Catherine Olsson and Nelson Elhage: Anthropic, Understanding Transformers
The Gradient: Perspectives on AI
00:00
You've Got a Program on Mechanistic Interpretability
The hope is that if our program sort of succeeds in the best possible version, you could use it to generate those much more specific answers. If we know how it works in general, we can also answer how it worked in some particular case. Like we generally don't know if it'll work, but we're hopeful that it will and we think there's good reason to believe that it will. And so your first work we can dive into here is a mathematical framework for transformer circuits. The word circuits comes from the earlier work in this mechanistic interpretability space.
Transcript
Play full episode
