
19 - Mechanistic Interpretability with Neel Nanda
AXRP - the AI X-risk Research Podcast
The Modeling of Attention Heads in Image Models
The attention head is the bit of the model that's due with rooting information between the different token positions and in part because this is just much more legible you can literally look at the attention patterns in the model. Attention pattern being which previous positions does the head think is most relevant to the current position okay yeah it is very easy to like beam misled by these but it does give you quite a lot of information sorry why is it easy to beam misled? The final answer to the question is we're just much better at interpreting the cognition around attention heads that we are about neurons.
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.