The two core units of mechanistic interpretability: features (what the model represents) and circuits (how those representations get computed and combined).
Features
A feature is a property the model has learned to detect — a direction in its activation space that "lights up" for a concept: this text is in French, this is a curve, this is sarcasm. Early-layer features are simple (edges); deeper ones are abstract.
Circuits
A circuit is a connected set of features that implements a computation — the model's learned "code." A famous example is an "induction head" circuit that learns to continue repeated patterns (AB…AB → A then B).
features = the variables / concepts the model represents
circuits = the wiring that turns input features into output featuresSuperposition — the wrinkle
Models pack more features than they have neurons by overlapping them (superposition). So one neuron isn't one concept — it's a blurry mix, which is exactly what sparse autoencoders try to untangle.
Finding clean features and tracing circuits is how researchers turn a black box into something legible.
Related: Mechanistic Interpretability · Probing and Sparse Autoencoders · Attention Visualization