The Pulse
Goodfire’s Nick Cammarata Recasts Superposition as Generalization
Goodfire’s Nick Cammarata argues that neural networks are more interpretable than the superposition hypothesis suggests. His essay reports preliminary audits of Qwen3-32B and calls for studying neurons and circuits directly.
AI.info Team · Written with AI ·
“Neural networks are already clean.”
Nick Cammarata, Goodfire
Goodfire researcher Nick Cammarata’s essay, published October 15, argues that a leading explanation for apparently mixed-up neural-network neurons has been misread. Rather than treating many neurons as arbitrary combinations of unrelated concepts, he says, researchers should look first for ordinary generalization: a neuron may respond to several related patterns, or to a pattern that broadens as a model learns. Cammarata calls the essay an informal working hypothesis, not a finished account of the question.
Why Cammarata revisits the cat-and-car neuron
Superposition has helped explain how networks with limited representational capacity can encode more features than they have neurons. Under that account, unrelated features share directions in a network’s internal space, and later computations read those features back out. The idea helped motivate methods such as sparse autoencoders, which try to separate a network’s overlapping activations into more interpretable features.
Cammarata focuses on a famous InceptionV1 example: a neuron that appeared to respond to both cat faces and car fronts. He now argues that the shared pattern may be a V shape, visible in the features the neuron responds to and in its circuit connections. In his reading, the neuron is not hiding two unrelated concepts in one representation; it is responding to a common visual form that occurs in both.
He also revisits an earlier circuit presented as evidence of superposition, in which a car feature seemed to spread through neurons associated with dogs. Cammarata says the downstream network did not read out the alleged combination, making the circuit a poor example of features packed together for later recovery. He writes that he knows of no second published circuit with the same proposed structure.
Qwen3-32B supplies a preliminary check
The essay’s numerical centerpiece is a preliminary audit of Qwen3-32B. Cammarata estimates that roughly 90% of neuron poles have one main job, often with related offshoots; about 7% have two related jobs; and 2–4% are not yet understood. Those are his estimates, not a standardized score: the essay says the shares change with choices such as how to count weak activations, how to treat a neuron’s depth and subcircuits, and what qualifies as one job.
He says the same broad pattern appears in models up to trillions of parameters, including Kimi K3, but points to a forthcoming technical companion paper for fuller detail on the language-model experiments. The essay does not present those large-model results as a completed, independently assessed study. Cammarata cautions against treating a single cleanliness percentage as a precise target, describing it instead as a rough guide to the model regime.
“Recursive generalization” replaces arbitrary mixtures
Cammarata’s alternative explanation is what he calls recursive generalization. A neuron can begin with a pattern and acquire related offshoots as the model learns—for example, a detector for the letter fragment “iy” extending to “diy” and then to “do it yourself.” The jobs may look unrelated if the investigator sees only the final examples, he argues, but their development reveals a chain of associations rather than arbitrary packing.
He also points to architectural weaknesses, limited analysis tools and small or simplified models as factors that made early neurons look less coherent. In his account, those conditions compounded: tools sometimes rewarded explanations that reproduced activations without capturing the mechanism, while researchers lacked a clear catalogue of the motifs neurons could learn. Better models and closer tests of individual neurons and circuits, he argues, can expose patterns that broad labels obscure.
A case for studying networks directly
The practical disagreement is not simply whether one interpretation method works better than another. Cammarata argues that researchers can investigate a model’s own neurons, circuits and information flow rather than first training a separate decomposition to reconstruct its features. He says that process has worked for him in modern convolutional networks and, so far, in language models, while acknowledging that the latter remain far from fully understood because of their size.
The essay’s claim is ambitious, but its status is deliberately provisional: Cammarata presents a working hypothesis and says a companion paper will provide more detail on the language-model experiments. For now, the specific evidence he puts forward includes a reinterpretation of the InceptionV1 cat-and-car neuron and preliminary Qwen3-32B estimates whose definitions depend on judgment calls about what counts as one job.