Goodfire: 90% of Qwen3-32B Neuron Poles Have One Main Job

Nick Cammarata of Goodfire, an AI interpretability research lab, argued in an essay published on 15 October 2026 that neural networks are far more readable than the field has assumed. In preliminary audits of Qwen3-32B, he writes, “we estimate that roughly 90% of neuron poles have one main job, often with small related offshoots; another ~7% have two related jobs; and only ~2-4% are not yet understood.”

The essay, titled “Neural Networks Are Already Clean,” sets that against a 2023 OpenAI project he says he started, which found about 1.7% of neurons interpretable by its most generous measure. He describes his own conclusion as a working hypothesis, and the fourth step of his argument as “highly speculative.”

Business Pill · TRANSFORMERS

A one-minute explainer of transformers and attention: how a language model reads every word at once. It teaches the general idea only and says nothing about any organisation in this story.

What the Essay Claims

Cammarata writes that, after many months studying neurons and circuits in image and language models, he has come to believe “a network’s own mechanisms are its legible building blocks.”

He argues that the canonical evidence of polysemanticity and superposition, the idea that networks pack unrelated concepts into the same neurons, “was misread.” What looked like arbitrary packing, he writes, was mostly a motif he calls recursive generalization.

“This means we can study networks directly, instead of first untangling them. Neural networks are already clean,” the essay says.

What Was Posted on X

Nick Cammarata’s own post announcing the essay on 15 October 2026 (15:40 UTC). He writes that he co-authored early mechanistic interpretability work at OpenAI and came away pessimistic, that he later started OpenAI’s first large-scale automated interpretability project, and that he now thinks ambitious interpretability will be solved by the time we need it most. These are his views; the figures in this piece come from the essay.

The Numbers in the Essay

For Qwen3-32B, the essay gives the 90%, ~7% and ~2-4% figures above and notes that the share of neurons with one job varies with how thresholds are set for counting weak activations, depth and what counts as “one thing.”

It says dense and mixture-of-experts models seem roughly similar, and that Goodfire found similar results in models up to trillions of parameters, including Kimi K3, which a companion paper will cover in more detail.

For the field’s earlier model organism, InceptionV1, the essay says architectural changes and no change to model size or training data left 99-100% of neurons in the last layer easily interpretable, where the original study found almost none.

It also cites Anthropic’s 2023 Towards Monosemanticity study of a small one-layer model, which it says gave the median neuron a 0 on its rubric out of a possible 14.

Who Is Making the Argument

Cammarata writes that in 2019 he and Chris Olah found the circuit that was originally introduced as the superposition phenomenon in the paper Zoom In, and that he helped propose both explanations for polysemanticity.

In his X post on 15 October 2026, he wrote that he “co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic,” and later started OpenAI’s first large-scale automated interpretability project.

The post had more than 1 million views when we read it. In it he wrote: “I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving.”

Why It Matters for Labs and Their Customers

Goodfire describes itself on its website as a research lab and offers Silico, which it calls “your interpretability agent,” to explain, debug and control model behaviour.

The essay contrasts two approaches: studying a network’s own neurons and circuits directly, or training a second model, such as a sparse autoencoder, to approximate the network unfolded and studying that instead. Cammarata argues the field fell back to the second because networks seemed to be tangled knots.

The essay says agents can now map out neurons and circuits, test explanations and track hundreds of thousands of attribution graphs, and that this process basically worked on modern convolutional networks.

On the same day, as we reported, Anthropic said it had embedded six METR evaluators inside the company to assess its models.

The Structural Read

As we read it, the headline number rests on definitions: the essay says the share of neurons with one job varies with the thresholds for weak activations, depth and what counts as one thing.

As we read it, the contrast with 2023 is about tools as much as models: the essay attributes the old pessimism partly to explainer models that couldn’t run the loop of conjecture and test, weak metrics and small or simplified models.

As we read it, the commercial stake is visible on Goodfire’s own site, which offers Silico, an interpretability agent to explain, debug and control model behaviour.

Three Implications

INTERPRETABILITY RESEARCH The essay argues for scaling ordinary investigation of a network’s own mechanisms, which it calls a possible bitter lesson for interpretability.

MODEL DEVELOPERS Cammarata writes on X that he now thinks ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving.

READERS OF THE NUMBERS The essay says a technical companion paper will give a more detailed account of the language-model neuron experiments.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework argues that “the transformer is a commodity now; what changes between paradigms is where intelligence accumulates and who can capture it.”

As we read it, Goodfire’s argument adds a question to that map: if a network’s own neurons are as readable as the essay claims, the ability to read them becomes something a lab or a tools company can build on, as Goodfire does with Silico. The framework predates the essay and does not discuss interpretability.

What Is Not Established

The essay describes the Qwen3-32B figures as preliminary audits and says the precise number depends on where thresholds are set. The companion paper with the language-model experiments is described as still to come.

The 1.7% and 90% figures come from different projects, models and measures, so they are not a like-for-like comparison. The essay itself calls the fourth step of its argument highly speculative.

The Bottom Line

Nick Cammarata of Goodfire argues that neural networks are “already clean”: in preliminary audits of Qwen3-32B, Goodfire estimates roughly 90% of neuron poles have one main job, against about 1.7% of neurons found interpretable in a 2023 OpenAI project.

He argues that interpretability may be more tractable than assumed, possibly in the near term, while calling that step highly speculative and leaving the detail to a companion paper.