In Neural Networks are Already Clean, @nickcammarata argued that neurons were far more legible than previously thought – and that this was great news for interpretability and safety.

Goodfire researchers released a technical companion and public descriptions for 1,500 neurons in Qwen3.6 35B-A3B, arguing that neurons are far more legible and predictable than previously thought. The team found that neurons typically have few computational roles, and that ablating the weakest 92 percent of activations has modest effects on model capabilities. Open article →