Peaked at #1 Now #2
Interpretability Called Most Important Science
Nick Cammarata says interpretability may be the most important scientific problem, recalling his pessimistic early mechanistic interpretability work at OpenAI and later large-scale automated interpretability project.
Key points
- Nick Cammarata says he co-authored early work in mechanistic interpretability at OpenAI and later started OpenAI's first large-scale automated interpretability project, which found about 2% of neurons legible with the primitive tools available at the time.
- He says that with modern agentic tools the vast majority of models up to trillions of parameters are legible, and that many famous examples of tangled neurons turned out to be misunderstood.
- He says he now thinks ambitious interpretability will be solved by the time it is needed most, and possibly quite soon.
Key points and the reaction summary are written by AI from the posts on this page. Check the original post. How we use AI
Original post
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
https://t.co/zHij5zwKGT
Top replies on X
Reaction on X: The replies are largely supportive and positive, with one user asking whether the approach will hold for new non-transformer architectures; the original poster did not respond.
Sign in to see 5 top replies from X
From @simonsarris, @robinhanson, @bayeslord and others. Spam removed, with English and Chinese translations.
Discussion 0
Sign up Sign in to join the discussion
No comments yet. Start the conversation.