Cammarata essay argues neural networks are already organized
The former OpenAI researcher, now at Goodfire, estimates about 90% of analyzed neurons in Qwen3-32B perform a single dominant function.
Nick Cammarata, a former OpenAI researcher now at the interpretability company Goodfire, published an essay titled "Neural Networks Are Already Clean" on October 15, arguing that neural networks are already organized and that earlier findings of mixed, unrelated concepts inside single neurons were misreadings of the data, according to aitimes.com.
Cammarata revisited the well-known cat-face and car-front neuron example. Rather than a blend of unrelated concepts, he argued the neuron responds to a shared pointy V shape. Based on preliminary positive and negative value analysis, he estimated that roughly 90% of the neurons he examined in the Qwen3-32B language model perform a single dominant function.
In the experiments, one AI agent wrote a description of a neuron and a second agent used only that description to generate sentences. Some of the generated sentences produced stronger activations than examples drawn from large datasets.
Cammarata described the essay as a working hypothesis and said a technical paper with the details of the language model experiments will be published separately.
Praise and pushback
roon (@tszzl) on X called the work possibly a major moment in scientific history that could turn alignment into a normal science. Neel Nanda, writing on LessWrong, said the work is interesting but not a big-picture advance for interpretability, noting that circuit diagrams may leave out information and that weakly active neurons are not sufficiently explained.
Cammarata co-authored early interpretability papers including "Zoom In" (2020) and "Multimodal Neurons" (2021) with Chris Olah and others, and began automated interpretability research at OpenAI in 2023, using AI to explain other AI neurons.
Quick answers
What did Nick Cammarata argue in his essay?
He argued that neural networks are already organized and that reports of mixed, unrelated concepts in single neurons were misinterpretations. He estimated about 90% of analyzed neurons in Qwen3-32B perform a single dominant function.
Is the essay a peer-reviewed paper?
No. Cammarata called it a working hypothesis and said a technical paper with the details of the language model experiments will be published separately.
How did other researchers react?
roon (@tszzl) on X praised it as possibly a major moment in scientific history, while Neel Nanda said on LessWrong that it is interesting but not a big-picture advance for interpretability.