Safety & Alignment

Mechanistic Interpretability Primer

How sparse autoencoders extract human-interpretable features from model activations, what circuit-level analysis buys you for safety, and where the science is still contested.

advanced · 10 min read · Premium