Induction head — copying what came next last time
A two-layer attention circuit that, at the current token, finds an earlier occurrence of it and predicts the token that followed.
It explains how trained models copy patterns from their own context, and is thought to underlie much of in-context learning (section 14.3).