features and circuits

This is the core vocabulary of mechanistic interpretability. A feature is a concept the model has learned to represent — a direction in its activation space that lights up for, say, the idea of a capital city, a Python function argument, or a sarcastic tone. A circuit is the wiring: the set of features plus the components that read some features and write others, forming a small computation that turns inputs into intermediate concepts into outputs.

The analogy is a logic diagram. Features are the meaningful signals on the wires; circuits are the gates and connections that transform them. A famous example is the induction circuit, where one head finds where the current token appeared before and a second head copies whatever followed it — two features and a link that together implement a copy-the-pattern rule. The bet behind the framework is that complex behavior decomposes into many such reusable parts.

The framework is powerful but leaky. Features are not always crisp, single-meaning directions — many are entangled by superposition, and the same neuron can participate in several circuits. So a circuit drawn cleanly on a whiteboard is usually an idealization of something more tangled inside the weights, and validating it requires causal tests, not just a plausible diagram.