Topic
Mechanistic interpretability
Mechanistic interpretability: reverse-engineering neural networks.
Mechanistic interpretability is the effort to reverse-engineer neural networks, working out how their internal parts produce their outputs. In March 2025, Anthropic published a study reporting that Claude 3.5 Haiku works out rhymes ahead of time and sometimes gives explanations that do not match the steps it took. Our Golden Gate Claude story recalls that in May 2024 Anthropic found one feature inside Claude 3 Sonnet, and the model talked about the Golden Gate Bridge for 24 hours. A story from November 2026 explains what happens inside an AI answer, covering features, circuits and research from 2024 and 2025. Another November 2026 story reviews the March 2025 study on tracing the thoughts of a language model.