Geoffrey Irving @geoffreyirving
Research Director at the UK AI Safety Institute (AISI). Previously DeepMind, OpenAI, Google Brain, etc. @[email protected] naml.us/blog London Joined September 2009-
Tweets3K
-
Followers8K
-
Following259
-
Likes11K
What’s the best class of O(1)-parameterized distributions that decently model *trained* weight matrices in neural networks (either generally or for transformers specifically)?
Lovely news to get on the morning of my first day at the UK AI Safety Institute. :)
Mech interp has been very successful in tiny models, but does it scale? …Kinda! Our new @GoogleDeepMind paper studies how Chinchilla70B can do multiple-choice Qs, focusing on picking the correct letter. Small model techniques mostly work but it's messy!🧵arxiv.org/abs/2307.09458
I’m surprised people need to hear this, but: If you’re considering whether to join a company, you should not sign a statement preventing you from talking to people with concerns about that company.
Part of AI alignment is picking tasks on which, if you do really well, the outcome is good.
What is the Mastodon instance that is simultaneously canonical for ML, EA, and Math Twitter?
Something related to one of @littmath's tweets.
Richard Ngo @RichardMCNgo
35K Followers 1K Following What would we need to understand in order to design an amazing future? Figuring that out @openaiGoogle DeepMind @GoogleDeepMind
944K Followers 275 Following We’re a team of scientists, engineers, ethicists and more, committed to solving intelligence, to advance science and benefit humanity.Miles Brundage @Miles_Brundage
43K Followers 10K Following Policy research at @openai. I mostly tweet about AI, animals, and sci-fi. He/him. Views my own.Michael Nielsen @michael_nielsen
96K Followers 6K Following Searching for the numinous 🇦🇺 🇨🇦, home in 🇺🇸 Research @AsteraInstitute https://t.co/maezekzRUbAmanda Askell @AmandaAskell
26K Followers 653 Following Philosopher & ethicist teaching models to be good @AnthropicAI. Personal account. All opinions come from my training data.Rob Miles (✈️ Tok.. @robertskmiles
18K Followers 790 Following Explaining AI Alignment to anyone who'll stand still for long enough, on YouTube and Discord. Music, movies, microcode, and high-speed pizza deliveryRob Bensinger ⏹️ @robbensinger
8K Followers 302 Following Comms @MIRIBerkeley. RT = increased vague psychological association between myself and the tweet.Jack Clark @jackclarkSF
68K Followers 5K Following @AnthropicAI, ONEAI OECD, co-chair @indexingai, writer @ https://t.co/3vmtHYkaTu Past: @openai, @business @theregister. Neural nets, distributed systems, weird futuresStefan Schubert @StefanFSchubert
28K Followers 2K Following Philosophy, psychology, and effective altruism.Jan Leike @janleike
44K Followers 322 Following ML Researcher, co-leading Superalignment @OpenAI. Optimizing for a post-AGI future where humanity flourishes.Neel Nanda @NeelNanda5
13K Followers 89 Following Mechanistic Interpretability lead @DeepMind. Formerly @AnthropicAI, independent. In this to reduce AI X-risk. Neural networks can be understood, let's go do it!David Krueger @DavidSKrueger
13K Followers 4K Following Cambridge faculty - AI alignment, deep learning, and existential safety. Formerly Mila, FHI, DeepMind, ElementAI, AISI.Kelsey Piper @KelseyTuoc
27K Followers 544 Following Senior writer at Vox's Future Perfect. [email protected]Catherine Olsson @catherineols
15K Followers 1K Following Hanging out with Claude, improving its behavior, and building tools to support that @AnthropicAI 😁 prev: @open_phil @googlebrain @openai (@microcovid)Sam Bowman @sleepinyourhat
35K Followers 3K Following AI alignment + LLMs at NYU & Anthropic. Views not employers'. No relation to @s8mb. I think you should join @givingwhatwecan.Peter Wildeford @peterwildeford
10K Followers 367 Following Pro forecaster w/ good track record. Seeking to understand + manage risks from advanced AI systems. - Co-CEO @RethinkPriors - Chief Advisory Executive @iapsAIkenshin_the_great @KenshinThe1337
0 Followers 88 FollowingPaylz @paylza
144 Followers 2K Following The best online market for digital downloads with best prices.Joe Skinner @joecskinner
5 Followers 57 FollowingNikita @nikitavoloboev
4K Followers 7K Following Make @LearnAnything_ Learn in public: https://t.co/GbFvuErkYn macOS course: https://t.co/JdbJWru6zG https://t.co/94R8ER7K2h https://t.co/ROkqhyhpEKAlphOmega @AlphOmegaTk
40 Followers 636 Following Πolitics / Crypto / Programming / Gaming / WhatevsKarolina Stanczak @karstanczak
515 Followers 446 Following NLP & ML PhD candidate @uni_copenhagen @CopeNLU✾ @acity_cap
2 Followers 118 FollowingAlexandru Tifrea @alexandrutifrea
169 Followers 322 FollowingRohan Gupta @ggrohdg
0 Followers 80 FollowingMatt Clifford @matthewclifford
25K Followers 2K Following Co-founder @join_ef; Chair @ARIA_Research; co-led AI Safety Summit at Bletchley ParkSoroush Ebadian @SoroushEbadian
132 Followers 228 Following In-between Computer Science, Innovation, and Violin.Louis Matha @loulouAI0662
4 Followers 45 FollowingIlia @IliaTeimouri
3 Followers 79 FollowingAnonymous Founder @anonymfounder
386 Followers 7K Following My startup diary. From startups to marketing, finance to entrepreneurship........and cryptocurrency.aj @awsedrftaj
10 Followers 108 FollowingTom Bouthillet 🇺�.. @tbouthillet
8K Followers 4K Following Co-Survivor • Business Development Manager • Battalion Chief of EMS (Retired) • Aspiring Screenwriter • Citizen of U.S., Canada, Ireland • ECGs • YouTube 👇🏻Amanda Cercas Curry @CurriedAmanda
576 Followers 739 Following Postdoc @MilaNLProc | Philosopher of Swift | Cohost of @letschatethicsSamuel Pyeng(GoDeihPi.. @SamuelPyang23
117 Followers 562 FollowingPatrick Dillon @mpdillon
20K Followers 5K Following "Daad!" Big fan @jomalleydillon/@jod46. @ObamaWhiteHouse, @Georgetown, @GUPolitics, etc. Texan in DC (El Paso forever). Tweeting is a bad idea but mine alone.Volodymyr Volkov @lepricon85
21 Followers 416 FollowingHorizon Events @HorizonEvents9
11 Followers 270 Following Events consultancy dedicated to advancing R&D in AI safetyAnaïs Moutot @AnaisMoutot
5K Followers 5K Following Journaliste @LesEchosWeekEnd / ex-correspondante à San Francisco (2016-2020) / [email protected]Smart Cities Connect @smartcityc
17K Followers 18K Following We provide meaningful content and connect a thoughtful community of decision-makers to empower smart cities at all stages of growth. Powered by @techconnect360.Kevin Nejad @kevin_nejad
293 Followers 2K Following PhD ML & Comp.Neuro @UniofOxford prev:@UofBristol, prev. @nyuniversity, MSc Applied Maths @EdinburghUni ,BSc CompSci @KingsCollegeLon✦✦✦ @not_infinite___
38 Followers 379 FollowingAkash Bajwa @AkashBajwa96
2K Followers 1K Following Investing in B2B software & fintech @EarlybirdVC. Writing about SaaS & fintech @ https://t.co/K2Ge60PyQQ.Vincent Zhang @centzh
17 Followers 690 Following policy enthusiast. recovering CS student. writing (mostly sharing stuff) on tech + society | All views are my own.sabrinaprovenzani @sabriprovenzani
925 Followers 2K Following Vice President @FPALondon bylines @fattoquotidiano @allthecitizens @espressonline @primaonline @AMINaOdv. Views my own. Pro kemmer. #VatiLeaksFAITH lN JESUS CHRIST @dabaalwayswinn1
226 Followers 3K Following Evangelist, Researcher and ConsultantReza Sayar @iamRezaSayar
168 Followers 673 Following 👨🏻🎓Life-long Learner👨🏻🎓 Kindness❤️, Helpfulness🫂 , AI🧠 & Reggaetón💃🏻Auventic, Inc. @auventic
19 Followers 75 Following Auventic stands at the forefront of Al safety, ensuring technology enhances, not overshadows, human potential.Vincent Manancourt @vmanancourt
7K Followers 2K Following Senior tech reporter @POLITICOEurope in London. [email protected]. DM for phone numberKrueger AI Safety Lab @kasl_ai
257 Followers 51 Following We are a research group at the University of Cambridge focused on avoiding catastrophic risks from AI.rupalim Sarma @rupalims
15 Followers 262 FollowingM.L.S @_mlspace
24 Followers 72 FollowingJoern Stoehler ⏹️ @JStoehler
78 Followers 193 Following math, physics, AI alignment alignment is too hard, we should do governance instead leave me anonymous feedback at https://t.co/A1Prj0teYXByeRose @byerose365
0 Followers 520 FollowingChris Wood @C_H_Wood
746 Followers 4K Following Building in AI & Consumer @earlywormapp Prev worked as SWE @ Meta building FB ads and LLM codegen tools. @HopkinsNanjing @UVAChris Kihereko @CKihereko
98 Followers 457 FollowingRuth Kaufmann Wolfe @rkaufmannwolfe
26 Followers 259 FollowingEliezer Yudkowsky ⏹.. @ESYudkowsky
175K Followers 89 Following The original AI alignment person. Missing punctuation at the end of a sentence means it's humor. If you're not sure, it's also very likely humor.Google DeepMind @GoogleDeepMind
944K Followers 275 Following We’re a team of scientists, engineers, ethicists and more, committed to solving intelligence, to advance science and benefit humanity.Michael Nielsen @michael_nielsen
96K Followers 6K Following Searching for the numinous 🇦🇺 🇨🇦, home in 🇺🇸 Research @AsteraInstitute https://t.co/maezekzRUbAmanda Askell @AmandaAskell
26K Followers 653 Following Philosopher & ethicist teaching models to be good @AnthropicAI. Personal account. All opinions come from my training data.Anthropic @AnthropicAI
262K Followers 26 Following We're an AI safety and research company that builds reliable, interpretable, and steerable AI systems. Talk to our AI assistant Claude at https://t.co/aRbQ97uk4d.Rob Bensinger ⏹️ @robbensinger
8K Followers 302 Following Comms @MIRIBerkeley. RT = increased vague psychological association between myself and the tweet.Jack Clark @jackclarkSF
68K Followers 5K Following @AnthropicAI, ONEAI OECD, co-chair @indexingai, writer @ https://t.co/3vmtHYkaTu Past: @openai, @business @theregister. Neural nets, distributed systems, weird futuresJan Leike @janleike
44K Followers 322 Following ML Researcher, co-leading Superalignment @OpenAI. Optimizing for a post-AGI future where humanity flourishes.Kelsey Piper @KelseyTuoc
27K Followers 544 Following Senior writer at Vox's Future Perfect. [email protected]Catherine Olsson @catherineols
15K Followers 1K Following Hanging out with Claude, improving its behavior, and building tools to support that @AnthropicAI 😁 prev: @open_phil @googlebrain @openai (@microcovid)Sam Bowman @sleepinyourhat
35K Followers 3K Following AI alignment + LLMs at NYU & Anthropic. Views not employers'. No relation to @s8mb. I think you should join @givingwhatwecan.Robert Wiblin @robertwiblin
34K Followers 643 Following Exploring the inviolate sphere of ideas one interview at a time: https://t.co/2YMw00bkIQHabiba @FreshMangoLassi
4K Followers 523 Following Co-founder @SpiroTB - new TB screening and prevention charity focused on children https://t.co/sBf6ONGMSLAlexander Berger @albrgr
11K Followers 2K Following Enjoys a good applied micro paper. CEO of @open_phil. Views my own, tweets self-destruct every once in a while.Katja Grace 🔍 @KatjaGrace
8K Followers 798 Following Thinking about whether AI will destroy the world at https://t.co/pMilDvd4ya. DM or email for media requests. Feedback: https://t.co/zGAm1i7SKHOriol Vinyals @OriolVinyalsML
167K Followers 82 Following VP of Research & Deep Learning Lead, Google DeepMind. Gemini co-lead. Past: AlphaStar, AlphaFold, AlphaCode, WaveNet, seq2seq, distillation, TF.Chris Olah @ch402
91K Followers 173 Following Reverse engineering neural networks at @AnthropicAI. DMs open! Previously @distillpub, OpenAI Clarity Team, Google Brain. Personal account.Henry de Zoete @HZoete
3K Followers 3K Following Adviser to the PM on AI, angel investor, NED @cabinetofficeuk & @oaknational Exited founder - Look After My Bills (@ycombinator W18)Ethan Mollick @emollick
211K Followers 553 Following Professor @Wharton studying AI, innovation & startups. Democratizing education using tech Book: https://t.co/CSmipbJ2jV Substack: https://t.co/UIBhxu4bgqIan Hogarth @soundboy
23K Followers 3K Following investor @pluralplatform; chair UK AI Safety Institute; co-author @stateofaireport; co-founder @songkick; chair @PhasecraftLtdVerena Rieser @verena_rieser
4K Followers 1K Following Researcher @DeepMind working on safer Conversational AI | Honorary Professor @HeriotWattUni | Co-founder @helloalana | mother of dragons | own opinions onlyLean @leanprover
4K Followers 35 Following Lean is a dependently-typed programming language and theorem prover.Dr. Nahema Marchal @nahema_marchal
2K Followers 1K Following research scientist @deepmind | 🔎 tech governance, online harms, socio-technical ai | previously @prodigi_erc @oiioxford | flibbertigibbet, views her ownKevin Buzzard @XenaProject
9K Followers 0 Following Mathematician learning Lean and trying to teach it to others. Now gone to Mathstodon (March 2023). No longer reading or replying to mentions.Alicia Parrish @AliciaVParrish
556 Followers 675 Following Research scientist at Google. I like CogSci & NLP. PhD from @nyuling. She/her.Álvaro Lozano-Robled.. @MathAndCobb
4K Followers 729 Following Prof. of math @UConn, number theorist, author, Hagoromo Ambassador. On *Twitter* hiatus. You can find me in bluer celestial pastures instead.Sebastian Riedel (@ri.. @riedelcastro
15K Followers 470 Following Researcher in NLP/ML @deepmind, @ucl_nlp, @[email protected] on MastodonTamar Shinar @ttshinar
31 Followers 225 FollowingSome theorems @CihanPostsThms
24K Followers 6 Following Posting some theorems, and occasionally other stuff. By @bahran_cihanRichard Chappell @RYChappell
1K Followers 114 Following Academic Philosopher. Blogs at https://t.co/d4D6CfLwuBNora Belrose @norabelrose
8K Followers 124 Following Working toward a free and fair future powered by friendly AI. Head of interpretability research at @AiEleuther, but tweets are my own views, not Eleuther’s.depths of wikipedia! @depthsofwiki
889K Followers 4K Following Hello I am @anniierau Please take away my blue check! I did not ask for it!Casey Newton @CaseyNewton
214K Followers 909 Following Writing @platformer. Co-hosting Hard Fork @nytimes. Posting good tweets to Instagram stories @crumbler. [email protected] | https://t.co/9KuJb8XCrrCari Tuna @CariTuna
3K Followers 83 FollowingVijay Bolina @vijaybolina
3K Followers 5K Following Hacking AGI. CISO @Google @DeepMind. Former @Mandiant @BoozAllen. Tweets my own.Ethan Porter @EthanVPorter
2K Followers 1K Following Associate professor at @GWtweets, @SMPAGWU, @GWIDDP, formerly @demjournal, @uchicago.Metaculus Prediction .. @MetaculusAlert
3K Followers 5 Following A bot that tweets whenever the Metaculus (@metaculus) community prediction for a question (currently Ukraine and Monkeypox) changes significantly.ninell oldenburg @nellsn1
420 Followers 608 Following PhD student in philosophy of AI @ ucph; tweets on linguistics, artificial "intelligence", the awful German language, & mapsStevie Bergman @tvr2c
548 Followers 976 Following Tech and society outside the global north. Senior research scientist at DeepMind. Past: Meta, Princeton, TVR2C at WPRB, Peace Corps, USDOJalex lawsen @lxrjl
3K Followers 745 Following AI Grantmaking @ Open Philanthropy Previously 80,000 Hours, teaching, forecasting, poker. Views my 🐒'sFredrik Johansson @hypergeometer
678 Followers 145 Following fredrikj @ mathstodon Computer algebra & Arbitrary-precision arithmetic. Researcher at @Inria.Casey Muratori @cmuratori
41K Followers 120 Following I want all my garmonbozia. https://t.co/Bdh1Xj2PpVQuinta Jurecic @qjurecic
56K Followers 3K Following senior editor @lawfare, fellow @BrookingsInst, contributing writer @TheAtlantic. views my own, RTs = @infinite_scream.Javier Blas @JavierBlas
297K Followers 1K Following Energy and commodities columnist at Bloomberg. Co-author of the 'The World for Sale' https://t.co/GAcVleqiqp Any views expressed are my own. [email protected]Freja Kirsebom @freja_kirsebom
7K Followers 689 Following Senior epidemiologist - COVID-19 vaccines @UKHSA. PhD in the immunology of respiratory viral infections @imperialcollege.Zvi Mowshowitz @TheZvi
24K Followers 283 Following Blogger world modeling, now mostly AI and AI x-risk, at Don't Worry About the Vase (https://t.co/O9LbMQjKoo or WP/LW), founding Balsa Research to fix policy.Amelia (Mia) Glaese @mia_glaese
76 Followers 38 FollowingDorothy Chou @dorothychou
914 Followers 451 Following Learning things @DeepMind & @UniofOxford | Angel Investor @Atomico | Formerly @Uber @Dropbox @Google, Board @simplysecureorg | She/her 🇹🇼🇺🇸📍🇬🇧Jacob Steinhardt @JacobSteinhardt
7K Followers 67 Following Assistant Professor of Statistics, UC BerkeleyJohn Aslanides @john_aslanides
608 Followers 1K Following Aussie battler trying to make it in the Big Smoke. Building on the critical path to AGI since '16. Currently: RLHF that doesn't suck @GoogleDeepMindNat McAleese @__nmca__
3K Followers 306 Following Superalignment by models helping humans help models help humans at OpenAI. Previously @DeepMind. Views my own.Howie Lempel @HowieLempel
1K Followers 497 FollowingBenjamin Todd @ben_j_todd
12K Followers 143 Following Founder @80000Hours Writing about what to do about AI, doing good, and using research to have a nice life 🦑Liv Boeree @Liv_Boeree
254K Followers 497 Following Looking for the win/wins in life. Not a fan of Moloch traps. Brand new podcast out now, link below👇Seán Ó hÉigeartaig.. @S_OhEigeartaigh
2K Followers 1K Following Director of https://t.co/gCEDoKdKBT at Uni of Cambridge | Researching Big Risks, and impacts of AI & emerging tech. Opinions ownAbby Novick Hoskin @CorpusCalosseum
662 Followers 825 Following PhD in Psych/Neuro from Princeton. 80,000 Hours Advisor. Mother of two multimodal multitasking neural networks. Views are my own.Allan Dafoe @AllanDafoe
3K Followers 565 Following AGI governance: navigating the transition to beneficial AGI (Google DeepMind)Jacob Menick @jacobmenick
4K Followers 266 Following Researcher @OpenAI. PhD candidate @UCL. previously @DeepMind 🇺🇸/🇬🇧Do models need to reason in words to benefit from chain-of-thought tokens? In our experiments, the answer is no! Models can perform on par with CoT using repeated '...' filler tokens. This raises alignment concerns: Using filler, LMs can do hidden reasoning not visible in CoT🧵
I believe we have discovered the flaw in the 'post message that you will be deactivating your account removing all of your Tweets' plan.
gonna deactivate for a while. nothings wrong i just need to detox from the site a bit
Rose: The idea is extremely simple and well-motivated, and the effect sizes are large. Thorn: p=0.05 :( (Tbc, I am very confident we would have reached statistical significance for Gated SAEs being more interpretable, if we had a large enough N.) x.com/sen_r/status/1…
New @GoogleDeepMind MechInterp work! We introduce Gated SAEs, a Pareto improvement over existing sparse autoencoders. They find equally good reconstructions with around half as many firing features, while maintaining interpretability (CI 0-13% improvement). Joint w/ @ArthurConmy
@geoffreyirving @KLdivergence @sindero And I found the Jensen Huang reference from GTC keynote 2016. Look at the slide:
@geoffreyirving @KLdivergence @sindero I don’t have an issue with that. I have an issue with using “inference” as a verb, as in “inferencing”. That drives me nuts. x.com/fhuszar/status…
@CellTypist @KLdivergence My issue is grammatical: using inference as a verb. I.e instead of to infer something they say to inference something. Or the word “inferencing” Others might take issue using the word “infer” instead of forecast or predict I don’t know.
@geoffreyirving @KLdivergence @sindero phi-3 paper is the last time I saw it:
I will never get over how AI/ML people use the word “inference”
We are looking for an AGI Safety Manager to support @GoogleDeepMind 's AGI Safety Council: please encourage excellent people to apply! This role will work closely with my team, Scalable Alignment and Safety, and Responsible Development and Innovation. boards.greenhouse.io/deepmind/jobs/…
Most of the time you don't really notice the world changing. Then one day you're sitting in the back of a driverless car, listening to music on your phone while asking an AI something, when suddenly you're struck by a memory of childhood and you realize you now live in Star Trek.
@_NicT_ In chapter 1 of deeplearningbook.org we say that data representations are crucial for not just machine learning or even computer science but daily life (e.g. try dividing numbers by hand with Roman numerals).
@goodfellow_ian I blame myself for failing to cite this passage in your book!
No smooth curve that lies on a sphere can contain an inflection point (a point whose curvature is zero). But on a surface of constant negative curvature, like this tractroid, no such obstacle exists, and every non-meridian geodesic on the tractroid has 2 inflection pts (black).
@RokoMijic @ESYudkowsky @robinhanson I think you're over-extrapolating the success of RLHF (which I was worried people would do). Remember why people came up with "scalable alignment" ideas like IDA and Debate. Those solutions aren't coming online fast enough. @geoffreyirving was worried about this when I asked him.
This is fantastic news for NIST 🎉 I'm biased because he's a friend, but Paul Christiano is both technically brilliant and holistically super thoughtful - not to mention a pioneer in exactly the kind of frontier testing work he'll be leading at AISI. commerce.gov/news/press-rel…
The Future of Humanity Institute has shutdown 🙁 (2005-2024): futureofhumanityinstitute.org
I've been wanting to do this story for years and am thrilled it's finally out: inside the surprisingly small, highly specialized industry that repairs the internet cables on the bottom of the ocean theverge.com/c/24070570/int…
One unexpected benefit of sending out an email to every contact about an updated email address is that I suddenly hear from a lot of friends and acquaintances I have not interacted with for ages.
Unfortunately it seems that the universe has been fine-tuned in such a way that this tweet would receive some very annoying replies. Wishing the fundamental constants of the universe were very slightly different right now.
I just flipped a coin ten times, resulting in the sequence HTHHTHTTTH. The universe was fine-tuned for this outcome—if the fundamental constants of the universe had been even 0.1% different, I never would have observed this sequence of flips.
I just flipped a coin ten times, resulting in the sequence HTHHTHTTTH. The universe was fine-tuned for this outcome—if the fundamental constants of the universe had been even 0.1% different, I never would have observed this sequence of flips.
You flip a coin 200 times. The first 100 flips, it lands on heads; the second 100, on tails. As a proud Bayesian, you conclude you are most likely in the middle of a logic puzzle.