Goodfire researcher claims 90% of Qwen-32B neurons have one job, challenging polysemanticity

Nick Cammarata of Goodfire claims that agent-driven analysis resolves many known cases of polysemanticity, and that roughly 90% of neurons in Qwen-32B have a single function. The claim comes from his essay "Neural Networks Are Already Clean," published October 15, 2026. It has not been independently replicated.

What was claimed

Cammarata's stated working theory is that neural networks are not the tangled "knots" that polysemanticity and superposition imply. Many documented cases of those phenomena, which are cited as the reason individual neurons resist human interpretation, reportedly "fall to further agentic analysis."

Investor and commentator Deedy Das summarized the essay on X:

• Neural networks are not tangled knots, and polysemanticity is wrong.
• Neurons can be studied directly, without first "unfolding" them with a second model.
• 90% of neurons in Qwen-32B "have one job."

The 90% figure comes from that third-party summary of Cammarata's work. The figure, and the definition of "one job," should be checked against the primary essay.

Who is making the claim

Cammarata says he co-authored early mechanistic interpretability work at OpenAI. He is now at Goodfire, which describes itself as a mechanistic interpretability frontier lab. Goodfire's Myra Deng and Mark Bissell discussed the company's approach on the Latent Space podcast in February 2026.

Reaction

The thread drew 319K views and 5.76K likes in the post quoted by OpenAI researcher roon, and 135K views and 1.5K likes in the post quoted by Das. Roon wrote that the result "may be remembered as a major moment in scientific history" and that it could let researchers "make alignment into a 'normal science.'" Das called it "one of the most important results in AI, period." These are opinions, not findings.

What is not yet known

The source material does not include:

• The agent setup, including which models drove the analysis and what tools they used.
• The criteria for deciding that a neuron has a single function.
• Whether the result held across other models or only Qwen-32B.
• Any peer review or outside replication.
• Comparisons against existing methods that decompose activations with a second model, such as sparse autoencoders.

What this means

If the result holds, it changes the economics of interpretability. Dictionary-learning approaches add a second trained model and its own error to every analysis. Reading neurons directly, with agents doing the labor, would remove that layer and scale with agent capability, which is consistent with the "bitter lesson" framing that Anthropic's Emmanuel Amiesen discussed in 2025.

The claim is strong, though, and several questions decide whether it survives scrutiny. A single-function label that an agent assigns can be too coarse to count against superposition. A 90% figure depends heavily on how much behavior a "job" is allowed to cover. The result also needs testing on larger and differently trained models.

The near-term test is whether outside groups reproduce the Qwen-32B analysis and whether Goodfire releases enough methodology and data to allow it. Until then, this is a significant claim from a credible team, not an established finding.