1. We will present HypotheSAEs at #ICML2025, Wednesday 11am (West Hall B2-B3 #W-421). 2. Let me know if you'd like to chat about: - AI for hypothesis generation - why SAEs are still useful - whether PhD students should stay in school

I've resolved this positively: 2 papers convincingly show sparse autoencoders beating baselines on real tasks: Hypothesis Generation & Auditing LLMs SAEs shine when you don't know what you're looking for, but lack precision. Sometimes the right tool for the job, sometimes not.

The mainstream view of AI for science says AI will rapidly accelerate science, and that we're on track to cure cancer, double the human lifespan, colonize space, and achieve a century of progress in the next decade. In a new AI Snake Oil essay, Arvind Narayanan and I argue that

🌟 HypotheSAEs update: open LLMs now supported for the full hypothesis generation pipeline! Labeling SAE neurons and annotating concepts works very well with Qwen3-8B and larger models ⬇️ (notably, other models didn't work as well). Brief 🧵

One paragraph pitch for why sparse autoencoders are cool: Text embeddings capture tons of information, but individual dimensions are uninterpretable. It would be great if each dimension reflected a concept (“dimension 12 is about cats”). But text embeddings are ~1000 dimensions

How do we reconcile excitement about sparse autoencoders with negative results showing that they underperform simple baselines? Our new position paper makes a distinction: SAEs are very useful for tools for discovering *unknown* concepts, less good for acting on *known* concepts.

1/10. In a new paper with Hamidah Oderinwale and Jon Kleinberg, we mapped the family trees of 1.86 million AI models on Hugging Face — the largest open-model ecosystem in the world. AI evolution looks kind of like biology, but with some strange twists. 🧬🤖

🚨 New postdoc position in our lab UC Berkeley EECS! 🚨 (please retweet + share with relevant candidates) We seek applicants with experience in language modeling who are excited about high-impact applications in the health and social sciences! More info in thread 1/3

Many thanks to Open Philanthropy for supporting our work on sparse autoencoders for hypothesis generation (arxiv.org/abs/2502.04382) - in particular, using these techniques to build safer and better-aligned LLMs! openphilanthropy.org/grants/uc-berk…

Excited to share our paper on zero-shot LLM evaluation was accepted to Findings of #EMNLP2025! We address a practical problem by analyzing LLM predictions on 316 prediction tasks: developing methods to predict LLM performance without labeled data. arxiv.org/pdf/2509.15356

Being Divya's labmate (and fellow ferry commuter) has been a real pleasure, and I've learned a ton from both her research itself and her approach to research (and also from the other random things she knows about).