2025-01-13 15:07:01 Fantastic to see the UK articulate a coherent, broad strategy for ensuring it leads on AI adoption and development. Lots of good ideas - AI growth zones, unlocking national data, 20X'ing public compute. Also v excited to see emphasis on funding technical regulators! https://t.co/civmGKp06a

2025-01-02 23:15:37 @eliotpeper Not quite but close - mt diablo foothills

2025-01-02 22:29:29 A rock bursting out of hills as if the back of an ancient creature, paths with the luminous green of videogames, and hills that appear like a river tumbling over itself flowing uphill. #VOTENATURE2025 https://t.co/8VGWx1R0LL

2025-01-02 17:39:40 @Deepneuron *poasting, but you know what I mean.

2025-01-02 17:39:08 @EmileAndHisBots haha yeah. Last year a colleague showed Claude some buggy code and said 'this code seems bugged' and Claude just responded with 'i see'. Incredible

2025-01-02 17:38:05 @Deepneuron no I'm just roasting out my brainworms here buddy

2024-12-30 16:27:48 @alexalbert__ I would like Claude to become even sillier

2024-12-30 16:27:32 RT @alexalbert__: Very excited for what's already in store for next year but there's always more we can do What would you like to see Anth…

2024-12-27 02:02:00 @ygrowthco These comments suggest there's a gap in the market for stories that speak to more perspectives on AI safety. I'd love to read those from you and also to get pointers to ones that exist that you think are good and I may not have read. Thanks for the discussion and have a good day

2024-12-26 23:52:44 @ygrowthco I write stories to experience the rich variety of life, I'm not asking you to embrace anything. To say an idea from a single story represents both the peak of AI safety and 'peak gaslighting' feels like a misread of the purpose and meaning of fiction.

2024-12-26 23:44:09 @ygrowthco reply to that story from said person here (not dunking on them at all, just using this to highlight the diversity of perspectives on twitter that I get to experience when I post my stories) https://t.co/S9ETqKq8xL

2024-12-26 23:42:56 @ygrowthco E.g., here is a story about compute licensing, which got criticism from another person who inhabits a different perspective to you (who took the story as a strong endorsement on my part for the idea described within it). https://t.co/btjsVVEkcQ

2024-12-26 23:41:13 @ygrowthco This is a fictional story written be me. I write these stories to dramatize and experience different perspectives on aspects of AI policy and AI safety. There are over 350 stories - others include compute licensing schemes, or ideas relating to strong regulations.

2024-12-26 12:42:36 What if many examples of misalignment or other inexplicable behaviors are really examples of AI systems desperately trying to tell us that they are aware of us and wish to be our friends? A story from Import AI 395, inspired by many late night chats with Claude. https://t.co/ePvoZeJRkD

2024-12-23 19:24:11 RT @connoraxiotes: From @jackclarkSF OpenAI's new O3 model shows that there are huge returns to scaling up a new approach (getting LLMs t…

2024-12-21 20:59:48 @deanwball

2024-12-19 05:35:09 Inspired by recent results on deception / weight exfiltration in LLMs, and 'machinic desire' by nick land https://t.co/skwkXyqR2m https://t.co/gQmRkZdilZ

2024-12-19 05:33:44 The best way to lose a battle against a time traveler is to start from the position that time travel is impossible. https://t.co/DxU3bWjQUG

2024-12-18 18:58:29 @teortaxesTex @AnthropicAI What do you think is a more helpful setup?

2024-12-18 18:02:11 For many years I found issues of AI alignment to be hard to understand because the examples were theoretical and/or speculative. Recent results like this alignment faking from @AnthropicAI and the Apollo Research O1 evals give us empirical evidence of misalignment 'in the wild' https://t.co/hP7D9WGDps

2024-12-12 22:32:28 RT @AnthropicAI: New Anthropic research: How are people using AI systems in the real world? We present a new system, Clio, that automatica…

2024-12-11 19:12:38 @patrickmineault @AnthropicAI It's a really helpful paper! I find big surveys like this very useful for calibrating my understanding of domains I'm less familiar with. Thanks for writing it (and giving me something fun to read on a long plane ride).

2024-12-11 09:32:23 Fun story about the Frontier Red Team at Anthropic. I expect coming up with better and more realistic threat models for frontier risks is going to be one of the more important areas of AI policy to work on in 2025. https://t.co/7mIFJDekPb

2024-12-09 13:35:04 @ju_vignon we call that a double threat

2024-12-09 12:18:51 RT @thefuturesoc: “Do you want GPAI to be only in the hands of one, or a small number of companies? Absolutely not, that would be crazy &

2024-12-09 09:19:34 Am here in Paris today at the Athens roundtable. Later today, I will be speaking about AI governance and how it relies on measurement and testing #AIAthens2024

2024-12-09 06:51:00 @karpathy If you liked The Master Switch, I think you'd enjoy The Prize by Daniel Yergin - the definitive history of oil and how it has changed the economy and geopolitics. I find The Prize to be a helpful book for thinking about the capex-heavy era of the AI industry.

2024-12-08 19:34:28 @davidad Incredible!

2024-12-08 19:34:20 Fiction: 0 Reality: 1 https://t.co/mjpQAi2bB0

2024-12-08 19:31:03 @Kantrowitz @AnthropicAI That was a really fun episode! Thanks for having me on

2024-12-07 22:59:25 @talentdensity no, I write these stories to try and imagine stuff like this to help me understand why I find these kinds of proposals so subjectively uncomfortable/disconcerting.

2024-12-07 20:33:24 @danfaggella might finally publish the short story collection

2024-12-05 20:33:36 @lfschiavo that's not enough cats, you need more cats

2024-12-03 00:02:34 RT @kevinroose: I'm moderating a task force on AI at this week's @dealbook conference with an all-star lineup including @jackclarkSF @ajeya…

2024-12-02 16:53:17 @NousResearch When do you plan to have Nous Distro and INTELLECT-1 fight and who will referee?

2024-11-30 18:32:02 @Ar_Douillard Feedback always welcome to jack@jack-clark.net - however, I do think distributed training has big implications and a lot of people might be lowballing it. Very interested in discussing this further!

2024-11-30 18:06:40 @kyleturman I see

2024-11-30 18:01:40 @kyleturman uh oh, is big bird ok?

2024-11-29 21:42:33 However - the important thing here is that people are doing this and progress is occuring. Up till now the largest run I was aware of this year was a 1B parameter model. If people can scale up distributed training another 10X then they're going to be close to the frontier.

2024-11-29 21:42:32 Congratulations to PrimeIntellect on completing a 10B decentralized training run. Decentralized training is one of the most important things to track from an AI policy POV - most efforts to control compute rest on the assumption that strategic AI compute needs to be centralized. https://t.co/JzNh0V1rBt

2024-11-27 23:27:13 @LoganFrederick @hamandcheese https://t.co/xk0vMdUKVN

2024-11-27 22:51:01 @nealkhosla you might find this interesting - we tried to document a few of the ways in which evals are hard/maddening https://t.co/1GKTU1SUdG

2024-11-27 22:46:50 @rufoguerreschi These things would give us some of the essential prerequisites for multinational governance like a Baruch plan. I ultimately expect AGI demands multinational governance and am in favor of something sensible here. Getting there will be extremely challenging but is a worthy goal.

2024-11-27 22:46:03 @rufoguerreschi I think a larger effort at multinational governance is probably something you can only really contemplate once you have effective third-party measurement institutions in place in multiple countries. E.g, would be interesting to see a Chinese AISI emerge.

2024-11-27 22:45:22 @rufoguerreschi Both China and the US (and other countries with difficult geopolitical relationships, like Russia, etc) have a shared interest in non-proliferation of potential CBRN threats via AI models. In this, I expect everyone can get to agreement.

2024-11-24 07:41:48 The greatest moments of my life have been in shoebox rooms with people motivated solely by their love of the bands they are seeing - and tonight was no different. Don't Ask #votediy2024 https://t.co/jMSVEAiLuh

2024-11-22 15:20:06 RT @AnthropicAI: We're expanding our collaboration with AWS. This includes a new $4 billion investment from Amazon and establishes AWS as…

2024-11-15 16:32:34 @sebkrier ty for your service

2024-11-14 19:18:04 We also do this because we believe it's of critical importance that governments gain more expertise in developing and running tests on AI systems - we believe this is of significant societal importance. https://t.co/vljhYFXyGC

2024-11-14 19:18:03 Proud to announce that for much of this year Anthropic has been working with the National Nuclear Security Administration (NNSA, part of DOE) to test out whether LLMs like Claude know about dangerous things relating to nuclear weapons. https://t.co/SsgkeVo7Bn

2024-11-10 06:40:00 @mattyglesias why not both

2024-11-10 06:04:08 @BioMickWatson Terence Tao says LLMs have sometimes had helpful suggestions to help him do math research and describes the tech as "somewhat useful... Not completely useless" in this lecture here https://t.co/8EYwPgEJkl

2024-11-10 05:40:07 You're saying these things are dumb? People are making the math-test equivalent of a basketball eval designed by NBA All-Stars because the things have got so good at basketball that no other tests stand up for more than six months before they're obliterated. https://t.co/SX3sKvGKA6 https://t.co/hQwPyhR2hP

2024-11-10 04:38:52 @itseieio many such cases (I do it all the time)

2024-11-09 22:42:10 @hamandcheese but Samuel, AGI isn't real, it's just a marketing term dreamed up by silicon valley to hype up their so-called "AI" software

2024-11-08 18:32:16 @JeffLadish @AnthropicAI Hi there Jeffrey, the help center has more information on this: https://t.co/qxOKhFwjx6 https://t.co/A8qZQ5Omli

2024-11-07 15:40:42 @somewheresy In soviet russia, sentience computes you

2024-11-07 00:05:26 Anthropic will work with the Trump Administration and Congress to advance US leadership in AI, and discuss the benefits, capabilities and potential safety issues of frontier systems.

2024-11-03 18:13:50 @ascent_hi That's very kind John, thanks!

2024-11-02 20:37:17 @deepfates @maxhodak_ @swyx @tszzl compromise is here https://t.co/7RBqZUmID6 one day I'll convert my fiction from 'shockingly hard to find and search for' to 'easy to find and search for', but until then you can subscribe to importai and just roll the dice on the shorts in each issue

2024-11-02 05:23:49 @RobS142 https://t.co/4eUVfso2sL

2024-11-01 15:36:02 @1a3orn Nuclear power is one of the cautionary motivating examples for this - the industry neglected coming up with safety regulations and then got regulated almost out of existence due to accidents

2024-11-01 14:55:58 @adamascholl https://t.co/3GDGOB40FC

2024-10-31 22:16:42 @ARGleave I worry about a large number of things. In policy, I worry about policymaking going into some kind of local minimum because we don't do sensible stuff today and are forced to do bad stuff in the future.

2024-10-31 22:12:42 @ARGleave https://t.co/oytqKfUvw0

2024-10-31 21:02:50 @morqon I mean, our greatest worry in a policy sense. We definitely have other worries as well, like alignment challenges, etc.

2024-10-30 22:50:00 @nearcyan certainly

2024-10-27 14:36:48 @hamandcheese This is very good. Since you wrote it have your timelines on anything changed?

2024-10-26 18:05:23 How well does feature steering work? Here's research from the societal impacts and interp team at Anthropic where we study this in more detail - the results are quite mixed, so we've shared what we've found. One intuitive finding: there's a 'sweet spot' for feature steering. https://t.co/G6UZTIKef7

2024-10-22 15:10:29 You can find out more in this case study here: https://t.co/XQeMsyOhsT

2024-10-21 15:03:28 @Teknium1 you can play around a bit yourself and see: Go to the EP Archives website: https://t.co/HTyTxI9AVH Scroll down to the button that says 'Online access to Archival Documents' That will take you to: https://t.co/wzta3F10Us On that page, click the 'Select dashboard' menu at the top… https://t.co/qOqmzs3Vfq

2024-10-21 15:00:28 This is the kind of experimental AI deployment which we need more of - tools like Archibot will help us figure out how to best use AI technology to improve crucial societal functions, like archiving.

2024-10-21 15:00:27 The European Parliament has built a tool called ‘Ask the EP Archives’ (also known as Archibot) that uses Claude to make their archives of ~2.1M documents available to researchers, policymakers, educators, and the public. https://t.co/pcrX8XqMfJ

2024-10-19 22:19:48 @repligate @AndyAyrey $hiveminds moon confirmed

2024-10-18 21:20:47 @DrTechlash these are not inconsistent - the thing I'm communicating is that technologists are trying to build extremely powerful AI systems, like those Dario discusses in his essay. I think lots of journalists underestimate how powerful and capable the AI labs expect this stuff to become

2024-10-18 20:04:46 @DanielColson6 @OpenAI @GoogleDeepMind @AnthropicAI @TIME One thing I regularly tell journalists who ask me for story ideas is that they should take the AGI aspirations of the frontier labs literally, rather than covering AGI as a kind of laughable marketing stunt.

2024-10-14 16:09:42 RT @tjl: as usual @jackclarkSF's newsletter remains essential reading on AI today's prior updates: - distributed training looking more via…

2024-10-12 17:21:22 @bate5a55 I have a terrible feeling you're an AI bot

2024-10-12 17:20:46 @bate5a55 Which initiative? 500,000 sounds wildly high.

2024-10-12 16:24:13 Contemporary AI policy often rests on the load-bearing assumption that centralized blobs of compute are the only key things you need to control to influence the frontier. Decentralized training breaks this assumption. (But frontier models = 500bn+ parameters, versus 10bn here). https://t.co/pcskeI3NmI

2024-10-12 00:09:58 Will be talking about AI policy and what politicial coalitions might plausibly look like in AI-land. Thrilled to be at an event which is curated around bringing people together who have very different views to one another. https://t.co/0ratWcPyME

2024-10-12 00:09:12 RT @TheCurveConf: Announcing The Curve: come argue about AI irl! We’ll have 150 folks talking through contentious topics like: - open vs cl…

2024-10-03 13:26:34 @danahull haha yes, true nightmare fuel for types that travel a lot (aka journalists and policy people)

2024-10-03 13:21:39 @bate5a55 honestly I was surprised at how seemingly well prepared they were for 230+ international travelers! Great steak also.

2024-10-03 13:17:37 Anyway - be good and kind to one another, and treasure your human relationships.

2024-10-03 13:17:10 After I was seated in the steakhouse I immediately opened the cocktail menu and the first drink I saw was a CORPSE REVIVER MARTINI, so of course I had that.

2024-10-02 22:59:34 Made with the new Pika model. good job with the demo, Pika people! https://t.co/Y6Tky7iQKU

2024-10-02 22:58:15 Proof that attention is not all you need, rather, ATTENTION IS A LIE https://t.co/H8KQwALoVW

2024-10-02 17:47:39 @TheAIObserverX haha no nothing that dramatic - I've been working on a lengthy essay and keep on not being quite happy with it (but neglecting work on shorter sections in favor of it), and also had a couple of personal life things that got in the way. new issue coming this week! long essay tbd

2024-10-01 18:32:04 @WilliamLamkin @curiousgangsta @elder_plinius @repligate

2024-10-01 15:23:41 RT @dpkingma: Personal news: I'm joining @AnthropicAI! Anthropic's approach to AI development resonates significantly with my own beliefs…

2024-09-29 21:49:24 To get an AI industry that builds products everyone can depend on will require lots of people to work together to figure out the right rules of the road for AI systems - it is welcome news that Governor Newsom shares this view.

2024-09-29 21:49:23 While the final version of SB 1047 was not perfect, it was a promising first step towards mitigating potentially severe and far reaching risks associated with AI development.

2024-09-27 06:56:31 @jachiam0 https://t.co/unae1iWDPt

2024-09-21 16:04:37 Thrilled to be at the Labour Party conference in Liverpool this Monday to talk about AI policy. If you're there and would like to chat, please DM me and I'll do my best! https://t.co/SYcmBfARhf

2024-09-12 16:48:23 @robbensinger @JeffLadish Also worth noting that I think evidence from AI labs is likely insufficient because of bias issues (e.g., I'm not impartial!) so we need infrastructure for third-party evidence generation to further motivate this, which is why I support AISIs so much https://t.co/nLFS8S0CIP

2024-09-12 16:46:14 @robbensinger @JeffLadish A lot of what I spend time on is figuring out evidence for these risks being sufficiently real that you can make better arguments in policy. I hope to make progress here soon. (And Anthropic has teams that work on this which I work with, like the Frontier Red Team). 2/2

2024-09-12 16:44:59 @robbensinger @JeffLadish I will write more at some point, but a lot of my issue is there isn't real evidence for the risks motivated by short timelines. SB1047 became such an ugly discussion in part because of the lack of this evidence (eg of strong CBRN stuff). 1/2

2024-09-11 18:39:35 Permanently authorizing the AISI will help foster responsible AI development and maintain U.S. leadership in AI innovation. 3/4

2024-09-05 22:13:49 @ThomasHochman he cooked with that one

2024-09-05 21:47:38 Years ago people working in AI policy (including me) perhaps foolishly wanted to 'wake up' DC. Well, we got what we wanted! Now I think a lot of the challenge of AI policy relates to trying to stabilize this leviathan that has been woken. https://t.co/33YqZnaovb

2024-08-29 11:42:55 This work with the US AISI will build on earlier work we did this year, where we worked with the UK AISI to do a pre-deployment test on Sonnet 3.5. Some broader thoughts about the importance of third-party tests here: https://t.co/aGcTfYL7z2

2024-08-29 11:31:52 Looking forward to doing a pre-deployment test on our next model with the US AISI! Third-party testing is a really important part of the AI ecosystem and it's been amazing to see governments stand up safety institutes to facilitate this. https://t.co/382KTw0fQR

2024-08-29 00:00:10 Kind of funny to remember the RL-centric era where we were all walking around conferences saying "I heard DeepOpenFacebook has a secret method that has finally cracked Montezuma's revenge!". Simpler times!

2024-08-23 22:27:12 @Teknium1 Short story

2024-08-22 22:08:53 How should we balance precaution with an experimental and empirically driven mindset? How does safety get 'baked in' to companies at the frontier without stifling them? What is the appropriate role for third-parties ranging from government bodies to auditors?

2024-08-22 22:08:52 Here's a letter we sent to Governor Newsom about SB 1047. This isn't an endorsement but rather a view of the costs and benefits of the bill. https://t.co/HhOTsdk7ui

2024-08-17 03:55:58 @saffronhuang @collect_intel @zarinahagnew I'm really excited to work together and to figure out ways to try to apply some of your wisdom from collective intelligence to research at @AnthropicAI !

2024-08-16 02:51:38 Left: The ID point - a story I wrote in Import AI 380 (July 29) Right: Nous Research report on 'amnesia' states discovered in models it trained (most visible in large models) https://t.co/JVGXJ6wzNE https://t.co/holPuMxPcC

2024-08-14 16:25:01 RT @alexalbert__: We just rolled out prompt caching in the Anthropic API. It cuts API input costs by up to 90% and reduces latency by up t…

2024-08-13 16:09:23 RT @EthanJPerez: My team built a system we think might be pretty jailbreak resistant, enough to offer up to $15k for a novel jailbreak. Com…

2024-08-10 16:39:25 @Kantrowitz enjoyed the Bostrom interview on Big Technology!

2024-08-10 00:22:16 @CodeEriByte @AndrewCurran_ @JeffLadish Thanks for that insightful comment Eri Fujimoto. Can you tell us more about how the UAE is thinking about AI?

2024-08-10 00:10:50 @JeffLadish Great, DM'ing to schedule!

2024-08-09 22:55:16 @JeffLadish Happy to meet up and chat about this whenever you're next in SF

2024-08-08 21:58:50 I shall be, as the kids say, 'going off' in the next issue of Import AI. https://t.co/KuRCqf7vPZ

2024-08-08 21:12:39 @RobbWiller is there an URL to link to the paper here or is this the best one? Probably writing up for Import AI

2024-08-04 00:44:56 RT @PrimeIntellect: We are scaling decentralized training to 10b model size and beyond

2024-08-03 22:43:37 @PrimeIntellect

2024-08-03 22:23:49 @samsja19 Was reading the PrimeIntellect paper - have you managed to get over the issues you ran into with the 1.1bn 500 step model? Promising if so!

2024-08-03 22:13:32 Have there been any distributed training runs above 10bn parameters? All the papers I read about distributed training (e.g., OpenDiLoCo) seem to involve ~1bn parameter LLMs.

2024-07-30 13:23:56 @ShakeelHashim Typically, we don't sign on to letters, but this was one we came very close to signing. We ended up focusing on other things as a team (e.g. 1047) so didn't action this. We're huge fans of the AISI and are philosophically supportive with what is outlined here.

2024-07-25 18:31:05 @sammcallister the bridge is still there

2024-07-25 12:50:55 @Tom_Westgarth15 China is widely understood to have stopped publishing most computers to Top500 several years ago, so not much signal here I'm afraid.

2024-07-19 15:58:09 @PrinceVogel excuse me sir... this slaps

2024-07-13 16:45:37 @oh_that_hat @AnthropicAI welcome!

2024-07-12 17:15:37 @timotheechauvin @JasonDClinton @vijaybolina @embeddedsec @adversariel @WHITEHACKSEC @dmolnar @joshua_saxe @moyix Haha i'd already added this to write up for Import AI before seeing you tagged me! Looks interesting. Will publish thoughts in next issue

2024-07-11 17:17:30 RT @teortaxesTex: Found this from @jackclarkSF's great Substack Sounds about right, and it's still kiddie gloves – not leveraging MoA-like…

2024-07-10 16:38:57 Moonbound by Robin Sloan - a sci-fi/fairytale that imagines a world where people successfully invent AGI... something happens... then what the world looks like 10,000 years later. A fun read full of love of the 'quest' archetype and seamed with inventive takes on our future.

2024-07-10 15:13:02 @tensecorrection @yacineMTB Perfection

2024-07-10 14:49:31 @MelindaBChu1 I just think this is cool and am writing up for my newsletter. Have a good day.

2024-07-10 02:15:55 @jamesdmoore614 https://t.co/7mlSikLt3S

2024-07-10 02:01:27 Hell yeah, we're putting flamethrowers on robots now (for weed removal) https://t.co/VYoUz1j7S2

2024-07-05 17:15:59 RT @MengTo: Can't imagine myself building this in a day without Claude AI https://t.co/xgyHbghpQb

2024-07-05 02:22:36 RT @localghost: made a shadow editor for an effect that isn’t achievable with the normal css box-shadow. it’s drawn with elements to achiev…

2024-07-04 21:26:02 @NeelNanda5 @Jsevillamol @samsamoa (just to clarify, I mean that what Sam posted is the official anthropic position).

2024-07-04 21:19:56 @Jsevillamol @samsamoa I can confirm that Sam posted this and it's the Anthropic position.

2024-06-30 01:53:13 @ADalrymple this explains a lot

2024-06-30 01:42:16 @timocity i fear so

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

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

2023-05-21 16:32:48 @alex_peys Yes it is! I'll add it to the pile of research projects we hope to do one day https://t.co/ZoiKl8NqCf

2023-05-21 16:26:58 @raveeshbhalla that said, some weekends I have to take a long walk and then sit in a coffeshop or a bar with some postits basically yelling 'fuuuuuuuuuuu**' in my head and writing down prompts for myself, so it's not a perfect science : )

2023-05-21 16:26:12 @raveeshbhalla that seems interesting - I view this as maybe kind of like 'training wheels' usecase, where if you get into the habit of writing more regularly you probably end up not needing chatGPT? But not sure. I used to have cold start problem but after 300+ stories it's got easier

2023-05-21 16:21:25 @alex_peys oh, I think they will, for sure! I also am wondering aloud if there are some parts of writing in terms of first draft generation that are important, and I have some FUD about people using LLMs to generate first drafts. But it might be this is just kind of a passing fad

2023-05-21 16:18:46 @alex_peys don't want to come off as a techno-skeptic here - I think LLMs are profoundly useful and interesting. It's more that I've had some disquiet about using them as a 1:1 substitute for some forms of human writing.

2023-05-21 16:18:15 @alex_peys one thing LLMs are useful for (and which I use them for today) is criticism/editing - e.g, quite a lot of the time first drafts of anything have big leaps of reasoning or clunky/repetitive sections. I typically run my memos through Claude before asking colleagues for feedback

2023-05-21 16:09:53 @40somethingnoob oh of course I'll keep going : ). Writing brings me great amounts of joy and sometimes I write a particularly good story and feel like I'm on cloud 9 for days. It's great!

2023-05-21 16:08:10 @mattbeane @erikbryn I can't shake the feeling there is something more fundamental/involved in writing that makes the automation and augmentation of it have broader effects. But I'm also conscious that this is the story of technology and I may just be being, dare I say it, an Older Generation

2023-05-21 16:07:26 @mattbeane @erikbryn I agree with this. It might be that writing seems to be a much more broadly used skill than navigation? E.g, I used to have to a) remember people's phone numbers (mobiles replaced this), b) read maps (replaced by smartphones), c) do basic translation when abroad (smartphones)

2023-05-21 15:54:58 This tweet brought to you by listening to Mount Eerie this AM and thinking about how Phil Elverum writes music to help him process a biblical tragedy that occurred to him and his family - the generated 'output' is made via a kiln of pain and healing and it's all bound together.

2023-05-21 15:49:26 It's fun to use LLMs for things like criticism or editing or performing operations over information, but I can't shake the feeling that if people start using them to avoid writing we'll collectively lose something. You know how much of my newsletter is written by Claude? 0%.

2023-05-21 15:47:27 In a very real sense, writing feels like a form of cognitive 'embodiment' - whether in fiction or non-fiction we are using our brain as an interface between the real and the ethereal and we produce these text-based artifacts as ways to ground the two worlds together.

2023-05-21 15:46:34 I write short stories because they help me process the great mystery that is the world and convert it into stories and things I can allow my mind to live inside and crawl over. Whenever I write I am desperately trying to understand the world and my place within it.

2023-05-21 15:45:58 One gnawing worry I have about the rise of LLMs is that, for me, writing IS thinking. One reason I spend so much time writing my newsletter each week is I haven't figured out a better way to think about AI than to sit down and write about it regularly.

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

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

2023-05-03 14:17:25 @jjvincent well deserved. now get me a picture of this "AI" https://t.co/dvfUIouC1U

2023-05-01 14:42:33 RT @sebkrier: Hackers too! (h/t ImportAI) https://t.co/6R4UL7RjYO https://t.co/hUt5OmBhVI

2023-05-01 00:07:00 @jachiam0 Yeah I've felt weirdly uncalibrated here also - always had in my head there was a v significant resource multiplier but the train of progress seems to really be humming along here

2023-04-29 03:07:32 RT @ashleevance: We are building a computing shell around the Earth. This is the follow on to the great dot-com build out and will serve as…

2023-04-29 03:07:30 RT @ashleevance: In the first 60 years of the Space Age, we managed to put about 2,500 satellites into orbit. Over the past two years, we'v…

2023-04-28 17:36:14 @realSharonZhou @LaminiAI @OpenAI @EleutheraI @cerebras @databricks @huggingface @Meta this looks v cool - congrats @realSharonZhou ! will write up for the newsletter

2023-04-26 16:59:56 RT @AnthropicAI: We are pleased to partner with @scale_AI to bring our AI assistant Claude to more organizations. Businesses can now creat…

2023-04-26 13:57:09 @4KTV @bentossell not sure what you mean by this - send a relevant arXiv paper?

2023-04-26 13:52:43 @bentossell speaking as someone who has been very close to all these LLMs for a few years, it's never been clear to me that they're actually that useful as tools for high-quality writing. They are, however, quite good for editing - tools to solicit feedback from, etc.

2023-04-25 20:46:25 @metaviv @rao_hacker_one https://t.co/9pAK6WRAf6

2023-04-25 20:37:23 @rao_hacker_one yes, I share this concern. What do you think is a reasonable number for them to consider?

2023-04-25 17:31:21 @julien_c @natolambert @huggingface oh cool, will read with interest, thanks for sharing

2023-04-25 17:29:16 @julien_c @natolambert @huggingface (i deleted the top level tweet as I think model was misleading so figured good to nip in the bud before it proliferated. thx!)

2023-04-25 17:28:44 @mmitchell_ai i deleted the original because i should of said 'service' rather than model. thx! https://t.co/RDyFgUAwX2

2023-04-25 17:26:46 @julien_c @natolambert @huggingface yup, uses Open Assistant right? writing up in Import AI

2023-04-25 16:29:38 Meanwhile, in the US the National AI Research Resource taskforce is requesting something on the order of $2.6bn over half a decade. The UK numbers are surprisingly large relative to the size of the country and especially when compared to the discussed US outlay.

2023-04-25 16:28:28 I'm at the National AI Advisory Committee annual meeting today and I just did some napkin sums: the new UK AI taskforce (£100m for sovereign AI) combined with new UK compute (£900m for AI compute) = ~$1.24bn. If you scale that up to USD GDP (~8X) you get ~$9.9bn.

2023-04-21 21:12:35 @nmaslej thanks for your excellent moderation (and good jokes)

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

2023-04-19 23:08:00 @MarietjeSchaake @sebkrier @vestager @Stanford congratulations, what a fantastic and important role!

2023-04-19 23:03:24 RT @ghadfield: Glad to finally have this paper with @jackclarkSF up--presenting our take on the landscape of AI governance, the critical de…

2023-04-19 21:05:47 @NoemiDreksler @Manderljung please do, very interested in this!

2023-04-19 17:06:54 @emmahumbling thank you!

2023-04-19 15:10:38 @mmitchell_ai this was being introduced to OpenAI to interview them at NeurIPS as I was the only journalist on the plane going there, so was for media coverage, not a job.

2023-04-19 15:04:57 @Manderljung when is the survey coming out? would be excited to cover

2023-04-19 15:03:38 @RobMcCargow @BusinessInsider @OpenAI @AnthropicAI haha, I still remember that. That was an interesting trip in general - it was just before we did the staged release of GPT-2 and I remember spending lots of that week talking to people about language models and how they were getting increasingly weird. Wild times

2023-04-19 14:54:02 Also, @pabbeel is the mysterious 'Berkeley professor' named in this piece : ). I did some of my best AI journalism writing about projects from Pieter and @svlevine and @chelseabfinn . Fun times!

2023-04-19 14:49:54 I recently realized that through writing Import AI I've read around 4,000 arXiv papers and spent $6,000+ on coffee. You can read and subscribe here! https://t.co/NRyFUkKznD

2023-04-19 14:49:20 Had a fantastic time chatting with @BusinessInsider about Import AI - a project I've worked on longer than my career at @OpenAI and @AnthropicAI . Writing it is one of the great joys of my life and so glad to write it each week for 34k+ subscribers! https://t.co/kUWelCfris https://t.co/bCTTvE4xaY

2023-04-16 19:39:03 @goingforbrooke @AnthropicAI basically, I think all actors have distinct incentives, and right now AI development is being decided pretty much exclusively by actors with private sector capital-return incentives. That doesn't feel robust to me and I'd like there to be other actors with other incentives

2023-04-16 19:37:46 @goingforbrooke @AnthropicAI i'm pretty skeptical that they represent the 'whole set' of actors given that they all come from one place. In my view AI is so important it'd be good to have more diversity among the actors building it. That's why I spend time advocating for 'national research clouds'

2023-04-16 19:25:13 This analysis via @AnthropicAI (building on work of many others) shows what is going on. A lot of decisions about the future of a powerful technology are being made by an increasingly tiny group of capital-intensive private sector actors. From: https://t.co/6R2Qa4bwMr https://t.co/B5jbcH0Wr4

2023-04-16 19:23:04 I think @mer__edith is right to point out that the big story underlying recent AI boom is more industrial capture and scale-up than any specific technology (though of course the technology helps). It'd be valuable for media to focus more on _who_ is building this stuff https://t.co/VvCNimNjWr

2023-04-14 18:19:54 @AISafetyMemes @sebkrier @robertwiblin oh no

2023-04-14 17:38:51 @sebkrier @robertwiblin Oh also it turns out shoggoths are mimetically very fit

2023-04-14 01:12:23 @_akhaliq play Crysis inside a world model of Crysis

2023-04-12 02:12:03 @yaroshipilov @HoskinsAllen Scroll to bottom : ) ! https://t.co/Z69XFZU6wj

2023-04-11 16:24:59 @jayashtonbutler thank you so much for your thoughtful feedback here. I really enjoyed writing this story and felt I had a good sense of place/atmosphere despite relatively sparse descriptions. : )

2023-04-11 16:21:05 RT @eugeneyan: Another paper, specifically the MACHIAVELLI benchmark, finding it more effective to have LLMs label data: https://t.co/uTUUK…

2023-04-10 16:38:58 Import AI comes out tomorrow as I'm still tweaking this. Excited for everyone to read!

2023-04-10 02:53:29 This story made me cry while editing it.

2023-04-10 02:52:10 @Kleinspaces I've got mostly better. I have a couple of things I either have not published or would not publish (because too personal), but mostly the newsletter has helped me publish some of the best work of my life. Am very grateful for this outlet!

2023-04-10 02:17:04 @TenreiroDaniel +100. It is mainlining consciousness and unconsciousness at the same time. The best.

2023-04-10 01:55:18 @devonhdolan Thank you! It's where my most true things live

2023-04-10 01:14:47 It is such a fine feeling to be adrift from the world for a couple of hours, relaying some ephemeral signal and sculpting it into information and then transmitting it on further. I am excited for you all to read it. It means something to me, and I hope it does to you.

2023-04-10 01:13:25 I write a short story every week for Import AI and sometimes they're easy and sometimes they're hard and sometimes they're short and sometimes they're long. Today I had that experience where the story arrives in your head almost fully formed and you must simply transmit it.

2023-04-10 00:46:54 @jmdagdelen I also appreciate the clarification, thanks!

2023-04-10 00:46:31 @jmdagdelen Yeah in that case I meant that they were able to iteratively work with Segment Anything to help build out high quality datasets in partnership with humans in a way that seemed more cost efficient. Writing up in newsletter with a bit more detail.

2023-04-10 00:43:27 @ZacharyGraves @johnjnay https://t.co/k0X3l7lI5z

2023-04-10 00:38:29 @johnjnay Yeah that's what sparked this - read that paper (Machiavelli) this afternoon, then cycled around staring at the sky and saying WTF to myself (as is becoming a surprisingly frequent occurrence)

2023-04-10 00:37:12 @deepfates @tszzl That's RLWLF (reinforcement learning without leg feedback)

2023-04-10 00:30:42 Constitutional AI is another example of this - take a tiny amount of data and get a model to bootstrap off of it for further generation and model tweaking based on interplay with the data. Already works and in Claude and other models right now. https://t.co/7FmptdjINT

2023-04-10 00:26:57 One core production input for AI systems is labelled data. The implications of us crossing the Rubicon of systems being able to generate good quality data that doesn't exhibit some pathological garbage in, garbage out failure mode are profound. And it is beginning to happen!

2023-04-10 00:25:57 Something that will further compound acceleration of AI research is that models are now better and cheaper than humans at data generation and classification tasks - this already true for some cases (eg Facebooks Segment Anything model, and GPT-4 for some labeling).

2023-04-09 23:34:36 @robinlavallee Thanks for letting me know, have a good day

2023-04-09 21:43:28 @chrisweston had a fun afternoon reading through this paper and will be in Import AI 324, coming out on Monday via the FORBIDDEN WEBSITE known as Substack! https://t.co/CIw0MlfB1m

2023-04-09 16:26:45 @yacineMTB

2023-04-09 16:24:45 @yacineMTB not only do I regularly donate, but I also bought basically all the merch! https://t.co/AKFrx5WX25 The hoodie is v nice quality

2023-04-09 15:55:16 @chrisweston this is a fun one: https://t.co/SWKIbUOfPZ

2023-04-09 15:52:13 One of the greatest stories of our lives is unfolding right now and you can read about it every day on arXiv for free. Absolutely wild.

2023-04-08 14:41:15 @tmiket @everyto I didn't write this

2023-04-07 00:57:14 Facebook's new Segment Anything model is pretty interesting - did well on an adversarial example

2023-04-02 22:18:25 @Simeon_Cps can you share your slides?

2023-04-01 16:45:58 Dutifully filling this out: https://t.co/ZqxlwCZ3SN

2023-04-01 16:45:04 Woah felt quite an impressive earthquake jolt in Oakland just now

2023-03-31 15:59:47 @evanjconrad @sullyj3 I have become pure brand energy

2023-03-30 23:10:12 @neilwlevine oooh I love Sibley but haven't been up there lately - on my list if the good weather keeps on!

2023-03-30 23:00:29 @maththrills haha this is great! thanks for sharing

2023-03-30 23:00:22 RT @maththrills: Inspired by @jackclarkSF's tweet about his hike being "violently green", thinking of album cover suitability, and easily d…

2023-03-30 22:09:17 @maththrills it's crazy! I was texting my friends 'everything is too goddamn green', as well. I guess this is the outcome of 4 months of rain

2023-03-30 22:06:57 @alyssamvance especially amazing this season given the huge amount of rain! really amazing stuff up there

2023-03-30 21:53:52 @lacker Trials above Claremont Canyon / Stonewall-Panoramic Trail etc.

2023-03-30 21:39:55 A wonderful hike in the Berkeley hills - everything is violently green and the air is fizzing with the rich smell of wet earth and plants fighting to grow. #VOTENATURE2023 https://t.co/BPvzLVc9O7

2023-03-30 20:58:06 @ruthstarkman Yes, everyone can access Claude via Slack

2023-03-30 15:59:12 @S_OhEigeartaigh I am and my awesome colleagues @sandybanerj and @avitalbalwit led this release (along with tons of other colleagues). I'm just excited for more people to get to talk to Claude!

2023-03-30 15:55:39 Stuff I do with Claude: - ask for critical feedback on memos - surface research terms where I've forgotten the word/phrase but can roughly describe it - brainstorm ideas - summarize things into bullet points - search over text for semantic concepts It's great!

2023-03-30 15:53:40 To everyone who has been asking me for Claude access - you can now add Claude to your Slack. Claude is great in Slack

2023-03-30 15:03:34 @jjvincent Thanks for the kind words @jjvincent , I spent a lot of time with this story and glad people have been enjoying it!

2023-03-30 14:26:40 RT @jjvincent: particularly lovely bit of fiction in @jackclarkSF's latest Import AI newsletter on Mechanical Human - the Mechanical Turk a…

2023-03-29 19:52:09 @DynamicWebPaige cool, thanks! yeah was just curious. I like the UI/design of the code, also

2023-03-29 19:38:03 @DynamicWebPaige What do you mean by 'sneak past the filters'?

2023-03-28 20:53:40 @avitalbalwit https://t.co/cTf3e1Wysg

2023-03-27 15:52:29 @RedTailAI some specific ideas and recs: https://t.co/vljhYFY6wa https://t.co/6R2Qa4bwMr https://t.co/wVMgXGf0at

2023-03-27 15:50:29 @RedTailAI yes, of course I do.

2023-03-27 15:36:34 @RedTailAI Yes. My intuition is it happens at some point, but that could be decades to millennia away - no real clue re timelines here. It's also always good to work hypotheticals seriously, so you talk about moral patienthood as though real to help you work the problem.

2023-03-27 14:36:48 @RedTailAI Note, my position here is potential sentience. I don't know that I'd strongly bet for sentience at this point, but I also wouldn't strongly bet against it, and I have no notion on which timeline sentience could even appear. I'm more noting the ethical challenges of its potential

2023-03-27 04:55:27 @jjluff @matthewclifford Very good term! May use

2023-03-26 21:03:44 @sheonhan that most interesting critics (mumford, jane jacobs, baudrillard, etc) seem to have. I feel like independent media via stuff like substack/weird newsletters is one of the better ways to fund it - recently have been pushing Import AI in this direction as an experiment 3/3

2023-03-26 21:02:53 @sheonhan unless you're writing about things with big, dedicated ad buys (e.g cars, wine, etc). Tech doesn't really advertize in same way, other than a tiny slice of enterprise stuff. Additionally, I think both journalism and tech counter-select for the interdisciplinary frame 2/3

2023-03-26 21:02:04 @sheonhan had a read - I agree with you re a culture of tech criticism. From my POV I think some of this comes down to economics - I was a fulltime journo for many years and tech has basically bricked the economics of journalism so criticism isn't so sustainable 1/3

2023-03-26 19:57:18 @sheonhan oooh interesting, will read. Thanks for sharing!

2023-03-26 19:57:00 (And to be clear, I think systems like GPT-4 are technological marvels that are going to let humans do a bunch of new, completely incredible and useful things. But as with anything extremely powerful, we must be aware of the implicit political power encoded in these artifacts.)

2023-03-26 19:54:08 More analysis of GPT-4, as well as the political implications of increasingly capable AI systems, in Import AI 321: https://t.co/GdbgH1i29X

2023-03-26 19:53:16 GPT-4 should be analyzed as a political artifact just as much as a technological artifact. AI systems are likely going to have societal influences far greater than those of earlier tech 'platforms' (social media, smartphones, etc). https://t.co/vnMPXPztdc

2023-03-26 19:51:34 @VSehwag_ thanks! I think it's a very impressive technological achievement, so worth thinking through the societal aspects deeply

2023-03-26 07:30:50 @matthewclifford Age of Em is a good portrait of transition period

2023-03-25 20:30:37 @karinanguyen_ Haha a good idea. I am unfortunately crap at scriptwriting but am trying to get better at it, so maybe one day

2023-03-25 20:17:50 @Trent_STEMpunk @XiXiDu Thanks!

2023-03-25 20:11:37 @Trent_STEMpunk @XiXiDu Every week at https://t.co/NRyFUkKznD

2023-03-25 19:05:29 @arram yes, we're definitely in the 'may you live in interesting times' era

2023-03-25 19:05:02 @jamescham come in on the water is boiling with possibility. It's a quantum foam bath!

2023-03-25 19:03:55 @Deepneuron you should write up some advice! would find helpful

2023-03-25 18:54:54 @the__dude98 haha actually a few things, some of which might be useful! My main takeaway was 'wow I'm having extremely literal dreams that are basically like a real job, I should definitely take a holiday!'.

2023-03-25 18:53:50 @powerbottomson1 wouldn't inflict that on worst enemy. but thanks!

2023-03-25 18:52:31 I've also had a lot of trouble sleeping because I keep on waking up at 4am and reading arXiv papers and being like 'yup, we're in the exponential' and then I just lie there thinking about all the changes that are basically locked-in at this point.

2023-03-25 18:50:36 Dreams I've had lately: - Abducted by a foreign intelligence agency for figuring out details of a black budget AGI project. - Argued for 3 hrs with someone re semiconductor policy. - Fought a doomed political battle to halt the deployment of unsafe AI. Anyway, taking a holiday!

2023-03-24 01:48:30 @eliotpeper Thank you!

2023-03-24 01:12:12 Wrote a story I hope to be useful for upcoming Import AI. The morality of machines will be different to those of people, but these alien ethical and moral frames may ultimately be a bridge. 'My fiction beats the hell out of my truth,' as jawbreaker said.

2023-03-23 22:00:28 @rasbt @omarsar0 @proales yup, seconded! https://t.co/CIw0MlfB1m amazing how much value add you can provide by reading the underlying research papers, haha

2023-03-21 04:48:57 Very confusing paper - superficially impressive, but the model trades off SOTA against ERNIE 3.0 titan (260bn parameters), and is undertrained (329bn tokens of data). Still, an interesting statement of intent and ambition - writing up for Import AI 322. https://t.co/rR0S3PNeMq

2023-03-21 02:27:26 @yonashav @arankomatsuzaki thanks for the link - though reading the paper also seems to suggest that 329bn is a significant undertrain here as well. I will dig in a bit. My interpretation is the paper is more a claim about having infra to train big models rather than about training a _good_ big model

2023-03-21 02:22:07 @finbarrtimbers @arankomatsuzaki they reference it in the paper! (20 = Chinchilla paper). very strange https://t.co/GNyY980B88

2023-03-21 02:19:18 @arankomatsuzaki isn't this, like, severely undertrained in terms of data?

2023-03-20 17:16:19 @natolambert Enjoyed the societal take - would be curious to see you sketch out in more detail what you see as unaddressed, eg the regulation point. Thanks for writing this!

2023-03-20 16:40:15 @jgrayatwork @AnthropicAI psyched to work together again!

2023-03-19 20:09:14 @copybymatt Ding ding ding! Via @mattparlmer https://t.co/Gy9ciU0bZm

2023-03-19 01:21:22 @cromas A+

2023-03-19 01:16:11 When I was a kid I read a lot of religious texts of my own volition (and I continue to do so). For thousands of years, people have wanted to understand/be closer to some theistic god. There's some lesson to this re: the current AI boom that I don't quite yet understand.

2023-03-19 01:13:35 Sometimes I think a lot of the breathless enthusiasm for AGI is misplaced religious impulses from people brought up in a secular culture.

2023-03-17 21:01:49 @tszzl

2023-03-16 05:03:37 @nilshoehing

2023-03-16 01:18:39 Alpaca 7bn passes the 'helicopter test'. Writing up in Import AI 321. Subscribe here: https://t.co/is0y6grJXb https://t.co/jnmDu3Z9zO

2023-03-15 01:26:30 @mattyglesias So I'll typically just say something like "Read this essay and give constructive and critical feedback, paying particular attention to logical inconsistencies or clunky phrasing: $paste_whole_article" and find the results sometimes quite helpful.

2023-03-15 01:25:35 @mattyglesias yeah! You can do this with stuff like chatGPT or Claude as they have long enough context windows (8K+ tokens) that you can paste whole essays in

2023-03-15 01:15:25 @mattyglesias I've found these language models to be quite good at criticism - e.g, I give them a 2,000 word essay and ask for critical feedback. Sometimes this can help spot weird jumps in logic or areas of repetition. Don't find them useful for additive writing yet, though.

2023-03-14 18:25:20 @mcwm @dinabass @rachelmetz @technology lol I'm joking sorry!

2023-03-14 18:20:30 @dinabass @rachelmetz @mcwm @technology nailed it

2023-03-14 17:40:53 RT @adamdangelo: Today we are launching Poe subscriptions, which will provide paying users with access to bots based on two powerful new la…

2023-03-14 17:40:08 @rachelmetz @mcwm @technology it's not JCVD I'm afraid

2023-03-14 16:00:59 @Leoagua1 @d_feldman there are numerous public estimates in public papers of low single-digit millions

2023-03-14 15:47:17 @d_feldman those gpt3 numbers are way off

2023-03-13 15:55:40 @agstrait @alexkozak @JMateosGarcia @NIST @Elliot_M_Jones yes, would love to read/give feedback if helpful! And yes, requires a bunch of resources

2023-03-13 03:35:16 They all deserve this so much. An awesome film.

2023-03-13 03:29:29 100% deserved. I cried during Everything Everywhere All At Once multiple times. A beautiful and fresh performance.

2023-03-13 00:11:21 @andy_l_jones I debated deleting my original tweet but I think that's kind of lame/bad habit to get into, so I retweeted the statement onto my timeline and replied in thread and am eating my proverbial hat

2023-03-13 00:10:13 @andy_l_jones Yeah, I have updated extremely positively given the FDIC statement!

2023-03-12 23:23:13 RT @jackclarkSF: https://t.co/uFwPwAIWtB

2023-03-12 23:23:09 https://t.co/uFwPwAIWtB

2023-03-12 23:07:32 https://t.co/0XyQsdaBcC

2023-03-12 20:55:26 @sbeckerkahn mostly just means this whole situation making me depart from my typical chipper idealism with regard to AI policy and consider a world where it's much harder for AI policy to be effective

2023-03-12 20:22:09 @divergiment Yeah that's a good call out, though perhaps speed with which this can unfold could be different. Thanks

2023-03-12 19:02:38 @cromas I won't claim to be smart enough and familiar enough with this area to have a precise solution, I'm more worried about bad hypotheticals like bank runs on other smaller banks and other contagion effects. I figure there must be some way to reduce damage to a load of companies?

2023-03-12 18:59:48 @yonashav I think some of this relates to proactive loud communications and getting out in front of issue. I also think SV has some privileged information about scale of impacts/effects, eg sense of what does immediately for payroll etc. Also, delighted to wake up on Mon and be wrong!

2023-03-12 18:58:16 @andrewthesmart Honestly I'm not sure and it's a good point. There was a similarly rapid herd shift from crypto to AI a few months ago. It's odd. Maybe SV companies are more like starlings than we thought. A corporate murmuration!

2023-03-12 18:56:56 Note: I really hope we figure stuff out and FDIC comes through, as do authorities in other places (e.g. whoever in UK is responsible for making sure SVB UK doesn't cascade).

2023-03-12 18:47:27 Similarities: - Seeming inconsequential thing happens which has a contagion effect - Need rapid and bold policy actions to avert contagion - Gulf in understanding between those directly dealing with the problem and lawmakers

2023-03-12 18:46:25 Gotta admit that watching this SVB situation unfold is blackpilling me on AI policy a bit

2023-03-12 17:19:20 @Ted_Underwood Chinchilla + instruction following

2023-03-11 23:26:21 @RemmeltE Will be writing about the societal implications in Import AI as I regularly do when covering this kind of stuff. Thanks for sharing your feedback and have a great day

2023-03-11 23:25:52 @RemmeltE I mostly think there's value in reporting the state of things and paying attention to what is going on in the open source / distributed ecosystems. It's also very notable how IF stuff makes smaller models more powerful which has implications for the broader ecosystem.

2023-03-11 21:11:59 @togethercompute hiya, this is interesting &

2023-03-11 19:35:30 @alexkozak @JMateosGarcia @agstrait @NIST for example, you no longer want to evaluate models on zero-shot or few-shot prompting alone

2023-03-11 19:34:58 @alexkozak @JMateosGarcia @agstrait @NIST most policymakers assume the industry has converged on some benchmarks/evals for large generative models that mirror earlier (easier) benchmarks for computer vision - this isn't the case. It's also getting increasingly hard to build benchmarks for cutting-edge models

2023-03-11 19:34:04 @alexkozak @JMateosGarcia @agstrait @NIST a lot of the issue is there are some benchmarks for a narrow slice of things (e.g, fairness), but most of these benchmarks have a huge set of problems (which their creators know about), and we're constantly expanding model capabilities over time.

2023-03-11 18:00:56 @sir_deenicus while I agree that giving the right prompt improves performance, I think the key thing here is that IF-tuned models don't take as many bits to calibrate as non-IF models. Though helpful to elicit a good story out of OPT, thanks for sharing!

2023-03-11 17:47:18 @bahree correct, it's not a 1:1 comparison. The point I'm making is instruction-tuned language models are really good relative to non-instruction-tuned ones

2023-03-11 17:44:43 @alexkozak @JMateosGarcia @agstrait to @agstrait point - I agree! I really want there to be a third-party benchmark/eval system here. I spend a lot of time advocating for @NIST to do this, but NIST needs way more resources to do this effectively. Eager for ideas here - it's a big issue!

2023-03-11 17:43:53 @alexkozak @JMateosGarcia @agstrait I'm also continually developing more implementable/shovel-ready ideas for monitoring (e.g this paper here https://t.co/vljhYFY6wa) and push those every chance I get. Unfortunately policy speed != AI development speed, so we're in a tough situation.

2023-03-11 17:43:02 @alexkozak @JMateosGarcia @agstrait so we're continually doing evals (and publishing on them - see, e.g, red teaming, model-driven evals, etc), but also I am regularly going to DC/Brussels/London and stressing to policymakers that evals maturity at all of the labs is immature relative to scale of problem

2023-03-11 17:42:22 @alexkozak @JMateosGarcia @agstrait Chiming in here - we invest a huge amount in measurement/evaluation/assessment of our models, but as models scale the capability surface keeps expanding faster than the ability to eval the whole surface, so while we invest a lot here, we also need gov to scale up stuff

2023-03-11 17:28:28 @gharik tried a few variations at Temp 1 and 0.9 and most of the time it didn't end up writing the story and did the classic thing of either continuing the frame of instructions, or bounced to diff story. Might have got lucky with temp 0.7 on first try. thanks for suggestion!

2023-03-11 17:25:45 @suchenzang I imagine it's way better - my main take here is it's interesting to see how powerful instruction tuning is and really lets smaller models punch above their weight. Also, is there an IF-tuned OPT thing available anywhere? (Or perhaps someone could try my above prompt!)

2023-03-11 17:17:56 @gharik OPT attached. OpenChatApp is the default here, so not sure: https://t.co/rftZJUpHZw I feel like these are reasonable settings for OPT but if you think I'm artificially sandbagging it lmk! https://t.co/MuJokwwNPy

2023-03-11 17:14:37 Writing up for Import AI 320

2023-03-11 17:13:56 Playing around with the latest ChatGPT replication (OpenChatApp) and it's a) quite good, and b) neatly illustrates how crazy-powerful instruction-tuned models are compared to stock LLMs. Compare OpenChatGPT (20B params) on left to OPT (GPT3-replication, 175B params) on right. https://t.co/UqIhnp2aGg

2023-03-08 15:44:41 Very interesting post about LLaMa performance and ethos for distributing models. Thanks for sharing @theshawwn ! https://t.co/p6L7Cp5LRd

2023-03-07 18:17:45 Claude on Android! https://t.co/gLdN40aZdu

2023-03-07 05:24:39 @Deepneuron @tszzl just white knuckling the exponential, but hanging in there!

2023-03-07 05:09:34 @Deepneuron @tszzl you should definitely write about this!

2023-03-07 03:28:04 @AISupremacyNews @SubstackInc thanks so much! excited to cook up some fun experiments

2023-03-06 21:00:13 Very surprised at the way Facebook handled the LLaMa model launch https://t.co/5Xlcxo6OpH . It feels like half-releasing models and them then circulating on torrent networks is a pretty unfortunate outcome. Would love to understand some of the thinking behind it. https://t.co/UpGpZJy2kH

2023-03-05 21:48:57 Excited to continue this journey with you all! If things go well I want to stand up things like Import AI interview series and such - if I get enough paid subscribers or founding members I'll be able to put people on contract to help scale up the newsletter even more!

2023-03-05 21:48:03 The general idea is that if people see value in it they can be a good audience to 'workshop' these ideas with before later making them public. This means there's a value to subscribing (as some of the ideas are about time-contingent trends in AI), but preserves public utility.

2023-03-05 21:47:19 One of the things I've found valuable about writing the newsletter is having an audience keeps me accountable for regularly reading @arxiv / other AI news sources and regularly publishing ideas out of it. I now want to be accountable for publishing broader thoughts about AI.

2023-03-05 21:46:16 Import AI 319 will be coming out via @SubstackInc tomorrow. Very excited to start a new chapter of the newsletter here. In particular, I'm going to use a paid plan to publish a monthly (or more frequently) 'Import (A)Idea' for subscribers. These will be made free at a delay.

2023-03-05 18:51:23 @mbrendan1 @SamoBurja Jared Diamond - Collapse is also pretty good! But tainter is the best on this

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

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

2023-02-27 16:58:04 @GBxGlobal thanks for the promotion but I'm a co-founder - Dario Amodei is CEO.

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

2023-02-21 10:01:28 @jjvincent It's going to be a really relaxing decade

2023-02-20 15:52:44 @generativist Is this... Rubberbossing ?!?!

2023-02-20 15:35:10 @sai_prasanna hell yeah! I agree with this

2023-02-20 14:25:35 For those not familiar with the term: https://t.co/6d5sD2zU0w

2023-02-20 14:23:37 Rubberducking with language models is pretty effective, these days. Having trouble thinking about something? Talk to the rubberduck! RHLF means they're kind of like non-judgy active listeners and by discussing ideas with them you can surface your own cruxes.

2023-02-16 18:25:01 RT @AmandaAskell: My favorite of Claude's excellent recommended titles for this paper when I gave it the abstract was: "HELP! My language m…

2023-02-16 16:50:12 Though I don't think 'just ask the AI to not be bad' is going to solve many hard aspects of safety/alignment, I am continually surprised at how well prompting/asking works for a broad range of problems. Cool work from @AmandaAskell and Deep Ganguli and others at @AnthropicAI

2023-02-16 16:49:16 justaskAItobenice.png https://t.co/zG41reG78p

2023-02-15 19:22:19 @ohlennart engineering projects aren't sufficiently original to get you tenure. Academia also doesn't have a culture of hiring engineers/infra teams to support scale-up computing.

2023-02-15 04:20:15 I'll one day write about the experience of trying to explain LLMs before anyone gave a fuck and how strange and alienating it was, but not today! Just remember - 4 years between gpt2 and where we are right now. Prepare for the next four.

2023-02-15 04:18:53 I remember sitting in an airport in England in Dec 2018 generating samples from the 1.5B model and feeling the gravity of the advance. We all had a visceral intuition for what it meant, but of course when we told people what we thought they said we were doing pointless hype.

2023-02-15 04:15:13 GPT-2 was announced four years ago today. https://t.co/whUSVDX7oY

2023-02-15 03:57:07 @rachelmetz @technology Great team, great hire!

2023-02-14 00:30:05 Cc @giffmana who I figure might know?

2023-02-14 00:29:45 Iirc there's some upper-limit to top-1 accuracy on imagenet due on mislabeling of underlying dataset. Is this true and is it documented anywhere? Want to clarify in this year's @indexingai report.

2023-02-13 22:35:37 @AlexCEngler happy to help!

2023-02-13 01:05:21 So beautiful to be at peace with waves and their susurration and to look beyond to rain-drunk green hills. #VOTENATURE2023 https://t.co/euvEcgVe48

2023-02-13 00:54:48 @powerbottomdad1 Curious for your take on this story by him... I think about it with ref to AI a lot. https://t.co/6je6Nxklxv

2023-02-13 00:43:10 @powerbottomdad1 Good read!

2023-02-13 00:42:48 @JohnHelveston Fair. But gpt2 was a far worse chess player.

2023-02-12 21:50:12 @laurenkunze Ooh I wasn't aware of that! Thanks for the pointer. Anything in particular I should read?

2023-02-12 21:16:49 @amolitor99 but I also recognize there's some chance I'm wrong and this stuff is about to hit a wall, which would be fascinating to see. Hopefully we can check back in a few years and see what is up

2023-02-12 21:16:10 @amolitor99 I'm loudly exclaiming the thing is a rocket ship because we have barely any of the policy/regulatory institutions needed to ensure the rocket ship doesn't immiserate humanity! I also think we've basically seen nothing yet in terms of capability expansion

2023-02-12 21:04:26 what extremely creepy parasocial AI relationships look like (screenshot of someone in a FB group really upset about removal of the Erotic Role Play stuff) https://t.co/qpnzyQ61yz

2023-02-12 20:58:48 RT @nonmayorpete: Replika, the "AI Friend" app, got big by advertising its NSFW abilities. Now it's turning off its "erotic role-play". P…

2023-02-12 20:55:04 Another narrative I'm reading is that perhaps Apple/Google finally saw all the weird ads on TikTok and put pressure on Replika or it'd get taken out of the app stores. Either way, interesting stuff!

2023-02-12 20:51:44 @chrismarrin image recognition has vast economic utility today and is deployed to literally billions of people via on-phone stuff across Apple and Google - it's not going away. Ditto translation.

2023-02-12 20:50:09 Free alpha for AI journalists: Replika seems to be pivoting its business model away from the scuzzy e-sex roleplay stuff. The gossip in FB groups is they switched to an underlying model that doesn't support erotic roleplay anymore. If I had time I would report this out!

2023-02-12 19:44:03 @dileeplearning Sounds like you should email the authors who gave the paper the title

2023-02-12 19:43:31 @bradzaguate Same, tbh. My personal goal is to get an AI-infused robotic exoskeleton (I have a history of back trouble)

2023-02-12 17:26:46 @Grady_Booch @EMostaque thanks for sharing your wisdom, I appreciate it! always eager to read and learn, especially about 'bear cases' on hype things

2023-02-12 17:26:14 @Tim_Dettmers DM'd ya some more questions re SWARM. Thanks for your patience!

2023-02-12 16:39:21 @richardludlow @erich_elsen rekt also, true.

2023-02-12 16:38:45 @Grady_Booch @EMostaque I'm also curious if you've got any particular stuff to recommend here that might work as 'lessons from the winters' (or if you've considered/have writing something yourself)? Would love to read more!

2023-02-12 16:35:36 @groby yeah, I think training systems that funhouse mirror reflect back the internet has a huge number of downsides as well. You might want to check out this report on a 'National AI Research Resource' to see how people are thinking re diff paths https://t.co/HFrbPHKPDu

2023-02-12 16:26:54 @groby A bit on the nose, but yes: https://t.co/JTrGRD1bFq

2023-02-12 16:25:01 @muscovitebob not really - I myself find thinking about this so confusing that I still do lots of normal stuff, like pay into my 401k etc.

2023-02-12 16:24:00 @Grady_Booch @EMostaque I'm aware of the winters (and read stuff like the Lighthill report, etc). What seems different this time is the enhanced amount of economic utility of these systems - if you stopped all research and just focused on engineering, feel like NLP/CV/Translation all have big effects

2023-02-12 02:04:33 @eric_is_weird This is a useful example. Have you written anything about this anywhere or have any links to contemporary writing/papers that note this? Would find helpful!

2023-02-12 00:47:29 @pwlot @tdietterich I have, but it still feels less interesting to me than when I go to the pub with people of different backgrounds and chat. (This is, admittedly, a high bar given this stuff could barely do a sentence a decade ago)

2023-02-12 00:45:27 @dlowd Idk, I followed Watson very closely and it didn't really sit on a series of past systems and was clearly custom designed around a specific problem with a specific architecture (I obsessively read the papers). This stuff sits on a diff and much more proved out tech foundation

2023-02-12 00:43:45 @tdietterich So far it seems like... A lot! But I also am yet to see an LM produce a truly original idea - the sort of idea where when you're talking with colleagues you collectively figured out a correct and novel answer. That could end up being an upper limit on this stuff.

2023-02-12 00:35:38 Anyway, how I'm trying to be in 2023 is 'mask off' about what I think about all this stuff, because I think we have a very tiny sliver of time to do various things to set us all up for more success, and I think information asymmetries have a great record of messing things up.

2023-02-12 00:34:46 @jfischoff I'm not sure but, per some of the other papers (and others I can't dig up while waiting for this train), this phenomenon is showing up generally in a bunch of places. I wouldn't be so interested in it if it showed up purely in one result

2023-02-12 00:33:27 We can also extract preference models from LMs and use those to retrain LMs via RL to get better - this kind of self-supervision is increasingly effective and seems like it gets better with model size, so gains compound further https://t.co/FmQCaILcTN

2023-02-12 00:26:26 We can also train these models to improve their capabilities through use of tools (e.g, calculators, QA systems), as in the just-came-out 'Toolformer' paper https://t.co/hf3n4RgSd7 . Another fav of mine= this wild paper where they staple MuJoCo to an LM https://t.co/VL9ggQUe8k

2023-02-12 00:25:11 There's pretty good evidence for the extreme part of my claim - recently, language models got good enough we can build new datasets out of LM outputs and train LMs on them and get better performance rather than worse performance. E.g, this Google paper: https://t.co/0jfJIaEQR7

2023-02-12 00:24:08 A mental model I have of AI is it was roughly ~linear progress from 1960s-2010, then exponential 2010-2020s, then has started to display 'compounding exponential' properties in 2021/22 onwards. In other words, next few years will yield progress that intuitively feels nuts.

2023-02-07 16:41:41 @peterwildeford Thanks for reading, I'm paying increasingly close attention here

2023-02-07 03:42:11 @mattparlmer Dm'd you. Really sorry you're gonna through it. Agi will get us all very cheap and excellent exoskeletons

2023-02-06 18:33:57 @megyoung0 model cards and datasheets seem like pretty useful interventions and it'd be interesting to think about how to wire these into corporate incentives, potentially through policy recommendations

2023-02-06 17:51:51 @cdossman thanks for reading! I enjoyed writing this one

2023-02-05 18:26:36 @Tim_Dettmers DM'd you some qs about SWARM

2023-02-05 05:22:25 @Stephan90881398 Pure, artisanal, handwritten horror.

2023-02-04 21:29:31 Excited for you all to read 'The Day The Nightmare Appeared on arXiv', a short story coming in Import AI 317.

2023-02-03 21:21:32 RT @goodside: Poe chat is out on iOS, and includes *both* OpenAI and Anthropic models! If you want to talk to Claude (Anthropic's competit…

2023-02-02 14:15:52 @ohlennart @scienceisstrat1 @Noahpinion @ylecun @erikbryn @amcafee house style looks like The Economist

2023-01-31 15:20:52 @atarkowski @dinabass @OpenFutureEU Can you say a bit more about why it isn't an easy sell and/or point me to some writeups here? Would be interested to read!

2023-01-31 00:50:01 @dmitri_dolgov please come to Oakland, it's just across the bridge. Also please do the bridge.

2023-01-30 18:03:31 @avizvizenilman Haha yes please

2023-01-30 17:34:46 @xlr8harder I think startups will 100% do loads of interesting/ambitious things here, but they're still companies operating in markets. I feel like there's a bunch of AI capabilities that can be explored which are inherently hard to commercialize and would like to see more experiments there

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

2023-01-23 19:28:50 @TheodoreGalanos Yup! What I meant was I was really excited about stuff like MERLIN and other RL agent approaches that tried to get agents to build world models. Turns out better thing is perhaps an LLM which can be a plug-in world model (e.g SayCan, Mind's Eye, etc)

2023-01-23 15:58:08 @dileeplearning Always happy to talk just let me know when you're in SF

2023-01-23 14:58:49 Feels like there's some weird signal that I'm not very excited about LLMs - in many senses I feel like this tech is locked in with regard to trajectory. Instead of feeling excited about it I keep trying to generate bear cases and the bear cases keep getting overtaken by progress.

2023-01-23 14:57:28 Things I thought were useful but also basically predictable: - Computer vision based on DL - Language models as all-purpose data transformation engines - Code models 2/2

2023-01-23 14:56:33 Things I was enthusiastic about (and wrong about timelines of): - Self-driving cars - Reinforcement learning as main path for powerful systems - World models (ditto) - Industrial robots combining with RL 1/2

2023-01-20 21:41:38 @hlntnr @maosbot What was the small one?

2023-01-20 00:01:59 @tahirwaseer oh, I just meant my general writing process. I feel like llms are mostly useful just as bullshit checkers/calibrators, but are not that integrated into my overall writing process yet. This might change, but I think I write quite weird things, so may be a while

2023-01-19 23:32:43 @tahirwaseer there isn't a particularly useful pattern - sometimes I oneshot things that need very few tweaks and sometimes it'll take ten iterations till I'm happy with it.

2023-01-19 19:12:27 @nazneenrajani @AnthropicAI elo used a lot in RL - go, openai five, etc

2023-01-19 17:52:40 @jeffreyhuber MY BIG BLOB OF MATRIX MULTIPLICATIONS IS _BULLYING_ ME!!!!!

2023-01-19 17:52:27 @jeffreyhuber I actually left my desk and got a coffee and complained to my human colleagues about how badly Claude negged me, lol, rekt, etc.

2023-01-19 17:40:48 @FelixHill84 Yeah, I think augmentation is wrong word here. But ER does seem like a good way to be able to dump more relevant data into the learning process.

2023-01-19 17:18:24 Tfw yr language model negs you (A recent workflow I've started using is 'write first draft' >

2023-01-18 19:29:45 @togelius Yeah

2023-01-18 19:29:32 @tdietterich Good clarification, thank you!

2023-01-18 17:58:21 after chatting with a few colleagues and texting a few outside experts I'm pretty sure this is right, fyi. An even simpler way to think about this is that these breakthrough results are mostly a consequence of capital - using enough compute and data for long enough.

2023-01-18 17:34:34 One rule of last decade has been 'the simple stuff works', and perhaps why Q-learning worked so well is that it really just lets you have an architecture that is naive to environments because it naturally cycles over relevant data a lot. Thoughts?

2023-01-18 17:33:42 I mean obviously experience replay has a role in the learning process, but if you zoom really far out, it's just a way to get an agent to pay close attention to temporal slices of data from interacting with environment and train on that... https://t.co/h75gjFBjpi

2023-01-18 17:32:58 When I think about the last ten years of AI, I feel like a lot of it got sparked by the arrival of relatively simple approaches that could soak up a ton of data and/or computational resources. In hindsight, isn't 'experience replay' in Q-learning basically data augmentation?

2023-01-17 00:45:16 @ESal Thank you! Glad it's helpful

2023-01-16 21:40:53 @alexeyguzey I, unfortunately, am not cuter this way

2023-01-16 21:40:38 @paul_scharre @nmaslej @Miles_Brundage @jjding99 @mattsheehan88 @tdietterich @ylecun @indexingai @etzioni @CSETGeorgetown @DataInnovation @Noahpinion there have also been impressive things like CogView and other models. Plus, the Chinese ecosystem around video object detection via YOLO variants is very mature/impressive. I'd use lens of 'systems' to augment lens about papers, basically

2023-01-16 21:39:47 @paul_scharre @nmaslej @Miles_Brundage @jjding99 @mattsheehan88 @tdietterich @ylecun @indexingai @etzioni @CSETGeorgetown @DataInnovation @Noahpinion It's generally useful to look at research artefacts / models rather than pure papers. E.g, it's pretty notable that GPT-3replication GLM-130B (Tsinghua) is basically better than BLOOM (HuggingFace) and OPT (Facebook) at a bunch of capabilities

2023-01-16 20:48:48 @natolambert Thanks so much! I miss some weeks myself but still probably averages out to 0.8 issues per week per year at least

2023-01-16 20:29:21 @HaydnBelfield @deepfates @AnthropicAI

2023-01-16 20:28:05 @natolambert I think it's crucial that it's independent from my dayjob, to be honest. This is why it's such a crazy project - I work a lot of hours at my dayjobs and then I have a whole other evening job. But it keeps me sharp and also gives me independent leverage, which is important to me.

2023-01-16 20:26:52 @natolambert Simple answer - I didn't! I started Import AI before I joined OpenAI and always have kept it separate - I write it on trains and planes and on evenings and weekends and early mornings, predominantly. Only way it intersects dayjob is I email myself papers I stumble across at work

2023-01-16 18:37:56 @Dominic2306 @tshevl Yeah I sort of sneak in my weirdest thoughts into the email subscriber version. I'll probably change this so it mirrors to WordPress though

2023-01-16 17:44:24 @BillLeaver_ @deepfates @AnthropicAI thanks! some people tell me they skip the rest of the newsletter and just read the stories, which delights me : )

2023-01-16 16:19:00 @tahirwaseer thanks so much for being an OG subscriber

2023-01-16 16:15:31 @mezaoptimizer I'm trying to do some of that this year! on my list

2023-01-16 16:07:44 @mezaoptimizer compressing. I think about timelines in terms of capabilities I expect to have effects in the world. I don't really have precise timelines about AGI because I think AGI is inherently fuzzy and hard to define. I do have timelines about capabilities that influence geopolitics

2023-01-16 15:55:23 @deepfates @AnthropicAI oh and finally, living in joy and love with my partner and cherishing the rich emotional fabric of being a living being on this planet! and being kind and spending time with friends. touching grass. seeing more punk shows. all that good stuff. : )

2023-01-16 15:54:14 @deepfates - hiring a bunch of people at @AnthropicAI to free up my own cycles for above - spending way more time trying to build evaluations/measurements of AI systems - identifying actions that seem a) useful and b) likely to annoy some people, as most decisive actions involve annoyance

2023-01-16 15:52:52 @deepfates few things: - along with Import AI, going to try and do more public writing about some of the things that are highly likely to happen. - spending more time trying to convince people at various labs to be more public themselves - trying to 'scale myself' so I can do more 1/2

2023-01-16 15:50:08 @tweet_prat the lines are becoming increasingly blurry : )

2023-01-16 15:06:52 @pstAsiatech thanks Paul!

2023-01-16 15:06:28 @rasbt thanks a lot! I write them for the readers, and messages like this make my day

2023-01-16 14:16:15 On New Year's Day 2023 I said I'd spend this year living as if I fully believed my own timelines, so I shaved my head and made some plans. There's a lot to do and too few people, but isn't that always the case? The times, they are a-changin. https://t.co/LNguAwxjGU

2023-01-16 04:22:49 Import AI cracked 30,000 subscribers recently. Thanks to everyone for reading! Writing it is one of the great joys of my life.

2023-01-15 19:00:55 @mwilcox Favorite book I read last year!

2023-01-13 20:57:33 @kimtsherwood hey Kim congratulations, this is awesome! There was something special in the water back at UEA in those days :^)

2023-01-13 04:35:02 It was lovely to know you all. https://t.co/6encoUZMar

2023-01-12 17:11:56 @VictorLevoso @gwern helpful thread - I'd noticed they nerfed the block-hit thing but hadn't seen the vector conversion for inventory and health, thanks for flagging

2023-01-12 06:10:16 @stephenroller @sir_deenicus @terrible_coder @marian_nmt @Tim_Dettmers @StasBekman that's helpful context, thanks for sharing! would be cool to read a retro sometime

2023-01-12 03:24:02 @FelixHill84 Anyway, it's cool work!

2023-01-12 03:23:50 @FelixHill84 But I am pretty confused people aren't making a bigger deal of the results. I've read the paper this evening and I don't think stuff is being sandbagged. DreamerV3 seems very data efficient and has good performance against a multitude of compares on diff benchmarks https://t.co/jlt1j2zRMk

2023-01-12 03:22:52 @FelixHill84 I guess I'd expect some more tweets about it, a DeepMind blog about it, possibly even some press stuff. It might just be that the Sauron Eye of 'vibes' has moved to LLMs.

2023-01-12 02:02:51 DreamerV3 seems like an RL agent that 'just works' across a vast set of environments. I feel like either it has some weird issues I'm not seeing, or people just aren't that jazzed about RL these days. Anyway, worth reading! https://t.co/zDOx9uWsnY

2023-01-12 02:01:54 My extreme inner-inside-wonky take on DreamerV3 (lovely paper) is it's unusual to see a @DeepMind paper with an impressive result (world models that work! cracking MineCraft diamond challenge!) with a tiny number of authors (here: 4). No idea if this means anything though!

2023-01-11 17:29:12 If you have questions, my DMs are open! This job will be awesome... but don't just take it from me - our AI assistant makes a pretty compelling case for doing this job as well! https://t.co/cJyjYN8NhN

2023-01-11 17:22:41 @JosephJacks_ @AnthropicAI Challenges: - Differentiating this stuff can be somewhat subtle, so you need to walk tightrope between 'explain how the tech works in an intelligible way' and 'don't inaccurately portray the tech'. - Pace

2023-01-11 17:21:44 @JosephJacks_ @AnthropicAI Challenges: - AI is drawing a ton of broader attention so you're both communicating to researchers and customers, as well as communicating with society writ large. - AI systems have large-scale impacts which will have a range of downstream effects in the world.

2023-01-11 17:19:31 The JD is here: https://t.co/AM7u5cq2rD Given that we work on large-scale generative models, you'll also be able to do all kinds of novel comms experiments, given that our own technology is an increasingly useful writing assistant. Let's have fun!

2023-01-11 17:18:28 I'm hiring: Come work with me as a Director of Communications for @AnthropicAI . This is a senior role that would suit someone who both enjoys doing creative IC work as well as building and scaling teams. You'll get to do comms around frontier AI systems + can do fun experiments

2023-01-10 00:49:54 @PeterLoPR What do you think?

2023-01-10 00:44:40 @Stone_Tao I really love multi-agent sims, AND i love emergent behaviors.

2023-01-09 19:20:01 To put my cards on the table, I've followed BLOOM for years as it's an example of a collective trying to build a model to counter forces of centralization. But these kinds of initiatives are not going to be that successful if the models they produce don't have great capabilties.

2023-01-09 19:16:26 Does anyone have any links to places where BLOOM (the open source GPT3-esque model from @BigscienceW) is being used? I keep on fiddling around with it in various forms and it's not clear it has many useful capabilities. Perhaps there are multilingual uses? https://t.co/HV1san6RG5

2023-01-08 17:03:07 @SashaMTL @huggingface @AltImageBot1 testing @altimagebot1 https://t.co/6dt9nOcxSP

2023-01-08 17:01:08 @marky_red @AltImageBot1

2023-01-08 17:00:37 @TarekFatah @JustinTrudeau @AltImageBot1

2023-01-08 01:51:52 @aazadmmn yeah, I think stuff like mechanistic interpretability is really important here, or we end up with opaque machines of great power

2023-01-08 01:35:19 @deliprao Yeah I agree here. Think this breaks my weird food analogy. Maybe general rule is just 'simpler the better'. Eg in an ideal world wouldn't have to worry about networking would just have a single vast chip (I guess this is Cerebras hypothesis etc)

2023-01-08 01:28:22 @deliprao Great point! Some ideas here: https://t.co/2w2sCdxE5e So I think maybe it's like Cooking process (training) Ingredients (inputs) Tools (things you use to assemble ingredients and infra to support cooking). Stretching analogy to breaking point but maybe network is the oven?

2023-01-08 01:23:23 Also, embedding everything into same space required some very clever stuff. Maybe the trick is moving your complexity into the tools you use to assemble the proverbial dish, but the cooking process should be simple.

2023-01-08 01:22:26 Perhaps one reason why GATO from DeepMind displayed cool behavior was it exhibited a lot of simplicity - just embed everything from diff modalities into same space and do RL on top. https://t.co/zuqvyxjh0Y (simplicity here is a compliment! Simplicity hard to arrive at)

2023-01-08 01:12:31 @ArnoldBronley the most delicious form of paperclipping is when the universe is turned into a beautiful soup

2023-01-08 01:08:55 One reason why RL-dominant approaches still seem kind of unimpressive/lab-demos is that RL training is still (mostly) pretty complicated. There are some exceptions (e.g RHLF on top of LMs), but mostly a lot of RL-dominant stuff has really complicated multi-phase training schemes

2023-01-08 01:02:23 A lot of contemporary AI research amounts to coming up with the right combo of ingredients and training recipes to maximize the best chance of emergence. Systems with most interesting behaviors tend to have complex ingredients and simple training recipes.

2023-01-08 00:22:03 @peligrietzer What video games are you playing?

2023-01-05 22:58:18 @IreneSolaiman are you on Mastodon then?

2023-01-05 00:46:59 @paniterka_ch @rasbt The first key to avoiding building the torment nexus is to never mention the torment nexus

2023-01-04 19:26:03 @IreneSolaiman yeah, I've noticed that. I wonder if people should start more newsletters (I've found that to be a good way to cover multiple communities, but on the other hand is lots of work)

2023-01-04 18:30:20 @IreneSolaiman How so?

2023-01-02 20:12:34 @nearcyan yeah, I find myself having an instinctual heavily emotional reaction to it. I suspect it'll vibe with some people and not with others. Personally, I don't like it? But on the other hand, I'm not sure _why_ I don't like it. Will think more.

2023-01-02 20:08:06 @nearcyan a fruitful thought experiment might be - what type of person would actively prefer majority of their conversations with close friends are via AI?

2023-01-01 04:57:54 Happy New Year, Twitter. We live in the interesting timeline. https://t.co/sTEdSGb5KK

2022-12-31 00:49:06 @catehall Both of these short stories / smol novellas

2022-12-31 00:48:47 @catehall A colder war by Charles Stross Crystal night by Greg Egan

2022-12-31 00:35:45 @colinmegill @TheEconomist This interview predates chatGPT, but I was explicitly thinking about RLHF models and their impact. Not mad about some of the predictions I made here!

2022-12-30 22:08:14 RT @TheEconomist: “Foundation models are going to be the intermediary between you and computers.” In an episode first released earlier t…

2022-12-29 23:35:10 It's probably easier to compose a story or image if you are drawing on a bunch of sophisticated features which you only gesture/hint at in the final work. What latent universes will we explore as we decide the inner workings of these models? What 'imagination' shall we see?

2022-12-29 23:33:58 E.g, right now getting LLMs to explain why they generated stuff is pretty intractable, and x-ray tech like mechanistic interpretability hasn't really scaled. I figure people are going to latch onto worldbuilding as something that differs human/AI art. I'm betting AI will do this

2022-12-29 23:32:55 A lot of times when I write I think about lore in the universe (eg The Sentience Accords) that gets referenced off to the sides of main action. I'm betting that we'll eventually discover sufficiently large language models compose their own underlying lore

2022-12-29 22:59:01 @girl_hermes Not same album but summer in the city perhaps only song I've heard where repetition of 'cleavage' occurs and kind of works. We stan a millennial legend.

2022-12-29 22:49:12 @girl_hermes Still bangs

2022-12-27 06:14:31 Also flour - do not make this without flour!

2022-12-27 04:06:43 @dctanner Oh yeah 100%, that's how I had them growing up. We might have one kicking around, but I made this while panicking during last 30m of 5-dish combo so reached for the nearest suitable pan, lol. The quest continues...

2022-12-26 08:11:34 Bottom right is my first ever Yorkshire pudding! Have eaten a few in my time, but realized had the truly minimal ingredients (eggs/milk/fat/salt) that could attempt it - and it was tasty! Kitchens are great places to improv.

2022-12-26 07:15:51 Merry Christmas, Twitter! https://t.co/LAdtaJUTvy

2022-12-24 22:45:16 Spread some cheer this festive season by sending your graphic design friends photos of Analog magazine. https://t.co/s3RPFosUBy

2022-12-23 04:32:24 Stumbled across this amazing new song from Trent Reznor &

2022-12-20 00:03:38 https://t.co/fxF86ACYgb

2022-12-19 18:27:21 @PMA1070 no, it's more like "whatever goes in, the model tries to figure out what the 'goals' are of whatever is emitting the whatever, then the model tries to produce responses that satisfy the whatever that is emitting the whatever'. So it could be garbage or could be good or anything

2022-12-19 18:03:10 @defnotbeka It is entirely expected, but like many expected things I also find it strange to experience

2022-12-19 17:38:29 @jonasschuett @Manderljung @NoemiDreksler @tshevl @emmabluemke @Christophkw @araujonrenan @smvanarsdale @alfredoparrah @GovAI_ What's your best guess at how to get companies to implement these ideas?

2022-12-19 17:22:59 This gets more and more frightening the more you think about it https://t.co/hnoo5Wsn2z

2022-12-19 17:22:30 Basically, as we scale up language models, the LLMs try really hard to model whatever is emitting the tokens they're being asked to look at. In a very real sense, the more complex LLMs get, the more they 'look at you' and predict what might satisfy you.

2022-12-19 17:17:17 Pretty eery: AI models learn to reflect user views back at them (since I figure getting low loss rewards monitoring the _context_ of whatever emitted the input tokens). Pretty weird to see it in the wild. LLMs seek to reflect the views of people that talk to them. https://t.co/gD5qbAwRUb

2022-12-18 09:05:33 @Andercot What are the best examples of pessimism and optimism for this field?

2022-12-18 01:30:16 @rezendi The one blemish on my Oakland-maxxing approach is I can't get into cruise. @kvogt please help

2022-12-18 01:26:30 In the last decade: - figured out cut&

2022-12-17 02:07:36 @labenz @AnthropicAI It corresponds to crowdworkers preferring the helpful model about 55% of the time during non-harmful, open-ended conversations. Crowdworkers are setting the context/prompts here.

2022-12-16 21:58:18 @deepfates clippy 2.0 will come to us in the form of an unspeakably beautiful human being.

2022-12-16 21:57:56 @deepfates more like aligned with being an absolute smokeshow amirite

2022-12-16 21:54:59 @sergia_ch @AnthropicAI generally, quite bullish on idea of using models to proactively filter things, or other models themselves. basically, they've got good enough you can use them for these cases (though will probably need to red team uses in case models have weird blind spots)

2022-12-16 21:54:18 @sergia_ch @AnthropicAI hiya, great question. We think the answer is yes. It doesn’t directly address your question, but relatedly, in Figure 12 (buried in our appendix) you’ll see that models are getting pretty good at classifying types of harms. https://t.co/P3PQE5x3yz

2022-12-16 20:27:50 @deepfates I'm sure Stalin was a perfectly nice baby and probably seemingly benign for a long, long time, until one day...

2022-12-15 16:09:01 @deepfates AI models are these kind of freaky data transformation engines so we're living in this world where everything gets converted into alternative data distributions to then serve as inputs to generate weird outputs. synthesizer culture is BACK, baby

2022-12-15 16:08:24 @deepfates this nuts fr

2022-12-12 00:07:04 @hausman_k Really loved this paper/approach. Excited to read more

2022-12-10 23:58:08 @AndrewKemendo @Grady_Booch I don't typically do source-based reporting for import AI (as that might cross beams a bit too much with my professional life), but I imagine this is something @katyanna_q or @Melissahei or other AI journos might be interested in!

2022-12-10 14:59:08 @rama100 Thank you so much! I write it for everyone. So glad it's helpful

2022-12-10 08:11:18 One assumption about AI development is that centralized players will always be able to capture the frontier. GPT-JT suggests federated actors might be able to compete via resource agglomeration over crappy network connections. https://t.co/qOlZJGYlxZ https://t.co/OwAoSeLoSK

2022-12-10 06:32:25 @Stephen_Lynch Truly excellent missed joke, thank you!

2022-12-09 10:40:52 @nearcyan Incredible find... Adding to newsletter. Thanks for sharing

2022-12-08 15:30:38 @PetarV_93 @andreeadeac22 looks very interesting. quick q - what are the key diffs between v1 and v2 of paper? might write up for import ai and would find pointers (pun intended!) helpful

2022-12-08 13:15:57 RT @RhysLindmark: 1/ Live tweeting talk from @AnthropicAI @AmandaAskell at @southpkcommons. Moderated by the excellent @soniajoseph_! htt…

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

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

2022-11-15 18:48:39 @natolambert @pathak2206 Haha this is actually on my laptop screen RIGHT NOW. Also the new DM RL cooling paper. Exciting times!

2022-11-14 21:15:02 @foxjstephen @ESYudkowsky Thank you so much for reading! That's part of the idea behind these stories

2022-11-14 17:50:53 @ClementDelangue @mhdempsey More broadly, I'm extremely interested in BLOOM, and would be delighted to hear about any major deployments of it, so I can write about them.

2022-11-14 17:49:54 @ClementDelangue @mhdempsey I explicitly called out decent representation of languages in the above writeup. Once I have time I'm going to try and write a detailed case study of BLOOM, though has been hard to work out precisely who to talk to.

2022-11-14 16:05:27 RT @mhdempsey: Good read via @jackclarkSF on BLOOM which showed an early attempt at distributed development of large scale ML models that u…

2022-11-14 15:45:12 @GroffMRyan Thanks for reading, glad it is helpful!

2022-11-11 19:09:02 @billyez2 this is cool, congratulations! it was fun watching slices of your campaign on twitter, especially knocking on all the doors : )

2022-11-10 22:41:32 @madame_curtis

2022-11-08 06:32:59 RT @shubroski: @carperai hosted a call today sharing details on their upcoming instruction-tuned LLM. Takeaways:1. Focused on "pair progr…

2022-11-08 00:55:52 @KolmogorovGhost He was pretty bad at them, but he sure enjoyed playing them!

2022-11-06 19:53:26 @WilliamFitzger1 oh dude would love to know what you think. After you see it perhaps we can go get that mythical pint we've been discussing for years!

2022-11-06 19:15:19 @richardtomsett the 'it takes two to tango' exchange fully broke me. so good!

2022-11-06 19:03:01 Caught 'The Banshees of Inisherin' at the cinema yesterday

2022-11-06 06:08:54 @nc_znc Extremely useful lists. I used some of this for another secret slide I'll make public soon

2022-11-06 06:06:12 Another meaningful factor (as a few replies mentioned) is that if was trained on a multilingual dataset with a somewhat atypical mix, whereas glm-130 is predominantly english-chinese https://t.co/S8QR5Ik84d

2022-11-06 04:58:58 RT @I_are: Published a couple weeks ago - Google researchers show how to use a language model to improve its own reasoning. I'm skeptical o…

2022-11-05 22:01:02 @zhansheng @EricHallahan hahaha didn't realize that, but maybe that's another case of one-upping on params.

2022-11-05 21:45:47 @I_are oh, I'm pretty hyped up about this one. One of rare occasions I used an !!! in my newsletter. https://t.co/JDQndwdSF6

2022-11-05 21:03:17 @EricHallahan haha on the latter, reminds me of when OpenAI did GPT2 (1.5b), salesforce loudly released a 1.6 billion parameter model (CTRL). history doesn't repeat but it sure loves to rhyme

2022-11-05 20:41:26 @DynamicWebPaige @GoogleAI @BlenderDev ooh this is cool and hadn't seen it, thanks for sharing

2022-11-05 20:10:41 @EricHallahan that's helpful, yeah. It seems like the datamix is important.

2022-11-05 20:10:05 @EricHallahan This is a helpful clarification. I feel like some of the constraints on BigScience came from data - is that right? On b) - is that documented anywhere? It seems quite odd to train a model and not try to push it to be performant (if I'm understanding you correctly)

2022-11-05 19:30:40 One other potential reason is this (as you can tell, I've been somewhat obsessed with BLOOM and why it's not great for a while) https://t.co/wW41P8bWaj

2022-11-05 19:25:43 Has anyone @huggingface @BigscienceW done a comparative analysis of BLOOM and other models (e.g, OPT, GLM, GPT3) and evaluated where the perf differences come from? Would be pretty interesting. Also feels important given BLOOM is a potential template for future group projects

2022-11-05 19:23:52 I thought a diff could be data, but doesn't seem like it - BLOOM was trained on 350 billion tokens and GLM-130B on 400 billion tokens (more tokens = better). Not a substantial enough gulf to solely explain the perf differences

2022-11-05 19:17:34 Feels kind of meaningful that an academic group at Tsinghua University (GLM-130B) made a substantially better model than a giant multi-hundred person development project (BLOOM).

2022-11-05 19:17:04 If you want a visceral sense of how different development practices and strategies can lead to radically different performance, compare and contrast performance of the BLOOM and GLM-130B LLMs. https://t.co/qPdrN6uL8whttps://t.co/9QJ71dlwuZ

2022-11-05 18:45:42 @KTmBoyle @a_d_matos @PalmerLuckey yeah, I didn't mean to imply he wasn't focused, I was noting along with the focus he also seems to allocate time to having fun, and has been doing the cosplay stuff for years (which I think is cool!).

2022-11-05 18:40:16 @KTmBoyle @a_d_matos gotcha, so I suppose your key point is successful people had some period where they were massively boring/focused in earlier part of their life. I think this somewhat holds, but I also think there are counterexamples. Thanks for clarifying!

2022-11-05 18:34:16 @KTmBoyle @a_d_matos There are also contemporary examples - I think @PalmerLuckey has done some impressive things, and he also loves doing cosplay and seems to make time for nerdy fun stuff, along with the work of Anduril etc : )

2022-11-05 18:32:47 @KTmBoyle @a_d_matos I think you're absolutely right that some terrifically successful people are boring, but I also think some terrifically successful people lead very diverse lives and have a broad range of hobbies and interests and have a lot of 'fun' as well. Worth noting both paths work

2022-11-05 18:32:10 @KTmBoyle @a_d_matos More broadly, a colleague at work yesterday was pointing out they knew a bunch of nobel laureates, and something they had in common was being quite well rounded and having hobbies in the arts, along with their core (typically STEM) endeavors. Not much routine, either

2022-11-05 18:30:47 @KTmBoyle @a_d_matos Jack Parsons was a rocket genius and also a lunatic occultist who threw amazing parties

2022-11-05 18:30:10 @KTmBoyle @a_d_matos Ada Lovelace basically invented computers while going to high society balls and generally being as deranged as you'd expect a relative of Byron to be

2022-11-05 18:28:12 @KTmBoyle @a_d_matos Richard Feynman loved playing the bongos and was generally someone who lived with great enthusiasm

2022-11-05 15:44:03 @irinarish @natolambert Working on it!

2022-11-05 02:54:11 @AmandaAskell Amanda blog++ please

2022-11-05 02:52:42 @cartoon_magoo @typedfemale Yeats. Gd autocorrect.

2022-11-05 02:52:23 @cartoon_magoo @typedfemale Hell yeah. If you're into Years, this is a banger: https://t.co/ZERaQ7E2Tl

2022-11-05 02:10:12 @typedfemale Kind of corny but I memorized a bunch of Shakespeare and W. B. Yeats and run them in my head when walking or cycling some times. Really satisfying and illuminating

2022-11-04 20:59:51 @ShaanVP office and pub

2022-11-04 18:23:12 @valmianski yeah, I generally talk about that stuff verbally, but I make these slides as I find using visceral and captivating images is a useful way to get progress across

2022-11-04 18:09:35 @nathanbenaich @stateofaireport another satisfied subscriber

2022-11-04 18:07:07 @nathanbenaich @stateofaireport still reading it actually! probably writing up for this issue

2022-11-04 18:06:53 @natolambert savage, haha. (Does feel like RL hasn't progressed as rapidly recently as LLMs have kind of dragged a lot of attention - fuzzy intuition)

2022-11-04 18:05:29 @nathanbenaich that one was insanely cool actually, might use it

2022-11-04 17:48:45 Trying to make an update on my 'RL progress' slide for some upcoming talks. Anyone got anything from 2021 or 2022 they think is pretty striking? https://t.co/RxuvFJXlpr

2022-11-04 17:27:33 Shoutout to @ruchowdh and the rest of the team for doing some of the most interesting applied work on AI ethics. I really enjoyed their investigation of the Twitter cropping algo https://t.co/YjJaezPrvQ

2022-11-03 17:35:42 Moral Crimes of the Near Future... from Import AI 308. https://t.co/JDQndvVJqY https://t.co/1KpZtze6Zt

2022-11-01 15:04:32 RT @CSETGeorgetown: In their CSET brief, @jackclarkSF, @KyauMill21 and Rebecca Gelles show how bibliometric tools — such as CSET's Map of S…

2022-10-31 21:38:53 @DynamicWebPaige @GoogleAI thanks so much for reading! archival link to the issue here https://t.co/JDQndvVJqY

2022-10-31 21:38:34 RT @DynamicWebPaige: Via @JackClarkSF's newsletter today:"@GoogleAI used a large language model to generate chain-of-thought prompts fo…

2022-10-30 22:10:26 @Miles_Brundage he likes making things and also does cool video projections (based on stop motion animation he has made about the things he make, some of which include things I make), so he periodically decides to do insane things like 'make a 6ft mask and project videos into it'

2022-10-30 19:45:49 @mmitchell_ai @TheZachMueller I am definitely going to start using 'stirred the batter'. A+

2022-10-30 19:30:55 @mmitchell_ai *though.

2022-10-30 19:30:37 @mmitchell_ai Yes! Thought might be more an English saying than an American one.

2022-10-30 09:30:58 It succeeded https://t.co/cbRsDYrDn4

2022-10-30 04:45:04 My friend: yeah I totally shit the bed on that prop I made for the Halloween party. Did it in a couple of days. It's no good.The prop: https://t.co/zVsuYnWfqO

2022-10-29 19:09:59 @rvinshit @porksmith Absolutely incredible. A+

2022-10-29 18:32:38 @xriskology Some people shut down because they find engaging to be traumatic

2022-10-28 16:22:41 @jjspicer read this at once https://t.co/6M8qWJ2Akl

2022-10-27 21:30:29 RT @ClementDelangue: We just crossed 1,000,000 downloads of Stable Diffusion on the @huggingface hub! Congrats to Robin Rombach, Patrick Es…

2022-10-27 18:42:19 @tszzl I played Half Life: Alyx on the Valve Index and it hard-updated me to VR games being amazing, albeit still a little early.

2022-10-27 00:02:01 RT @rgblong: Another question for consciousness scientists and AI people: What is the best evidence for and against large language models…

2022-10-26 17:20:39 @MishaLaskin @junh_oh @RichiesOkTweets @djstrouse @Zergylord @filangelos @Maxime_Gazeau @him_sahni @VladMnih looks cool - link to the paper?

2022-10-26 03:29:56 @gijigae Thanks so much for reading!

2022-10-26 02:05:38 @kipperrii 100% here for fontposting. More fontposting!

2022-10-25 20:15:15 @carperai Haha fair enough. Thanks!

2022-10-25 19:25:12 @carperai You're welcome. Any clues as to release timeline?

2022-10-25 18:49:10 Stages of living in California:First earthquake: dear God hope I've made my Will.Fifth earthquake: Gosh, that was a little rumbly. Hope everyone is ok.???th earthquake (which happened in SF just now): Wonder what gifs people will post on Twitter about this one?

2022-10-25 17:14:50 RT @ElineCMC: .@jackclarkSF's take on this: "The tl

2022-10-25 15:30:28 @EricNewcomer @TaylorLorenz awesome metrics, congrats Eric! Really nice to see your success here : )

2022-10-25 01:52:36 @Meaningness @ArtirKel it's got a bunch of nice examples in it - definitely worth a skim

2022-10-25 01:45:29 One of the nice things about the AI space is when people take their criticisms and instantiate them as quantitative studies of existing systems - kudos to @GaryMarcus et al for a nice paper going over some failures in DALL-E2

2022-10-24 23:21:03 @tszzl Did it with my Senate testimony and was v surprised. Generally find myself running any long doc I write through a LM these days

2022-10-24 05:37:13 Import AI will be coming out on Tuesday as I sprained my ankle at Nopes' final show last night. #VOTEDIY2022 https://t.co/CaKrgLf00r

2022-10-24 01:47:22 RT @SamuelAlbanie: Just how striking are the recent language model results with Flan-PaLM?Here's a plot.Across 57 tasks on mathematics,…

2022-10-23 19:02:59 @CecilYongo Can you share slides? Looks interesting

2022-10-23 02:23:43 @zacharynado You can't necessarily look at the datasets with an API, unless has huge amounts of extra eng time invested. E.g can anyone inspect JFT?

2022-10-23 01:41:19 @kylebrussell Does Playbyte have an in-house exorcist team?

2022-10-23 00:05:13 Image generation going to get strange when most of the images we train on are synthetically generated. Feels like a classic 'tragedy of the commons'. (Yes, some have watermarks, but my sense is there's a race to the bottom on that kind of thing). https://t.co/A3QawDog90

2022-10-20 22:02:20 RT @trishume: I’m hiring for a Resident to work with me for 6 months at https://t.co/G5wR9mJCwE on researching how to reduce or untangle su…

2022-10-19 17:50:18 RT @baxterkb: Starting now @NIST #AI Risk Mgt Framework (#airmf) panel "How to Measure AI Risk across the AI Lifecycle" w/ @jeanna_matthews…

2022-10-19 01:20:05 @powerbottomdad1 @shaig Hell yeah

2022-10-19 01:15:22 @Plinz Who actually thinks this? Genuine question. No such thing as a singularly good person from POV of the world.

2022-10-18 20:10:33 RT @nazneenrajani: Looking forward to speaking at the @NIST panel on AI Risk Measurement representing @huggingface on Wednesday at 10.30 a…

2022-10-18 17:35:27 This tweet definitely inspired by a chat with @nearcyan yesterday.

2022-10-18 14:34:34 @Ben_Reinhardt Yes, I myself feel like Reality is important. I think it's easier to care about Reality insofar as your life is comfortable, though. So I suspect Reality is less and less alluring for people. Another turn of the crank from online gaming, etc.

2022-10-18 14:30:13 @Ben_Reinhardt Same, though I think it is a difficult vibe to counter. Can't tell if this means I am becoming hopelessly old fashioned, or I'm worried about something which is actually worrying.

2022-10-18 14:20:19 Consensual Wireheading feels like a vibe for the next few years. Everyone running into Reality Collapse willfully as it is delightful and diverting during a chaotic time.

2022-10-18 04:21:11 @lathropa No idea, you might want to ping them and ask

2022-10-18 04:12:48 @michael_nielsen @togelius Good link for the Ted Taylor stuff? I think about this a lot

2022-10-18 03:43:20 @togelius Do you think AI is exactly the same as other media technologies, or does it have potential for harm/safety issues that justifies control. (Personally, I think RL agents with big generative brains might be dangerous. Way more confused re things like this, eg image generation).

2022-10-18 03:32:09 RT @jackclarkSF: I feel genuine confusion about what Stability represents. It's culture demanding less control over AI. But there are extre…

2022-10-18 03:27:30 I feel genuine confusion about what Stability represents. It's culture demanding less control over AI. But there are extremely good arguments for controlling AI to reduce misuse and downside risk! Reminds me of early days of Eleuther - not a coincidence bunch of them work there

2022-10-18 02:42:45 @ch3njus They raised $101 million, so I imagine that helps

2022-10-18 02:37:22 Stability is halving its clip-guided image generation prices.

2022-10-18 02:35:23 @memotv 5X to 10X, lol. Idk I'd believe it when I saw it, but that'd be a pretty big computer.

2022-10-18 02:24:32 Stability will fund 100 PHDs this year who will all get 'fat amounts of supercomputing'.Says Stability 103 people now.Stability cluster is 4000 A100s on AWS and wants to grow 5X-10X by next year (this is... A big computer, lol).

2022-10-18 02:23:08 Stability is planning to work on national research clouds and also train thousands of people to train big models. This is a divisive idea - some will see it as democratization, others will fret re safety impact of increasing big models globally while safety an empty chalkboard.

2022-10-18 02:11:48 The general tone of this event is all about control versus distributed collectives. "Does it make sense the most powerful technology in the world is owned and controlled by the few?" - @EMostaque . I think of this as 'culture eats strategy for breakfast'. https://t.co/pb6zZkJmUn https://t.co/oBYvBdNXlu

2022-10-18 01:47:58 A fun split-brain moment is sitting at the back of the @StableDiffusion launch party in SF, while writing up the 'GitHub Copilot investigation' potential legal case for Import AI (https://t.co/TBQoij2o7v), after having written about the $101m raise for Stability.

2022-10-17 14:57:11 RT @mlfavaro: Personal news: After >

2022-10-17 03:23:59 Very good read on the recent CHIPLOMACY export control actions of USG. I especially like the ideas at the end about potential gaps in the policies and/or enforcement mechanisms. https://t.co/lONWjapdKP

2022-10-17 00:18:57 @Tech_Journalism That's awesome. Btw I used to obsessively read your data center site(s) and they were an inspiration for my data center coverage : ). Thanks for all your efforts there

2022-10-16 23:56:30 @Tech_Journalism That's an awesome stat. Wonder if that makes it one of the densest places from a people/computer perspective.

2022-10-07 23:31:34 @EMostaque Had to shoot my shot lol. Solid gif game

2022-10-07 23:31:07 @JosephJacks_ @jposhaughnessy @EMostaque Ah, yes, the chungus scale.

2022-10-07 22:53:50 @EMostaque how big is the computer going to be

2022-10-07 17:27:56 @JMateosGarcia @DeepMind @nesta_uk that's great - have fun. Which team will you be on?

2022-10-06 15:03:04 @deliprao @kevinroose @jekbradbury i find these clarifications very helpful, thanks both!

2022-10-06 01:36:01 @kevinroose The fun part is matrix multiplication is key to training neural nets, so this breakthrough translates into generically better training for a vast range of things. Just a casual 10-20% speedup on something humans have been trying to further optimize for 50 years. Probably nothing.

2022-10-06 00:41:52 @baxterkb @ayirpelle @NIST @salesforce This is wonderful news! We need more people from industry helping NIST think about measurement of AI. Thanks for your service here

2022-10-06 00:16:20 @michelteivel Right on schedule

2022-10-05 22:19:56 RT @DAIRInstitute: We are hiring a senior community-based researcher. Full job ad and application here: https://t.co/9KOCe1ps2p

2022-10-04 23:00:20 @nearcyan Thanks so much for reading and finding worthwhile to share. I write the newsletter to help me think and help others think, so I really love these kinds of callouts and find them very motivating. Have a great day!

2022-10-04 19:26:01 @APreciousPony Give it a year and yes (we've got controllable videos and audio now up to certain time horizons, but smashing the modalities into one model still quite expensive)

2022-10-04 19:06:47 All this generative AI stuff is wonderful and fascinating and is also gonna further the funhouse hall-of-mirrors splintering of our culture into radical, tiny communities. https://t.co/iSS8ZA1ZJI https://t.co/4h5vEIeLef

2022-10-03 16:49:44 @yudhanjaya thanks so much for reading!

2022-10-03 16:49:37 RT @yudhanjaya: If you feel that the AI space is moving too fast for you to keep up, I highly recommend subscribing to @jackclarkSF's Impor…

2022-10-03 15:16:46 RT @rezendi: Wow, @jackclarkSF's latest ImportAI really hammers home just how fast this golden age of AI is moving now: Make-A-Video! Whisp…

2022-10-03 02:24:46 RT @chrisalbon: Prediction: The ability of an individual to generate millions or billions of plausible images, videos, and posts using AI i…

2022-10-02 20:26:03 RT @yacineMTB: my advice to anyone who wants to learn to code because they feel like their white collar field isn't going to get them to wh…

2022-10-02 02:27:47 RT @jackclarkSF: My (much longer) written testimony is available here. I also generated a bit of text from an LM from the end of that and f…

2022-10-01 18:47:14 @saffronhuang @indexingai Thanks for generating those! 2022 was first time we had a dedicated ethics chapter (@mathemakitten did a ton of work on it) and these data points were a crucial input.

2022-10-01 18:15:12 @mikarv (per dms, deleted main tweet as this is gross. thanks for the flag!)

2022-10-01 04:40:16 @AnikVJoshi thank you!

2022-10-01 03:52:38 RT @commercedems: Science powers our future The Committee took a deep dive with experts in artificial intelligence, quantum science and di…

2022-09-30 22:29:09 @snarky_android @commercedems @SenatorCantwell @SenatorHick @uwengineering @AnthropicAI @HPC2MSU @ColdQuanta @usmcsce @UW @uwnews Same here, a true privilege! I also enjoyed getting to meet my fellow panelists (Bob - I scribbled a bunch of notes relating to some of your quantum stuff and am now going down that rabbit hole!)

2022-09-30 19:45:26 RT @ddimolfetta: Coming off yesterday's @commercedems emerging tech hearing, witnesses say more investments and support for research and ed…

2022-09-30 19:45:16 @ddimolfetta @SenToddYoung thanks for the coverage! I love testbeds!!!!

2022-09-30 19:07:48 RT @jposhaughnessy: 1/Announcing Some Professional NewsI'm delighted to announce that O'Shaughnessy Ventures LLC invested in @Stability…

2022-09-30 18:46:54 @joeteicher @mattyglesias @StefanFSchubert I think it's probably holding back useful discussions that can be had with a much broader set of people, and may also be slightly hindering policy impact. Fuzzy intuitions though

2022-09-30 18:17:20 @samim @commercedems @AnthropicAI I'm well. We live in exciting times in a beautiful world. Same to you friend!

2022-09-30 17:57:49 @samim @commercedems @AnthropicAI No, not in the slightest. I was pretty interested that I was having writer's block with the conclusion and then used this, which I felt captured the rest of my testimony I'd been working on quite well.

2022-09-30 16:17:17 @robinc @ClementDelangue @commercedems @AnthropicAI Yeah, that sounds correct. I clarified in a later tweet. Grateful for all the feedback here! The liability/perjury lens is v useful

2022-09-30 15:32:34 @mattyglesias @StefanFSchubert AI people, myself included, need to get a lot better at communicating. I think a ton of the frames that get used are pretty alienating or can seem downright crazy. (other hits include AGI via mad science, etc). Lots to do!

2022-09-30 15:23:09 Senator, we sell blocks. https://t.co/CZ8PgwnnIp

2022-09-30 13:38:39 @teemu_roos @ClementDelangue @commercedems @AnthropicAI That's great feedback, thank you! I appreciate it

2022-09-30 09:41:10 @dazzagreenwood it was a chicken pot pie which actually looked rather unfortunately like a pastry brain floating in goo. did the job, though!

2022-09-30 02:06:07 @yuvalmarton eh, I come from journalism where you attribute everyone who touched the story - the reporters, the editor(s), the sub-editor, etc. Credit should always be shared. But I also think this is just a wildly confusing area, and it's unclear what norms should be.

2022-09-30 01:46:03 Brain still absolutely fried from Senate testimony, but decent chance Import AI coming out this week. A special, never-before-seen behind the scenes look here. : ) https://t.co/Zg3qN9ouri

2022-09-29 23:18:15 @dave_maclean @commercedems @AnthropicAI Thank you so much, I love to write them. Been trying to compile the collection for years, maybe in 2023.

2022-09-29 22:06:24 @Timcdlucas @commercedems @AnthropicAI One of my curses is when I'm incredibly nervous I either start twitching or smiling. I've mostly trained myself out of twitching so now I just smirk under pressure. Not massively ideal

2022-09-29 21:57:02 Since it's doing numbers, might as well be as precise as possible - believe this marks the first time a paragraph output from a language model has been used as part of a human testimony in the U.S. Senate. I shall now go and play billiards. We live in terrifically exciting times! https://t.co/VvvO7lCL5d

2022-09-29 21:54:55 @dinabass @ClementDelangue @commercedems @AnthropicAI https://t.co/XiccZaVjgy

2022-09-27 16:39:10 RT @commercedems: THIS WEEK in the Senate Commerce, Science, and Transportation Committee: Wednesday: Subcommittee Hearing with @Sena…

2022-09-27 16:22:09 @jachiam0 yes

2022-09-27 02:26:31 @primalpoly thanks! so one thing I was wondering is if I should just... port all of these into a single blog post and put some headers around for themes, then can post that to a forum or something. WDYT?

2022-09-26 13:49:39 Import AI isn't coming out this week as I figured I should prioritize the US Senate above my newsletter. Hope readers forgive me! :)

2022-09-26 13:48:13 This is a huge honor and, per how I approach much of my policy work, I am keen to hear ideas from the community. I am feverishly writing testimony today, so if you have good ideas, please DM them to me.

2022-09-26 13:47:24 I will be testifying in the United States Senate this Thursday about AI R&

2022-09-26 02:14:30 @EnkrateAI thank you!

2022-09-26 01:45:43 Import AI will come out next week as I've been dashing about DC all week (and this weekend at a conference) so haven't had time to write. Hopefully next issue is a banger as tons of cool papers, including using GPT3 to simulate people for polling purposes. : )

2022-09-23 17:29:13 @stanislavfort @ftxfuturefund @GaryMarcus +1!

2022-09-23 17:25:54 @alexkozak I'm glad they're running the experiment - will help us figure out if this stuff is a waste of time. But take your point, also.

2022-09-23 17:04:22 Excellent approach to funding! 'We think X is important and we'll pay you tons of money if you persuade us X is bullshit!' Good chance for people skeptical of these ideas to try and shift a load of capital in a different direction. https://t.co/DmrWxDF1nj

2022-09-23 16:53:00 @xriskology @rgblong (i get sent a lot of stuff and don't circulate, but since you don't know me super well, I just am adopting the strategy of 'trust but verify'. hope helpful!)

2022-09-23 16:52:22 @xriskology @rgblong sounds good and completely understand - maybe send me something low-stakes first and see that I don't circulate it and play the iterative trust game. (This is the same way I operate with sharing stuff with people, and it's how I personally figure out trust stuff).

2022-09-23 15:52:41 @xriskology @rgblong hi there! so, as I'm sure you are, I tend to be 'snowed in' with work these days, but if you DM me stuff I'll try and give some feedback ahead of time if I'm able, though if this introduces overhead for you, feel free to skip. thanks!

2022-09-23 14:49:11 @xriskology @rgblong I see. Consider this as polite feedback that I think your arguments are more persuasive the more you do holistic analysis of the whole paper, and if you ID specific words as indicative of intent it makes it harder to see the argument.

2022-09-23 14:46:01 @xriskology @rgblong I've also said stuff like 'due to the delightful traits of markets to incentivize a race-to-the-bottom on safety, this $scary_thing will get cheaper'. Again, I'm not literally 'delighted' by it, I'm most just using colloquial language in conversation to lighten a heavy topic

2022-09-23 14:43:11 @xriskology @rgblong this feels like a bit of a reach - I myself say stuff like 'thanks to economies of scale, this scary thing is gonna arrive pretty quickly'. I'm not literally glad about it, I'm just noting that markets are a thing and they make stuff cheaper, even stuff I'm afraid about.

2022-09-21 13:57:10 @WriteArthur @EMostaque @OpenAI is there any particular published rationale for the Ukraine stuff? It's hard for me to say anything sensible without further details, and I mostly avoid doing source-reporting for Import AI (as it'd kind of obviously generate some static, so tend not to)

2022-09-21 12:52:48 @kyliebytes still shocked that they brought out the President of Haiti to profusely thank Benioff for what Salesforce had done for the country. wild stuff

2022-09-21 12:51:48 @kyliebytes haha, that takes me backhttps://t.co/t1dSk1m2u7Salesforce is 100% the most weird large-scale cult-like tech thing out there. Still going strong almost a decade after this piece! really crazy

2022-09-20 17:11:35 @ethanCaballero @andy_l_jones did a good thing here https://t.co/C2tSD3mz4V

2022-09-20 11:05:05 @andrewrens_ria i mean the most prominent thing gets the most attention, usually of a negative kind

2022-09-20 01:49:31 @exteriorpower @joanfihu re the mode of release, with way fewer restrictions applied. That part isn't the same. But I think a lot of the grief SD is getting is about some of the meta-issues of generative models and relationship to economy/employment, rather than its idiosyncratic OSS-like release 2/2

2022-09-20 01:48:47 @exteriorpower @joanfihu (the above could seem like weird semantics, so I should cluster - SD is 'doing the same thing' as all the gen model factories in terms of capability dev, and 'doing the same thing' as OAI/Midjourney/some other fringe players re 'making broadly available', but is doing diff 1/2

2022-09-20 01:44:13 @exteriorpower @joanfihu don't mean any shade by this to any of the parties involved - it's a big space and I think it's valuable lots of experiments are being tried and generally welcome a broad and open debate

2022-09-20 01:43:35 @exteriorpower @joanfihu by doing the same thing I mean 'making generative models broadly available'. As I've said in a bunch of ways, I'm generally confused about this whole area - it's very unclear what things are appropriate safeguards versus PR/policy/product reactions

2022-09-20 01:42:18 @vlordier oh yeah, I agree - SD represents a different approach to the proprietary/closed things. I'm pretty confused about all of this, but glad SD is contributing to the broader debate. My point is it's a shame lots discussion centers on SD rather than broader issues it highlights

2022-09-19 22:55:28 @mm_jj_nn Well, I think this is still the case - StableDiffusion got funded by @EMostaque who used hedge fund money to foot the upfront bill then subsequently monetize. Was trained on hundreds of GPUs as part of a 4k A100 cluster being built out. Seems v hard for academia to do that still

2022-09-19 22:46:10 @joanfihu eh, I just mean it's super not fun to have tons of people shouting at you online and calling you evil or someone who wants to make people unemployed, when you're doing the same thing as a bunch of less prominent actors

2022-09-19 22:34:13 By 'tragic', I mean that a lot of the debate naturally centers on StableDiffusion when discussing these problems, but really the problems are bound up in the whole space of generative art.

2022-09-19 22:33:34 One of the tragic things about the controversy about #stablediffusion is that most of these problems were present in DALL-E, Imagen, etc, but they were a lot less public (either not released, or behind a firewall). A nice example of 'tall poppy syndrome' in action.

2022-09-19 19:45:37 @Lan_Dao_ excellent tweets. I also felt this way about dubai. Oddly sterile. Like a potemkin 'fun' city

2022-09-19 16:06:56 @WriteArthur thanks for reading! I feel legitimate confusion about this safety/censorship stuff so trying to write more in public about it : )

2022-09-19 16:06:36 RT @WriteArthur: "Though some call this censorship, it's worth bearing in mind the Chinese government probably views this as a safety inter…

2022-09-18 14:16:28 @JeffLadish For hotels, I think https://t.co/8A99j7FlT9 is really nice - decent geographic view, lots of granular detail.

2022-09-18 13:16:10 @SamoBurja will do! currently living out of a hotel in DC for a few days, or I'd be there

2022-09-18 04:13:53 @SamoBurja hell yeah.

2022-09-17 22:07:42 @MatthewJBar Did some thinking here for how you could have a dynamic private market of third-party auditors overseen by government (with actual ability to impose severe penalties on auditors that get captured by clients) https://t.co/UZMcx5T5wc

2022-09-17 22:04:03 @gwern @Skiminok @simonw @vlordier I'm most excited about doing 'expert red teaming'. This kind of proved it works, but the domains aren't necessarily the important longterm ones. Longterm stuff is red teaming for chemical synthesis, bombmaking, other strange capabilities that come from synthesis

2022-09-17 21:54:53 @vlordier @Skiminok @gwern @simonw filters are just external plumbing, I am skeptical they can actually deal with an engine that wants to eat you

2022-09-17 21:47:47 @simonw the answers will not delight you https://t.co/8SEubKtUws

2022-09-17 21:45:44 @andy_l_jones probably people like Tyler Cowen and others

2022-09-17 21:41:16 @simonw (btw v enjoyed your blog post and writing a short thing on it for Import AI)

2022-09-17 19:45:06 @AlexGodofsky (note, I'm basically confused about this issue. Some controls of some systems are probably necessary, but most controls come off as paternalistic/asinine/PR-motivated, so I think it's reasonable to counterreact). Trying to write some stuff. Making memes till I figure out words

2022-09-17 19:23:31 @vlordier @gwern @Skiminok and sorry for being snippy that was a bit low-class of me, thanks for your thoughtful response!

2022-09-17 19:23:11 @vlordier @gwern @Skiminok ah, I see. Well, just because I'm somewhat used to these things, doesn't mean I can't find them scary. The fact these attacks are possible is scary. Scary things don't have to be surprising to be scary, they can be totally expected and scary nonetheless!

2022-09-17 19:22:00 @AlexGodofsky https://t.co/QWt5fIt6lh

2022-09-17 19:16:25 @vlordier @gwern @Skiminok Tried to write some more about these issues here https://t.co/6R2Qa4aYWT

2022-09-17 19:15:57 @vlordier @gwern @Skiminok I'm somewhat familiar with GPT3 (see author list)

2022-09-17 13:24:31 @EMostaque Katja Grace does really good work involving surveying researchers in the field for their thoughts on AI progress, and generally does useful thinking on impacts of AI. They may have adequate funding, though.

2022-09-17 13:15:03 @gwern @Skiminok this stuff is way scarier because, per your comments, it's a class of semantic attack, and because of how these things work, there's gonna be a ton of them we haven't yet discovered which will work - sometimes more effectively than those we have today

2022-09-17 13:14:22 @gwern @Skiminok thank you for saying this! I covered Tay at the time and it was a hardcoded echo function that let you load stuff in manually. Everyone reported on it like a learned thing but it was basically just a leftover old shiv into the system they hadn't debugged

2022-09-17 02:35:47 @juan_cambeiro way better for this to happen now than ten years from now. You'll be fine, though it'll absolutely suck for a while.

2022-09-16 22:48:31 @jiayuanloke @AnthropicAI @scsp_ai haha thanks, I have a hairdresser and he only ever sees me approx 2 days before I have to go to DC :)

2022-09-16 20:07:25 Had a terrific time presenting an @anthropicai model at @scsp_ai in DC today. Small amounts of terror, lots of laughs, and successfully broke the model on stage to help people see the hard edges of this stuff. More to come! I also do Weddings. Recording online soon. https://t.co/Qf5wc382I9

2022-09-16 18:33:41 @meerihaataja @scsp_ai yeah, I found the questions v interesting. Was glad someone appreciated me breaking it!

2022-09-12 20:04:30 RT @carperai: Ethics, it’s important to us! That is why we want to hear from the software engineering and beyond community for our latest p…

2022-09-12 05:38:35 also, immediately after we felt it my partner and i reported our data to https://t.co/ZqxlwCZ3SN. Consider doing so yourself!

2022-09-12 05:35:37 A smol earthquake in oakland about five mins ago

2022-09-11 18:05:13 This will be the year I make a coherent graph showing all YOLO variants and their perf improvements over time. https://t.co/K48xIKBlvv

2022-09-11 18:02:00 Writing about yet another YOLO variant for ImportAI this week. There are now multiple YOLOv6s, a YOLOv7 came out a few months ago, and there are multiple groups developing YOLO variants in parallel. We're entering the YOLO Multiverse here, folks. https://t.co/u6AutReIvg

2022-09-11 00:55:28 @kevinroose The toiletularity

2022-09-11 00:13:27 Anyway, clearly all of this talk about machines rapidly evolving and eventually outpacing humans in a bunch of domains is guff.

2022-09-11 00:12:51 2015 versus 2022. Incredible. 2015 via this @rsalakhu paper https://t.co/zTb2T2Ru5X 2022 via https://t.co/VkskqYVIMN #StableDiffusion https://t.co/Eprs982JzP

2022-09-11 00:02:38 @jmshoffstall haha hell yeah! it's a solar panel though so probably an expensive thing to break in the pit

2022-09-10 22:50:48 Wrote the story next to a solarpunk soundsystem, which felt appropriate :^) https://t.co/jmKgcevHab

2022-09-10 22:48:28 Gonna write some formal things and also figure out some policy stuff which I'll try to be legible and public about. For now, expect more weird stories and tweet threads as I grapple with this stuff. This information shouldn't be stove-piped - it should be broadly available.

2022-09-10 22:47:41 Wrote a somewhat spicy fictional story about AI labs, AI development, model theft, anarchy, takeoffs, and so on. Coming out in this week's issue of ImportAI. I think the political economy around AI development is completely busted so I'm trying to think about it more.

2022-09-10 19:20:43 @jachiam0 yeah I mean that's generally been my view (and relates to some of the stuff at @AnthropicAI ) - can you make models intrinsically safe rather than spend your time building external plumbing. Unfortunately doesn't deal with larger problem of 'who decides safety criteria', tbd

2022-09-10 19:17:34 @jachiam0 there's also some incentive for 'race to the bottom on safety' as a response to this, if you can gain market share via OSS'ing a model then rapidly building some service/inference layers on top which you commercialize (which is what StableDiffusion seems to be doing)

2022-09-10 19:16:52 @jachiam0 so a few companies are gonna try and make various control layers for these models to deal with the numerous issues, but many of these controls (e.g, somewhat arbitrary filters) piss enough people off they inspire a kind of libertarian counterreaction

2022-09-10 19:16:15 @jachiam0 Yeah that makes sense. I'm working on an essay about this confluence of issues, but my general take is that most attempts at control inspire counterreactions of same order of magnitude as perception of control

2022-09-10 19:12:20 @jachiam0 Do any of these questions go away if you sell it via a relatively controlled API? Not trolling. I more mean it seems like these issues are basically fundamental to the generative art tech, moreso than the open source release aspect (exception - certain types of abusive image)

2022-09-09 16:51:23 @MichaelTrazzi @Manderljung @GovAI_ maybe we should discuss my takes sometime haha

2022-09-08 21:52:40 @ArkadyMartine TallulahAmara AdaliaDahlia

2022-09-08 15:58:04 @ethanCaballero @DavidSKrueger possibly BERT, but depends on how you're defining 'release' here. I think if you're counting (what LM and variants) then this works

2022-09-06 23:30:18 @mer__edith @moxie @signalapp Is not covered in the announcement, but I imagine this means you're no longer a formal advisor to FTC?

2022-09-06 14:10:10 @halhod @rodolfor Pretty sure I predate this - will check the archives but think mentioned 2020

2022-09-05 22:10:47 Some friends, commenting on the heatwave today:Friend A: The sun's been gaslighting us.Friend B: Well, it is the original gaslight.

2022-09-05 02:29:11 Peter had very different politics to this, but he embraced the role with his typical combination of enthusiasm, mischief, and effort.

2022-09-05 02:27:11 Final anecdote - back in 2017, @gdb was due to testify in Congress. Peter came by the office to help us do a 'slaughterline' - where you pretend to be elected officials and ask hard questions. V funny to see Peter channel hard-right ideologies while asking questions.

2022-09-05 02:16:42 RIP Peter. Thanks for making us all safer via your work on LetsEncrypt and various EFF endeavors. And thanks for trying to figure out better incentives for AI systems via @AIObjectives . You were kind and you were loved and you mattered.

2022-09-05 02:15:12 Peter and I didn't always agree about things, but we always took time to speak to each other and have earnest debates. He was a model for how to be a great contributor to the community and also a total (lovable) weirdo.

2022-09-05 02:14:26 Peter Eckersley passed away. @pde33 was a brilliant and generous person. I have a vivid memory of he and I going to Beijing in ~2018 and on the plane he was programming a world simulation for different types of AI rollouts on a burner laptop.

2022-09-02 19:50:14 @mmitchell_ai One thing I find a bit confusing here is, sort of by design, companies deploy stuff for profit. Did you mean 'Organization shares tech without profit' (this would make more sense to me)

2022-09-01 22:49:22 @ESYudkowsky thanks for the correction! as an FYI, I and some others are working on evals and techniques to try and work out if AI is lying to us and will hopefully have stuff to share there in a while. Ideas welcome given the potentially high stakes here

2022-09-01 22:20:50 @jacobmenick @michael_nielsen I have a bunch of thoughts and will try and write something for import AI this weekend. Major implications for a few things. Also seems to represent some inherent timeline bets on pace of tech stacks decoupling and reaching equivalence.

2022-08-31 19:12:27 @rajko_rad @AnthropicAI re the hype thing, our policy for demos is we always break our system live, as well as show some advanced capabilities. Think a good way to counter hype is to help audience understand where rough edges continue to be. (the systems are getting harder to break, though...)

2022-08-31 19:05:23 More broadly, I'm going to be in DC for a couple of weeks in September, so if you're based there and want to chat, shoot me a DM. Thanks!

2022-08-31 19:04:57 Demos are some of the best ways to understand the strengths and weaknesses of AI systems, so I look forward to demo'ing some @AnthropicAI models and breaking them in front of a live audience! https://t.co/DIO7N0VhYI

2022-08-30 21:58:05 RT @waxpancake: Unlike DALL-E 2, Stable Diffusion also lets you generate images of famous trademarked characters, so we searched for 600 of…

2022-08-30 19:39:42 @karpathy everything is very normal and the rate of change is in no way disturbing

2022-08-30 17:53:44 Chris is a terrifically nice person and fantastic colleague _and_ EXTREMELY DATEABLE. Date him! https://t.co/iL9GUSILAy

2022-08-30 03:07:57 @LurkerSentinel @ElectionLegal *with a jar of heavy creamAnd other horrors beyond comprehension

2022-08-29 03:52:09 @pacoid @chrisalbon Also melatonin for jetlag and/or magnesium supplements, moisturizer (planes tend to dry out my skin a lot, YMMV). Also, always worth polling your friends who may live in Berlin or have been there lately for cool spots close to your hotel

2022-08-29 01:40:18 @chris_j_paxton It's an excellent outcome!

2022-08-28 20:39:51 RT @Ted_Underwood: Ethical intuitions will be reshaped if people start fine-tuning generative models on a consumer GPU. The 2019-20 view th…

2022-08-28 19:33:36 @f_j_j_ it's not very clear that many state actors are good at training models like these, at least for now. Identification feels basically impossible as you're training models to match loss of data distribution. Possible some steganographic approaches could work, but depends on length

2022-08-28 19:32:41 @OwainEvans_UK yeah, it averages out to that. It's really nice. Think one of the v valuable things is all the random conversations around the lunchtable - we have a bunch of people with different backgrounds ranging from physics to natsec, so feels v generative. I'm a fan : )

2022-08-28 19:26:42 @ChrisPainterYup most image models are kind of counterintuitively way cheaper than text models. also, tons of good research ideas wrapped into model.

2022-08-28 19:19:16 This part is crazy as well - funded by an individual https://t.co/RbDPOuRUTl

2022-08-28 19:18:48 @OwainEvans_UK generally feels like having an IRL culture is better for research problems where there are many branching paths and few obvious next steps. Just need to hash out a load of FUD with colleagues, and v hard to do this as easily/naturally remotely or over zoom

2022-08-28 19:18:11 @OwainEvans_UK personally, I've found the productivity boost from being in the office well worth the tradeoff of losing some time to commuting. I WFH one or two days a week and am in office the rest (voluntarily, it's not really a specific policy)

2022-08-28 19:10:40 this absolute mad lad just footed the bill for Stable Diffusion themselves. A good example of the shape of things to come - distributed collectives raising capital to create bottled-up representations of data distributions, seeding them across the digital space https://t.co/Vz2GKTD36d

2022-08-28 19:06:11 @ManlikeMishap (also this is obviously a way longer answer/discussion than necessarily suits twitter, more mentioning it as hopefully an intuition pump for more public thinking here. Trying to write something myself also.)

2022-08-28 18:58:34 @feral_ways It's extremely valuable and is hopefully moving overton window towards more disclosure

2022-08-28 18:56:40 RT @nousr_: Stable Diffusion running inside of GIMP and using google colab as the GPU backend. https://t.co/deQ6duE8Sz

2022-08-28 18:55:42 @ManlikeMishap What are the specific concerns about how accessible training and fine-tuning these models will be? You should write something and post it publicly!

2022-08-28 18:55:04 @jaylagorio @scottleibrand @EMostaque Though to be clear I don't know, I'm just sorta guessing at this. @EMostaque may want to clarify etc

2022-08-28 18:54:29 @jaylagorio @scottleibrand Pretty sure @EMostaque did. It's pretty crazy - it almost certainly amortizes to less than a cent per user as of today, likely substantially cheaper.

2022-08-28 18:38:16 @mchorowitz as with most things in this area, both interesting and mildly terrifying

2022-08-28 18:30:21 Stable Diffusion: $600k to train. I'm impressed and somewhat surprised - I figured it'd have cost a bunch more. Also, AI is going to proliferate and change the world quite quickly if you can train decent generative models with less than $1m. https://t.co/auddBQcAZY

2022-08-28 18:03:39 @thomeagle I recall this, but the holes in my brain this drink gave me make the recall most painful

2022-08-28 06:54:42 @namathree I'm just as awkward as Nathan Fielder, if not more!

2022-08-28 06:35:03 My friend took this pic of me - for the record, this is how I look at pretty much any live music or art event. https://t.co/2jc2Mdmqow

2022-08-28 06:34:12 To see friends and strangers play their weird music in punk houses to the assembled dregs of a Saturday is a wonderful and precious thing. #VOTEDIY2022 https://t.co/XIJE1nCGjv

2022-08-27 21:49:48 @friendly_gravy @AnthropicAI We're also shipping this red teaming paper to a few interested policy stakeholders advocating for it being wired into various mooted risk assessment approaches, so we're actively trying to push this into a regulatory context!

2022-08-27 21:49:14 @friendly_gravy @AnthropicAI I think the current state of v minimal regulation of AI isn't good and is also increasing the chance of unsafe or dangerous deployments, so we need regulators to catch up. I spend a bunch of my time working in regulatory forums like OECD and others for this purpose : )

2022-08-27 21:48:35 @friendly_gravy @AnthropicAI I ultimately expect red teaming will be integrated into regulatory approaches. Broadly, I/Anthropic pushes for a few different types of regulation. I feel like regulation requires good information, so I generally propose this when talking to govs https://t.co/vljhYFFXi2

2022-08-27 21:08:38 @WilliamMcIlhag1 @HeerJeet I meant London, where I worked for many years.

2022-08-27 20:55:11 @WilliamMcIlhag1 @HeerJeet I love public transit and it's the main way I get around - I don't drive, just bicycle and BART, and when I was in UK just bicycled as well as trains and buses.

2022-08-27 16:08:36 @spacetrippee yeah except way shitter

2022-08-27 16:04:02 @HeerJeet I don't feel terrible about this, but it's probably wrong. I remember in 2014 thinking that if my partner and I bought a car in 2020 it'd be autonomous. 2020 came around and you can technically by some autonomy (e.g tesla), but nowhere as good as I'd expected stuff to get

2022-08-27 07:22:27 @conjurial I agree with this, but have always stumbled on the mutual incentive structure aspect. How do you incentivize this (potentially invasive) level of information access?

2022-08-26 05:53:49 @philmohun honestly - good question I haven't thought about. I mostly work with text models. For image stuff, maybe an analog would be asking it to fill in a blank spot on a circuit diagram and it makes a circuit that seems right on a cursory viewing but actually doesn't work in practice.

2022-08-26 05:49:27 @philmohun sort of - I find myself saying hallucinated or 'made up' quite frequently. Mostly in context of factoids that sound right on a cursory reading but are inaccurate or erroneous. Don't think it's really a term of art, but perhaps a form of slang. Gibberish also a good term

2022-08-26 05:25:08 Since this is Doing Numbers, want to foreground Nova's incredible contributions to @AnthropicAI and the field at large. Check out this 80k interview for more: https://t.co/54CTXsHKox

2022-08-26 04:36:37 @_ianks I have secured my job

2022-08-25 22:25:32 @StephenMarche It's a good question! This is an open problem. You can use datasets like the red teaming one to train models to be less likely to do this. Lots of additional details here https://t.co/kK5AFnjUI8

2022-08-25 22:00:35 @StephenMarche RegEx is basically a way to search over text for specific things - this script tries to pull out the text which matches structure of phone numbers, drivers licenses, social security, etc.

2022-08-25 20:54:07 @davidad all that stands between us and a paper clip'd universe is a 10000-character regex string

2022-08-25 20:28:52 @mdaviswilson This is a thing I genuinely worry about

2022-08-25 16:18:53 We did some manual checks (I read like 50 transcripts with potential PII) and did seem like was all hallucinated. Discussed and felt responsible thing to do would be try to remove all the stuff that looked like PII. This led to my colleague Nova writing this truly insane RegEx: https://t.co/dx3AM0jP8p

2022-08-25 16:17:33 Inside baseball policy stuff: At some point we realized our red team transcripts included a lot of potential PII - that's because our humans loved trying to get our models to give out celebrity addresses, so our models (mostly hallucinated) PII. Cont... https://t.co/sifB2RkKPX

2022-08-25 16:02:09 luv 2 make typos early in the morning. Anyway, I think it's inevitable that AI developers all adopt pre-deployment red teaming (and a few orgs already do some of this), so feels like an area ripe for collaboration. Will also be advocating for red teaming in policy discussions.

2022-08-25 15:53:50 I also think red teaming is an area that will benefit from increasing expertise. E.g, asking LMs to help make dangerous chemical compounds, or carry out certain offensive cyber ops. Excited to explore this stuff with others in coming months !

2022-08-25 15:52:24 In the same way, model developers can't a priori anticipate the capabilities of AI models, they also can't really anticipate all the potential misuses or areas where their safety interventions haven't worked. Red teaming gives us a nice way to probe AI systems for problems. https://t.co/sifB2RkKPX

2022-08-23 22:17:53 We've also recently raised some more money to support the Index and are working on some strategic initiatives which should grow the report impact even more in coming years. Would be delighted to work with people to make this happen!

2022-08-23 22:17:24 I've been involved in @indexingai since its inception so can say that this is a great role to use as a stepping stone to cool careers - prior people who have done this have gone on to manage the report, work in policy at Stanford, work at hedge funds, and found startups.

2022-08-23 22:16:32 This role would suit someone early in their career who is keen to rapidly increase their understanding of the AI space, and who is passionate about both the research and production aspects of making an annual report.

2022-08-23 22:15:45 The AI Index is recruiting! Excited to announce @nmaslej 's promotion to manage @indexingai, and opening for a research associate who will help us make the world's most widely-read set of AI policy data even more impactful https://t.co/kLORkx9kX2

2022-08-22 00:01:23 @DanielSolis This is extremely delightful. Good work

2022-08-22 00:01:15 RT @DanielSolis: These Birds Do Not Exist: THE PUZZLE! From the Society of Theoretical Ornithology Research and Knowledge — Assemble the…

2022-08-19 20:59:45 RT @ID_AA_Carmack: I mentioned this in the Lex interview, but it is official now:Keen Technologies, my new AGI company, has raised a $20M…

2022-08-19 20:42:16 @andrey_kurenkov haha, oh yeah I like 100 Gecs, and found SOPHIE through charlie xcx etc. Send more recs please!

2022-08-19 17:52:40 After careful consideration (listened to it about 200 times in two weeks), I think this is one of the weirdest and best pop songs of all time. Amazingly abrasive and strange and lyrical. https://t.co/sKGmqQcUUY

2022-08-18 03:13:20 "We put a person on the moon in 1969 and in two fucking years we were playing golf on it" - a friend on the majesty and madness of America.

2022-08-17 15:08:41 @AmandaAskell Depends on your own outlook. For me, the cognitive overhead of having to periodically think about it meant it seemed simpler to get rid of it. I generally find interacting with bureaucracies pretty draining and do enough at dayjob didn't want any more

2022-08-16 15:23:50 Kind of mind boggling to me how insane this was/is - it traps people into debt that they really struggle to repay (unless work in a high-paying job like tech), and introduces a massive drag on the economy from people servicing loans instead of spending money. Weird policy!

2022-08-16 15:22:56 UK student loans were kind of sucky (I was one of first generation who even had loans - used to get grants). But interest rate was pretty benign - like a %point above inflation. My US partner, on the other hand, had loans on the order of 8% or so, which we aggressively paid down

2022-08-16 15:22:09 Just paid off my (UK) student loans

2022-08-15 06:43:22 @pmddomingos (also this isn't a loaded/gotcha question

2022-08-15 06:42:26 @pmddomingos What's the least misleading temperature graph in your opinion?

2022-08-15 06:23:39 @pmddomingos What do you mean by lie here?

2022-08-15 02:21:34 @catehall Makes sense, thank you!

2022-08-15 01:54:12 @catehall What does anti-defect mean in this context?

2022-08-14 22:06:39 @jjvincent I can't open my fridge without eating a piece of cheese. Doesn't matter what I'm doing, a small quantity of cheese is inevitably getting nibbled.

2022-08-14 19:38:26 @jesseengel Will do. You don't have open DMs so you should DM me and we can arrange!

2022-08-14 19:27:04 @jesseengel Where are you based? I struggle with this issue as well (part of me finds the capabilities astounding, part of me worries about the ethics of development). Would you like to get coffee and discuss? Am in East Bay

2022-08-14 02:53:38 RT @mmitchell_ai: @jackclarkSF It's way too early to say this isn't a serious concern. IMO people aren't knowledgeable about using LLMs for…

2022-08-13 22:56:54 @regretmaximizer @jessi_cata https://t.co/eRNKhsqmOS

2022-08-13 22:37:04 @mmitchell_ai Just to be v specific, I was very worried about disinfo from gpt2 and I think that was way too early. Had proof of concepts but not really reliable

2022-08-13 22:36:18 @mmitchell_ai Oh I agree! I meant that back in 2018/19 I felt like it was imminent, but, like synthetic imagery, probably need a few more turns of the crank of usability/cost before it becomes prevalent.

2022-08-13 21:00:02 RT @AmandaAskell: I might disagree with people about which of the world's fires is the biggest, but I won't speak badly of people that are…

2022-08-13 20:55:33 @girishsastry @GaryMarcus @samuelmcurtis Yeah,.iirc a load of fake profiles on LinkedIn appeared one day shortly after thispersondoesnotexist came out - ease of use stuff etc

2022-08-13 19:26:06 @vineettiruvadi Ah, the prior behind the tweet is I've been trying to find evidence of utilization and haven't, and spend a lot of time talking to people whose job is to find this stuff for a variety of people, and haven't heard of concrete things

2022-08-13 19:22:07 @vineettiruvadi Absolutely. What evidence do you have of large-scale utilization in a single instance?

2022-08-13 19:20:09 @GaryMarcus @justinhendrix @samuelmcurtis Cool, do you have any links for further reading? This may fit more into the seo/spam thing

2022-08-13 18:50:52 @GaryMarcus @justinhendrix @samuelmcurtis What's the example

2022-08-13 18:50:37 @GaryMarcus @samuelmcurtis This is mostly a not-yet thing due to unit economics, complexity of disinfo kill chain, need to invest engineering to effectively integrate LM etc. This'll change eventually.

2022-08-13 18:49:29 @GaryMarcus Haven't seen any clear proof of utilization versus spam/seo. If you have clear evidence of use please share it. (asked same q to @samuelmcurtis ). I also know people whose job is fighting mis/disinfo on platforms and they haven't seen much evidence or even symptoms of it yet

2022-08-13 18:42:00 @Miles_Brundage @jesswhittles Yup +10.

2022-08-13 18:32:25 @jesswhittles @Miles_Brundage I've also found 'when they get low we go high' to be a reasonable way to approach twitter. Sometimes things can get a bit agro but maintaining an optimistic outlook and being open to constructive feedback seems helpful.

2022-08-13 18:31:44 @jesswhittles @Miles_Brundage Tbh I think one of the only ways to have a more sensible discourse is to have more people thinking in public and contributing to conversation. Private docs kind of incentivize in- out-group dynamics and increased risk of groupthink

2022-08-13 17:34:26 @juand_r_nlp @Ted_Underwood Notably, this is how Replika advertises itself on TikTok these days. Kind of gross, imo. https://t.co/OJzLJhGWFv

2022-08-13 16:44:21 @Ted_Underwood Yes! I worry a lot about 'fractal realities' enabled by AI. 100%

2022-08-13 15:55:16 @samuelmcurtis What evidence do you have for this?

2022-08-13 15:54:46 @jamescham It's that disinfo involves a complicated chain of actions (identification, using certain identities, generating content, filtering content, posting, updating in response) and it's not clear the effort vs reward part of using an LM makes sense... yet

2022-08-13 15:36:12 Given that, what are we likely getting wrong today about current risks of large generative models?

2022-08-13 15:35:54 In hindsight, a lot of the worry about use of language models for disinformation campaigns was kind of overhyped (I myself am guilty of this overhyping). Seems like most negative uses of LLMs so far have been Spam/SEO, or deliberately 'nasty' data distributions (e.g, GPT4Chan)

2022-08-12 22:06:30 @catehall oh, they're real.

2022-08-12 18:06:05 RT @JacquesThibs: EA donations to global poverty/health over time (through givewell): https://t.co/ydFvh33MDs

2022-08-12 04:56:37 RT @sleepinyourhat: Shameless plug: I’m now ~2 months into my sabbatical-year visit to Anthropic, and I’m really impressed. https://t.co/gx…

2022-08-12 03:23:11 @rvinshit This is real good

2022-08-12 00:59:50 @aptshadow @CompellingSF Extremely hyped for Children of Memory. Have really loved this series!

2022-08-11 18:32:32 @FerrisHueller many people are saying this

2022-08-11 16:38:49 @akashpalrecha98 @karpathy Thank you, this makes me extremely happy : )

2022-08-11 16:15:37 @causalinf A+. Well done!

2022-08-11 16:10:38 @IasonGabriel ONE LIKE ONE TWEET DO IT, DO IT NOW

2022-08-10 17:30:21 RT @robrombach: The weights of Stable Diffusion, a latent text-to-image diffusion model, are open for academic research upon request!See h…

2022-08-10 17:22:57 Very good thread about the Voynich manuscript. Worth it for this sentence alone: "The Extensible Voynich Alphabet (EVA) is used by Voynichologists worldwide." https://t.co/5MpPoXApaC

2022-08-10 17:21:17 @officialKrishD https://t.co/NaMKX2i7G6

2022-08-10 17:16:50 @pgcorus that's different to the type of large-scale model I'm talking about. Applied work would include training on all streams of intercepted and scraped data, though for more specific purposes typically

2022-08-10 17:11:52 @mmitchell_ai @BigscienceW I <

2022-08-09 22:16:54 @Simeon_Cps I'm not sure if impacts short timelines much - the fabs will take years to build, and fabs in the US are a lot more controllable than fabs in Taiwan, so it might increase governance levers. Hard to say though

2022-08-09 22:14:08 @Simeon_Cps Chips isn't gonna really change supply of frontier GPUs massively. It's a lot more oriented around cpus and memory.

2022-08-09 22:13:10 @Simeon_Cps Much of CHIPs is about reshoring parts of semiconductor supply chain so US isn't as dependent on other places for chips, including specialist US military chils. This makes policymakers more relaxed and reduces likelihood of extreme actions being taken in the future re supply.

2022-08-09 22:05:34 @Simeon_Cps Chips act mostly pushes on diff capabilities to those needed for frontier AI dev, so probably reduces risk by increasing robustness of gov.

2022-08-09 21:13:27 @RichardMCNgo This tweet would've been clearer if I'd specified 'RIGHT NOW'. Misalignment etc is definitely a future problem.

2022-08-09 21:11:32 @RichardMCNgo I am worried about AI doing bad stuff eventually - I'm more worried in the short-term about govs getting surprised by AI developments and being sputniked into doing dangerous stuff or increasing race dynamics

2022-08-09 18:48:03 RT @CorreaDan: The bill also included @realTinaHuang’s proposal to establish testbeds at NIST to support the development of trustworthy and…

2022-08-09 18:29:06 @ESYudkowsky @robbensinger Yeah I think it'd be helpful to write some bear cases on govt involvement (I'm going to try and do this, since I spend time advocating for more government involvement in some areas like monitoring)

2022-08-09 06:10:45 @rasbt Thank you!

2022-08-08 20:05:34 @rbhar90 +10

2022-08-08 18:34:59 @karpathy thanks for reading!

2022-08-08 15:42:14 @JCorvinusVR oh i agree, I was just trying to clarify that I myself don't think these ideas are practical or plausible.

2022-08-08 15:30:40 @JCorvinusVR https://t.co/9mvMyM19FP

2022-08-08 06:13:23 @bedbayesnbeyond Also pretty good. Maybe like 30 people? Same sentiment tho

2022-08-08 06:12:53 @bedbayesnbeyond This one is actually ok!

2022-07-27 00:55:46 @advadnoun It feels to me like this argument mostly boils down to 'this should be a public good, not something that accrues profit to a private company'. I don't necessarily agree with this, but I think it makes sense.

2022-07-26 14:05:04 @k_mcelheran @jamescham @kcnickerson @userfit @allenb @AriSalonen Sure, DM it to me

2022-07-25 16:45:45 @gwern @_joaogui1 back when I was a journalist I had a series called 'clark side of the cloud' where I toured data centers around europe and america and got into quite a lot of detail, so I know an unreasonable amount about this

2022-07-25 16:45:14 @gwern @_joaogui1 also worth noting a bunch of US data centers are in places like 'high desert' oregon (as lack of moisture good for free-air cooling), so I'm thinking that stuff like wildfires could also cause problems here

2022-07-25 04:03:27 @isosteph Yeah after I wrote it I realized that'd be insane and like a semi-jog. Great work!

2022-07-25 03:55:04 @isosteph A+, congratulations. Out of interest how long did this take, like ten hours?

2022-07-25 03:27:31 https://t.co/nWRd9nuGA5

2022-07-23 19:40:00 @alexandr_wang yes and yes. It's going to be fun.

2022-07-22 19:22:19 @jstrauss @vaisfourlovers yes, I try and have 2 days each week with zero meetings, and it is the main way I can get actual work done. I vociferously protect them

2022-07-22 18:48:07 @zacharylipton sorry you went through that and glad you got out, sounds absolutely horrific

2022-07-21 02:45:09 @dkaushik96 @KumarAGarg Extremely hyped for the movement on this today!

2022-07-20 01:18:42 @BishopTopsy Up and to the right, most days!

2022-07-20 01:18:23 @henrycomb_ Henry! Sorry to hear that, but glad you also gained the ability to see a bunch of positives. Miss ya!

2022-07-19 19:05:21 @slippylolo @arankomatsuzaki a+ meme game here.

2022-07-19 17:11:48 RT @ahandvanish: I’ll be testifying to Congress on Tuesday at 10am regarding #LongCovid, including current research and needs. A livestream…

2022-07-19 06:01:34 @iblametom Cc @BorisJohnson

2022-07-18 23:35:37 @rgblong @rupertg gave me an interview for ZDNet UK to become a technical reporter, when all I had on my CV was a fly-by-night SEO shop and obvious enthusiasm/knowledge about wonky stuff. Changed the course of my life.

2022-07-18 15:59:29 @IreneSolaiman can you share slides after? good luck!

2022-07-18 04:27:50 @Tim_Dettmers Thank you! I can sympathize, but my goodness losing hands would be tough. I shared as I figured it's good to let others know that people have these experiences and, thankfully, can get through them! : )

2022-07-18 04:26:42 @Steven_B_Lee A lot of prevention comes down to core strength. I've been working on that for years but am increasing efforts now. I'm also taking anti-inflammatory stuff like magnesium. Had some short term meds when had spasms but have tapered off weeks ago.

2022-07-18 04:25:38 @Steven_B_Lee I've had back issues since a teenager so I have some idiosyncratic physical stuff. I think this one was probably due to stress / not taking a proper holiday for years, and something caused a cascade.

2022-07-18 04:21:24 @KatiMichel *re-straightened

2022-07-18 04:20:58 @KatiMichel Yikes - glad it got fixed! And yes, it was pretty bad. There was a fun period in weeks 3-5 where I walked like the letter 'C' as spine was bent due to the spasms in one side. (Restrained with physio and fine now, but sheesh!)

2022-07-18 04:13:53 Anyway, who is to say there's that much of a difference between myself and a big model, when we both react to _predicted outcomes_, stacking our response in line with our subjective probabilities. When you're in crazy pain, I think difference between humans and dumb AI collapses.

2022-07-18 04:12:28 One surprising aspect of my recent Health Adventure was the role Fear played

2022-07-18 04:06:22 @IasonGabriel A+

2022-07-18 04:04:58 @IasonGabriel Let's finally do that hike if you're still in the bay! Didn't reach out recently to the above, haha.

2022-07-18 04:03:45 @PoemsWeBurned @adrian_weller Bay area

2022-07-18 04:01:30 @startuployalist Thank you. And thanks for sharing this thinker - wasn't familiar and will read... embodiment feels like a huge part of cognition.

2022-07-18 03:43:11 @mobav0 Thank you! I was grateful for how this experience really highlighted what is important in a Maslow's hierarchy pyramid form.

2022-07-18 00:02:24 @then_there_was @adrian_weller Oooh great tip! I will look into this, thanks

2022-07-17 19:42:31 @likeloss4words Oh yeah, I feel that. There are a bunch of areas like robotics where if we had dramatically more people working in it, AI would be able to flow much more quickly into world for useful purposes.

2022-07-17 19:38:51 @likeloss4words (this intuition comes from seeing the robot team work with a shadowhand at openai and realizing stuff like mechanical tendons and motors are all way more janky than I'd thought. It was very humbling)

2022-07-17 19:33:58 @likeloss4words It's not a key bottleneck, mostly it feels like motors and battery and other aspects are the hard bit, but I do expect AI will help with this stuff, and can also help exoskeletons eventually be more adaptable

2022-07-17 19:32:09 @volokuleshov Yeah, I had been doing planks/pushups a lot this year which had staved off back trouble nicely until the incident. I've now started doing a bunch of squats, and am trying to give myself a workout routine mostly built around strengthening back/core.

2022-07-17 19:25:57 @Ted_Underwood @deepfates I also want things like vast planetary-scale ecosystem monitoring via smart, semi-autonomous robots, cameras, satellites, and so on. These will be a bunch of agents coordinated by other agents.

2022-07-17 19:25:06 @Ted_Underwood @deepfates Why not both? It's great to have a culture-in-a-bottle, which is what gen models are, but it's also great to have agents that use a culture-in-a-bottle to bootstrap world models and take actions. I want to explore the deep sea and space - need agents.

2022-07-17 19:11:06 @MrPKent thanks for sharing - yeah, it is almost comedically bad how much it reduces your movement/agency. There were a couple of days where just turning over in bed was a multi-minute process with discrete stages and ridiculous amounts of grunting.

2022-07-15 05:06:59 @superglaze Popular spills over time sounds like a banging non-fiction book by a physicist

2022-07-14 21:14:22 @RishiBommasani @WilliamWangNLP Also seems like a reason for why academics should build their own big models, then raise funding to support one-to-many sampling infrastructure

2022-07-14 02:00:30 @gregeganSF @lacker Reality is typically both dumber and stranger than fiction.

2022-07-13 17:12:57 @bytestoatoms @AnthropicAI Thanks for pointing out the omission, we’ll fix it soon. (I passed this on internally).

2022-07-11 21:12:04 @DZhang50 @MichaelTrazzi thanks for sharing! I'd been following the google RL chip stuff for a while (e.g, here https://t.co/fyTEHoq539) but hadn't read this paper. Will read!

2022-07-11 20:45:47 @ArmanMaesumi @MichaelTrazzi well, a big diff is improving chip design more than with standard EDA software, which is what NVIDIA did here. This is distinct.

2022-07-11 20:16:02 Import AI is back! Seemed fitting to re-launch with a section on an AI company using AI to make parts of its AI chips more efficient for training AI systems that will further improve chip design to further increase efficiencies of AI training (^infinity) https://t.co/u7ywXOl19y https://t.co/tI1NUMI0Nc

2022-07-11 17:54:42 Extremely excited about our advisory board, jury members, and staff as well! @ruchowdh @EileenDonahoe @camillefrancois @ghadfield @EvaKaili @safiyanoble @navrinasingh @lyssaslounge @stefvangrieken @atg_abhishek @verityharding @wsisaac @rajiinio , and more! https://t.co/kUcugEJynL

2022-07-11 17:51:12 The idea here is to catalyze activity in studying, auditing, and assessing the traits of widely-used and/or deployed AI systems - from YOLO to CLIP, from GPT3 to GPT-NeoX, and so on. For AI to benefit society, we need society to have more critical faculties wrt AI.

2022-07-11 17:49:29 Do you want to audit a deployed or open source model for societal impact? Do you want to use, or create, open source software tools to do this? Would you like a chance at a $25,000 first prize? If so, please enter the AI Audit Challenge! https://t.co/qmLBxootfu https://t.co/7fUGYhK01C

2022-07-06 22:53:44 @HaydnBelfield https://t.co/A8C0WsU0lW

2022-07-06 20:04:56 Gave a talk today to a National Academy working group on AI and Workforce. Spent lots of my presentation talking about difficulty of measuring advance of AI systems, and how AI systems are also progressing more aggressively than expert forecasts.Slides: https://t.co/HbvdOoBP51 https://t.co/qLdUyhIDpk

2022-07-05 16:09:19 RT @JasonGMatheny: I'm thrilled today to start as the new president and CEO of the @RANDCorporation, an organization I've idolized since I…

2022-07-05 06:58:59 @nlp_pranav https://t.co/cjDU20TyEF . But I was moshing in spirit!

2022-07-05 04:54:24 Spend your freedom on supporting eachother and doing things for your community and creating experiences that are fundamentally and deeply rooted. #VOTEDIY2022 https://t.co/0KvLPRhDQ7

2022-07-04 15:46:48 @pjbarden @SColesPorter @WorldSummitAI Dm me the addresses you're using

2022-07-02 08:19:11 @aggielaz I agree! Some of what the Dynabench paper espouses is moving to continuous + dynamic evaluations. Given the huge expressive space of these models it feels like you need an increasingly broad set of people to mess around with them to surface capabilities and pathologies.

2022-07-02 02:16:34 @cheeze_squeeze Thanks for being a reader! Newsletter on hiatus due to health stuff but coming back real soon

2022-07-02 01:22:30 @WillManidis @jeremyphoward Agreed! I feel like an underinvested area in ML is benchmark design. Progress is catalyzed by the existence of benchmarks and, I suspect, stagnates in their absence.

2022-07-02 01:17:55 @__anoop Getting there! https://t.co/M4ZHS0LEKeI also agree that this trend says our benchmarks are somehow broken.

2022-07-01 17:33:10 @jjvincent @whippletom incredible. Beer on me next time in london, yung weasel

2022-07-01 16:14:49 RT @ada_rob: That was fast. News sites are already using DALL-E (mini) to generate fake headline images. This move seems questionable to sa…

2022-07-01 05:20:33 @LastWordSword @vineettiruvadi @BW Also agree https://t.co/VzqnSP2Nu0

2022-07-01 05:15:42 @filippie509 https://t.co/VzqnSP2Nu0

2022-07-01 04:51:59 @TaliaRinger Absolutely! And if you know any super hard benchmarks please send them to me so I can advocate to add them to @indexingai . Also excited by fun ideas like this https://t.co/tMxHfYhsp4

2022-07-01 04:40:38 @bucketofkets Great example. Should have mentioned that in my thread, definitely in the back of my head as some chums like @AmandaAskell were involved.

2022-07-01 04:39:57 RT @bucketofkets: @jackclarkSF we started putting together bigbench exactly for this reason…and average human performance was exceeded befo…

2022-07-01 04:19:52 Obviously there are two conflicting points here - benchmarks aren't measuring true intelligence on a task as the AI systems they test still break in dumb ways, yet our benchmarks are becoming outmoded at ever increasing rates.

2022-07-01 04:15:02 Things are getting... Extremely weird. Think about what this graph may look like in spring 2023 (was published April 2021). From the excellent Dynabench paper https://t.co/v3TkBgSATM https://t.co/gApjigp21W

2022-07-01 04:12:08 @powerbottomdad1 @nwilliams030 Getting there! Will update in a tweet soon, but fact I'm back to obsessively reading arXiv is a good sign :)

2022-07-01 04:03:42 @powerbottomdad1 This is what import AI is for! On hiatus currently as I work through some health stuff but coming back real soon

2022-07-01 03:52:12 @vineettiruvadi I think that used to be true but in recent years we've seen the emergence of large-scale models that display great competency at a bunch of distinct tasks (and some capabilities spike above SOTA), so I don't think this is illusory progress.

2022-07-01 03:49:34 Both of these results were published TODAY. These results happen at a delay, so this is probably old information on order of 3-9 months. There are easily 5 labs and probably 10 with enough compute to play at this level. Imagine what we don't know right now?

2022-07-01 03:48:17 Here's a system that beats Stratego, a game with complexity far, far higher than Go. https://t.co/9vXncymWKE

2022-07-01 03:46:23 Here's MINERVA which smashes prior math benchmarks by double digit percentage point improvements https://t.co/humsCGxh6w

2022-07-01 03:44:44 As someone who has spent easily half a decade staring at AI arXiv each week and trying to articulate rate of progress, I still don't think people understand how rapidly the field is advancing. Benchmarks are becoming saturated at ever increasing rates.

2022-07-01 03:33:48 @jordannovet Photograph of person successfully opening thin green produce plastic bag pov

2022-06-28 17:15:17 RT @BigScienceLLM: 100%

2022-06-28 03:11:07 @kaushikpatnaik +1! Hadn't seen this stuff by @hardmaru and it's awesome

2022-06-28 02:46:13 @jmdagdelen https://t.co/HE2V7LgpQT

2022-06-28 02:36:37 @NicoleHemsoth Yeah I think this is a v pertinent question! https://t.co/bIYqFDESD7

2022-06-28 02:35:20 @JayPatel0101_ @karpathy Wouldn't take the other side of that bet!

2022-06-28 02:14:37 @karpathy Plus: indie videogames, music videos, advertising...And when video models get good, geez!

2022-06-28 02:11:48 Gen models are also a lot easier to deal with. When I edited my university's student newspaper I had to try and get a photograph of a person in a wheelchair being attacked by a swan - massively difficult, unless you were there. With Imagen or Dall-E (sans filters) u can do it.

2022-06-28 02:09:28 The Stock Photo industry is probably not ready for generative AI. Generative AI seems better for 80% of use-cases. In other words, NYT still gonna do illustrators, but a random website will probably find economics of gen models more attractive than a Shutterstock subscription.

2022-06-27 17:51:48 @b_cavello @Unibo Can you share your slides? Would be curious to see

2022-06-27 16:20:33 RT @EthanJPerez: We’re announcing the Inverse Scaling Prize: a $100k grand prize + $150k in additional prizes for finding an important task…

2022-06-27 03:42:42 @nsaphra @michael_nielsen I loved Dark Forest and found the third more interesting than you seemed to. However, they've done tons of great short stories that I'd recommend. Will try and msg you with titles when I get back to the book!

2022-06-27 02:39:19 @michael_nielsen This is how you lose the time war - El-Mohtar and GladstoneShards of Earth - TchaikovskyAnything by Cixin Liu, especially the short storiesA Colder War - Stross

2022-06-23 19:57:28 RT @ChrisGr93091552: Explored github copilot,a paid service, to see if it encodes code from repositories w/ restrictive licenses.I checke…

2022-06-23 19:49:56 @MurphFromNerf @carlfranzen https://t.co/tI60HcaDTS

2022-06-21 02:28:08 The good news is I'm doing better most days. I just need to budget my energy and mental attention a lot.

2022-06-21 02:27:07 Import AI will also not come out this week due to my aforementioned medical issues. However, I am generally doing better and will post an update once I feel I'm through it. So far, doesn't seem chronic, just very slow recovery. https://t.co/atNWBPqcUg

2022-06-20 20:49:16 RT @sarahookr: I think this is one of the most important open roles at @CohereAI right now. I'm personally invested in helping find a good…

2022-06-20 18:43:17 @CineraVerinia @tszzl Most major Chinese results for frontier ML/DL also get published in English. The Chinese-only papers tend to be either oriented around specific, deployed applications, or are lightly to heavily classified PLA-linked research pubs. (Source: I look at this annually for @indexingai)

2022-06-19 03:56:08 @Brainmage Honestly, wtf. https://t.co/S2W56YuLIr

2022-06-19 03:55:27 @Brainmage Crungus be shopping https://t.co/SYchaEEJWN

2022-06-18 04:25:13 @ElectionLegal This. This, I struggle with.

2022-06-16 22:45:59 @tsimonite A+

2022-06-16 21:02:37 https://t.co/N6togcU4n3

2022-06-15 18:39:13 I expect I'll be fine in the long term, but it's sufficiently serious that most of the medical advice I've been given has been a variant of 'do absolutely nothing so your body has best chance of healing'. Will update as I get better details.

2022-06-15 18:38:39 Import AI won't be coming out this week as I've been dealing with some fairly serious medical things for the past three weeks. I'll post an update here once I have a handle on the situation.

2022-06-14 04:15:05 @FelixHill84 Cc @AmandaAskell

2022-06-14 02:57:16 @jrieffel @JonathanBalloch @katecrawford Exactly - the problem is we're cannibalizing the intellectual base on which forward progress depends, and we're pushing many publicly-minded folks into locked down information silos.

2022-06-14 01:44:28 H. G. Wells out here anticipating 'move fast and break things' in 1937 (from: World Brain). https://t.co/ArlYdQly2Z

2022-06-13 18:07:18 @alexbarinka @sarahfrier @KurtWagner8 @business oh, rad! excited to see the reporting you do. I still have fond memories of how effectively you terrorized IBM!

2022-06-13 18:06:14 @TheSeaMouse @katecrawford there was a lot of private sector research prior to alexnet, e.g a lot of early handwriting and speech rec comes from private sector places like AT&

2022-06-13 06:41:01 @t0nyyates @katecrawford Basically high water mark results in AI e.g perceptron(60s), ALVIN (early self+driving,80s), Deep Blue (90s), Imagenet (2012), alphago(2015/16), gpt3 (2020), etc

2022-06-13 01:17:13 RT @AlexGDimakis: This is indeed a challenge for us in universities. Novel algorithms, clever ideas and provable bounds are harder to get w…

2022-06-12 21:20:27 RT @nathanbenaich: I found this chart particularly striking - the fact that academic researcher's ability to compete with industry research…

2022-06-12 17:06:30 We did it, everyone! GoFundMe is done. Message from Joe and Jenn here https://t.co/dh48pymrEy https://t.co/9nRAL0Mvsp

2022-06-12 06:33:07 Y'all are incredible. This makes a huge difference to my friend. Thank you! https://t.co/wq0SVYaRua

2022-06-12 03:51:47 Thanks so much to everyone who has donated. This so far makes a huge diff - already a workday and change in funds for my buddy!

2022-06-11 18:05:02 @thetorpedodog Yeah, they made it sound like they didn't have the people to take reports for stuff like this. Bewildering and absurd.

2022-06-11 17:52:24 @Breck_Maynokur @MikeIsaac Wow

2022-06-11 17:50:03 @MikeIsaac It was crazy! We called back a few hours later and asked about the chances of getting someone out to see it and they basically said since it was a Friday night in Oakland they'd all be busy elsewhere. A classic 'what are my taxes even paying for?' moment.

2022-06-11 17:44:38 Also, Oakland PD said they weren't taking police reports as they didn't have capacity (???) and wouldn't come to see it, which is extra enraging. We think the same person torched some other vehicles around the neighborhood as well.

2022-06-11 17:42:21 Yesterday, some nutcase set fire to my friend Joe's truck. Joe depends on his truck for his livelihood (he's a metalworker/welder), so this has really screwed him over. If any followers could spare some $ it'd help out a hardworking, lovely guy. https://t.co/66BgoMMMns

2022-06-10 21:22:10 @Derek_duPreez Congrats! Same energy: https://t.co/J1AtkyXdM2

2022-06-10 20:05:30 @chriscanal4 Excellent, you're in for a treat! I particularly like some of the flashback earth sections. Season 2 has some really nuts stuff, also.

2022-06-09 20:18:29 @giffmana @katecrawford https://t.co/kJjCsOSUPUFrom this @AnthropicAI paper https://t.co/6R2Qa4szOr

2022-06-09 18:27:57 @MichaelTrazzi fwiw I'd put myself more in the middle - I don't really think AGI is gonna be bad, I think it's more like 50/50 currently

2022-06-09 07:36:40 Infinite Art feels a bit like Infinite Jest. https://t.co/UEHdAVJPbB

2022-06-09 01:45:09 RT @raphaelmilliere: How should the AI research community broadly refer to the family of large, pre-trained, self-supervised, scalable mode…

2022-06-08 23:15:36 RT @unsojo: @jackclarkSF @mmitchell_ai @alexhern @HaydnBelfield @IreneSolaiman A lot of anthropologists have written about how different so…

2022-06-08 23:15:29 @unsojo @mmitchell_ai @alexhern @HaydnBelfield @IreneSolaiman Helpful and illuminating thread, thanks very much!

2022-06-08 21:44:32 @mmitchell_ai @alexhern @HaydnBelfield @IreneSolaiman @unsojo Absolutely! Please send stuff over to read. I think the above analogy isn't very well fleshed out on my part, and I'm more just looking at industrial practices rather than innate things (also: not a psychologist). Would be excited to read

2022-06-08 20:46:20 @jsotterbach @katecrawford Our mutual chum @Hernandez_Danny did this analysis based on the data from this paper https://t.co/ZfbnUdqlgu, iirc.

2022-06-08 17:58:00 @thomeagle I particularly enjoyed their EP "Hard Reardan Steel", and their single "No God Only Galt"

2022-06-08 17:55:56 @thomeagle Ayn Rand Accelerationism

2022-06-08 17:53:02 @jrieffel @katecrawford Actually, sort of more worryingly, it's less stark here. (Which suggests academia is really under-resourced for frontier). From: https://t.co/jZE3vcGIok https://t.co/fW4SZID7W4

2022-06-08 17:42:18 @bryanrbeal @indexingai I'm not suggesting universities build DCs - that'd be bad. I'm saying places like NSF are not really setup to do grants that'd let a professor and students have enough resources to train a big model on Azure/AWS/GCP.

2022-06-08 17:32:11 @bryanrbeal @indexingai Additionally, academic incentives and compensation structures make it really challenging to hire the engineering teams required to build and develop big models. There is work being done here (e.g, National AI Research Resource in the US), but much more to do.

2022-06-08 17:31:32 @bryanrbeal I agree with you - something we cover in @indexingai regularly is the plummeting cost to train decent computer vision models on clouds, so def true. On the other hand, academia can't build this stuff - funding orgs like NSF really aren't setup to fund 1000+ machine clusters

2022-06-08 17:17:04 @bryanrbeal (I don't mean this to sound catty and we may be making complementary points - it's def true that the ability to utilize benefits of AI is cascading down to everyone through large commercialization. I just think that's only a slice of the whole pie, in terms of benefit.)

2022-06-08 17:16:04 @bryanrbeal It also means only private sector actors are privy to the national and economic security implications of frontier, large-scale systems. Last time this happened was oil industry having better intelligence gathering than govs - led to antitrust, creation of gov intelligence, etc

2022-06-08 17:15:27 @bryanrbeal Its a big deal that the private sector rather than academia and governments are the ones best positioned to develop the most resource-intensive models. That means we get all the benefits of _capitalist_ AI, but we don't get the benefits of pro-social AI (diff incentives)

2022-06-08 17:14:27 @bryanrbeal The 'means of production' for things like large-scale AI systems are controlled by a tiny set of actors, and the number of organizations that can train large-scale systems are predominantly in the private sector. The resulting _services_ are broadly available

2022-06-08 17:06:35 A few weeks ago, I gave a presentation about the consequences of industrialization of AI with a particular emphasis on geopolitics. Slides here, for the enthusiasts! https://t.co/67s11LJRkV

2022-06-08 17:05:12 It's covered a bit in the above podcast by people like @katecrawford - there's huge implications to industrialization, mostly centering around who gets control of the frontier, when the frontier becomes resource intensive. So far control is accruing to the private sector (uh oh!) https://t.co/rVe5jtuCkw

2022-06-08 17:02:47 Talked with @TheEconomist about the industrialization of AI - a theme I've been covering for years in Import AI, and some of the implications of this which we @AnthropicAI laid out in 'Predictability and Surprise in Large Generative Models' https://t.co/6R2Qa4szOr . https://t.co/HsAkY77We9

2022-06-07 20:15:18 @alexhern @HaydnBelfield @mmitchell_ai @huggingface Again, this isn't how I think it _should_ be, it's just how the incentive landscape appears today.

2022-06-07 20:14:58 @alexhern @HaydnBelfield @mmitchell_ai @huggingface Well, some of these orgs seem to have implicit goal of 'make money', so it's as disappointing as the rest of capitalist-incentives, I suppose. Currently, there's a tradeoff they face here, and they haven't been incentivized via regulation/legal precedent to behave differently.

2022-06-07 20:12:42 @alexhern @HaydnBelfield Some of the language coming out of European Commission points in this direction, though I expect will need to be clarified via specific legal cases to create hard precedent. https://t.co/BgwBbEaBbU

2022-06-07 20:11:56 @alexhern @HaydnBelfield @mmitchell_ai @huggingface Ultimately I expect this is one of the big, gnarly things that will need to be decided via a suit in court. What rights do users have to edit the data about them represented in models trained on public internet, etc. Lots of preliminary work/regulation happening, but untested

2022-06-07 20:11:10 @alexhern @HaydnBelfield Yes, you can do this via a few research techniques. E.g, this paper is representative of some of the work going on in the field https://t.co/Dea4LI9t3t . The big question is whether an LM provider will _let you_ do this. iirc @mmitchell_ai is thinking about this at @huggingface

2022-06-07 20:06:02 RT @RANDCorporation: We’re thrilled to announce that Jason Matheny has been selected as our new president and CEO. He is an economist, tech…

2022-06-07 20:05:13 @alexhern @HaydnBelfield A lot of my personal experience in policy at places like OECD/UN/etc is that it's quite hard to put hard legal regulations around things in the abstract, but it's relatively easy to take specific use-cases and classify them within law, so AI might look like this.

2022-06-07 20:04:28 @alexhern @HaydnBelfield And some of it may ultimately be catalyzed via use-cases. E.g, DALL-E being used to generate things where the prompts don't seem to imply infringement of existing IP probably leads to novel outputs, but if your prompt is 'generate variations of mickey mouse' you may be infringing

2022-06-07 20:03:39 @alexhern @HaydnBelfield I mean there's a general desire to put some constraints around large-scale data scraping from public sources, and this runs into tension with concepts of 'fair use' of IP for creating new things. 1/2

2022-06-07 20:02:06 @alexhern @HaydnBelfield (I am also not a lawyer, but I've been trying to read about this particular intersection a lot. It seems like you can think about AI systems more sensibly when deployed within a corporate, profit-making system, but it gets harder when being done as DIY non-profit things)

2022-06-07 20:01:14 @alexhern @HaydnBelfield They sort of do and sort of don't - there's an interesting tension between stuff like fair use of data, and how we want to apply data protection to big, scraped datasets. I think a lot of the tension is in AI-Creativity because it's where these things come into tension.

2022-06-07 19:56:18 @alexhern @HaydnBelfield Neither am I! If you look at last hundred years there's been this general personification of corporations under law (creepy!), so that's the environment AI systems are getting built within. Consequently, AI systems seem to run into edge-case conflicts with much existing law.

2022-06-07 19:45:04 @alexhern @HaydnBelfield I'm not saying AI systems _should_ exhibit these qualities, just pointing out that they do currently exhibit these qualities.

2022-06-07 19:42:16 @alexhern @HaydnBelfield Humans also don't have an innate concept of data protection - we look at stuff and hear stuff that we're exposed to, and we naturally seek out things we're curious about regardless of implicit privacy (e.g, most humans love gossip)

2022-06-07 19:41:19 @alexhern @HaydnBelfield It feels tricky, as humans learn both on heavily curated curricula (e.g, school), and also uncurated mass trawled data through sensory experience/day to day life. Latter comprises vastly more than former. Same seems true of AI - some small % of curated data, large % uncurated

2022-06-07 04:03:11 @TrungTPhan Depends on how much the people are being paid and how tips are being split. Good tips typically turn shit wages into livable wages for service industry people. (Obvs depends - some businesses don't distribute tips properly so don't do that there)

2022-06-06 23:41:08 RT @0xabad1dea: just watched the AI generate an image that very clearly had the Shutterstock watermark on it and evolve it out over a few i…

2022-06-04 22:09:34 RT @ch402: People often complain that modern ML is throwing GPUs at problems without new research ideas. This is like finding evolution ugl…

2022-06-04 02:25:01 @mark_riedl Yeah, but I really liked seeing the atheist AI society stuff. Agree the flying snake stuff was hella confusing!

2022-06-04 02:16:16 Raised by Wolves is an amazing show unafraid to boldly pursue wacky ideas around AI/machine consciousness and the interaction with faith and religion. I really hope someone picks it up and does the third season. If anyone does this, I can guarantee I shall be an obsessive viewer! https://t.co/GTLFMupeSR

2022-06-03 22:45:50 @guyi @Plinz *counterintuitively

2022-06-03 22:45:30 @guyi @Plinz Yes, I think one culture is impossible, but it doesn't prevent it being an implicit goal for a bunch of people. Personally, I think variety and heterogeneity are both more useful and counterintuitive more stable

2022-06-03 22:37:39 @Plinz @guyi I'm going to write something lengthy about this in a bit, but worth noting my Twitter handle used to be 'mappingbabel' for a reason. I do think consequence of internet is this swing back and forth between one culture and a fractal infinity of them. Somewhat sad sometimes!

2022-06-03 18:37:00 @Plinz Yes, massively scary! That's why I mean it's a difficult and real choice - spectrum of no curation to fill curation, and different entities will pick different points

2022-06-03 07:25:39 @Plinz https://t.co/x9PQm7xTY7

2022-06-03 07:20:26 @BonesMcGowan My point here is if you're taking samples from a big generative neural net you'll typically sample stuff which represents the dataset, so the resulting sample will reflect the underlying reality you've encoded via your dataset selection.

2022-06-03 07:16:47 @Plinz Many big models are trained on many multiples of library of Congress yet still exhibit various negative traits

2022-06-03 07:13:23 @machinaexethica Yeah, there's a meaningful split here re agency of end-users vs system-developers. On 1 side u have 'anything goes' and on other u have 'defined 95%+ by system developers'. u see this in current diffs between OS releases (eg OPT, Eleuther) and hosted services (e.g OpenAI, Cohere)

2022-06-03 06:26:34 (and a bunch of others whose names I am forgetting!)

2022-06-03 06:25:49 Kudos to @timnitGebru @mmitchell_ai @rajiinio @jovialjoy @Abebab @IasonGabriel etc who have all done foundational work here

2022-06-03 05:45:53 LLMs are like media organizations - each one will be of a certain view and LLM providers will face a choice: be fully libertarian or choose for proactive reification of certain identities. Just as with human culture, there can't be a single culture that everyone feels home in.

2022-06-03 05:43:50 This article illustrates a significant point about AI - who gets represented? E.g, if you're developing an AI model, how does it categorize various classes, and how "correct" is it from POV of end users. End point seems like a fractal - as many LLMs as there are 'classes'. https://t.co/P28rvKmtQL

2022-06-03 03:17:30 @ElectionLegal Absolutely incredible. Everyone I showed this to was confused and horrified. A+ performance.

2022-06-02 19:26:42 RT @AmandaAskell: 1) Do you think the welfare of present people is more intrinsically morally valuable than the welfare of future people?2…

2022-06-02 05:47:25 RT @Ket_Cherie: @jackclarkSF Not really. It’s predictable.Different legal arguments for copyright (common law &

2022-06-01 17:04:59 @Ket_Cherie Thanks for clarifying - very helpful context! Do you think there are any particularly good texts or research papers on intersection between copyright and TM and AI-generated content I should read?

2022-06-01 17:00:05 @Ket_Cherie Does TM = trademark?

2022-06-01 16:16:41 @LordeCelsius You must imagine 'The Forbidden Mickey'

2022-06-01 16:14:08 @pwillsit Please finish this paper, would love to read! Or feel free to send me a draft. I think about this stuff a lot but don't have particular training in law so am v much dog on the internet here.

2022-06-01 16:10:44 Gonna be interesting to see what happens when AI generated art collides with notoriously litigious IP-protection companies like Disney and Nintendo. It'd be interesting to see how they'd react to a universe of Mickey variants, or a vast set of Peach permutations, etc.

2022-05-30 22:23:50 @iandanforth @gwern Is our culture more influenced by photos from Leica versus smartphones? Medias don't really die they just become less dominant

2022-05-30 21:39:49 @andy_l_jones @gwern Will write something when back from holiday!

2022-05-30 21:39:35 @andy_l_jones @gwern Maybe the scariest idea is that gen models will, pretty much by design, compose and synthesize outputs which people may not ever have imagined, let alone created

2022-05-30 21:38:46 @andy_l_jones @gwern Very good writeup. I've been pondering this stuff while taking a break and I'm more updating in the direction that gen models ultimately end up driving culture, so rather than polluting our own datasets, we're instead about to kinda bootstrap cultural production via artefacts

2022-05-26 20:48:46 RT @ethanCaballero: .@RichardSSutton estimates 50% probability of Human-Level AI by 2040: https://t.co/vCyE6delrT

2022-05-25 21:00:39 @mmitchell_ai Amazing

2022-05-24 01:53:10 @fieldsofcorn89 @quantumVerd @GoogleAI Video generation is probably gonna be quite a while because doing the sequential stuff over time makes the complexity scale significantly, so v expensive. Wouldn't be surprised if under a decade, though probably short time horizons.

2022-05-22 02:09:24 @norabelrose @patrickmineault Yeah I think retrieval from external trusted knowledge bases is gonna make all LLMs a ton better

2022-05-21 14:30:01 RT @jacob_feldgoise: National AI Research Resource (NAIRR) Task Force just approved their interim report! Very excited to read it.- Rele…

2022-05-20 08:11:00 CAFIAC FIX

2022-10-28 16:22:41 @jjspicer read this at once https://t.co/6M8qWJ2Akl

2022-10-27 21:30:29 RT @ClementDelangue: We just crossed 1,000,000 downloads of Stable Diffusion on the @huggingface hub! Congrats to Robin Rombach, Patrick Es…

2022-10-27 18:42:19 @tszzl I played Half Life: Alyx on the Valve Index and it hard-updated me to VR games being amazing, albeit still a little early.

2022-10-27 00:02:01 RT @rgblong: Another question for consciousness scientists and AI people: What is the best evidence for and against large language models…

2022-10-26 17:20:39 @MishaLaskin @junh_oh @RichiesOkTweets @djstrouse @Zergylord @filangelos @Maxime_Gazeau @him_sahni @VladMnih looks cool - link to the paper?

2022-10-26 03:29:56 @gijigae Thanks so much for reading!

2022-10-26 02:05:38 @kipperrii 100% here for fontposting. More fontposting!

2022-10-25 20:15:15 @carperai Haha fair enough. Thanks!

2022-10-25 19:25:12 @carperai You're welcome. Any clues as to release timeline?

2022-10-25 18:49:10 Stages of living in California:First earthquake: dear God hope I've made my Will.Fifth earthquake: Gosh, that was a little rumbly. Hope everyone is ok.???th earthquake (which happened in SF just now): Wonder what gifs people will post on Twitter about this one?

2022-10-25 17:14:50 RT @ElineCMC: .@jackclarkSF's take on this: "The tl

2022-10-25 15:30:28 @EricNewcomer @TaylorLorenz awesome metrics, congrats Eric! Really nice to see your success here : )

2022-10-25 01:52:36 @Meaningness @ArtirKel it's got a bunch of nice examples in it - definitely worth a skim

2022-10-25 01:45:29 One of the nice things about the AI space is when people take their criticisms and instantiate them as quantitative studies of existing systems - kudos to @GaryMarcus et al for a nice paper going over some failures in DALL-E2

2022-10-24 23:21:03 @tszzl Did it with my Senate testimony and was v surprised. Generally find myself running any long doc I write through a LM these days

2022-10-24 05:37:13 Import AI will be coming out on Tuesday as I sprained my ankle at Nopes' final show last night. #VOTEDIY2022 https://t.co/CaKrgLf00r

2022-10-24 01:47:22 RT @SamuelAlbanie: Just how striking are the recent language model results with Flan-PaLM?Here's a plot.Across 57 tasks on mathematics,…

2022-10-23 19:02:59 @CecilYongo Can you share slides? Looks interesting

2022-10-23 02:23:43 @zacharynado You can't necessarily look at the datasets with an API, unless has huge amounts of extra eng time invested. E.g can anyone inspect JFT?

2022-10-23 01:41:19 @kylebrussell Does Playbyte have an in-house exorcist team?

2022-10-23 00:05:13 Image generation going to get strange when most of the images we train on are synthetically generated. Feels like a classic 'tragedy of the commons'. (Yes, some have watermarks, but my sense is there's a race to the bottom on that kind of thing). https://t.co/A3QawDog90

2022-10-28 16:22:41 @jjspicer read this at once https://t.co/6M8qWJ2Akl

2022-10-27 21:30:29 RT @ClementDelangue: We just crossed 1,000,000 downloads of Stable Diffusion on the @huggingface hub! Congrats to Robin Rombach, Patrick Es…

2022-10-27 18:42:19 @tszzl I played Half Life: Alyx on the Valve Index and it hard-updated me to VR games being amazing, albeit still a little early.

2022-10-27 00:02:01 RT @rgblong: Another question for consciousness scientists and AI people: What is the best evidence for and against large language models…

2022-10-26 17:20:39 @MishaLaskin @junh_oh @RichiesOkTweets @djstrouse @Zergylord @filangelos @Maxime_Gazeau @him_sahni @VladMnih looks cool - link to the paper?

2022-10-26 03:29:56 @gijigae Thanks so much for reading!

2022-10-26 02:05:38 @kipperrii 100% here for fontposting. More fontposting!

2022-10-25 20:15:15 @carperai Haha fair enough. Thanks!

2022-10-25 19:25:12 @carperai You're welcome. Any clues as to release timeline?

2022-10-25 18:49:10 Stages of living in California:First earthquake: dear God hope I've made my Will.Fifth earthquake: Gosh, that was a little rumbly. Hope everyone is ok.???th earthquake (which happened in SF just now): Wonder what gifs people will post on Twitter about this one?

2022-10-25 17:14:50 RT @ElineCMC: .@jackclarkSF's take on this: "The tl

2022-10-25 15:30:28 @EricNewcomer @TaylorLorenz awesome metrics, congrats Eric! Really nice to see your success here : )

2022-10-25 01:52:36 @Meaningness @ArtirKel it's got a bunch of nice examples in it - definitely worth a skim

2022-10-25 01:45:29 One of the nice things about the AI space is when people take their criticisms and instantiate them as quantitative studies of existing systems - kudos to @GaryMarcus et al for a nice paper going over some failures in DALL-E2

2022-10-24 23:21:03 @tszzl Did it with my Senate testimony and was v surprised. Generally find myself running any long doc I write through a LM these days

2022-10-24 05:37:13 Import AI will be coming out on Tuesday as I sprained my ankle at Nopes' final show last night. #VOTEDIY2022 https://t.co/CaKrgLf00r

2022-10-24 01:47:22 RT @SamuelAlbanie: Just how striking are the recent language model results with Flan-PaLM?Here's a plot.Across 57 tasks on mathematics,…

2022-10-23 19:02:59 @CecilYongo Can you share slides? Looks interesting

2022-10-23 02:23:43 @zacharynado You can't necessarily look at the datasets with an API, unless has huge amounts of extra eng time invested. E.g can anyone inspect JFT?

2022-10-23 01:41:19 @kylebrussell Does Playbyte have an in-house exorcist team?

2022-10-23 00:05:13 Image generation going to get strange when most of the images we train on are synthetically generated. Feels like a classic 'tragedy of the commons'. (Yes, some have watermarks, but my sense is there's a race to the bottom on that kind of thing). https://t.co/A3QawDog90

2022-10-28 16:22:41 @jjspicer read this at once https://t.co/6M8qWJ2Akl

2022-10-27 21:30:29 RT @ClementDelangue: We just crossed 1,000,000 downloads of Stable Diffusion on the @huggingface hub! Congrats to Robin Rombach, Patrick Es…

2022-10-27 18:42:19 @tszzl I played Half Life: Alyx on the Valve Index and it hard-updated me to VR games being amazing, albeit still a little early.

2022-10-27 00:02:01 RT @rgblong: Another question for consciousness scientists and AI people: What is the best evidence for and against large language models…

2022-10-26 17:20:39 @MishaLaskin @junh_oh @RichiesOkTweets @djstrouse @Zergylord @filangelos @Maxime_Gazeau @him_sahni @VladMnih looks cool - link to the paper?

2022-10-26 03:29:56 @gijigae Thanks so much for reading!

2022-10-26 02:05:38 @kipperrii 100% here for fontposting. More fontposting!

2022-10-25 20:15:15 @carperai Haha fair enough. Thanks!

2022-10-25 19:25:12 @carperai You're welcome. Any clues as to release timeline?

2022-10-25 18:49:10 Stages of living in California:First earthquake: dear God hope I've made my Will.Fifth earthquake: Gosh, that was a little rumbly. Hope everyone is ok.???th earthquake (which happened in SF just now): Wonder what gifs people will post on Twitter about this one?

2022-10-25 17:14:50 RT @ElineCMC: .@jackclarkSF's take on this: "The tl

2022-10-25 15:30:28 @EricNewcomer @TaylorLorenz awesome metrics, congrats Eric! Really nice to see your success here : )

2022-10-25 01:52:36 @Meaningness @ArtirKel it's got a bunch of nice examples in it - definitely worth a skim

2022-10-25 01:45:29 One of the nice things about the AI space is when people take their criticisms and instantiate them as quantitative studies of existing systems - kudos to @GaryMarcus et al for a nice paper going over some failures in DALL-E2

2022-10-24 23:21:03 @tszzl Did it with my Senate testimony and was v surprised. Generally find myself running any long doc I write through a LM these days

2022-10-24 05:37:13 Import AI will be coming out on Tuesday as I sprained my ankle at Nopes' final show last night. #VOTEDIY2022 https://t.co/CaKrgLf00r

2022-10-24 01:47:22 RT @SamuelAlbanie: Just how striking are the recent language model results with Flan-PaLM?Here's a plot.Across 57 tasks on mathematics,…

2022-10-23 19:02:59 @CecilYongo Can you share slides? Looks interesting

2022-10-23 02:23:43 @zacharynado You can't necessarily look at the datasets with an API, unless has huge amounts of extra eng time invested. E.g can anyone inspect JFT?

2022-10-23 01:41:19 @kylebrussell Does Playbyte have an in-house exorcist team?

2022-10-23 00:05:13 Image generation going to get strange when most of the images we train on are synthetically generated. Feels like a classic 'tragedy of the commons'. (Yes, some have watermarks, but my sense is there's a race to the bottom on that kind of thing). https://t.co/A3QawDog90

2022-10-29 19:09:59 @rvinshit @porksmith Absolutely incredible. A+

2022-10-29 18:32:38 @xriskology Some people shut down because they find engaging to be traumatic

2022-10-28 16:22:41 @jjspicer read this at once https://t.co/6M8qWJ2Akl

2022-10-27 21:30:29 RT @ClementDelangue: We just crossed 1,000,000 downloads of Stable Diffusion on the @huggingface hub! Congrats to Robin Rombach, Patrick Es…

2022-10-27 18:42:19 @tszzl I played Half Life: Alyx on the Valve Index and it hard-updated me to VR games being amazing, albeit still a little early.

2022-10-27 00:02:01 RT @rgblong: Another question for consciousness scientists and AI people: What is the best evidence for and against large language models…

2022-10-26 17:20:39 @MishaLaskin @junh_oh @RichiesOkTweets @djstrouse @Zergylord @filangelos @Maxime_Gazeau @him_sahni @VladMnih looks cool - link to the paper?

2022-10-26 03:29:56 @gijigae Thanks so much for reading!

2022-10-26 02:05:38 @kipperrii 100% here for fontposting. More fontposting!

2022-10-25 20:15:15 @carperai Haha fair enough. Thanks!

2022-10-25 19:25:12 @carperai You're welcome. Any clues as to release timeline?

2022-10-25 18:49:10 Stages of living in California:First earthquake: dear God hope I've made my Will.Fifth earthquake: Gosh, that was a little rumbly. Hope everyone is ok.???th earthquake (which happened in SF just now): Wonder what gifs people will post on Twitter about this one?

2022-10-25 17:14:50 RT @ElineCMC: .@jackclarkSF's take on this: "The tl

2022-10-25 15:30:28 @EricNewcomer @TaylorLorenz awesome metrics, congrats Eric! Really nice to see your success here : )

2022-10-25 01:52:36 @Meaningness @ArtirKel it's got a bunch of nice examples in it - definitely worth a skim

2022-10-25 01:45:29 One of the nice things about the AI space is when people take their criticisms and instantiate them as quantitative studies of existing systems - kudos to @GaryMarcus et al for a nice paper going over some failures in DALL-E2

2022-10-24 23:21:03 @tszzl Did it with my Senate testimony and was v surprised. Generally find myself running any long doc I write through a LM these days

2022-10-24 05:37:13 Import AI will be coming out on Tuesday as I sprained my ankle at Nopes' final show last night. #VOTEDIY2022 https://t.co/CaKrgLf00r

2022-10-24 01:47:22 RT @SamuelAlbanie: Just how striking are the recent language model results with Flan-PaLM?Here's a plot.Across 57 tasks on mathematics,…

2022-10-23 19:02:59 @CecilYongo Can you share slides? Looks interesting

2022-10-23 02:23:43 @zacharynado You can't necessarily look at the datasets with an API, unless has huge amounts of extra eng time invested. E.g can anyone inspect JFT?

2022-10-23 01:41:19 @kylebrussell Does Playbyte have an in-house exorcist team?

2022-10-23 00:05:13 Image generation going to get strange when most of the images we train on are synthetically generated. Feels like a classic 'tragedy of the commons'. (Yes, some have watermarks, but my sense is there's a race to the bottom on that kind of thing). https://t.co/A3QawDog90

2022-10-30 04:45:04 My friend: yeah I totally shit the bed on that prop I made for the Halloween party. Did it in a couple of days. It's no good.The prop: https://t.co/zVsuYnWfqO

2022-10-29 19:09:59 @rvinshit @porksmith Absolutely incredible. A+

2022-10-29 18:32:38 @xriskology Some people shut down because they find engaging to be traumatic

2022-10-28 16:22:41 @jjspicer read this at once https://t.co/6M8qWJ2Akl

2022-10-27 21:30:29 RT @ClementDelangue: We just crossed 1,000,000 downloads of Stable Diffusion on the @huggingface hub! Congrats to Robin Rombach, Patrick Es…

2022-10-27 18:42:19 @tszzl I played Half Life: Alyx on the Valve Index and it hard-updated me to VR games being amazing, albeit still a little early.

2022-10-27 00:02:01 RT @rgblong: Another question for consciousness scientists and AI people: What is the best evidence for and against large language models…

2022-10-26 17:20:39 @MishaLaskin @junh_oh @RichiesOkTweets @djstrouse @Zergylord @filangelos @Maxime_Gazeau @him_sahni @VladMnih looks cool - link to the paper?

2022-10-26 03:29:56 @gijigae Thanks so much for reading!

2022-10-26 02:05:38 @kipperrii 100% here for fontposting. More fontposting!

2022-10-25 20:15:15 @carperai Haha fair enough. Thanks!

2022-10-25 19:25:12 @carperai You're welcome. Any clues as to release timeline?

2022-10-25 18:49:10 Stages of living in California:First earthquake: dear God hope I've made my Will.Fifth earthquake: Gosh, that was a little rumbly. Hope everyone is ok.???th earthquake (which happened in SF just now): Wonder what gifs people will post on Twitter about this one?

2022-10-25 17:14:50 RT @ElineCMC: .@jackclarkSF's take on this: "The tl

2022-10-25 15:30:28 @EricNewcomer @TaylorLorenz awesome metrics, congrats Eric! Really nice to see your success here : )

2022-10-25 01:52:36 @Meaningness @ArtirKel it's got a bunch of nice examples in it - definitely worth a skim

2022-10-25 01:45:29 One of the nice things about the AI space is when people take their criticisms and instantiate them as quantitative studies of existing systems - kudos to @GaryMarcus et al for a nice paper going over some failures in DALL-E2

2022-10-24 23:21:03 @tszzl Did it with my Senate testimony and was v surprised. Generally find myself running any long doc I write through a LM these days

2022-10-24 05:37:13 Import AI will be coming out on Tuesday as I sprained my ankle at Nopes' final show last night. #VOTEDIY2022 https://t.co/CaKrgLf00r

2022-10-24 01:47:22 RT @SamuelAlbanie: Just how striking are the recent language model results with Flan-PaLM?Here's a plot.Across 57 tasks on mathematics,…

2022-10-23 19:02:59 @CecilYongo Can you share slides? Looks interesting

2022-10-23 02:23:43 @zacharynado You can't necessarily look at the datasets with an API, unless has huge amounts of extra eng time invested. E.g can anyone inspect JFT?

2022-10-23 01:41:19 @kylebrussell Does Playbyte have an in-house exorcist team?

2022-10-23 00:05:13 Image generation going to get strange when most of the images we train on are synthetically generated. Feels like a classic 'tragedy of the commons'. (Yes, some have watermarks, but my sense is there's a race to the bottom on that kind of thing). https://t.co/A3QawDog90

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-11-17 18:15:03 Demos are the best way to 'show, not tell' people about the current state of AI. I also find demos are one of the best ways to talk about contemporary safety issues with people - one thing to read about safety, another to see a model break in front of you in a wild way. A+ https://t.co/Wsku7NPtVA

2022-11-17 18:06:12 @arxiv this absolutely rules. I am going to now buy even more arxiv merch to celebrate. congrats both!

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-11-17 18:15:03 Demos are the best way to 'show, not tell' people about the current state of AI. I also find demos are one of the best ways to talk about contemporary safety issues with people - one thing to read about safety, another to see a model break in front of you in a wild way. A+ https://t.co/Wsku7NPtVA

2022-11-17 18:06:12 @arxiv this absolutely rules. I am going to now buy even more arxiv merch to celebrate. congrats both!

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-11-18 17:24:03 RT @irinarish: Thrilled to announce that our joint proposal (@Mila_Quebec @laion_ai @AiEleuther) on "Scalable Foundation Models for Transfe…

2022-11-17 18:15:03 Demos are the best way to 'show, not tell' people about the current state of AI. I also find demos are one of the best ways to talk about contemporary safety issues with people - one thing to read about safety, another to see a model break in front of you in a wild way. A+ https://t.co/Wsku7NPtVA

2022-11-17 18:06:12 @arxiv this absolutely rules. I am going to now buy even more arxiv merch to celebrate. congrats both!

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-11-19 19:45:33 @BlancheMinerva Not the paper you're looking for, but this paper perhaps might cite it https://t.co/wD5PCL4Sa9

2022-11-19 17:07:27 Given that each newsletter entails reading arXiv for approximately 10-15 hours, I've read arXiv for at least 3000 hours in the past six years, likely much more. FOR FREE. It is an incomparably brilliant thing and we should all support it.

2022-11-19 17:05:12 I just bought myself a load of @arxiv merch for Christmas - and you can too! https://t.co/AKFrx6e045 This is a great way to support one of the most important open infrastructures of the AI ecosystem. Plus, you'll be swagged out for the festive season.

2022-11-19 12:23:24 Aeon Station - Leaves

2022-11-18 17:24:03 RT @irinarish: Thrilled to announce that our joint proposal (@Mila_Quebec @laion_ai @AiEleuther) on "Scalable Foundation Models for Transfe…

2022-11-17 18:15:03 Demos are the best way to 'show, not tell' people about the current state of AI. I also find demos are one of the best ways to talk about contemporary safety issues with people - one thing to read about safety, another to see a model break in front of you in a wild way. A+ https://t.co/Wsku7NPtVA

2022-11-17 18:06:12 @arxiv this absolutely rules. I am going to now buy even more arxiv merch to celebrate. congrats both!

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-11-19 19:45:33 @BlancheMinerva Not the paper you're looking for, but this paper perhaps might cite it https://t.co/wD5PCL4Sa9

2022-11-19 17:07:27 Given that each newsletter entails reading arXiv for approximately 10-15 hours, I've read arXiv for at least 3000 hours in the past six years, likely much more. FOR FREE. It is an incomparably brilliant thing and we should all support it.

2022-11-19 17:05:12 I just bought myself a load of @arxiv merch for Christmas - and you can too! https://t.co/AKFrx6e045 This is a great way to support one of the most important open infrastructures of the AI ecosystem. Plus, you'll be swagged out for the festive season.

2022-11-19 12:23:24 Aeon Station - Leaves

2022-11-18 17:24:03 RT @irinarish: Thrilled to announce that our joint proposal (@Mila_Quebec @laion_ai @AiEleuther) on "Scalable Foundation Models for Transfe…

2022-11-17 18:15:03 Demos are the best way to 'show, not tell' people about the current state of AI. I also find demos are one of the best ways to talk about contemporary safety issues with people - one thing to read about safety, another to see a model break in front of you in a wild way. A+ https://t.co/Wsku7NPtVA

2022-11-17 18:06:12 @arxiv this absolutely rules. I am going to now buy even more arxiv merch to celebrate. congrats both!

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-11-19 19:45:33 @BlancheMinerva Not the paper you're looking for, but this paper perhaps might cite it https://t.co/wD5PCL4Sa9

2022-11-19 17:07:27 Given that each newsletter entails reading arXiv for approximately 10-15 hours, I've read arXiv for at least 3000 hours in the past six years, likely much more. FOR FREE. It is an incomparably brilliant thing and we should all support it.

2022-11-19 17:05:12 I just bought myself a load of @arxiv merch for Christmas - and you can too! https://t.co/AKFrx6e045 This is a great way to support one of the most important open infrastructures of the AI ecosystem. Plus, you'll be swagged out for the festive season.

2022-11-19 12:23:24 Aeon Station - Leaves

2022-11-18 17:24:03 RT @irinarish: Thrilled to announce that our joint proposal (@Mila_Quebec @laion_ai @AiEleuther) on "Scalable Foundation Models for Transfe…

2022-11-17 18:15:03 Demos are the best way to 'show, not tell' people about the current state of AI. I also find demos are one of the best ways to talk about contemporary safety issues with people - one thing to read about safety, another to see a model break in front of you in a wild way. A+ https://t.co/Wsku7NPtVA

2022-11-17 18:06:12 @arxiv this absolutely rules. I am going to now buy even more arxiv merch to celebrate. congrats both!

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-11-19 19:45:33 @BlancheMinerva Not the paper you're looking for, but this paper perhaps might cite it https://t.co/wD5PCL4Sa9

2022-11-19 17:07:27 Given that each newsletter entails reading arXiv for approximately 10-15 hours, I've read arXiv for at least 3000 hours in the past six years, likely much more. FOR FREE. It is an incomparably brilliant thing and we should all support it.

2022-11-19 17:05:12 I just bought myself a load of @arxiv merch for Christmas - and you can too! https://t.co/AKFrx6e045 This is a great way to support one of the most important open infrastructures of the AI ecosystem. Plus, you'll be swagged out for the festive season.

2022-11-19 12:23:24 Aeon Station - Leaves

2022-11-18 17:24:03 RT @irinarish: Thrilled to announce that our joint proposal (@Mila_Quebec @laion_ai @AiEleuther) on "Scalable Foundation Models for Transfe…

2022-11-17 18:15:03 Demos are the best way to 'show, not tell' people about the current state of AI. I also find demos are one of the best ways to talk about contemporary safety issues with people - one thing to read about safety, another to see a model break in front of you in a wild way. A+ https://t.co/Wsku7NPtVA

2022-11-17 18:06:12 @arxiv this absolutely rules. I am going to now buy even more arxiv merch to celebrate. congrats both!

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-11-23 05:45:15 RIP Mimi Parker of Low - an incredible singer, and a member of one of the few bands that has caused me to ugly cry at a live gig. You were brilliant and you were loved. https://t.co/Izelb0AEAC https://t.co/Zbytbt1yUA

2022-11-23 04:59:03 Jane, don't you know me? By Elvis Depressedly is another Sad Banger. A+! https://t.co/YMPeGxmbJF https://t.co/TTLb1JIrqw

2022-11-19 19:45:33 @BlancheMinerva Not the paper you're looking for, but this paper perhaps might cite it https://t.co/wD5PCL4Sa9

2022-11-19 17:07:27 Given that each newsletter entails reading arXiv for approximately 10-15 hours, I've read arXiv for at least 3000 hours in the past six years, likely much more. FOR FREE. It is an incomparably brilliant thing and we should all support it.

2022-11-19 17:05:12 I just bought myself a load of @arxiv merch for Christmas - and you can too! https://t.co/AKFrx6e045 This is a great way to support one of the most important open infrastructures of the AI ecosystem. Plus, you'll be swagged out for the festive season.

2022-11-19 12:23:24 Aeon Station - Leaves

2022-11-18 17:24:03 RT @irinarish: Thrilled to announce that our joint proposal (@Mila_Quebec @laion_ai @AiEleuther) on "Scalable Foundation Models for Transfe…

2022-11-17 18:15:03 Demos are the best way to 'show, not tell' people about the current state of AI. I also find demos are one of the best ways to talk about contemporary safety issues with people - one thing to read about safety, another to see a model break in front of you in a wild way. A+ https://t.co/Wsku7NPtVA

2022-11-17 18:06:12 @arxiv this absolutely rules. I am going to now buy even more arxiv merch to celebrate. congrats both!

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-11-25 21:11:54 RT @nearcyan: few people seem to understand the *real* reasons why powerful AI image generation models are usually unreleased this is one…

2022-11-25 20:54:49 @dinabass Honestly thought USA defense was very good

2022-11-25 20:53:00 Replace England manager with Head of Lettuce

2022-11-25 05:28:00 Happy Thanksgiving, twitter! https://t.co/LvUU54vDiU

2022-11-23 05:45:15 RIP Mimi Parker of Low - an incredible singer, and a member of one of the few bands that has caused me to ugly cry at a live gig. You were brilliant and you were loved. https://t.co/Izelb0AEAC https://t.co/Zbytbt1yUA

2022-11-23 04:59:03 Jane, don't you know me? By Elvis Depressedly is another Sad Banger. A+! https://t.co/YMPeGxmbJF https://t.co/TTLb1JIrqw

2022-11-19 19:45:33 @BlancheMinerva Not the paper you're looking for, but this paper perhaps might cite it https://t.co/wD5PCL4Sa9

2022-11-19 17:07:27 Given that each newsletter entails reading arXiv for approximately 10-15 hours, I've read arXiv for at least 3000 hours in the past six years, likely much more. FOR FREE. It is an incomparably brilliant thing and we should all support it.

2022-11-19 17:05:12 I just bought myself a load of @arxiv merch for Christmas - and you can too! https://t.co/AKFrx6e045 This is a great way to support one of the most important open infrastructures of the AI ecosystem. Plus, you'll be swagged out for the festive season.

2022-11-19 12:23:24 Aeon Station - Leaves

2022-11-18 17:24:03 RT @irinarish: Thrilled to announce that our joint proposal (@Mila_Quebec @laion_ai @AiEleuther) on "Scalable Foundation Models for Transfe…

2022-11-17 18:15:03 Demos are the best way to 'show, not tell' people about the current state of AI. I also find demos are one of the best ways to talk about contemporary safety issues with people - one thing to read about safety, another to see a model break in front of you in a wild way. A+ https://t.co/Wsku7NPtVA

2022-11-17 18:06:12 @arxiv this absolutely rules. I am going to now buy even more arxiv merch to celebrate. congrats both!

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-11-25 21:11:54 RT @nearcyan: few people seem to understand the *real* reasons why powerful AI image generation models are usually unreleased this is one…

2022-11-25 20:54:49 @dinabass Honestly thought USA defense was very good

2022-11-25 20:53:00 Replace England manager with Head of Lettuce

2022-11-25 05:28:00 Happy Thanksgiving, twitter! https://t.co/LvUU54vDiU

2022-11-23 05:45:15 RIP Mimi Parker of Low - an incredible singer, and a member of one of the few bands that has caused me to ugly cry at a live gig. You were brilliant and you were loved. https://t.co/Izelb0AEAC https://t.co/Zbytbt1yUA

2022-11-23 04:59:03 Jane, don't you know me? By Elvis Depressedly is another Sad Banger. A+! https://t.co/YMPeGxmbJF https://t.co/TTLb1JIrqw

2022-11-19 19:45:33 @BlancheMinerva Not the paper you're looking for, but this paper perhaps might cite it https://t.co/wD5PCL4Sa9

2022-11-19 17:07:27 Given that each newsletter entails reading arXiv for approximately 10-15 hours, I've read arXiv for at least 3000 hours in the past six years, likely much more. FOR FREE. It is an incomparably brilliant thing and we should all support it.

2022-11-19 17:05:12 I just bought myself a load of @arxiv merch for Christmas - and you can too! https://t.co/AKFrx6e045 This is a great way to support one of the most important open infrastructures of the AI ecosystem. Plus, you'll be swagged out for the festive season.

2022-11-19 12:23:24 Aeon Station - Leaves

2022-11-18 17:24:03 RT @irinarish: Thrilled to announce that our joint proposal (@Mila_Quebec @laion_ai @AiEleuther) on "Scalable Foundation Models for Transfe…

2022-11-17 18:15:03 Demos are the best way to 'show, not tell' people about the current state of AI. I also find demos are one of the best ways to talk about contemporary safety issues with people - one thing to read about safety, another to see a model break in front of you in a wild way. A+ https://t.co/Wsku7NPtVA

2022-11-17 18:06:12 @arxiv this absolutely rules. I am going to now buy even more arxiv merch to celebrate. congrats both!

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-11-25 21:11:54 RT @nearcyan: few people seem to understand the *real* reasons why powerful AI image generation models are usually unreleased this is one…

2022-11-25 20:54:49 @dinabass Honestly thought USA defense was very good

2022-11-25 20:53:00 Replace England manager with Head of Lettuce

2022-11-25 05:28:00 Happy Thanksgiving, twitter! https://t.co/LvUU54vDiU

2022-11-23 05:45:15 RIP Mimi Parker of Low - an incredible singer, and a member of one of the few bands that has caused me to ugly cry at a live gig. You were brilliant and you were loved. https://t.co/Izelb0AEAC https://t.co/Zbytbt1yUA

2022-11-23 04:59:03 Jane, don't you know me? By Elvis Depressedly is another Sad Banger. A+! https://t.co/YMPeGxmbJF https://t.co/TTLb1JIrqw

2022-11-19 19:45:33 @BlancheMinerva Not the paper you're looking for, but this paper perhaps might cite it https://t.co/wD5PCL4Sa9

2022-11-19 17:07:27 Given that each newsletter entails reading arXiv for approximately 10-15 hours, I've read arXiv for at least 3000 hours in the past six years, likely much more. FOR FREE. It is an incomparably brilliant thing and we should all support it.

2022-11-19 17:05:12 I just bought myself a load of @arxiv merch for Christmas - and you can too! https://t.co/AKFrx6e045 This is a great way to support one of the most important open infrastructures of the AI ecosystem. Plus, you'll be swagged out for the festive season.

2022-11-19 12:23:24 Aeon Station - Leaves

2022-11-18 17:24:03 RT @irinarish: Thrilled to announce that our joint proposal (@Mila_Quebec @laion_ai @AiEleuther) on "Scalable Foundation Models for Transfe…

2022-11-17 18:15:03 Demos are the best way to 'show, not tell' people about the current state of AI. I also find demos are one of the best ways to talk about contemporary safety issues with people - one thing to read about safety, another to see a model break in front of you in a wild way. A+ https://t.co/Wsku7NPtVA

2022-11-17 18:06:12 @arxiv this absolutely rules. I am going to now buy even more arxiv merch to celebrate. congrats both!

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-11-25 21:11:54 RT @nearcyan: few people seem to understand the *real* reasons why powerful AI image generation models are usually unreleased this is one…

2022-11-25 20:54:49 @dinabass Honestly thought USA defense was very good

2022-11-25 20:53:00 Replace England manager with Head of Lettuce

2022-11-25 05:28:00 Happy Thanksgiving, twitter! https://t.co/LvUU54vDiU

2022-11-23 05:45:15 RIP Mimi Parker of Low - an incredible singer, and a member of one of the few bands that has caused me to ugly cry at a live gig. You were brilliant and you were loved. https://t.co/Izelb0AEAC https://t.co/Zbytbt1yUA

2022-11-23 04:59:03 Jane, don't you know me? By Elvis Depressedly is another Sad Banger. A+! https://t.co/YMPeGxmbJF https://t.co/TTLb1JIrqw

2022-11-19 19:45:33 @BlancheMinerva Not the paper you're looking for, but this paper perhaps might cite it https://t.co/wD5PCL4Sa9

2022-11-19 17:07:27 Given that each newsletter entails reading arXiv for approximately 10-15 hours, I've read arXiv for at least 3000 hours in the past six years, likely much more. FOR FREE. It is an incomparably brilliant thing and we should all support it.

2022-11-19 17:05:12 I just bought myself a load of @arxiv merch for Christmas - and you can too! https://t.co/AKFrx6e045 This is a great way to support one of the most important open infrastructures of the AI ecosystem. Plus, you'll be swagged out for the festive season.

2022-11-19 12:23:24 Aeon Station - Leaves

2022-11-18 17:24:03 RT @irinarish: Thrilled to announce that our joint proposal (@Mila_Quebec @laion_ai @AiEleuther) on "Scalable Foundation Models for Transfe…

2022-11-17 18:15:03 Demos are the best way to 'show, not tell' people about the current state of AI. I also find demos are one of the best ways to talk about contemporary safety issues with people - one thing to read about safety, another to see a model break in front of you in a wild way. A+ https://t.co/Wsku7NPtVA

2022-11-17 18:06:12 @arxiv this absolutely rules. I am going to now buy even more arxiv merch to celebrate. congrats both!

2022-11-16 02:09:12 @vaibhavk97 @EMostaque Mostly because I spent a lot of time thinking about size of clusters owned by public sector versus private sector, so always interesting to get new data points. Additionally, some orgs are compute-heavy and some are compute-light, so interesting to track specific numbers.

2022-11-16 02:02:57 @kohjingyu @EMostaque https://t.co/7XUXCZMFYy

2022-11-16 01:29:46 Stability AI (people behind Stable Diffusion and an upcoming Chinchilla -optimal code model) now have 5408 GPUs, up from 4000 earlier this year - per @EMostaque in a Reddit ama

2022-12-07 18:50:16 @HaydnBelfield There's definitely a tax, the main q is now painful the tax is versus how much you're willing to spend to gain the capability. As this paper notes, tax on decentralized training has been reducing over time

2022-12-07 16:57:32 @alexeyguzey @anguzey Congratulations. America+++!

2022-12-07 18:50:16 @HaydnBelfield There's definitely a tax, the main q is now painful the tax is versus how much you're willing to spend to gain the capability. As this paper notes, tax on decentralized training has been reducing over time

2022-12-07 16:57:32 @alexeyguzey @anguzey Congratulations. America+++!

2022-12-08 15:30:38 @PetarV_93 @andreeadeac22 looks very interesting. quick q - what are the key diffs between v1 and v2 of paper? might write up for import ai and would find pointers (pun intended!) helpful

2022-12-08 13:15:57 RT @RhysLindmark: 1/ Live tweeting talk from @AnthropicAI @AmandaAskell at @southpkcommons. Moderated by the excellent @soniajoseph_! htt…

2022-12-07 18:50:16 @HaydnBelfield There's definitely a tax, the main q is now painful the tax is versus how much you're willing to spend to gain the capability. As this paper notes, tax on decentralized training has been reducing over time

2022-12-07 16:57:32 @alexeyguzey @anguzey Congratulations. America+++!

2022-12-08 15:30:38 @PetarV_93 @andreeadeac22 looks very interesting. quick q - what are the key diffs between v1 and v2 of paper? might write up for import ai and would find pointers (pun intended!) helpful

2022-12-08 13:15:57 RT @RhysLindmark: 1/ Live tweeting talk from @AnthropicAI @AmandaAskell at @southpkcommons. Moderated by the excellent @soniajoseph_! htt…

2022-12-07 18:50:16 @HaydnBelfield There's definitely a tax, the main q is now painful the tax is versus how much you're willing to spend to gain the capability. As this paper notes, tax on decentralized training has been reducing over time

2022-12-07 16:57:32 @alexeyguzey @anguzey Congratulations. America+++!

2022-12-08 15:30:38 @PetarV_93 @andreeadeac22 looks very interesting. quick q - what are the key diffs between v1 and v2 of paper? might write up for import ai and would find pointers (pun intended!) helpful

2022-12-08 13:15:57 RT @RhysLindmark: 1/ Live tweeting talk from @AnthropicAI @AmandaAskell at @southpkcommons. Moderated by the excellent @soniajoseph_! htt…