Nick Cammarata argues neural networks are "already clean," and most neurons can be understood directly

Zach DavidsonOct 15, 2026

Nick co-authored early interpretability work at OpenAI that helped convince the field that neurons pack unrelated concepts together ("superposition"). He now says that read was wrong; the messy neurons came from weak tools and older architectures. In early audits of Qwen3-32B, about 90% of neurons have one main job, and the pattern seems to hold up to Kimi K3 at trillions of parameters. If he's right, interpretability is far more tractable than assumed.