Mechanistic Interpretability research @DeepMind. Formerly @AnthropicAI, independent In this to reduce AI X-risk. Neural networks can be understood, let's do it!