Cat-and-Car Neuron Revisited: Nine Years Later, Its Co-Discoverer Says Networks Are Already Clean
In 2017, researchers found a famous oddity inside the image-recognition network InceptionV1: a single neuron that fired on cat faces — and also on the front views of cars. Cats and cars, two unrelated categories, packed into one neuron.
The case became the creation myth of the interpretability field, giving rise to a theory: concepts in neural networks are stored in bundled form. With too few neurons, the network folds unrelated concepts together, in near-arbitrary ways. Like a tangled ball of knotted thread, studying the network first required training a second model to untangle the knots. Sparse autoencoders (SAEs), all the rage in recent years, do exactly this untangling work — Anthropic used them to claim that roughly 70% of a model's behavior had been interpreted.
This week, Nick Cammarata of Goodfire published a long-form piece titled *Neural Networks Are Already Clean*. His position is unusual: he was one of the authors of the 2017 paper *Zoom In*, which introduced the concept of "polysemanticity". A co-author of the creation story returning nine years later to re-examine the original evidence himself.
The Cat-Car Neuron's Truth: Pointy vs. Round
He went back and looked at that cat-car neuron again — this time not in isolation, but through all of its upstream and downstream connections. Once the connection graph was laid out, a pattern emerged on its own.
The neuron's positive inputs: pointy-eared fox faces, wolf faces, pointy-nosed dogs, van fronts. Its negative inputs — the ones pushing it in the opposite direction: round things — snails, golf balls, tennis balls, round clock faces. Its three strongest downstream outputs: flames, the V-shaped water jets from fireboats, and fox faces.
What do flames have to do with cats? With cars? Nothing. But they have something to do with "pointiness". Flames are pointy, water jets are pointy, fox ears are pointy, and a car's front grille is pointy. This neuron was never "cat plus car" — it's a "pointy" detector. Sharp things excite it; round things quiet it. Cat faces and car fronts just happen to be two things with pointy features.
A pointed rhetorical line from the original: if this were truly random bundling of cats and cars, why would it need such a strong connection to flames?
The author names this pattern *recursive generalization*. The neuron isn't shoving unrelated things together; the network, during learning, extends the rule "cat → pointy-eared animal" upward into abstraction: "things that are pointy". The source of the confusion isn't the network — it's the observer. We stared at the neuron itself instead of its connections.
Ninety Percent of Neurons Have Only One Job
The second line of evidence comes from language models. Goodfire ran a preliminary health check on Qwen3-32B: roughly 90% of neurons (by polarity) have a single primary role, usually with a few related side duties; about 7% have two related roles; only 2–4% remain genuinely unexplained.
This number gains weight in historical context. In 2023, OpenAI's automatic interpretability research (with Cammarata himself involved), using the most generous yardstick of the time, found only about 1.7% of neurons explainable. Same world, from 1.7% to 90%. The difference isn't just better tooling — the author's judgment is that a large part of why these neurons couldn't be explained back then is that we didn't know how to look.
The piece also takes a jab at SAEs, with a carefully designed experiment: they compressed one layer of InceptionV1 from 480 neurons down to 10, artificially creating the "too few neurons" scenario, then trained an SAE to untangle it. If the superposition theory held, the SAE should have recovered clean curve features. Instead, what the SAE found was a blurry mess. The author's interpretation: polysemanticity in early models stems mainly from architectural flaws (overflow from later layers), not some deep learning principle.
An IOU, Not a Check
At the end, the author prices his own optimism: "highly speculative". The team has studied thousands of neurons across models, including the trillion-parameter Kimi K3 — along the way releasing (to their knowledge) the first trillion-parameter attribution graph — but many experiments remain preliminary, and a formal companion paper is on the way.
The community reaction arrived within a day. The Hacker News thread scored 800+ points, but a measured hot comment resonated widely: this tone feels a lot like the excitement around Anthropic's first large-scale SAE results. Google DeepMind's Neel Nanda offered a fairer assessment: for the practical goal of "explaining why the model did a specific thing", this line of work has real evidence and may genuinely be useful; but claiming it's the great road to fully reverse-engineering neural networks — he disagrees.
That disagreement is worth holding onto. If "networks are already clean" holds, the impact is structural: interpretability research shifts from "untangle first, then study" to ordinary science done directly, and the investment logic behind the SAE untangling route is undermined. If it doesn't hold, we still gain a batch of high-quality neuron dissections. Either way, the cat-car story that has been misleading us for nine years has now been re-measured by its own discoverer. What science has never lacked less than new evidence is people willing to go back and re-examine the old evidence.
---
*Sources: Goodfire, "Neural Networks Are Already Clean" (Nick Cammarata, Oct 2026, 65-minute long-form, already-clean.goodfire.com); re-reading of the InceptionV1 cat-car neuron 4e:55 (pointy/round structure); preliminary Qwen3-32B audit (~90% single-role / 7% dual-role / 2–4% unexplained); OpenAI 2023 automatic interpretability (1.7%); Hacker News discussion (811 points, including Neel Nanda's comment). Working hypothesis, author-labeled "highly speculative".*