Reverse-engineer how language models think — features, circuits, and neurons — to explain and audit their behavior.
