a circuit
Knowing the words a model uses internally (its features) is like knowing the parts on a workbench. But understanding a machine means seeing how the parts connect to do a job. A circuit is exactly that: a small set of internal components — particular neurons, attention heads, and the features they read and write — wired together so that, step by step, they carry out one identifiable computation.
Concretely, a circuit is a partial wiring diagram of the network for a specific task. You trace how an early part detects something, passes it along, and a later part acts on it. The classic example is the induction circuit in transformers: one attention head finds where the current token last appeared, a second head looks at what followed it last time, and together they implement 'predict the same continuation again'. You can name each component's job and show that removing any one of them breaks the behavior — that is what turns a guess into a circuit.
Circuits matter because they are the level at which 'why did it do that?' gets a real answer rather than a vibe. If you could map the circuit a model uses to decide whether to refuse a harmful request, you could check whether it is doing genuine harm-reasoning or just keyword-matching that an attacker could dodge. The honest caveat is that fully mapped circuits so far cover narrow, mostly toy behaviors; scaling circuit analysis to a model's important, messy capabilities remains an open and hard problem.
In the indirect-object task 'When John and Mary went to the store, John gave a drink to ___', a transformer answers 'Mary'. Researchers mapped a circuit of cooperating attention heads — some that find the names, one that inhibits the repeated name 'John' — that together produce the right answer.
A circuit is a wiring diagram for one behavior: features plus the components that route them.
A circuit found by hand is a hypothesis until tested causally (for example by ablating each part); a diagram that merely looks sensible can still be an interpretation illusion.