callummcdougall / sae_vis

Create feature-centric and prompt-centric visualizations for sparse autoencoders (like those from Anthropic's published research).
MIT License
140 stars 27 forks source link