Advanced · Safety

Sparse Autoencoder (SAE)

Visual diagram · (in preparation) · Math · (in preparation) · Worked example · 3 difficulty levels.

TL;DR. An overcomplete autoencoder with an L1 penalty used in interpretability to decompose neural activations into monosemantic features.

Technical Definition

An overcomplete autoencoder with an L1 penalty used in interpretability to decompose neural activations into monosemantic features.

How it works

Anthropic's 'Towards Monosemanticity' (2023) and 'Scaling Monosemanticity' (2024) trained SAEs on Transformer residual streams with dictionary sizes 8–256× larger than the activation dimension and a sparsity penalty. The learned features are far more interpretable than raw neurons — corresponding to concepts like 'the Golden Gate Bridge' or 'unsafe code' — and can be steered by clamping their activations.

Related Concepts

• Autoencoder — A neural network that learns compressed representations by training to reconstruct its own input through a bottleneck layer.
• Representation Learning — The process of automatically discovering meaningful representations of data from raw inputs.
• Mechanistic Interpretability — A research program that reverse-engineers neural networks into human-understandable circuits, features, and algorithms.
• Activation Patching — An interpretability technique that swaps activations from one forward pass into another to causally localize where a behavior lives in a model.