Summary: Goodfire researcher Nick Cammarata published a long-form argument on October 16, 2026, that the dominant “superposition” theory of tangled neurons is likely wrong. In preliminary audits of Qwen3-32B, his team estimates about 90% of neuron directions handle one main job. If his working hypothesis holds, interpretability research could shift from untangling a knotted model to just reading it directly.
The theory under attack
For years, interpretability researchers have wrestled with “polysemantic neurons” — single units that seem to fire for unrelated things, like the famous neuron reacting to both cat faces and car fronts. The standard explanation is superposition: when a network runs out of neurons, it folds unrelated concepts into the same unit because they don’t interfere. Under that view, neurons reveal almost nothing, and researchers must first “untangle” the network, typically by training a second model like a sparse autoencoder to approximate a cleaned-up version.
The evidence once backed this pessimism. Cammarata cites Anthropic’s 2023 report “Towards Monosemanticity,” where the median neuron in a small single-layer model scored 0 out of 14 on interpretability. His own earlier automated interpretability project found only about 1.7% of neurons readable even under loose criteria. He admits he believed in superposition until recently.
Rerunning the classic cases
Cammarata’s core claim is that at least for InceptionV1, the mess came from architecture and weak tooling, not superposition. His numbers: early tools could interpret roughly 5% of the final layer; better tools, about 50%; retraining the same architecture with modern regularization, about 70%; and a parameter-matched modern ConvNeXt, about 99%. He also trained three ConvNeXts spanning a ~300x parameter range and found all readable — smaller networks covered coarser concepts (one neuron for many car types) rather than folding unrelated ideas together.
There’s a personal twist: in 2020, the “Zoom In” paper he co-authored with Chris Olah presented the car-in-dog-neurons circuit as the first example of superposition. He now calls that a misreading — ordinary generalization and composition. By his account, no one has published a second circuit matching superposition’s expected form since.
What “one neuron, many jobs” actually means
His replacement hypothesis is “recursive generalization”: neurons branch outward along related directions, like a detector for “iy” extending to “diy” and then to “do it yourself.” Generalization groups similar things, so you can still read what the model learned; superposition groups dissimilar things, so you can’t. His revised ordering treats superposition as a last-resort hypothesis — the most radical and hardest to test — behind poor tooling, wrong explanations, or training problems.
The language model evidence, and its caveats
On Qwen3-32B, splitting each neuron’s positive and negative directions, his team’s preliminary audit estimates ~90% of directions have one main task, ~7% have two related ones, and 2–4% remain unexplained. Redoing 500 neuron studies on GPT-2 XL, about 70% looked single-function but only ~5% purely so — a caveat Cammarata acknowledges, along with the fact that exact numbers depend on your threshold for “one thing.” He claims similar results in trillion-parameter models like Kimi K3, with details promised in a future paper. His team also published what he says is the first public attribution graph of a multi-trillion-parameter model using raw neurons only.
Why skip the autoencoder middleman? He argues raw neurons avoid reconstruction error, noting Anthropic’s largest transcoder predicts the same next token as the original model only half the time.
What builders should take from this
Treat this as a working hypothesis, not a settled result: the technical paper isn’t out, and the numbers lack third-party verification. Still, the direction matters for anyone thinking about model transparency and monitoring — Goodfire, where Cammarata works, launched an agent-monitoring tool built on Kimi K3 internals on October 8. If raw neurons are this readable, debugging and oversight tooling could get cheaper and more direct. That resonates with approaches I covered before, like Guide Labs’ Steerling-8B, which bakes interpretability in from the start rather than reconstructing it afterward.
Cammarata’s own framing echoes the “bitter lesson”: scaling general methods may beat clever disassembly tools. His boldest extension — that direct neuron study “may be most of what we need” — is one he flags as highly speculative. The practical next step is simply to wait for the technical paper and see whether independent researchers can reproduce the ~90% figure on other models.