Sam Altman amplifies Nick Cammarata’s interpretability claim

Sam Altman shared his enthusiasm on October 16, 2026, for updates from Nick Cammarata on progress in mechanistic interpretability of neural networks.

According to the post, early work at OpenAI had shown neural networks as tangled, with automated projects finding only about 2 percent of neurons legible. Modern agentic tools now indicate the vast majority of neurons in models up to trillions of parameters are legible, with beautiful circuits. Many previously cited examples of tangled neurons proved misunderstood, suggesting that simple scaling approaches at large compute levels can resolve interpretability challenges in time for advanced systems.