Why it matters

SAE shapes modern interpretability. Understanding shapes latest research.

Advertisement

The architecture

Train SAE on model activations.

Sparse + reconstructive.

Features often monosemantic.

SAE structureActivationstarget layerSAE encodersparse dictionaryFeaturesoften monosemanticAnthropic Golden Gate + Towards Monosemanticity + Scaling SAEs papers
SAE.
Advertisement

How it works end to end

Bricken et al. 2023 Towards Monosemanticity.

Templeton et al. 2024 scaling SAEs.

Gated + Top-K SAE variants.