Episode Description

For Episode 5 of AI Deep Dive, Goodfire co-founder and chief scientist Tom McGrath joins Rocket Drew to explore what it would take to open AI’s black box. Drawing on his work at Google DeepMind, where he co-founded the interpretability team, McGrath explains why written reasoning gives us an incomplete view of AI and how mechanistic interpretability could help us look deeper.

Could an AI “brain scan” catch cheating that a reasoning monitor misses? What happens when models do more thinking without writing it down? They explore the effort to reverse engineer language models, the challenge of changing what AI learns without hiding unwanted behavior, and how swarms of AI agents could help researchers understand other AI systems.