Nick's focuses a lot on interpreting individual neurons. Although this bottom-up understanding is definitely useful, I feel that it is an obvious missed opportunity to not e.g. see how a dog photo is interpreted by each neuron within the network and whether it explains anything non-obvious. Seems like an interesting demo opportunity that is within reach of current LLM coding tools.
Nick's focuses a lot on interpreting individual neurons. Although this bottom-up understanding is definitely useful, I feel that it is an obvious missed opportunity to not e.g. see how a dog photo is interpreted by each neuron within the network and whether it explains anything non-obvious. Seems like an interesting demo opportunity that is within reach of current LLM coding tools.