Cammarata argues AI neurons are more readable than believed, Nanda remains skeptical

Former OpenAI researcher Cammarata argues that famous examples of polysemantic neurons were misinterpreted, and that neurons in models like Qwen3-32B are mostly readable with one or two related jobs.

Google DeepMind interpretability lead Neel Nanda says the work is interesting but does not change the big picture and doubts that interpretability is close to being cracked. The technical companion paper has not yet been released.