Notas do episódio
A circuit can keep 99.5% of a model's right answers and still miss most of its mistakes. We walk through how the paper measures that gap, why it survives the obvious excuses, and the two missing computations that make GPT-2 pick the wrong name.

