Researcher Says 90% of Neurons in an AI Model Are Understandable

On October 15, Nick Cammarata published an essay called "Neural Networks Are Already Clean." Cammarata is a researcher at the Goodfire interpretability lab and a former OpenAI employee. The essay revisits a decade-old example from InceptionV1, an early image-recognition model. One neuron in that model was long held up as proof that neural networks mix up unrelated concepts. It fired on both cat faces and the fronts of cars. Cammarata argues the neuron is actually detecting a V-shape, a pattern that shows up in both cat faces and car fronts. He ran a preliminary scan of Qwen3-32B, a large language model. About 90% of its neurons do one clear job. Another 7% do two related jobs. Only 2-4% remain unexplained.

If this holds up, the question becomes a technical one, not a philosophical one. Right now, explaining why an AI agent made a given decision mostly comes down to guesswork from logs. Readability at the neuron level could eventually let engineers open up a model and see exactly what it reacted to. For now, the claim is unconfirmed. Cammarata himself calls the essay a working hypothesis, not peer-reviewed, and there is no technical paper yet. Stanford researcher Chris Potts says the conclusions are overstated. He argues the field understands only a small part of how language models actually work. For people building products on AI agents, a tool like this would matter more than one more model with a better benchmark score. It would explain why the model acted the way it did. That tool does not exist yet.