How it works

A sparse autoencoder compresses then reconstructs activations, forcing each feature to activate rarely. The result: one feature per concept, human-readable.