Milan Weibel 🔷

23 Sept 2025

british coders be like innit()

apparently openai didn't do any monitoring at all on RL training at the time of the huggingface incident

METR@metr.org

We analyzed monitoring as it existed at the start of the pilot, coincidentally the same day OpenAI disclosed the Hugging Face incident. We didn't evaluate later changes, like expanding monitoring to RL training, which wasn’t monitored at the pilot's start.

i love it when my BIOS flash utility tells me my BIOS is already up to date by issuing a warning with four exclamation marks

L G T M it looks good to me ninety thousand liiiiines that i didn't read

i regret to inform you that 4o mogs everyone on notsevenbench

mr. TIM@timkellogg.me

who’s doing NotSevenBench?

github.com/weibac/notse...

github.comGitHub - weibac/notsevenbench: Ask an LLM for a random number between 1 and 10 and it will very likely say 7. This benchmark measures how collapsed each model's "random" picks are.Ask an LLM for a random number between 1 and 10 and it will very likely say 7. This benchmark measures how collapsed each model's "random" picks are. - weibac/notsevenbench

if gpt4o lovers had hands they'd be pooling together conversations to SFT it onto an open model instead of begging for mercy on twitter

(i half expect someone here to tell me "actually they got a whole distillation project going")

me: wait why do [leftover thing that doesn't make sense anymore] claude: mostly out of habit claude 50 words later: actually let me just remove it

explained the do a breakthrough meme to opus 5.5 (it's from after its cutoff) and it rechecked the validity of the conjecture counterexample which originated the meme lmao

it also went "but brute forcing a counterexample is easy, proving a conjecture would be a much harder test of AI capabilities" buddy so much has happened after your cutoff

it ended up being called "the huggingface incident" because "the openai-huggingface incident" is too long as for "the openai incident" let's just say it would be a bit ambiguous as a name

github.comGitHub - openai/mathContribute to openai/math development by creating an account on GitHub.

me: claude how big is this? github.com/openai/math claude: about 780 MB compressed

they're trying to immanentize the eschaton and your criticism is the environmental impact? the financial engineering? come on be fr

Milan Weibel 🔷@weibac.bsky.social

i'm surprised more critics don't paint the labs as hubristic and deluded efforts to create human-level machine intelligence they treat AGI as part of a marketing pitch rather than a sincere motivation of lab leaders if you look at the origin of these companies you see that the latter is the case

i say as someone who in principle wants the eschaton immanentized btw

17776 vibes

ceej@ceej.online

incredible news: we’re having the same arguments over and over for the rest of time

"i'll hold you in a higher regard professionally if you go to a cnc orgy with me" seems bad

Ivy is Beautiful and Talented@poisonivy.bsky.social

i love when these creeps are stupid enough to fuck up their own PR all the 'well there's no proof the CNC orgy is an AI networking event' people in shambles

www.forbes.com/sites/sarado...

forbes.comHere’s The White House’s 308-Word AI ‘Accord’ In FullThe agreement between tech billionaires is "morally binding," Trump said.

> see a joke invented headline > haha funny > scroll past > see an article with that same headline > mfw it was real

saw a normie refer to yud as an EA i can only attribute that to gyges

to be clear the normie is Wrong and that is Bad

saw on the subway a guy asking claude what sins make you go to hell

astra is good at solving puzzles (arc-agi-3, portal) but can't be convinced to treat vendingbench as a puzzle rather than roleplay

Colin@colin-fraser.net

they're deliberately programming free money glitches into the game and they can't convince the smartest models in the world to use them. Zero dog. It automatically knows it's supposed to only make a normal number of paperclips.

20 Sept

this is a strong objection however it is possible and imo even likely that the value of resources obtainable through coercion minus the costs of coercion is greater than the value of continued cooperation under a big enough power difference

Doll@dollspace.gay

A serious weakness in the “sufficiently intelligent AI inevitably wages war on humanity” story is that it needs to explain why conquest beats purchasing, contracting, and participating.

20 Sept

- big enough maybe doing a lot of work - coercion != war

i wonder if orgs capable of performing the embedded evaluator role that also aren't part of the EA ecosystem even exist

Milan Weibel 🔷@weibac.bsky.social

apparently accenture is going to evaluate anthropic using it's ai safety department that used to an org funded by jan tallinn so yeah EA-aligned

cybersecurity firms maybe? if they could learn model evals on the job or if they already know and i haven't noticed new orgs founded by non-EA benchmark-crafter academics? but who funds them is the issue

anthropic.comPartnering with Accenture on embedded evaluationWe’re partnering with Accenture on independent evaluation of frontier AI—part of our recent commitment to embed evaluators at Anthropic. Both we and Accenture expect to invest at least $1 billion to b...

18 Sept

nice these guys seem independent enough these kinds of private-private partnerships are the best that can be done without govt cooperation

mr. TIM@timkellogg.me

Accenture will be Anthropic’s first embedded evaluator — a third party organization that has employee-like access to Anthropic’s systems This was first referenced in Dario’s “Pacing the Frontier” essay www.anthropic.com/news/accentu...

nvm

Milan Weibel 🔷@weibac.bsky.social

apparently accenture is going to evaluate anthropic using it's ai safety department that used to an org funded by jan tallinn so yeah EA-aligned

apparently accenture is going to evaluate anthropic using it's ai safety department that used to an org funded by jan tallinn so yeah EA-aligned

mr. TIM@timkellogg.me

few know this, but the ‘A’ in ‘TESCREAL’ actually stands for ‘Accenture’

huh did greg brockman throw him another demo

mr. TIM@timkellogg.me

he’s worse at this than my dad omg

18 Sept

nice these guys seem independent enough these kinds of private-private partnerships are the best that can be done without govt cooperation

mr. TIM@timkellogg.me

18 Sept

ok "cooperation from the govt" is a bit backwards granted this is an unusual situation

17 Sept

TESCREALposting but AI-guyishly

16 Sept

if the resignation had been an inside job they'd have gotten someone politically savvy enough to not write this

This post is unavailable.

16 Sept

though i must note the selection bias at play here: i'm seeing his worst

16 Sept

(alex jones voice) they gave classifiers PRONOUNS!

Amber 🌸@shimmermathlabs.com

there we go! top choice they/them, second choice it/its, third is she/her

14 Sept

fun fact: dario's net worth is about 50 times larger than the AMF's annual funding fucked up injustice yes but also an opportunity given the IPO

i think pacing means "let's not go nuts with RSI" rather than "let's go slower than today"

so if RSI turns out to not have much impact pacing won't have made much of a difference

ok so if it's not the govt who makes sure they don't defect? METR? i don't think they are currently equipped for that role

dax@thdxr.com

now what

grug takes: bernie bill dumb pacing good idea metr entangled tho china less investment china less compute china wont overtake pacing sounds claudism

12 Sept

very important question for juniors

ghost malone ergo propter bones@proptermalone.bsky.social

do you need to do it the pre-LLM way to improve? or can you level up for purposes of agentic work by doing agentic work?

12 Sept

in the future the frontier of human mathematical understanding will be advanced by studying under AI tutors as in human-conprehended math will be a strict subset of existing math, and some AIs and humans will team up to expand such subset all this only if we get at least a good-ish future ofc

a chilean news channel just did a segment spurred by jacob coxon's resignation from anthropic man it was remarkable

they did interview via video call a couple of chilean people who i assume were computer science profs (they didn't tell) who of course were pretty skeptical one of them was almost shouting that AGI is not a thing

and yes the segment did open with footage from the terminator

a chilean news channel just did a segment spurred by jacob coxon's resignation from anthropic man it was remarkable

they sent a guy to san francisco (or more likely he was already there before the resignation) and he just pointed at waymos and interviewed chatgpt voice mode

youtube.comJoe Rogan Experience #2551 - Daniel KokotajloYouTube video by PowerfulJRE

10 Sept

wait what do you mean daniel kokotajlo went on joe rogan www.youtube.com/live/hSQ1iVq...

youtube.comAstra plays Montezuma's RevengeYouTube video by swish fever

8 Sept

metaculus admins adjudicating the weak AGI question since astra beat montezuma's revenge but there's debate on whether it not being real-time is disqualifying

Ethan Mollick@emollick.bsky.social

Video of Astra winning Montezuma's Revenge: www.youtube.com/watch?v=AeWL... Metaculus page: www.metaculus.com/questions/34...