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.