AI Neurons May Be Easier to Read Than Thought
An unreviewed essay claims most neurons in large models have one main job, a hypothesis that remains untested.
Updated on Nov. 6, 2026 in Artificial Intelligence
Live Poll
Do you think AI systems are becoming understandable enough to trust?
Nick Cammarata, a researcher at AI start-up Goodfire and a former OpenAI staffer, has published an informal research essay arguing that neural networks are far easier to interpret than researchers assumed. The work is a working hypothesis that has not been peer reviewed, and South African academics have cautioned that its findings are unverified.
Why it matters
If most neurons in large models carry one main job, researchers could inspect a model directly rather than first training a second model to untangle its internals. That would matter for the interpretability field, which has long relied on the idea that neurons are tangled and hard to read.
The essay's central claim rests on a 2014 Google image model, Inception v1, where neuron 4e:55 responds to both cats and cars, which Cammarata attributes to a shared V-shaped feature. Preliminary audits of Qwen3-32B report about 90% of neuron poles with one main job, against roughly 1.7% of neurons judged interpretable in an earlier OpenAI project, but the Qwen figures are unreviewed and the two measures are not directly comparable.
The players
Nick Cammarata
A researcher at AI start-up Goodfire and former OpenAI staffer who co-wrote the 2020 Zoom in interpretability paper.
Goodfire
An AI start-up focused on interpreting how neural networks work internally.
The details
Neural networks are built from many artificial neurons whose activity is shaped by training, and interpretability researchers try to explain what each one responds to. Researchers have long pointed to polysemanticity, where a single neuron responds to unrelated concepts, and to superposition, where a model packs more concepts than it has neurons by sharing capacity among them. Cammarata argues that the canonical evidence for these ideas was misread: the cat-and-car neuron, he says, is held together by a shared V shape rather than being a tangle of unrelated meanings.
Timeline
2014: Google's Inception v1 image-recognition model is released.
2020: The Zoom in paper, co-written by Cammarata with work led by Chris Olah, is published.
2023: Cammarata says he began an interpretability project at OpenAI.
30 October 2026: The essay is revised to clarify the V-shape explanation.
6 November 2026: An article about the essay is published.
The Tech Race
The essay reopens a line of interpretability research that Chris Olah's group helped define through the 2020 Zoom in paper. If Cammarata's reading holds, it would push the field away from disentangling neurons toward inspecting them directly, though that remains an untested hypothesis.
The findings are a research hypothesis and cannot yet be used to change how any deployed model is built or audited. Readers should expect no practical change until the claims are tested and peer reviewed.
The takeaway
Watch for peer review of Cammarata's essay and any independent audit of the Qwen3-32B neuron figures. Confirmation on a second model family would make the one-job-per-neuron claim far more credible.
Further reading
For the wider context of how researchers try to read model internals, see Artificial Intelligence.
Source note: This article includes information reported by TechCentral.
Live Poll
Do you think AI systems are becoming understandable enough to trust?
