Nick Cammarata's AI interpretability essay challenging superposition theory

For six years, the cat-and-car neuron was the field's go-to proof that neural networks cram unrelated concepts into the same unit because they run out of room - the textbook case of superposition. Cammarata's essay says that reading was simply wrong: the neuron responds to a sharp, outward-flaring V-shaped geometric structure, not two unrelated ideas sharing a slot ^[2]. He goes further, calling his own original superposition finding an error that, to his knowledge, no one has replicated in a real network since ^[2].

In place of superposition, Cammarata proposes 'recursive generalization': neurons accumulate related meaning over training the way sediment builds up in layers, rather than being forced to fold unrelated things together ^[1][8]. That reframes what a neuron even is. Cammarata describes moving away from treating neurons as human-shaped nouns - 'one is a car, another is a curve' - toward thinking of them as control flow, more like routing logic than concept labels ^[8]. The original 2020 Distill paper that coined polysemanticity and superposition, co-authored by Cammarata, is the direct target of this reversal ^[9].

The headline statistic is a jump from roughly 1.7-2% of neurons explainable in OpenAI's 2023 automated-interpretability project to roughly 90% of neuron poles having a single clear job in a preliminary audit of the open model Qwen3-32B, with about 7% serving two closely related functions and only 2-4% left unexplained ^[3][7]. Cammarata attributes most of that jump to better agentic and automated tooling for generating and testing activation hypotheses, not to changes in the underlying models ^[5][6]. He also points to retrained, modernized versions of the older InceptionV1 vision model reaching 99-100% interpretability in their final layer across a 300x range of model sizes ^[4][6], framing the original polysemanticity observed in that model as stemming from fixable architectural and tooling issues rather than a fundamental shortage of neurons ^[4][8].

The caveats matter as much as the headline number. Cammarata explicitly labels his broadest conclusions 'highly speculative' ^[2][8]; the essay is not peer-reviewed, a companion technical paper has not yet been released, and the percentages are described as threshold-dependent rather than fixed ^[7]. Results on trillion-parameter models are, by his own account, not close to fully solved given the sheer scale involved ^[7].

The strongest counterweight to the essay's reception comes from inside the field. Neel Nanda, a mechanistic-interpretability researcher at Google DeepMind, argued publicly that superposition is still necessary to explain how models store far more facts than they have neurons to hold them, and that the essay is interesting work but not a significant update to the field's big picture. Other commentators took a middle position: Liv Gorton welcomed the more-interpretable-than-expected finding but argued that neuron legibility and superposition are not necessarily mutually exclusive, since the two framings may coexist.

Dean W. Ball raised a more methodological concern, noting that Cammarata's audits lean on subjective human and agentic judgment calls about what counts as 'one main job,' and that it remains an open question whether single-neuron legibility generalizes to understanding millions of neurons interacting across trillion-parameter models ^[8]. That distinction, between a neuron having a legible function and a concept or behavior being fully auditable, is the crux of the disagreement.

The public reaction has run well ahead of the evidence. Roon, an OpenAI researcher, called the essay a result that might be remembered as a major moment in scientific history, framing it as good enough news to lower AI doom estimates across the board ^[5]. That kind of reception is amplified by the fact that Goodfire, Cammarata's employer, is an interpretability startup that commercially benefits from a narrative in which interpretability is tractable and nearly solved ^[8].

Underneath the optimism sits an unresolved tension that cuts the other way: if refusal and safety-guardrail behavior really does live in a small number of legible, isolable neurons, that same legibility could make those circuits easier to identify and strip out, not just easier to audit. That double-edged read sits alongside, rather than inside, the headline dispute over whether superposition has actually been dethroned.