Catherine Olsson and Nelson Elhage: Anthropic, Understanding Transformers
The Gradient: Perspectives on AI
00:00
Is Mechanistic Interpretability a Good Idea?
The idea of mechanistic interpretability is a program really pioneered by Chris Olah. It's trying to look at neural networks at the sort of micro scale level and reverse engineer what individual weights or individual values in the network mean. And try to back out human understandable algorithms and interpretations from those components. So we don't know whether it's always possible, but this is a very exciting existence proof that for some networks, this approach really does work.
Transcript
Play full episode
