Stella Biderman

If you have more than even a passing interest in AI, watching this talk is more important than whatever it is you are currently doing. There’s a lot of cybersec jargon at times but I don’t think you need to know anything about security to understand what’s important. youtu.be/…

Black Hat USA 2026: The 'Breaking' News: The OpenAI–Hugging Face Incident youtu.be

Claude code is stenographically marking location information in its prompt to help Anthropic profile users by where they (appear to) be located thereallo.dev/blog/claude-...

Claude Code Is Steganographically Marking Requests thereallo.dev

Read the full paper: arxiv.org/abs/2606.06533 or come listen to our oral @icmlconf.bsky.social! Huge thanks to my co-authors @aflah02101.bsky.social Niloofar @catherinearnett.bsky.social @fbarez.bsky.social @nsaphra.bsky.social Stay tuned for a related workshop (hopefully) at …

Position: Don't Just "Fix it in Post": A Science of AI Must Study Training Dynamics arxiv.org

I had given Anthropic a lot of credit for turning down the DoD and its trillions of dollars, especially as it seemed to be the only example of any AI company making any financial sacrifices for moral principles. Of course, it turns out to be not really true. www.axios.com/2026…

axios.com

Note that the total openly licensed and public domain data seems to still be growing due to the fact that every year we generate more data even if many sites are restricting use. We discuss this in our recent paper on building open training corpora arxiv.org/abs/2506.05209

The Common Pile v0.1: An 8TB Dataset of Public Domain and Openly Licensed Text arxiv.org

See also our work on how previously open websites are becoming much more closed, documented “Consent in Crisis: The Rapid Decline of the AI Data Commons” arxiv.org/abs/2407.14933

Consent in Crisis: The Rapid Decline of the AI Data Commons arxiv.org

That talk also inspired a fun rule: if I go around saying “ someone should do X” for two years I have to do it or stop talking about it. This rule has caused the creation of several papers I’m very proud of including arxiv.org/abs/2508.06601 and arxiv.org/abs/2506.05209

Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs arxiv.org

Never forget that you live in a bubble and the world is full of people wanting to turn your every invention into a weapon to create dystopia. Opposing this is a fundamental ethical responsibility of any AI researcher. reason.com/2026/07/28/m...

Minority report: FBI seeks AI for political watch list reason.com

This is a crazy good article about AI and environmental impacts: blog.andymasley.com/p/a-cheat-sh... By @andymasley.bsky.social

Using ChatGPT is not bad for the environment - a cheat sheet blog.andymasley.com

I’m starting a blog! My first post is on how 3rd party embedded evaluators seem totally unsuited to addressing the problems we are currently facing, and what the real problem is. stellabiderman.ai/blog/embedde...

Embedded Evaluators Can’t Fix Companies That Choose to Be Bad — Stella Biderman stellabiderman.ai

These concerns are very much not hypothetical. In 2024 a private benchmark called FrontierMath developed by a third party called Epoch came out. We later learned that OpenAI funded the creation of the benchmark and while nobody else had the test Qs, OAI did. ai-frontiers.org/a…

Don’t Let AI Developers Hire Their Own Referees ai-frontiers.org

At @eleutherai.bsky.social we’ve worked to inform lawmakers in the US, Europe, and around the world about the impacts that their proposed regulations would have and how to tailor them more narrowly to accomplish their goals, including for the EU AI Act: blog.eleuther.ai/eu-aia/

EleutherAI's Thoughts on the EU AI Act | EleutherAI Blog blog.eleuther.ai

If you believe their story, OpenAI accidentally committed cyber warfare against Hugging Face while doing what they thought was an internal test of a model without internet access. In six months time will be obvious they’re faced effectively no sanction for this.

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. 🧵

Whenever I interact with an academic discipline outside of AI and Theoretical Computer Science, I am rapidly reminded how few problems I face for being trans in AI. Thank y’all <3

have never seen an AI model come up with an amazing idea that I hadn’t already thought of for a problem I’m working on. And the typical idea quality is garbage. If this isn’t your experience I’d love to see examples! Post them or DM me.

I’m walking back from an AI event to my hotel and I look up and see a rock climbing gym called “Benchmark.” Is this what living in SF is like?

The ability of tech co. to produce "intelligent" systems that are incompetent at doing any of the tasks I actually want them to do is mind-boggling. TIL that ChatGPT and Claude generally don't agree when you give them two papers and ask how many citations they have in common.

When I was in middle school (late 2000s), I was very interested in mathematics. I thought that the validity of the proof of the 4-color theorem was disputed in mathematics and was shocked when I learned about FLT, but I knew that humans would never beat an AI at chess ever again

Everyone else claims that AI is massively accelerating their research and meanwhile my ChatGPT can’t update a spreadsheet without both breaking the formatting and changing the convention I had been using for the past 100+ rows. How do I get the version that works :(