Instead of treating a model as a black box, mech interp identifies specific neurons, attention heads, and circuits that implement recognizable behaviors. Anthropic's work on sparse autoencoders is the current frontier.
GlossaryPart of the BotDoors AI Safety & Compliance Glossary · 37 terms.
Alignment
Mechanistic Interpretability
Mechanistic interpretability reverse-engineers neural networks into human-readable circuits and features.