The Core Claim
Nick Cammarata, a researcher at the AI startup Goodfire, published a 26,000-word essay titled "Neural Networks Are Already Clean" on October 15. In it, he argues that neural networks are far more interpretable than believed: the network's own mechanisms are readable building blocks, and the outlook for mechanistic interpretability is much better than he had feared.
This claim directly challenges the industry consensus around superposition and polysemanticity. Under that consensus, neurons store several concepts in a mixed, overlapping way — making it nearly impossible to isolate and read what each individual neuron "means." Cammarata pushes back: he says the field has misjudged how deep the problem runs, and that judgment is excessively pessimistic.
Cammarata bases his conclusion on preliminary audits conducted by the Goodfire team. In an audit of the Qwen3-32B model, roughly 90% of neuron clusters perform one main task and nearly 7% perform two related tasks; only 2–4% remain unexplained. The team found similar results in trillion-parameter models, including Kimi K3 — the patterns observed in small models hold in the largest systems as well.
The essay also reinterprets a classic example: the famous "machine and cat" neuron in InceptionV1 is not random superposition, he argues, but a detector of a general V-shape — a cat's face or the front of a car. With improved tools, 99–100% of the neurons in that model's final layer are now easily interpretable, the author writes.
The State of Interpretability Today
The goal of mechanistic interpretability is to read neural networks from the inside: to understand how each neuron or group of neurons contributes to a model's decisions. The dominant view holds that superposition gets in the way — a single neuron stores several unrelated concepts in a mixed form. Jacob Hilton of the Alignment Research Center estimates that researchers understand less than 0.01% of the internal mechanisms of the largest language models. Cammarata wants to redraw that pessimistic picture: the network's own mechanisms are readable building blocks, he says, and the situation is not as dire as feared.
Why This Debate Matters
MIT Technology Review covered the essay with an exclusive interview on October 19, presenting it as a claim that turned the AI world upside down. According to the publication, one researcher called it "a big moment in the history of science" and another "the most important AI article of the year." The praise itself shows how heated the debate has become.
The question at the center of the dispute is precise: are neurons naturally "clean," or is that cleanliness merely a fluke of small, old models? Cammarata defends the first answer; the critics defend the second. Both sides urge caution in drawing conclusions — one because the stakes of the claim are high, the other because the evidence has not yet been published openly.
The Critics Respond
Stanford professor Chris Potts judged the material Cammarata presented as "nothing more than conjecture" and said that "this blog post doesn't make the problems any easier." In his view, the essay offers conclusions drawn from personal observations rather than technical evidence.
Alignment Research Center researcher Jacob Hilton notes that scientists understand less than 0.01% of the internal mechanisms of the largest language models. In his view, the scale of the problem is so vast that a single essay cannot fundamentally change it.
Neil Nanda of Google DeepMind found the work interesting but does not consider it a serious new development for the interpretability field as a whole. What all three critiques share is a gap between the scale of the claim and the open technical evidence supporting it. The views of all three experts are collected in MIT Technology Review's exclusive piece from October 19.
The Author and Next Steps
Nick Cammarata was previously one of the pioneers who developed mechanistic interpretability tooling at OpenAI. He says that for much of last year he spent about ten hours a day examining neurons — the essay's conclusions are the product of that daily work.
The Goodfire team plans to publish technical results on modern large language models soon. Those results will test the strength of the claim: the critics are waiting precisely for that open data to arrive. The full text of the 26,000-word essay is published openly on Goodfire's official website.
"I think I have good news about the nature of neural networks." — Nick Cammarata, "Neural Networks Are Already Clean" (Goodfire)
