A neuron that responds to both cat faces and car fronts suggests that AI organizes information in ways humans cannot explain. Goodfire researcher Nick Cammarata offers another interpretation: the neuron recognizes a shared V-shaped pattern. The categories look unrelated to us, but their visual features give the neuron a coherent job.
This example anchors Neural Networks Are Already Clean, published by Goodfire on October 15, 2026. Cammarata presents the essay as a working hypothesis: researchers may have underestimated how much they can learn from a model’s existing neurons and connections.
If the findings apply to advanced AI systems, laboratories may devote more funding to studying neurons directly. However, researchers would still need to establish how those neurons interact to produce a model’s decisions.
What the cat-and-car neuron reveals
Mechanistic interpretability examines a model’s internal computations to explain how it produces an answer. Researchers study artificial neurons, the numerical units that process information, and trace how their connections influence the model’s output.
This work becomes harder when a single neuron appears to respond to unrelated concepts, a phenomenon called polysemanticity. The cat-and-car neuron became a prominent example because its responses seemed to combine two distinct categories.
Cammarata’s revised interpretation identifies a shared geometric feature: a V-shaped pattern. The neuron’s connections to pointed faces and flames supported this explanation, while its negative response to rounded shapes provided further evidence that it distinguished shapes.
Although people classify cats and cars as different objects, a model may detect visual similarities between them. Describing a neuron only through object labels can obscure the pattern it recognizes.
Cammarata proposes that some apparently unrelated responses develop through recursive generalization. As a neuron learns, its function expands to related patterns, which can eventually differ from the original feature. Tracing this development could help researchers explain why seemingly distinct concepts activate the same neuron.
Why researchers suspected superposition
Researchers had experimental reasons to suspect superposition, which lets a network represent more features than its available dimensions by overlapping representations. When those features rarely activate at the same time, the network can accommodate some interference between them.
Anthropic demonstrated this behavior in simplified networks in its September 2022 paper, Toy Models of Superposition. The experiments showed infrequently activated features could share limited representational space.
Goodfire’s findings leave those demonstrations intact but question how often superposition explains confusing neuron responses in modern models. Some observations previously attributed to overlapping representations may instead reflect shared properties that researchers overlooked.
To investigate overlapping representations, researchers use sparse autoencoders: auxiliary models that reconstruct a network’s activity from learned features, with only a small subset active at once. These features can be easier to interpret than individual neurons, as Anthropic demonstrated in its 2023 dictionary-learning research.
Studying neurons allows researchers to examine the model’s existing computations and test how changes to specific units affect its outputs. Comparing both approaches on the same models could establish which produces more accurate predictions about model behavior.
What Goodfire has demonstrated so far
Cammarata reports that architectural changes made 99–100% of neurons in the final computational stage of a modified InceptionV1 vision model readily interpretable, while model size and training data stayed the same. The result suggests that some earlier interpretability problems arose from how the model learned, rather than from capacity limits alone.
In preliminary audits of Qwen3-32B, approximately 90% of neuron poles had one main function, while roughly 7% had two related functions. Around 2–4% remained unexplained. A neuron’s poles represent the positive and negative sides of its response.
These estimates depend on how researchers classify functions and account for weak activations. They describe the audited neuron responses, rather than measure understanding of the whole model, and cannot be assumed to apply to other systems.
Goodfire reports similar patterns in models with trillions of parameters, including Kimi K3, although the accompanying technical paper on its language-model experiments is still forthcoming.
As The Atlantic reports, much of the evidence presented concerns comparatively simple vision models. Whether the approach can explain sophisticated language models remains uncertain, and the essay presents a working hypothesis rather than a peer-reviewed demonstration of complete interpretability.
What this could change about AI safety
Understanding which computations produce harmful behavior could help researchers develop more targeted safety tests. If they identify deception mechanisms, they could examine when those mechanisms activate and test whether changing them affects the model’s actions.
Business Insider’s coverage explores the research’s implications for understanding and controlling AI. However, Goodfire’s neuron findings have not yet demonstrated reliable prevention of harmful behavior in deployed frontier models.
Researchers would need to show that a neuron helps cause a harmful decision and establish how its contribution varies across situations. Suppressing its activity could disrupt other capabilities, while the model might still produce the same harmful behavior through different computations.
Examining internal mechanisms could also reveal failure conditions that prompt-based evaluations have missed. Testing those explanations would require access to the model’s internal activity, a concern also discussed in The Neuron’s coverage of independent evaluations.
The evidence needed next
Goodfire’s findings give laboratories a reason to devote more resources to neuron analysis and compare it with sparse-autoencoder methods on the same models. Reproducible procedures and consistent evaluation criteria would help establish whether either approach produces explanations researchers can use to guide model development.
A neuron description should predict its activity on unfamiliar inputs, while a circuit explanation should predict what happens when researchers alter it. Independent teams would then need to reproduce the language-model findings across different architectures and model sizes.
Testing should also examine rare failures, since even a small proportion of unexplained neurons could contribute to consequential behavior. The percentage of neurons researchers can describe does not establish how reliably they can anticipate a model’s actions.
To demonstrate a safety improvement, researchers would need to predict harmful behavior before it occurs and show that a targeted intervention reduces it across unfamiliar situations. Goodfire’s hypothesis could help direct those experiments, but its value for controlling advanced AI will depend on measurable changes in model behavior.
