Nick Cammarata, founder of Goodfire, published the essay “Neural Networks Are Already Clean” on October 15, 2026, at roughly 65 minutes of reading. His thesis: canonical examples of polysemantics and superposition were largely misinterpreted, and what looks like “random folding” of a neuron is usually recursive generalization — as in the case of the “iy” detector that grew into a “diy” and “do it yourself” detector. A preliminary audit of Qwen3-32B found about 90% of neuron poles have one primary function. If these results survive independent verification, interpretability may become much cheaper.

image

What happened

On October 15, 2026, Nick Cammarata published the essay “Neural Networks Are Already Clean,” and on the same day Yair Halberstadt posted a linkpost on LessWrong, which is drawing a hot discussion — at least 13 comments, including a quote from Ron: “this may go down in history as a major moment.” The core of the work is the preliminary audit numbers. In Qwen3-32B, about 90% of neuron poles have one primary function, another ~7% have two related functions, and only 2–4% are not yet understood. Re-analysis of GPT-2 XL found about 70% “clean” neurons, versus less than 2% by early automatic estimates, and an SAE on the final layer gave 88% activity as near-copies of single neurons.

Context

For a long time, polysemantics was considered fundamental: one neuron performs several unrelated tasks because the network uses superposition to pack more features than it has neurons. From this canon grew mechanistic interpretability in its current form — SAEs and other decompositions became the main tool, and examples like InceptionV1 became textbook classics. Cammarata, co-author of early OpenAI mechanistic interpretability work, proposes revisiting the foundation: feature folding often looks random only to the observer, and the actual mechanism is recursive generalization, where a simple pattern during training grows into derivatives. The numbers on tools support the revision: explainability of the final layer of InceptionV1 rose from 0–5% with early utilities to about 50% with the best ones, and a modern ConvNeXt of the same size shows 99–100%. Hence the argument in the spirit of Rich Sutton’s Bitter Lesson: interpretability may turn out to be solvable by the time it is needed most.

Why this matters for the industry

For the industry, this is a challenge to the dominant methodology: if most “confusion” in networks is explained by generalization and weak early tools, rather than fundamental superposition, SAEs and other decompositions stop being the only path, and direct study of neurons and circuits — including in trillion-parameter models — becomes the main front. Labs like Goodfire, OpenAI, and Anthropic will need to revisit tool priorities. Possible economics: reading neurons directly is cheaper than SAE pipelines, so if the numbers are confirmed, a standardized “cleanliness audit” as a de facto benchmark, first open-source mass explainability tools, and eventually an “interpretability as a service” layer — from dashboards for debugging and degradation control to post-fine-tuning verification and built-in traceability in large LLM platforms — are likely. There are no deployable artifacts today: the sources do not claim tools, APIs, prices, or latency metrics, so there is no reason to immediately rebuild pipelines.

Why this matters for users

The essay gives readers a reason to update their views on how “dirty” neural networks are: the question “is polysemantics everything?” is backed by concrete numbers, not a meme. The primary source can be read independently, and the LessWrong discussion shows counterarguments in real time — the community is already working out how much of this is real progress and how much is overestimation of old results. Engineers with R&D resources have a reason to run a small experiment to test the thesis, without waiting for ready-made tools. Others just need to watch attempts to replicate the audit: Qwen3-32B is an open model, so reproducing the numbers is accessible to any group. It is important to remember that this is one author’s working hypothesis, not a community verdict.

What is still unknown / limitations

This is a preliminary audit by one author: there are no independent replications in the provided sources, and the method for counting “poles” and “one primary function” has not yet been tested by other groups. The comparison of ConvNeXt and InceptionV1 cannot be presented as proof that progress was masked for years by weak tooling: these are different architectures and different training eras — a classic confounder, and there is no controlled comparison of one model across multiple generations of tools in the sources. The numbers for Qwen3-32B and GPT-2 XL need confirmation on at least a couple of architectures — otherwise this is not a paradigm shift, but an invitation to verify. The idea of recursive generalization is falsifiable through tracing across training checkpoints, but such tracing has not yet been shown.

Sources

Author

Look at AI, editorial team