OpenAI's Interpretability Founder Says 90% of Neurons Have a Single Function

Nick Cammarata, founder of the interpretability team at OpenAI, published a very long essay on Friday with a thesis that upends the familiar picture: about 90% of neurons in large models perform a single function, and a trillion-parameter model becomes "nearly readable" with modern agentic tools. The Chinese outlet Jiqizhixin called the text one of the most important results in AI. It has already sparked heated debate in the professional community.

What exactly Cammarata claims

In his words, the idea of a neural network as an impenetrable "black box" has turned out to be largely a misconception. Most of the obstacles researchers expected do not arise in practice. His conclusion: interpretability will be solved, and possibly quickly.

The limitation is worth stating up front. The available excerpt does not describe the methodology, the set of models, or the criteria by which a neuron counts as "performing a single function." The 90% figure comes from the headline of the piece. It cannot be verified without the full text of the essay and reproducible experiments.

Why the claim sounds bold

For the past several years, researchers have assumed the opposite. Neurons in large models are often polysemantic: the same neuron responds to several unrelated concepts. This picture is tied to the superposition hypothesis, which holds that a network packs in more features than it has neurons. To untangle that packing, labs have adopted sparse autoencoders and related methods. It is an expensive workaround, and it makes sense only if "clean" neurons are rare.

Cammarata himself took part in OpenAI's 2023 work in which GPT-4 wrote text explanations of GPT-2's neurons. Most of those explanations received low scores, and the method seemed to be hitting a ceiling. He now says the main limitation was not the subject of study itself but the tools.

By the essay's logic, agents change the economics of the problem. A single researcher cannot manually work through billions of neurons. An agent can formulate a hypothesis for each one, run an experiment, test it on new examples, and rewrite the description. For a trillion-parameter model, that turns unreachable manual labor into a computational task.

Key points

• Author: Nick Cammarata, founder of OpenAI's interpretability team.
• Format: a very long essay published on Friday.
• Headline figure: about 90% of neurons perform a single function.
• Scale: models of up to a trillion parameters.
• Tool: modern agentic systems instead of manual analysis.
• Forecast: interpretability will be solved, "possibly soon." The available excerpt gives no specific timeline.
• Status: the author's position, not a peer-reviewed paper.

How engineers and safety teams should read this

Skepticism here is a professional norm, not distrust of the author. First, you need to find out how "a single function" is defined: by behavior on examples, or by causal interventions, where a neuron is switched off to see what breaks. The second test is far stricter. Next, it matters whether the result reproduces on open models, where independent groups can run the same agentic pipelines. And finally, 90% of neurons does not mean 90% of model behavior. The remaining ten percent may carry the most dangerous and most complex capabilities.

If the thesis survives scrutiny, the consequences are practical. Auditing models for deception, hidden goals, and vulnerabilities would stop being a research experiment. It would become an engineering procedure that can be demanded from a model provider. Teams building products on LLMs should watch for open replications and for whether ready-made interpretability pipelines appear in agentic tools. For now, the essay is best read as a strong hypothesis from someone who has worked on this topic longer than most.