A sparse autoencoder (SAE) reconstructs residual-stream activations through an overcomplete bottleneck, with an L1 penalty pushing most latent units to zero on any given input. Under the superposition hypothesis this is exactly the right inductive bias: the decoder columns converge toward the model’s own feature directions, of which only a handful are active at a time, so each latent unit tends to correspond to one human-recognizable concept.
Anthropic scaled SAEs from a proof-of-concept on a one-layer model (Towards Monosemanticity, 2023) to millions of features on a production Claude model (Scaling Monosemanticity, 2024), surfacing features for concepts from the Golden Gate Bridge to code bugs and sycophancy. Because features are directions, they can also be steered: clamping a feature to a high value changes behaviour, as in the Golden Gate Claude demo.
