Stella Biderman

@BlancheMinerva

Ensuring that tech companies don't have a monopoly on being able to do research on cutting edge AI @AiEleuther. She/her

stellabiderman.ai Joined May 2019

928 Following 19K Followers 15K Posts

Latest posts 5

Pinned post

Stella Biderman @BlancheMinerva Aug 7

If you follow me and haven’t listened to this talk in its entirety, doing so 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.

Eric Wallace Verified account @Eric_Wallace_ Aug 6

Yesterday, my OpenAI collaborator and I gave a detailed talk on the Huggingface incident, our models creating "the message board", model misalignment, and more.

youtube.com/watch?v=87DyyM…

I hope it can answer a lot of the questions folks have, and we will release a full detailed postmortem at a later time!
View quoted post on X

Stella Biderman @BlancheMinerva 10h

A really interest deep dive on (apparently) eval awareness.

This reminds me of a lesson from working on T0: generalization in FT requires significant surface form variation. Too similar surface forms and models treat the presentation format as a causal effect on the output.

Daniel Yoo @_yooniel_ 1d

My first blog post: "Revisiting Verbalized Eval Awareness in OLMo and tracing it to the training data"

Verbalized eval awareness (VEA) might not be a reliable measure of how aware a model is. In OLMo 3, VEA is mostly copied from the SFT data.
View quoted post on X

Stella Biderman @BlancheMinerva 10h

Really interesting work.

Zhizhuo (Z) Zhou Verified account @zhizdev 2d

Why is AI slop so frustrating to read?

AI writing often obscures main ideas and invents awkward phrases, making sentences more complex than necessary.

Many of these slop patterns violate writing principles established in English writing books.

In our benchmark, AI writing breaks these principles far more often than human writing, with Claude Code replies by far the worst offenders.

We show examples of the most common slop patterns in this thread

View quoted post on X

Stella Biderman @BlancheMinerva 5d

We publish A at a workshop.

We collab w/ people to write B, which goes on arXiv

We update A and submit it to ICLR

Reviewer calls us out on claiming novelty for things that were in B. They were also in the workshop version and are in B b/c of us

How would you approach this?

Stella Biderman @BlancheMinerva Nov 4

If the Democrats do an above average job of being leaders and are willing to put being real liberals and making the world a better place as their #1 priority it would be the end of the Republican Party as a political force in the US.