Nick Cammarata, a researcher at interpretability startup Goodfire and formerly with OpenAI, argues that most neurons in large models can now be understood, and that our grasp of neural circuits has improved significantly.
Cammarata laid out the case in an essay published October 15, 2026, titled “Neural Networks Are Already Clean.” His central claim is that modern large networks have a clearer internal structure than earlier research suggested, with the majority of neurons performing distinct roles rather than being “polysemantic” — firing for a grab bag of unrelated concepts.
Using advanced agentic methods, recent analyses found that nearly 90% of neurons in models like Qwen3-32B serve a single primary function. Another 7% cover two related roles. Only 2–4% remain unexplained, according to the preliminary findings.
From 2% to 90% in three years
Cammarata is, in part, revising his own earlier conclusions. In 2023, he worked on a project at OpenAI that found only about 1.7–2% of neurons were legible with the tools available at the time.
He also revisits older examples of so-called “tangled” neurons, including cases noted in InceptionV1, an image recognition model studied in earlier interpretability work. His argument suggests that some of those apparent tangles came from inadequate analysis methods rather than from the networks themselves.
More broadly, he challenges the superposition and polysemanticity narratives that have shaped the field, arguing that many observations taken as evidence for those ideas stem from architectural flaws instead.
Cammarata co-authored papers on mechanistic interpretability in 2020 and 2023.
Not everyone is ready to declare victory
Reaction to the essay has been mixed. Some researchers see it as a potential breakthrough for AI alignment. Others caution against overstating the implications and point out that concepts like representational superposition remain central to the discussion.
Cammarata frames the shift as central to alignment, expressing optimism about what scaled automated methods are making possible.
What this means
The gains came from agentic, automated analysis rather than humans painstakingly inspecting neurons one by one. That suggests interpretability might scale alongside the models it studies.
What to watch next is replication. Independent groups testing these methods on other models, and checking whether the single-function labels survive deeper scrutiny, will determine whether “already clean” becomes the new consensus or a provocative essay that moved the debate forward without settling it.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.