"Your circuits make me nervous," goes the p(doom) song, and the annotation makes clear that these are circuits in the sense of mechanistic interpretability (osmarks, n.d.). The term comes from Olah et al. (2020), who, studying the InceptionV1 vision network, put forward three "speculative claims" modeled on Schwann's three theses about the cell (the third of which turned out to be false):
- Features: features are the fundamental unit of neural networks and correspond to directions in a layer's activation space.
- Circuits: features are connected by weights and form circuits, computational subgraphs that can be rigorously studied; for instance, curve detectors built from edge detectors in earlier layers.
- Universality: analogous features and circuits form across different models and tasks.
Elhage et al. (2021) carried the program over to transformers by recasting the architecture in a mathematically equivalent way. The residual stream works as a communication channel in which every component reads from and writes to subspaces; each attention head splits into a QK circuit, which decides where to look, and an OV circuit, which decides what information to move. In attention-only toy models, a zero-layer transformer amounts to a bigram table; a one-layer one, to bigrams plus "skip-trigrams" (A… B C); with two layers, induction heads appear, which complete sequences [A][B]…[A] → [B].
Buckner (2024) stresses the philosophical interest of that last finding: copying enriched with abstract similarity ([A][B]…[A] → [B′]) yields something rule-like, as Locke and Hume would have it (2024, p. 282). The question remains open: is a circuit a rule the network follows, or a regularity we describe in the vocabulary of rules?