a suicide note: I only logged in 200 times because your shitty system kept logging me out now a simple reviewing task has become endless I cannot do this again farewell @openreview.bsky.social

if you don't appreciate this joke, please unfollow me

why must millennials put emoji on everything

RewardBench 2 is here! We took a long time to learn from our first reward model evaluation tool to make one that is substantially harder and more correlated with both downstream RLHF and inference-time scaling.

I am furious about the news out of Colorado of Jews being lit on fire for the crime of calling for hostages to be released. I am furious at how dirtbag leftists on this website and elsewhere treat this as a game where violence against American Jews is justified if we aren't good tokens.

Congratulations to @yizhongw.bsky.social, who successfully defended his PhD thesis today! Advised by @hanna-nlp.bsky.social and me.

trying to decide whether to go to Vienna in late July (there's a formerly-AI conference happening there and I might want to see some friends who still go) maybe you can help me decide who's going to be there and wants to rent a place and play music together

(I don't know why this will never get old) *upload photo of anyone at all to Molmo, type "where is the scientist?"* Molmo: I don't see a scientist here

the economist called us losers and we ... rolled our eyes because their argument is behind a paywall Gen X (at its best) doesn't feed trolls

it's always glutes

used a LM to diagnose a medical problem, figure out steps to get treatment with minimal phone calls and time wasted waiting for healthcare providers to perform administrative tasks (1/n)

intellectually I get that his name is "Robert Prevost" but in my head I can't not hear it as "Ford Prefect"

before we get to their math, can someone explain how modern politicians choose to capitalize words? is there any rhyme or reason to it, or is it just a new version of boomer-random?

"When we erode the boundaries between the academic and the political, we ultimately harm both."

always hold these two thoughts in your mind when considering evaluation of AI systems: 1. rigorous evaluation is essential; we must not go back to the 1980s (anecdote world) 2. continuous improvement to eval methodology is essential; evaluations are usually broken in ways worth fixing

protip: the less you reply, the less email you get

True story: the NLP community elected the person holding this position to the presidency of our professional society. Let's take the greatest, most impactful thing to come out of our little world and shit all over it! It's the Tr*mp presidency in miniature

where's Johns Hopkins?

🔭 Science relies on shared artifacts collected for the common good. 🛰 So we asked: what's missing in open language modeling? 🪐 DataDecide 🌌 charts the cosmos of pretraining—across scales and corpora—at a resolution beyond any public suite of models that has come before.

new work on test-time alignment using MCMC

Meet Ai2 Paper Finder, an LLM-powered literature search system. Searching for relevant work is a multi-step process that requires iteration. Paper Finder mimics this workflow — and helps researchers find more papers than ever 🔍

This is just to say I have covered your Tesla in Kraft Singles which probably irritated and perplexed you Do not forgive me I will do it again and again and again

a small change to building your BPE tokenizer gets your pretrained LM 8 MMLU points (for example) and 27% inference-time efficiency boost ...

It’s unfair to compare the trashing of Teslas to the Boston Tea Party because tea is worth something

Humbled to make @fastcompany.com's 2025 most innovative companies list for making AI models that are truly open. "Ai2 is setting a strong benchmark for what transparency can look like in the entire AI industry."🎉

Senate Democrats who voted YES on cloture for the CR: Schumer Gillibrand Fetterman Schatz Durbin King Shaheen Hassan Peters Cortez Masto

nice try, everyone still knows you're a collaborator

I will not vote to weaken my power to fight for WA state. I will not vote to turn YOUR taxpayer dollars into Trump & Elon's personal slush funds. I will vote NO. We should pass my short term CR, avoid a shutdown, and do our jobs.

til people think of linkedin as a social network and post stuff there it's even more cringe than twitter

with democrats like these, who even needs republicans

Man Who Invested in Trump Shocked to be Losing Money

I'm really proud of our team for reaching this milestone! Let's put to rest the propaganda (1) that fully-open AI models are second-class, (2) that researchers can only have impact or gain understanding of current systems if they work at a soulless for-profit, and (3) that AI is sooo mysterious.

This is not fighting antisemitism, it IS antisemitism. Any Jew of any ideology who thinks that if they feed enough "bad Jews" to the antisemitic dragon that it won't eat them next is morally and politically deluded. Permission slips for attacking Jews are never used just once.

exactly. secular people have found this stuff abhorrent and uncouth for a very long time, nothing new about that

any hope that my employer, the University of Washington wouldn't "comply with the fascists in advance" is now dashed ... this is some serious bullshit www.washington.edu/provost/2025...

The rule of law is a good thing. I don't like this guy, I don't like his movement. I still expect the government to provide him due process and make the case with evidence to an impartial judge on why he should be deported. If the gov fails to make the case, I expect that ruling to be respected.

evidence accumulates that Elon Musk hates Jews more than is strictly necessary *reluctantly fetches space laser from the basement*

cold email via linkedin (lmao) inviting me to send my cv for consideration as an expert witness (lmao) the noise I made when I read the parenthetical that the cv should be in msword format considering sending them my cv.pdf and quote $100M but it wouldn't be enough #lawyers

the first rule of any academic editing job (e.g., program chair, editor) is to *never* send email through a system that doesn't make it clear (1) what human is responsible for the message and (2) includes their email address ... looking at you, openreview

mazl tov to the beings who made openreview ... truly pioneering in the effective use of a rarely mastered design pattern known as Gaslighting The User

I'm very excited we are finally hiring a POLICY LEAD for Ai2 / OLMo. This is going to be super useful to massively amplify the impact of our all our open efforts and have a positive impact on the world. You will work with people like me + Ai2 leadership to improve the AI ecosystem.

Vance mistook Ukraine for the Republican congressional caucus

me (on my bike, at a stop light): how do you like your nazi car tesla driver: I have a child in the car me: oh I guess nazis can have children Do people really use their kids as a device to defend against their shameful shit? I thought that was just an old movie trope. "But I've got a family!"

new OCR tools from the OLMo team open, as god would have intended, did it exist

this morning's regret that I had only two blocks to bike on my street at progressively slower speeds with the enraged horn-pounding driver stuck behind me welcome to my neighborhood, why dontcha stay awhile, friend

starting position: meetings with more than 50 researchers are bad, >150 painful, and >450 unbearable it's been >10 years since I went to any ML conference try to convince me to consider going to iclr