Your bookmarks

Loading your bookmarks

Latest updates

New courses, skills, book chapters, and site improvements will appear here.

Create free accountSign upDashboardDashboard

Circuit

  • Control
  • Research term
2 steps

A circuit is a proposed set of parts inside a model, together with the connections among them, that carries out or contributes to a specific behavior.

A model works in layers, each passing a list of numbers to the next, and one behavior is thought to use a small share of that traffic. A part may be one unit, the slot that holds one of those numbers, or a pattern spread over many units. One part often serves many behaviors, so circuits overlap and changing a part can disturb others.

Where you’ll meet it

A tech newsletter you read says researchers have "found the part of an AI that does addition." Its diagram shows about a dozen connected parts lighting up as a model adds two-digit numbers.

What to do

Read it as a claim about one model and one kind of problem. Parts lighting up shows only that they were active. Look for the stronger evidence, a test where the researchers switched those parts off and the model failed at new sums while its other skills kept working.

Easy to mix up

A feature is a pattern of activity that stands for one concept, such as a city. A circuit shows how parts like these connect to carry the model from an input to a behavior.

Check yourself

Researchers switch off the circuit they proposed for finishing a sentence with the right name. The model still gets most of those sentences right. What does that result show?

Go deeperCircuit research began with image models and moved to language models

Chris Olah and colleagues set out the approach in "Zoom In: An Introduction to Circuits" (Distill, 2020), arguing from an image model that its features connect into small, readable algorithms. In 2022 Kevin Wang and coauthors mapped about two dozen attention heads in the language model GPT-2 small that together pick which earlier name should finish a sentence. Each head is a small component that chooses which earlier words to draw on.

The switching-off test is called "ablation." "Activation patching" copies a part's values from a run on one input into a run on another, which shows whether that part carries the difference between them. A circuit's boundaries depend on the method and the cutoff the researchers chose, so two studies can draw different circuits for the same behavior.

Most detailed circuits come from small models or narrow tasks. Anthropic's 2025 attribution graphs trace circuits automatically in a much larger model, one prompt at a time, and the authors note that each graph accounts for only part of what the model computed.

Practice Modern AI

Or take the learning quiz on a mix of every set