Stella Biderman
@stellaathena
I make sure that OpenAI et al. aren't the only people who are able to study large scale AI systems.
Got a pretty hilarious warning from Google T&S today @defcon.bsky.social @aivillage.bsky.social There’s actually nothing that should flag this, it’s just screenshots, images, and text and none of it is about phishing. My talk was about model stealing though APIs.
How do you feel about the fact that the library was written by an AI?
An updated guide to where you can find @eleutherai.bsky.social at #ICML2025! We have a lot going on, including an oral presentation tomorrow (Tuesday) where I’m going to be talking about my scientific research agenda. Come say hi to us!
A common issue with position papers is that they leave the reader wondering “okay, but what should I actually do”? To address this we provide open problems on a wide variety of topics throughout to illustrate our perspectives and guide future research
We ground discussion in the history and philosophy of science. What did it take for other fields to move from cataloging phenomena to predicting and controlling them? AI can learn from that playbook.
In film, "we'll fix it in post" is what you say when something went wrong on set and you don't want to redo it. AI research has made it our entire methodology: train the model, then patch whatever comes out. Our new ICML oral argues this can't be the basis of a science of AI. 🧵
You have a moral imperative to refuse to work with these people or develop models for these purposes.
Has anyone else had Claude code become non-functional recently? Even with a test input it spins for minutes without doing anything. Same thing happens in terminal.
The only reasons I use social media platforms is to get eyeballs on research and to yell at people who are wrong online. X > Bluesky at both for me
Incredibly proud of my friend and colleague @storytracer.com. Two weeks ago he and his cofounders @sucho-org.bsky.social were honored for organizing a global network of volunteers to exfiltrate and back up endangered Ukrainian cultural heritage in the wake of the invasion by Russia.
In the original Pile paper we talked about various conceptions of consent (though I don't stand by everything I wrote about this topic 5 years ago). None of this data has EIC, though I think that the ones marked "author" in the table are ones where authorial objection would be unreasonable.
Thanks for the signal boost! I think you'll appreciate the serious shade we threw at the end of our blog post: blog.eleuther.ai/deep-ignorance/
"But wait," the skeptic cries. "Surely this is infeasible for frontier models! Their datasets are far too large to expect companies to meaningfully understand or document!" Actually our methodology is extremely cheap, mostly runs on CPU, and adds an overhead of less than 1%.
Preventing a model from going into an interaction with relevant knowledge isn't a panacea though: filtered models see minimal performance loss when it comes to reasoning about information provided in-context. The models are still smart, they just don't know the concepts innately.
Our results are competitive with circuit-breaking when it comes to out-of-the-box performance and substantially more robust to both adversarial and benign finetuning. These results lead us to believe that we are genuinely impairing the knowledge of LLMs, not just suppressing it
Our headline result is simple: data filtering can send WMDP-Bio scores to random chance without hurting general performance. And FT'ing these models to match the performance of the baseline model w/o FT'ing requires 300M tokens of finetuning, >10x more than any other approach.
Are you afraid of LLMs teaching people how to build bioweapons? Have you tried just... not teaching LLMs about bioweapons? @eleutherai.bsky.social and the UK AISI joined forces to see what would happen, pretraining three 6.9B models for 500B tokens and producing 15 total models to study
"It's unclear if this research matters because real users speak English and Chinese" has got to be up there for worst dismissive takes about how multilingual doesn't matter.
At #ICML2025 and a fan of EleutherAI? Come find me for laptop stickers!
Our deep-dive case studies show typical AI blind spots: models struggle with long-tail knowledge absent from web data and extremely long contexts; without fully spelled-out derivations, they misinterpret calculations and ignore domain-specific conventions, leading to student-like errors. 🔍
We also look at what happens when you run a model multiple times: across 8 trials models rarely discover the same errors and generally assign a near-zero confidence to their claims.
We benchmarked 10 top models, both closed and open, and the results are sobering – the best results are from o3 which has a precision of 6% and a recall of 21%. All other models score below 4% precision and 10% recall.
2️⃣ The dataset spans long documents (avg 12,000 tokens) with rich visuals (avg 18 figures) 📊. Challenging both long-context reasoning and image understanding capabilities of modern LLMs. To make sure these errors are genuine, we only include those that have been acknowledged by the original authors!
People keep plugging AI "Co-Scientists," so what happens when you ask them to do an important task like finding errors in papers? We built SPOT, a dataset of STEM manuscripts across 10 fields annotated with real errors to find out. (tl;dr not even close to usable) #NLProc arxiv.org/abs/2505.11855
I can't read the linked article, but is this representation is even vaguely accurate that is deeply humiliating for ATI. LLMs may had caught the public by surprise in 2023 but they've been among the sexiest things in AI research since at least 2020 (when I entered the field, can't speak to before)
@eleutherai.bsky.social succeeded in its original goal of ensuring independent research on LLMs would be possible. But the research landscape is still overwhelming dominated by corporate funding and corporate priorities and there's much more still to do to help non-profit research thrive.
In case you're curious how much of a hellscape X is. I opened it today to get greeted with a porn ad. The censored section is a 20 second video clip of a woman sucking a dick, with audio. It autoplays.
Obligatory "actually my lab invented test-time-compute" post. In "Stay on topic with Classifier-Free Guidance," we show that CFG enables a model to expend twice as much compute at inference time and match the performance of a model twice as large. arxiv.org/abs/2306.17806