Neel Nanda presents a tour of mechanistic interpretability, arguing that machine learning models develop human-comprehensible algorithms even without explicit guidance. He explains how techniques like sparse autoencoders help uncover hidden model str...

In this discussion, Chris Olah explains mechanistic interpretability, a field focused on understanding the algorithms inside neural networks by “growing” them rather than programming them directly. He walks through how features and circuits emerg...

This video explores how researchers use mechanistic interpretability—especially sparse autoencoders—to uncover hidden, human‐understandable features in large language models. It highlights the challenges of pinning down internal model behavio...