Mechanistic Interpretability explained | Chris Olah and Lex Fridman

3-Second Blueprint (English)

1. Neural networks are grown rather than programmed: architecture acts as a scaffold where gradient descent cultivates emergent circuits.
2. The universality hypothesis observes that distinct neural architectures often converge on the same natural feature abstractions.
3. Under linear representation, concepts map to activation directions like Word2Vec vectors, turning network weights into circuit edges.