Nick Cammarata's interpretability reversal

Nick Cammarata, who co-authored early mechanistic interpretability research at OpenAI, has reversed his earlier pessimism about the field. Where a large-scale automated interpretability project at OpenAI once found only about 2% of neurons legible using primitive tools, Cammarata now says the vast majority of neurons in models up to trillions of parameters are legible and their circuits comprehensible. He attributes the earlier pessimism partly to misunderstanding: many neurons previously cited as tangled or uninterpretable turned out not to be genuinely illegible. Using modern agentic tools and closer examination, the picture has changed substantially. Cammarata now believes ambitious interpretability will be solved before it is critically needed, and potentially soon as automated scientific tools continue to improve.