 Giving Galactica a try. This mainly tells me that the face recognition community use this term differently than I do. For me, open category == open set. https://t.co/jGYx8CbxJq
2022-11-16 03:45:41 @GaryMarcus But the great thing about proofs is that they can be checked. Think of it has an hypothesis generator, not a source of truth. Same applies to generating code.
2022-11-15 22:12:55 RT @MattBetts11: Excited for this week’s EECB talk from @karst_justine !
2022-11-15 21:43:17 @noop_noob @rao2z Yes, some theories of consciousness would consider that a consciousness mechanism.
2022-11-15 20:43:53 @AmanBitz @ylecun @DebasmitDas1 @roydanroy Current SOTA according to https://t.co/aLwWum6M4Z is 96.81, and it appears to be asymptoting.
2022-11-15 20:42:44 @AmanBitz @ylecun @DebasmitDas1 @roydanroy But suppose you could raise CIFAR100 accuracy to 97%, I wouldn't be interested unless the same techniques worked on ImageNet and COCO and were practical at scale. Working well on CIFAR100 alone doesn't tell us much.
2022-11-15 18:37:29 @ylecun @DebasmitDas1 @roydanroy Well, it depends on what the paper is about. A accuracy-oriented computer vision paper evaluated on MNIST, CIFAR10, and CIFAR100 is no longer interesting. But, a paper studying the open category problem using these data sets is interesting (IMO: https://t.co/0KUqdPhE3f)
2022-11-15 18:15:14 @INSMI_CNRS @BachFrancis @giry_claire @sup_recherche @ENS_ULM @INS2I_CNRS Congratulations!
2022-11-15 16:40:40 @falsalem76 @_jasonwei Not unless it is something that can be learned from data, which I doubt. It is an architectural property.
2022-11-15 15:54:24 @mhudack @AutismCapital No, it’s anarchist
2022-11-15 04:42:08 @DebasmitDas1 @roydanroy That’ll be obsolete tomorrow
2022-11-15 04:39:36 @andyroyle_pwrc Much cheaper in Spain. Maybe subsidized there?
2022-11-15 04:39:00 @MarkSorel8 @andyroyle_pwrc My pharmacist did the paperwork for me. Thanks @FredMeyerInc
2022-11-15 00:33:11 @rao2z I agree that an agent that can detect that the anticipated effects of actions were not observed would be more robust (and could detect novelty). But is this capability required for a system to be treated as an "agent"?
2022-11-14 23:59:23 @rao2z Similar models can be trained to predict the next action in a policy. So they could *execute* a policy, but not create one. If we take the "agent stance", we can predict their behavior by attributing goals to them. Are you wanting an agent to do more?
2022-11-14 16:20:50 @jacobpattywagon The later shots are indoors. Outdoors there would be hats (but different ones)
2022-11-14 06:32:09 @Raza_Habib496 If we can predict the actions of a system by attributing to it goals and beliefs, then it is useful to treat it as an agent. What are the goals, beliefs, and actions of an LLM?
2022-12-07 17:51:00 @rasbt @kchonyc I have found that people give more useful answers when asked to quantify than when asked to give a written assessment
2022-12-07 17:47:43 @gentschev @divergiment @michael_nielsen If I want to give it a try, obviously I’m going to ask it about something I know rather than something where I don’t the answer. Why should I trust it, especially given the failures of previous models?
2022-12-07 17:38:03 @jlopez_dl @percyliang Why are humans the measuring stick? We ask jets and photocopiers to do things that people can’t do, and yet we need a definition of correctness in those cases
2022-12-07 13:04:55 @julianharris What a nightmare that would be!
2022-12-07 12:57:11 How do you convince @amazon that they have slapped an incorrect UPC sticker on an item? They keep sending us the wrong thing and we keep returning it.
2022-12-07 17:51:00 @rasbt @kchonyc I have found that people give more useful answers when asked to quantify than when asked to give a written assessment
2022-12-07 17:47:43 @gentschev @divergiment @michael_nielsen If I want to give it a try, obviously I’m going to ask it about something I know rather than something where I don’t the answer. Why should I trust it, especially given the failures of previous models?
2022-12-07 17:38:03 @jlopez_dl @percyliang Why are humans the measuring stick? We ask jets and photocopiers to do things that people can’t do, and yet we need a definition of correctness in those cases
2022-12-07 13:04:55 @julianharris What a nightmare that would be!
2022-12-07 12:57:11 How do you convince @amazon that they have slapped an incorrect UPC sticker on an item? They keep sending us the wrong thing and we keep returning it.
2022-12-08 21:45:21 @ai__pub It was not the only reason for renaming to NeurIPS, but it was a contributing factor. The name was a source of discontent in prior years.
2022-12-08 14:08:19 @adjiboussodieng I agree, this requires some care to design it well and to establish the validity of the methodology. Even the simplest ideas turn into research projects
2022-12-08 13:34:14 @adjiboussodieng Maybe we could write descriptions of three students and then ask volunteers from different countries to write letters? I could do US letters, maybe as a seed example. Anyone out there interested in giving this a try? Or maybe there is a better approach using example phrases?
2022-12-08 13:02:15 @adjiboussodieng Yes! It would be great to have some examples of "the same" letter as it might be written by people from different cultures as a reference for those of us reading these letters. Maybe also at 3 levels of excitement?
2022-12-08 03:05:25 @deliprao The LLM can't determine what content to communicate
2022-12-07 17:51:00 @rasbt @kchonyc I have found that people give more useful answers when asked to quantify than when asked to give a written assessment
2022-12-07 17:47:43 @gentschev @divergiment @michael_nielsen If I want to give it a try, obviously I’m going to ask it about something I know rather than something where I don’t the answer. Why should I trust it, especially given the failures of previous models?
2022-12-07 17:38:03 @jlopez_dl @percyliang Why are humans the measuring stick? We ask jets and photocopiers to do things that people can’t do, and yet we need a definition of correctness in those cases
2022-12-07 13:04:55 @julianharris What a nightmare that would be!
2022-12-07 12:57:11 How do you convince @amazon that they have slapped an incorrect UPC sticker on an item? They keep sending us the wrong thing and we keep returning it.
2022-12-09 00:19:25 @rasbt @pfau People are doing de novo design, so it is not just refinement. It is an amazing success!
2022-12-08 21:45:21 @ai__pub It was not the only reason for renaming to NeurIPS, but it was a contributing factor. The name was a source of discontent in prior years.
2022-12-08 14:08:19 @adjiboussodieng I agree, this requires some care to design it well and to establish the validity of the methodology. Even the simplest ideas turn into research projects
2022-12-08 13:34:14 @adjiboussodieng Maybe we could write descriptions of three students and then ask volunteers from different countries to write letters? I could do US letters, maybe as a seed example. Anyone out there interested in giving this a try? Or maybe there is a better approach using example phrases?
2022-12-08 13:02:15 @adjiboussodieng Yes! It would be great to have some examples of "the same" letter as it might be written by people from different cultures as a reference for those of us reading these letters. Maybe also at 3 levels of excitement?
2022-12-08 03:05:25 @deliprao The LLM can't determine what content to communicate
2022-12-07 17:51:00 @rasbt @kchonyc I have found that people give more useful answers when asked to quantify than when asked to give a written assessment
2022-12-07 17:47:43 @gentschev @divergiment @michael_nielsen If I want to give it a try, obviously I’m going to ask it about something I know rather than something where I don’t the answer. Why should I trust it, especially given the failures of previous models?
2022-12-07 17:38:03 @jlopez_dl @percyliang Why are humans the measuring stick? We ask jets and photocopiers to do things that people can’t do, and yet we need a definition of correctness in those cases
2022-12-07 13:04:55 @julianharris What a nightmare that would be!
2022-12-07 12:57:11 How do you convince @amazon that they have slapped an incorrect UPC sticker on an item? They keep sending us the wrong thing and we keep returning it.
2022-12-09 00:19:25 @rasbt @pfau People are doing de novo design, so it is not just refinement. It is an amazing success!
2022-12-08 21:45:21 @ai__pub It was not the only reason for renaming to NeurIPS, but it was a contributing factor. The name was a source of discontent in prior years.
2022-12-08 14:08:19 @adjiboussodieng I agree, this requires some care to design it well and to establish the validity of the methodology. Even the simplest ideas turn into research projects
2022-12-08 13:34:14 @adjiboussodieng Maybe we could write descriptions of three students and then ask volunteers from different countries to write letters? I could do US letters, maybe as a seed example. Anyone out there interested in giving this a try? Or maybe there is a better approach using example phrases?
2022-12-08 13:02:15 @adjiboussodieng Yes! It would be great to have some examples of "the same" letter as it might be written by people from different cultures as a reference for those of us reading these letters. Maybe also at 3 levels of excitement?