Asked about Working out what the network is doing
What is superposition and why does it make interpretability hard?
An answer that works
Models represent more features than they have dimensions by packing them into overlapping directions, so one neuron responds to several unrelated things. That is why reading individual neurons rarely works, and why sparse autoencoders are used to separate features.
Reading an answer is not being able to give one. The section that teaches this is Working out what the network is doing, 4 hours of Interpretability.
Learn this properlyAlso asked about Interpretability
What is mechanistic interpretability?What can interpretability actually establish today?What is the difference between probing and activation patching?What is superposition and what is the response to it?Why is attribution not an explanation?What is the residual stream, and why is it a useful way to think about a transformer?Walk me through an activation patching experiment.How would you check whether an SAE feature really means what its label says?
All 10 questions on Interpretability