1 essay on this topic.
Superposition, sparse autoencoders, Golden Gate Claude and attribution graphs: how researchers find concepts inside a model, and what they still can't see.
