A slow read of Neural Networks Are Already Clean

Author

Mostly Claude and Brian Swingle

These are working notes on Nick Cammarata’s essay Neural Networks Are Already Clean (Goodfire, 2026). The audience is people with a physics background who are curious about interpretability but have not worked in it. The aim is to go through the essay slowly, one section at a time, filling in the background each section assumes and re-creating its claims in the smallest instructive setting that fits on a single consumer GPU.

The essay itself is long (about 27,000 words with its appendix) and is written as a working hypothesis rather than a paper. Its central claim is that the famous examples of “polysemantic” neurons, the ones that seemed to respond to unrelated things and that motivated the superposition hypothesis, were misread. On the author’s account the mess came from broken tools, weak architectures, incomplete explanations, and a growth pattern he calls recursive generalization, all of which are fixable or benign. True superposition, unrelated things folded into one neuron, should be the last hypothesis rather than the first.

How these notes are organised

Preliminaries fill in what a neuron, a feature, a residual stream, an attention head, a sparse autoencoder, and feature visualisation are, in the sense the essay uses them.
Background: Toy Models of Superposition reproduces the Anthropic toy model that made superposition precise, and the essay’s modification of it. This comes before the essay’s own “Superposition” section because the essay is largely a critique of that framing.
The essay, section by section then follows the essay’s own top-level headings. Each page gives a summary, the background needed, what the claim would look like if false, and a small reproduction where one is feasible. Appendix material is folded into the section it supports.

What is the essay’s and what is ours

Most of each page summarises or re-creates something the essay or the papers it cites actually say. Where we go further, running an experiment of our own or drawing a conclusion the essay does not draw, the material sits in a tinted box labelled Beyond the essay. Those boxes make no claim beyond what the experiment in them shows; they are there so the reader can tell what comes from the literature and what we added.

This is what the marker looks like. Anything inside a box like this is our own addition.

Ground rules for the reproductions

None of the essay’s models, checkpoints, or datasets are public, so nothing here is a replication in the strict sense. Each experiment is a re-creation at the smallest scale that keeps the phenomenon, run on one 12 GB GPU in at most a day. Where we had to substitute (a smaller dataset than ImageNet, a smaller language model than Qwen3-32B) the page says so.

Source

The plain-text version of the essay is kept in source/essay.md for reference. Quotations are from that file.

--- title: "A slow read of *Neural Networks Are Already Clean*" --- These are working notes on Nick Cammarata's essay [*Neural Networks Are Already Clean*](https://already-clean.goodfire.com/) (Goodfire, 2026). The audience is people with a physics background who are curious about interpretability but have not worked in it. The aim is to go through the essay slowly, one section at a time, filling in the background each section assumes and re-creating its claims in the smallest instructive setting that fits on a single consumer GPU. The essay itself is long (about 27,000 words with its appendix) and is written as a *working hypothesis* rather than a paper. Its central claim is that the famous examples of "polysemantic" neurons, the ones that seemed to respond to unrelated things and that motivated the superposition hypothesis, were misread. On the author's account the mess came from broken tools, weak architectures, incomplete explanations, and a growth pattern he calls *recursive generalization*, all of which are fixable or benign. True superposition, unrelated things folded into one neuron, should be the last hypothesis rather than the first. ## How these notes are organised - **Preliminaries** fill in what a neuron, a feature, a residual stream, an attention head, a sparse autoencoder, and feature visualisation are, in the sense the essay uses them. - **Background: Toy Models of Superposition** reproduces the Anthropic toy model that made superposition precise, and the essay's modification of it. This comes *before* the essay's own "Superposition" section because the essay is largely a critique of that framing. - **The essay, section by section** then follows the essay's own top-level headings. Each page gives a summary, the background needed, what the claim would look like if false, and a small reproduction where one is feasible. Appendix material is folded into the section it supports. ## What is the essay's and what is ours Most of each page summarises or re-creates something the essay or the papers it cites actually say. Where we go further, running an experiment of our own or drawing a conclusion the essay does not draw, the material sits in a tinted box labelled **Beyond the essay**. Those boxes make no claim beyond what the experiment in them shows; they are there so the reader can tell what comes from the literature and what we added. ::: {.beyond-inline} This is what the marker looks like. Anything inside a box like this is our own addition. ::: ## Ground rules for the reproductions None of the essay's models, checkpoints, or datasets are public, so nothing here is a replication in the strict sense. Each experiment is a re-creation at the smallest scale that keeps the phenomenon, run on one 12 GB GPU in at most a day. Where we had to substitute (a smaller dataset than ImageNet, a smaller language model than Qwen3-32B) the page says so. ## Source The plain-text version of the essay is kept in `source/essay.md` for reference. Quotations are from that file.