According to Sam Altman on X citing Nick Cammarata’s post and essay at already-clean.goodfire.com, modern agentic tools make most neurons in large models legible, suggesting ambitious interpretability could arrive soon.

According to The Rundown AI citing Nick Cammarata, his 2023 OpenAI project explained about 1.7% of neurons, and he now reports at Goodfire that most neurons in models up to trillions of parameters appear understandable.