Why it matters

Superposition frames interpretability challenges. Understanding shapes SAE motivation.

Advertisement

The architecture

Features as directions in activation space.

More features than dims possible via non-orthogonality.

Sparsity enables recovery.

Superposition hypothesisActivation spaced-dimFeatures > dnon-orthogonal directionsSparsity enablesrecovery via SAEElhage et al. 2022 toy models of superposition; motivates SAE approach
Superposition.
Advertisement

How it works end to end

Elhage et al. 2022 toy models.

Sparsity + limited dim drives.

Explains polysemanticity.