Definition
Plain language
A small group of parts inside an AI model that work together to carry out one specific job, like a little wiring diagram.
As stated in the literature
In mechanistic interpretability, a subgraph of model components — particular attention heads, MLP units, and the connections between them — that jointly implements an identifiable behavior; the level at which OV and QK circuits are analyzed.
Also called: circuits
Why it matters: Identifying circuits lets researchers explain why a model does what it does, turning an opaque network into something that can be inspected and debugged.
For example, researchers might find that a handful of attention heads working together are responsible for copying a name mentioned earlier in a sentence.
Heard on the show
“A lot of mechanistic interpretability proceeds by looking for localized causes — the salient token, the sparse feature, the identifiable circuit — because that's what makes attribution tractable.”Episode 243 — How a Hundred Meaningless Word Choices Add Up to Flip a Model's Answer