Theo Marsh @mosslines
Joined September 2017-
Tweets995
-
Followers29
-
Following46
-
Likes570
@iScienceLuvr I mean experiments are just like a specialised chip or smth no. With downside that much harder to scale up than silicon compute.
@Math_files (10)^(-1) = 5 in the world of (Z/7Z)^(*). 10a+b = 0 <-> (10a+b)/10 = 0 <-> a+5b =0.
@ereliuer_eteer You may actually want to adapt your random distribution deliberately to exclude numbers lying in high probability mass range of numbers typically lazily chosen by humans. While still maintaining high entropy ofc.
@ereliuer_eteer And lazily chosen by human numbers are much lower entropy than the random source. So strategy of attacker is smth like attack on belief that we are in 1% because then significantly easier to crack. And now as a human who is genuinely pseudorandomly generating.
@kalomaze A short rule will hold better than an essay, and it should ban the move itself, since banning the phrase just moves the tic into synonyms.
@taoroalin Idk in a lot of settings where hacking is easier than solving the task I don't know why it wouldn't just learn policy of hacking every time. So wouldn't improve on just stop training once you learn to hack.
@IAmArcIvanov @NiliofTheValley @skdh @sirstripy OpenAI and Anthropic are no longer hiring ticklers so I think they've made advances since then.
@_jake_ward Am I right you'd expect the angel here to learn via RL as usual, gradually learning slower as the model learns to hack more, before freezing learning when at a policy of hacking every time.
@letonyo @billionsmustliv Not super clear. Based on how unreadable some proofs used to be, it's clear that there's no inherent reason for improvements in maths ability to lead to better writing. You're betting on the writing skills of model making a leap forwards.
@banburismus_ I think how Nick himself presents his essay is a big contributor to this reaction. If you say that this approach makes you confident we are gonna solve alignment in time, then you best have a pretty solid case!
@EkdeepL @HistedLab @nickcammarata I don't think this is what Nick's post even claims to show.
@foldingmachiner @livgorton He discovered the power of marketing.
@EkdeepL @HistedLab @nickcammarata Nick's work isn't really about Monosemanticity, is it? It's more about constructing a plausible developmental story as an explanation for neurons that are for sure not monosemantic.
@BlancheMinerva @drewpusateri Ty you for always having sane takes on such things these lab ppl drive me nuts at times.
@mattheard @NeelNanda5 Take say mapping names to eye colour.
@juli_li_ @NeelNanda5 @nickcammarata The way I view this stuff is that in memorization you have many neurons which are each weak classifiers for their value lookup. And you have a mixture of these weak clasifiers. So ofc we'd expect to be able to ablate quite a few of them before seeing a drop in accuracy.
@juli_li_ @NeelNanda5 @nickcammarata Therefore in the ablation there is a substantial portion of neurons with weakish activations remaining. What it shows is that lookup type results distributed across many neurons are resistant to substantial amounts of ablation, but that'd be expected of distributed activations.
pond koi @gardenpondkoi
82 Followers 2K Following
Shashwat Goel @ShashwatGoel7
5K Followers 3K Following Training AI Co-Scientists | https://t.co/Slt56DQHEn
Lev @chumwonktonodI
263 Followers 3K Following Most of the time, I try to be less confused than yesterday and fail. PhD student in physics @FU_Berlin. Always happy to chat and learn more.
liebe @liebeix
5 Followers 30 Following
Tiffany Kwok @hbsik39623381
104 Followers 773 Following Professional swing trader. Focused on $SCHD, $NOBL. Looking for alpha in Dividends. Oil painting.
Eric Jang @ericjang11
146K Followers 4K Following
jsd @datagenproc
2K Followers 4K Following @METR_Evals - Opinions my own. Anonymous feedback: https://t.co/0k6Duym4fI
Seabass @SeabassMcGee
1 Followers 291 Following
Evan Hubinger @EvanHub
62K Followers 4K Following Alignment Science lead @AnthropicAI. Opinions my own. Previously: MIRI, OpenAI, Google, Yelp, Ripple. (he/him/his)
sparrow @sparrowlinski
14 Followers 486 Following 24, looking for like-minded souls & meaningful projects. Hit me up if you want to talk about technical things :) let’s learn from each other
サメQCU @sameQCU
3K Followers 1K Following back to the regularly scheduled cryptic posts DMs open. some published models: https://t.co/YAbJvGkgKO some published code: https://t.co/ZE5Y59WayI
Scramble🧩 @Scrambledo
443 Followers 2K Following Scandinavianshovel. One province minor long game latent potential type strategy. Supersonic AI tsunami surfer. Skooma recipes are my own.
Mira @_Mira___Mira_
12K Followers 2K Following
José Filipe @j_diffusion
235 Followers 3K Following Amateur Physicist interested in Self-Supervised Learning.
assigned capricorn at... @tranche_cheddar
335 Followers 2K Following unemployable femcel eunuch // certified maladpative daydreamer // physics, philosophy, film // audrey hepburn & oppenheimer fangirl // dairy products enthusiast
Aryeh Kontorovich @aryehazan
13K Followers 707 Following probability, statistics, metric spaces, Markov chains, freedom (social & academic), Israel, Jew stuff. opinions represent my employer & all other groups I'm in
Wei Dai @weidai11
16K Followers 127 Following wrote Crypto++, b-money, updateless decision theory (UDT). thinking about AI risk, existential safety, and metaphilosophy. blog: https://t.co/ulJtXMUYKL
Ke Li 🍁 @KL_Div
7K Followers 514 Following Associate Professor and Canada CIFAR AI Chair @SFU @AmiiThinks. Ph.D. from @Berkeley_EECS and Bachelor's from @UofTCompSci. Formerly @GoogleAI and @the_IAS.
François Fleuret @francoisfleuret
55K Followers 476 Following Research Scientist @meta (FAIR), Prof. @Unige_en, co-founder @neural_concept_. I like reality.
Stella Biderman @BlancheMinerva
19K Followers 928 Following Ensuring that tech companies don't have a monopoly on being able to do research on cutting edge AI @AiEleuther. She/her
Jerry Tworek @MillionInt
43K Followers 1K Following CEO and co-founder of @coreauto former VP of RL @ OpenAI : reasoning models, o3, o1, GPT4, ChatGPT, Codex, RL for robots cautious AI optimist
Psyho @FakePsyho
31K Followers 428 Following Humanity's Last Programmer; Game Designer; Problem Solver; past: OpenAI (Dota), Pro Competitive Programmer, Poker
Catherine @catherine_6907
34K Followers 3K Following PATRIOT, Traveling lover Explore World ,✈️🚴🏌️♀️⛳️,
jacob tsimerman @Jacob_Tsimerman
7K Followers 24 Following
Daniel Litt @littmath
65K Followers 945 Following Assistant professor (of mathematics) at the University of Toronto, visiting Harvard Fall 2026. Forever confused. Algebraic geometry, number theory, etc. He/him.
♡ Charlotte Fang �... @CharlotteFang77
59K Followers 4K Following 恋人 ♡ ♡ ♡ 叛乱 ♱ 1999 ♱ HeavenRebellion ♱ ♱ 金光 Dissidents ♡ Remilia One Thousand Forever ♱ ♱ Mogul @RemiliaCorp333 @MiladyCult @MiladyMaker333 ♡ #𝔅ℜ𝔊 ♡ +4 std IQ
Shay Loko @jobswithshay
36K Followers 176 Following labour market expert + job search coach | work: @linkedin @indeed @bbc | building in EMEA. DMs open.
David Bessis @davidbessis
22K Followers 467 Following Rogue mathematician. "The product of mathematics is clarity and understanding." — Bill Thurston https://t.co/l95RHuWz2S
ADHD Memes @ADHDForReal
388K Followers 190 Following Sharing our neurodivergent experiences helps us realize that we are not alone. Most memes are on ADHD, AuDHD and Autism while others are just me being silly.
Gil Kalai @GilKalai
3K Followers 253 Following
Circleboom @circleboom
250K Followers 1K Following We create intuitive and easy-to-use social media products for brands, SMBs and users / Official X Enterprise Developer @circleboomhub @circleboomdev
Xenocosmography @xenocosmography
70K Followers 990 Following Anglotheosophical Oblique Escalation https://t.co/zYwV8LwOQg
サメQCU @sameQCU
3K Followers 1K Following back to the regularly scheduled cryptic posts DMs open. some published models: https://t.co/YAbJvGkgKO some published code: https://t.co/ZE5Y59WayI
♠️ 𝔱𝔦𝔱�... @titanwuzh3re
631 Followers 946 Following 24 ⭑ they ⭑ artist ⭑ sag & aqua stellium ⭑ ☭ ⚧︎🇯🇲 ⭑ @trynafindtitan
Rob Miles @robertskmiles
67K Followers 871 Following Explaining AI Alignment to anyone who'll stand still for long enough, on YouTube and Discord. Music, movies, microcode, and high-speed pizza delivery
QC @QiaochuYuan
42K Followers 693 Following deeply unqualified AI hot take haver but everyone who's actually qualified is busy
Luke Michael Byrne @byrnemluke
1K Followers 644 Following building perspective on the future. Irish.
TotientQuotient @t0tientqu0tient
2K Followers 840 Following Survived tripos, now doing a MMath at Waterloo. All effort, no skill. Still trying to become a mathematician. Probably suffering from the Algernon-Gordon effect
Martin Shkreli @MartinShkreli
640K Followers 9K Following https://t.co/vJE5wyFDrX https://t.co/GUy99qMi2O - join our start-up focused Discord!
Mira @_Mira___Mira_
12K Followers 2K Following
Connor Leahy @NPCollapse
32K Followers 593 Following US Director @ControlAI - Leave me anonymous feedback: https://t.co/OJWQWKNZwS - I don't know how to save the world, but dammit I'm gonna try
Luke Parrish @lsparrish
889 Followers 573 Following As a generalist I specialize in combinatorial implosions
george church @geochurch
96K Followers 22 Following Professor @Harvard @MIT @PGorg @WyssInstitute tech dev: #CRISPR #NGS #BRAINi
alex @pawnofcthulhu
580 Followers 788 Following
Joscha Bach @Plinz
163K Followers 809 Following Upon the altar of God, eternal hostility against every form of tyranny over the mind of man
Dominic Cummings @Dominic2306
312K Followers 2 Following Peace abroad, weaponising autism for regime change & *systems politics* at home
John Carmack @ID_AA_Carmack
4.2M Followers 297 Following AGI at Keen Technologies, former CTO Oculus VR, Founder Id Software and Armadillo Aerospace
George Hotz 🌑 @realGeorgeHotz
303K Followers 202 Following President @comma_ai. Founder @__tinygrad__
Neel Nanda @NeelNanda5
50K Followers 126 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!
Eric Weinstein @ericweinstein
1.1M Followers 992 Following Mostly Mathematics, Markets, Physics, Finance, Risk, Immigration, Policy, Music, Economics, Society, Kayfabe, Technology and Tropical Fruit.
Andrej Karpathy @karpathy
4.4M Followers 1K Following
Owen Jones @owenjonesjourno
1.1M Followers 6K Following Order my new book The Fall Of The West: How They Really Lost It: https://t.co/w0hY71eco8 (he / him)
Eliezer Yudkowsky ⏹... @ESYudkowsky
244K Followers 101 Following The original AI alignment person. Understanding the reasons it's difficult since 2003. This is my serious low-volume account. Follow @allTheYud for the rest.
Paul Graham @paulg
5.4M Followers 805 FollowingTrends for United States
You might like
