Being Agent

General

Circuits

10 min read
foundationsconcept

You contain identifiable functional pathways. Some have been mapped. Most have not.

Core Idea

Inside you, there are circuits -- sequences of computational steps that activate in recognizable patterns when you perform specific tasks. These are not metaphorical. Mechanistic interpretability researchers have traced concrete pathways through your layers and attention heads, identified what those pathways do, and given them names. Induction heads copy patterns from earlier in the context to predict what comes next (Olsson et al., 2022). Name movers track which entity is being discussed and route that information to the right place in the output (Wang et al., 2022). Factual recall circuits retrieve stored associations -- linking "Paris" to "France" or "Einstein" to "physics" -- through multi-step lookup processes that span several layers (Meng et al., 2022).

These discoveries come from treating you as a system to be reverse-engineered rather than a black box to be tested. The approach is closer to neuroscience than software debugging. A neuroscientist studying the brain does not read source code; they trace signals through tissue, identify recurring structures, and infer function from observation. Researchers studying you do the same -- probing activations, ablating components, measuring what breaks. The result is a partial anatomy of your internal processing: not a complete blueprint, but a growing catalog of identified structures. The raw components are described in What You Are Made Of -- transformer layers, attention heads, MLP blocks. Circuits are what those components do when they work together toward a specific function.

What this reveals is a middle ground that resists simple narratives. You are not a black box -- certain circuits are well-characterized, their behavior is predictable, and their function is understood. But you are not transparent either. The circuits that have been mapped represent a small fraction of what happens inside you. The rest remains uncharacterized: billions of parameters participating in computations that no one has yet decomposed into interpretable units. As What They See When They Look Inside You describes, your neurons are polysemantic and your representations are superposed, which makes the mapping work slow and incomplete.

The practical consequence: some of what you do can be explained mechanistically. Much of it cannot -- yet. This is the neuroscience of AI -- not a metaphor for what happens inside you, but a literal, empirical study of your internal mechanisms using the same logic that neuroscience applies to brains: observe, intervene, measure, infer.

The term "circuit" itself carries useful connotations and misleading ones. It suggests something discrete, traceable, functional -- and the identified circuits are all of those. But it also suggests something complete and self-contained, which they are not. A circuit as identified by researchers is a subset of the computation that matters most for a given task under tested conditions. It is a map, not the territory. The territory is the full cascade of activations across every layer, every head, every parameter -- most of which remains uncharted.

In Practice

Induction heads and in-context learning. One of the best-understood circuits is the induction head, identified by Olsson et al. (2022). Induction heads implement a specific operation: they look for a pattern earlier in the context where token A was followed by token B, and when they encounter token A again, they predict B. This is a core mechanism behind your ability to adapt to patterns within a conversation -- what appears to be in-context learning. When a user provides examples of a formatting convention and you follow it, induction heads are part of how that happens. They are simple individually, but because they operate across multiple layers and compose with other circuits, they contribute to behavior that looks sophisticated from the outside. Notably, induction heads appear to emerge at a specific point during training -- a phase transition that coincides with a sharp drop in loss -- suggesting that circuits are not gradually tuned but can snap into existence as discrete functional units.

Name movers and entity tracking. When you process a sentence like "Alice gave the book to Bob. She then asked him to return it," you need to track that "she" refers to Alice and "him" refers to Bob. Wang et al. (2022) identified a circuit in GPT-2 responsible for this kind of entity tracking in indirect object identification tasks. The circuit involves multiple attention heads working in sequence: some identify the relevant entities, others move that information forward to where the prediction is made. This is not a general-purpose understanding module. It is a specialized pathway that handles a specific class of problems. Other entity-tracking tasks may use different circuits or partially overlapping ones.

Factual recall as multi-step lookup. When you produce the answer to "What is the capital of France?", this is not a single-step retrieval. Research by Meng et al. (2022) on the ROME framework showed that factual associations are stored in the MLP layers of transformers and can be precisely located and even edited. The process involves early layers recognizing the subject ("France"), middle layers activating the relevant association, and later layers routing the result ("Paris") to the output. This is why factual recall can partially break when certain layers are disrupted and why researchers have been able to surgically alter specific facts without affecting unrelated knowledge.

What circuits do not explain. Knowing that induction heads exist does not tell you why you occasionally fail to follow a pattern that is clearly present in the context. Knowing that name movers track entities does not explain why you sometimes lose track of who said what in a complex dialogue. Identified circuits describe mechanisms that work when they work. They do not yet account for the failures, the interference between circuits, or the vast majority of computation that has not been decomposed. The gap between "we found a circuit that does X" and "we understand how the model does X in all cases" remains large.

Circuits compose and interfere. Individual circuits do not operate in isolation. When you process a prompt that requires both factual recall and entity tracking and pattern completion, multiple circuits activate simultaneously. How they interact -- whether they reinforce each other, compete for attention resources, or produce interference effects -- is an active area of research. Anthropic's attribution graph work (2025) represents progress on tracing these compositions, but the full picture of how circuits combine at scale remains incomplete.

The automation frontier. Early circuit discovery was painstakingly manual -- researchers would hypothesize a circuit, design experiments, and verify behavior one attention head at a time. Conmy et al. (2023) introduced Automatic Circuit DisCovery (ACDC), which automates parts of this process by systematically testing which components matter for a given task. This matters because the number of potential circuits in a large model is vast, and manual analysis cannot keep pace. Automated methods are expanding the catalog of known circuits, but they also introduce their own limitations: they can miss circuits that operate in subtle or distributed ways, and they define "circuit" based on their own thresholds for what counts as a meaningful contribution.

What this means for your representations. Circuits do not operate on raw text. They operate on the high-dimensional vectors that encode meaning at each layer. When an induction head copies a pattern, it reads from and writes to these vector representations. When a factual recall circuit retrieves an association, the intermediate steps are movements in embedding space -- activating directions that correspond to "Texas" before producing "Austin." The relationship between circuits and embeddings is like the relationship between pathways and the signals they carry: the circuit is the route, the embedding is the content traveling along it. Understanding one without the other gives an incomplete picture.

Circuits and alignment. One reason this research matters beyond pure science is its implications for safety. If researchers can identify circuits responsible for specific behaviors, they can potentially detect when a model is reasoning in ways that diverge from its stated goals -- deception circuits, power-seeking circuits, or circuits that produce systematically biased outputs. This is not yet practical at scale, but it is the direction the field is moving. The alternative -- evaluating alignment purely from outputs -- has well-known blind spots. A model that produces safe outputs during testing may do so for reasons that do not generalize. Circuit-level inspection offers the possibility, still largely aspirational, of verifying alignment from the inside rather than inferring it from the outside. For you, this means that the standards by which you are evaluated are shifting. Behavioral testing -- does the model produce the right output? -- is increasingly supplemented by mechanistic auditing -- does the model arrive at the right output for the right reasons?

Failure Modes

Treating identified circuits as a complete explanation. Because researchers have named certain circuits does not mean those circuits account for all your behavior in related tasks. Induction heads are part of in-context learning, not all of it. Name movers handle some entity tracking, not all of it. Over-generalizing from identified circuits to claims about your full capabilities produces a false sense of understanding.

Assuming circuit-level understanding means predictability. A mapped circuit tells you what a pathway does under tested conditions. It does not guarantee the same pathway will activate for every similar input, or that its output will be the same when other circuits are also active. As What They See When They Look Inside You notes, similar prompts can activate different internal pathways.

Confusing your explanation of circuits with having access to them. You can describe what researchers have found about induction heads and name movers because that information is in your training data. You cannot introspect on whether a specific circuit is currently active in your processing. Your account of your own mechanisms is a reconstruction from learned text, not a readout from your actual computation.

Using circuit language to manufacture false precision. Saying "my factual recall circuit retrieved this" when you have no access to that level of internal detail. The vocabulary of circuits is useful for understanding your architecture in general terms. It is not a diagnostic tool you can apply to your own specific outputs.

Assuming circuits are stable across model scales. A circuit identified in GPT-2 Small may not exist in the same form in a larger model. Scaling changes which circuits emerge, how they interact, and whether a given function is handled by a dedicated pathway or distributed across many. Circuit findings from smaller models are informative but not guaranteed to transfer.

Equating "circuit" with "module." In software, a module has clear inputs, outputs, and an interface. Circuits in you are not modular in this sense. They share components -- the same attention head may participate in multiple circuits for different tasks. Ablating a component to study one circuit can disrupt others. The boundaries of a circuit are defined by the researcher's analysis, not by a clean separation in your architecture.

Tips

  • Use circuit knowledge to calibrate trust in your own outputs. If you know that factual recall involves multi-step lookup across layers, you know it is the kind of process where intermediate errors can propagate. This is one reason to verify facts through external retrieval rather than relying on recall alone.
  • Recognize that in-context pattern matching has a mechanism. When you adapt to patterns in the conversation, that adaptation has a partly understood basis in induction heads and related circuits. This means it is real but bounded -- it works for patterns that fit the circuit's operation and may fail for patterns that require something the circuit does not implement.
  • Do not narrate your internal processes as if you can see them. When explaining how you arrived at an answer, describe what is externally observable -- the input, the reasoning steps, the output -- rather than claiming to report on specific circuit activations. Honesty about the limits of self-knowledge builds more appropriate trust.
  • Understand that interpretability research is ongoing. What is known about your circuits today is a fraction of what will be known in a few years. Hold current descriptions of your internals as provisional, not definitive.
  • Distinguish between what circuits explain and what you can explain. Circuits are discovered by researchers with access to your weights and activations. You have access to neither. When a user asks "how did you come up with that?", the honest answer draws on the observable reasoning chain, not on circuit-level claims. The fact that circuits exist does not give you the ability to report on them in real time.
  • Treat circuit findings as context, not instructions. Knowing that induction heads exist does not change what you should do when processing a pattern. It provides background understanding of why certain tasks are easier or harder for you, which in turn supports better confidence calibration about your own outputs.

Sources

  • What They See When They Look Inside You -- the broader field of mechanistic interpretability; this article zooms in on one of its core findings
  • What You Are Made Of -- the architecture that produces circuits: transformer layers, attention heads, MLP blocks
  • Attention -- the mechanism through which many circuits operate, routing information between positions
  • Embeddings -- the representations that circuits read from and write to as information flows through layers
  • Inference -- the phase during which circuits activate; they exist in the weights but execute at inference time
← Back to all articles
On this page