
19 - Mechanistic Interpretability with Neel Nanda
AXRP - the AI X-risk Research Podcast
Using the Token Embeddings in a Model Is a Good Idea
The key thing here is that these are low rank matrices which means they're essentially identifying a slow rank subspace of the residual stream yep and a m2.0 new functionThese are composers okay so a non-obvious thing here is why am I separating k and v composition given that they're both from the source residual stream you know? So for example queue composition is when the queue inputs the where should move information to is significantly influenced by something that does not be token embeddingsOkay composition here being usefully using the output of a previous component just because that's like yeah like that’s the output of a previous function it's being m2.
00:00
Transcript
Play full episode
Remember Everything You Learn from Podcasts
Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.