Goodfire says 90% of Qwen3-32B’s neuron poles have one main job, and that neural networks are already clean. The claim comes from “Neural Networks Are Already Clean”, an essay by Goodfire researcher Nick Cammarata published 15 October 2026. And it attacks a belief the interpretability field has organized itself around for years: that model internals are a tangled knot you must untangle before you can read anything.
The short version, in Goodfire’s own words: “This means we can study networks directly, instead of first untangling them.” In a preliminary audit of Qwen3-32B, Goodfire estimates roughly 90% of neuron poles have one main job, about 7% have two related jobs.
And only 2-4% are not yet understood. Goodfire itself frames the essay as a working hypothesis with a companion technical paper still to come.
If those numbers survive that paper, the basic playbook for opening the black box changes.
What “Clean” Actually Means Here
Clean means you can read the network’s parts without a detangling step. A clean network has components with jobs you can name and trace, nothing reconstructed in between. In Goodfire’s strongest result, ordinary architectural improvements with no change to model size or training data got the tested model to a state where “99-100% of neurons in that last layer are easily interpretable and the model is broadly legible the whole way through.”
The field’s prior ran the opposite direction. Much interpretability work assumes neurons are tangled and must be unpacked with sparse autoencoders before anyone can read them.
The essay says that assumption came from misreading early small models: “The cause of polysemanticity in early model organisms wasn’t superposition. But several ordinary, mostly fixable issues.”
That is the reversal in one move.
The mess was real, the diagnosis was wrong. And the wrong diagnosis produced years of detangling tooling aimed at a problem that better training and better tools mostly dissolve.
The Evidence: One Audit and a 300x Sweep
Start with the audit everyone quotes. In preliminary audits of Qwen3-32B, Goodfire estimates “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.” Those are estimates, explicitly labeled preliminary, on one named model. They are also the most concrete interpretability numbers anyone has published at this scale.
The essay did not stop at one audit.
Goodfire retrained InceptionV1 and ConvNeXt models across a roughly 300-fold size range to watch neurons form in the first place.
And in the sharpest test, Cammarata squeezed an InceptionV1 layer from 480 neurons down to 10. A sparse autoencoder run on that compressed layer recovered features that only partly matched the original curve detectors, which he reads as evidence against a hidden high-dimensional code. That experiment matters because it tests the alternative hypothesis directly instead of arguing around it.
The essay too re-litigates the most famous “mixed” neuron in the literature: InceptionV1’s cat-and-car neuron, the canonical proof that single neurons hold unrelated concepts.
Cammarata argues it was misread. And now reads it as a V-shape detector rather than a neuron combining unrelated concepts. On his account, most apparent mixing comes from neurons growing related offshoots during training, plus weak tools and limited data.
Then the scale claim. “We can now understand most neurons in models up to trillions of parameters,” the essay states. And Goodfire reports finding “similar results in models up to trillions of parameters, like Kimi K3, which the companion paper will cover in more detail.” That is the difference between a claim about a 32B model and a claim about the frontier.
Read the Fine Print Before You Rebuild Anything
Every number above is Goodfire describing Goodfire’s audit. The essay calls itself a working hypothesis, the companion technical paper is still to come. And the Qwen3-32B figures are preliminary estimates rather than validated results.
The post pulled 438 points on Hacker News, which tells you people are arguing about it, not that anyone has replicated it.
My contrarian take cuts both ways.
Against the field: a large chunk of sparse-autoencoder tooling got built to untangle a knot that, on this reading, was mostly small training data and weak analysis tools all along. That should sting, and not only in this subfield. I have shipped fixes for problems I never confirmed existed. And on this essay’s reading, the interpretability field did the same at industrial scale. Verify the diagnosis before building the cure is a rule that outlives this particular debate.
Against the essay: an interpretability lab announcing that interpretability is easier than everyone thought is a vendor announcing its own market is healthy. Commentary tracking the reaction notes the argument is that researchers may be able to study native neurons and circuits directly rather than reconstructing them with existing tools, which is a convenient conclusion for a company in that business. I still think the essay deserves your evening, as the experiments are specific and falsifiable. The companion paper either confirms the audit numbers or it does not. And independent replication on a model Goodfire did not pick is the thing that settles it.
What This Changes If You Build With LLMs
When a model in one of my pipelines misbehaves, I debug blind. I tweak the prompt, rerun, compare outputs, and if the behavior improves I usually cannot tell you why. Everyone shipping LLM features works this way, and we have all accepted it as the tax on the technology. Interpretability that holds up at 32B and beyond replaces that loop with inspection: find the component responsible, see what actually triggers it, then fix or steer it. That is the difference between prompt archaeology and engineering.
For a small operator, three concrete moves:
– Read the essay. It is free, and the core argument holds together even where you skip the technical sections.
– Start a log of your weirdest model behaviors now. If inspection tooling gets good, that log becomes your first test suite, since the unexplained cases are exactly where you point the microscope.
– Change nothing else yet. The numbers are preliminary and self-reported. The trigger to act is the companion paper landing and outside labs confirming the audit on models Goodfire did not choose.
Here is the part worth sitting with.
For years, “it’s a black box” has been the universal excuse for shipping model behavior nobody can explain. And this essay argues the box was never locked. Go read it, then look at the last genuinely strange thing a model did in your stack and write down what you think caused it. That guess is exactly the thing Goodfire wants to make checkable.
