Understanding AI from the Inside: The State of Mechanistic Interpretability for AI Safety
A survey of mechanistic interpretability for AI safety that covers its foundations and methods, proposes an evaluation framework, and assesses how understanding internals supports editing, steering and monitoring.