// ⌖ Interpretability

Sparse Autoencoders (SAE)

Decomposes activations into many sparse, interpretable features.

稀疏自编码器 SAE 2023 Advanced core

IN PLAIN WORDS

A much wider autoencoder with sparse activations reconstructs internal activations; units often map to crisp concepts (Golden Gate Bridge, code bugs). Anthropic, OpenAI and DeepMind scaled it to frontier models.

Used in: Claude features, Gemma Scope, Neuronpedia

TIMELINE · prerequisites → this → what builds on it

2013 Variational A…2022 Superposition2023 Sparse Autoencode…2025 Circuit Traci…

01Original paper

Sparse Autoencoders Find Highly Interpretable Features in Language Models

Hoagy Cunningham, Aidan Ewart, Logan Riggs, et al. · 2023

arXiv ↗ ✓ Verified · arXiv API

02Where it sits

Learn the left column first; the right column builds on this one.

Learn first

Sparse Autoencoders (SAE)

Then you can learn

Circuit Tracing & Attribution Graphs

Related:Activation Steering

Left: prerequisites · right: builds on this · dashed: related. Full map →

03Learning resources

Official docs, explainers, courses and reference code. All links checked.

Papers

04Where the course teaches it

U91 Interpretability: from Attribution to Circuits Attribution, probes, the logit lens, superposition, SAEs, circuit tracing, knowledge editing and activation steering. 10 h9 algorithms4 paperslesson ready

On learning paths:Generative Models & Frontier Research

For AI / data

Verification: verified, checked 2026-10-21.
JSON: /api/learn/algorithms.json (id: interp-sae)