sparse-autoencoders

Mechanistic interpretability reverse-engineers neural networks into features and circuits. What it has found in GPT-2 and Claude, and what it cannot yet do.