Writing
A compact intuition for using sparse autoencoders to turn opaque activations into interpretable features.
