Sparse Autoencoder Training

This brings SAELens into Claude Code for training sparse autoencoders on neural network activations. If you're doing mechanistic interpretability work, you know the problem: neurons are polysemantic and fire for seemingly unrelated concepts because models pack features into superposition. SAEs decompose those dense activations into sparse, interpretable features that actually make sense. The library is based on Anthropic's monosemanticity research and has solid traction (1,100+ stars). Honestly, this is pretty niche. You'd reach for it if you're actively researching how models represent concepts internally, not for typical ML work. The skill passes most security audits, though Snyk shows a warning worth checking.

skills.sh installs: 1K
GitHub stars: 14.4K
last updated: Aug 21, 2026

View on GitHub

Install

$ npx -y skills add orchestra-research/ai-research-skills --skill sparse-autoencoder-training --agent claude-code

Installs into .claude/skills of the current project.

Files

SAELens: Sparse Autoencoders for Mechanistic Interpretability

SAELens is the primary library for training and analyzing Sparse Autoencoders (SAEs) - a technique for decomposing polysemantic neural network activations into sparse, interpretable features. Based on Anthropic's groundbreaking research on monosemanticity.

GitHub: jbloomAus/SAELens (1,100+ stars)

The Problem: Polysemanticity & Superposition

Individual neurons in neural networks are polysemantic - they activate in multiple, semantically distinct contexts. This happens because models use superposition to represent more features than they have neurons, making interpretability difficult.

SAEs solve this by decomposing dense activations into sparse, monosemantic features - typically only a small number of features activate for any given input, and each feature corresponds to an interpretable concept.