“Training a sparse autoencoder on a neural network's activations is an effective method of dictionary learning that reveals interpretable, monosemantic features.”