Finding recognizable subroutines
A useful reframing for interpretability work is to treat a trained network the way a reverse engineer treats unfamiliar compiled code: you don't have the original source or comments, but you can still identify meaningful substructures by tracing how information flows through the program. In a neural network, the analogous substructures are called circuits, specific, identifiable subgraphs of neurons and connections that together implement some recognizable piece of behavior.
A well-known example from transformer language models is the induction head, a small circuit that helps a model continue a repeated pattern: if the token sequence 'A B' has appeared earlier in the context, an induction head raises the likelihood of predicting B again the next time A appears. This isn't a hand-coded rule, it's a pattern of attention and computation that emerges from training and can be located, studied, and even manipulated once found.