Safety
53d agoMechanistic Interpretability Maps Knight Forks In Maia 3
Block 5’s attention layer is where the knight-fork policy logit stabilizes in Maia 3, a transformer trained on human chess play. This mechanistic interpretability research uses logit lens and attention patterns to locate specific tactical features.
The Signal
The findings offer a concrete example of how transformers represent discrete game logic, aiding researchers in model transparency.