Patterning: Shaping model structure via training data

Patterning is the dual of mechanistic interpretability: given a desired internal structure, it determines what training data produces it to move beyond "growing" neural networks.