Nick Cammarata's Superposition Challenge Rests on a 90% Neuron Claim

Nick Cammarata has challenged superposition theory with a striking estimate: roughly 90% of sampled Qwen3-32B neuron poles have one main job. His October 2026 essay argues that neural networks are far more organized than interpretability researchers assumed. Yet the number comes from a preliminary audit, and an early independent retest produced a lower result.

The dispute reaches beyond one percentage. Cammarata says researchers misread the examples that helped establish superposition as a central explanation for confusing neural behavior. He proposes recursive generalization instead, where related functions grow from a shared computational role rather than being arbitrarily packed together.

That claim directly challenges research strategies built around decomposing model activity into cleaner features. OpenAI and Anthropic have developed automated explanations, sparse autoencoders, and circuit-tracing methods for that task. Cammarata's alternative starts with the model's native neurons and asks whether the supposed mess came from limited tools.

The 90% Claim Changes the Starting Assumption

The important change is not that one interpretability score improved, but that Cammarata wants researchers to begin with native neurons again.

Cammarata calls his 65-minute research essay a working hypothesis rather than a peer-reviewed paper. It combines new experiments, interactive visualizations, historical reinterpretation, and a broader theory of how neural networks organize themselves.

His preliminary Qwen3-32B audit focuses on neuron poles. A pole is one directional side of a neuron, covering strongly positive or strongly negative activity. This distinction matters because the two directions can support different computational effects.

Goodfire's estimate assigns roughly 90% of examined poles one principal job. About 7% reportedly have two related jobs, while 2% to 4% remain unexplained. Cammarata explicitly says these percentages depend on thresholds, model depth, circuit density, and the definition of one job.

The analysis therefore does not establish that 90% of every model is fully understood. It does not show that researchers can predict all behavior from individual neurons. It also does not convert interpretability into a solved problem.

Instead, the figure supports a narrower proposition. When researchers examine strong activations with better tools, most sampled neuron poles appear to organize around a coherent function. Smaller related branches can exist without implying arbitrary combinations of unrelated concepts.

Cammarata reports examining 100 million tokens for one Qwen3-32B neuron. Only about 160 tokens reached the strong half of its activation range. He uses that distribution to argue that weak activity should not carry the same interpretive weight as a neuron's strongest responses.

That choice creates a methodological divide. A researcher focused on dominant causal behavior might disregard faint activations as low-impact residue. Another researcher might treat the same activity as evidence of hidden secondary functions.

The essay also presents 1,000 randomly selected neuron investigations across Qwen3-32B and Qwen3.6-35B-A3B. These reports inspect activation contexts, output effects, connections, and interventions. They are intended to move beyond labels based only on correlated text examples.

One displayed circuit connects several Qwen3-32B neurons to a diagnosis-related answer. Removing a small group reportedly changes the model's next diagnostic guess. Other examples connect neurons to answer hedging, user confidence, and culturally specific associations.

Those interventions are more informative than a label alone. A neuron that activates around a subject might merely correlate with another process. Altering it and measuring the output provides evidence that it participates in the computation.

Still, local causal evidence does not automatically validate the full 90% estimate. The gallery covers a tiny fraction of Qwen3-32B's internal units. Its explanations also depend partly on automated agents making judgments about semantic coherence.

Cammarata acknowledges another limit. A technical companion paper covering the language-model work has not yet supplied the full experimental account. The public essay offers considerable evidence, but it mixes documented findings with a larger conjecture.

The distinction should guide how readers interpret the headline. Nick Cammarata's superposition challenge is a serious research proposal backed by experiments. It remains a challenge to the prevailing picture, not a settled replacement for it.

Why the Famous Cat-and-Car Neuron Matters

Cammarata's case depends on showing that interpretability research misunderstood its best-known examples, not merely on finding cleaner neurons in newer models.

One important example is neuron 4e:55 in InceptionV1, an early image-classification network. Researchers observed responses to cat faces and the fronts of cars. The combination appeared arbitrary, making it a memorable example of polysemanticity.

Polysemanticity describes a unit that responds to several apparently unrelated concepts. Superposition offers one explanation: a network represents more features than it has available dimensions by overlapping their representations. Sparse features can share capacity when they seldom interfere.

Cammarata co-authored the 2020 circuits study that helped shape this interpretation. That history gives his reversal unusual weight. He is revisiting evidence that influenced his own earlier position.

His new reading begins with visual structure. Cat faces and car fronts often contain comparable V-shaped geometry. The neuron may therefore detect a useful configuration rather than storing two unrelated object categories.

This difference sounds small, but it changes what the neuron represents. A cat-and-car unit suggests arbitrary concept packing. A V-shape detector suggests ordinary visual generalization across examples with shared structure.

Cammarata applies similar scrutiny to another circuit from the earlier work. The circuit appeared to spread car information into several dog-related neurons. Researchers interpreted that pattern as information hidden across a high-dimensional combination.

The new essay argues that nothing downstream reads the supposed hidden code. Cammarata now treats the pattern as ordinary composition and generalization. He says he has not found another published real-network circuit with the expected superposition structure.

That is not a mathematical disproof of superposition. Absence of a second documented circuit might reflect limited observation. However, it weakens the rhetorical force of treating one early example as representative of modern networks.

Architecture also enters the argument. InceptionV1's final layer once appeared especially difficult to interpret. Cammarata attributes much of that confusion to architectural bottlenecks, overflow in later layers, and visualization methods that generated misleading images.

He retrained ConvNeXt image models with 92,000, 7.06 million, and 28.6 million parameters. That represents a size range of roughly 300 times. He says the models remained broadly legible across those scales.

The essay further reports that ordinary architectural improvements made 99% to 100% of the last-layer neurons readily interpretable in a revised setup. Model size and training data were held constant for the central comparison, according to Cammarata.

That result supports a different account of capacity pressure. With fewer neurons, models might learn broader categories, such as one neuron covering several kinds of cars. With more room, those broad categories can separate into narrower specialists.

This process does not require the network to hide unrelated concepts together. It resembles a taxonomy becoming more detailed as capacity expands. The categories remain connected through their computational history.

Cammarata calls that process recursive generalization. A neuron begins with a coherent role, then acquires branches connected through spelling, shape, context, or circuit function. An apparent mixture can therefore possess an internal lineage.

One language example begins with character fragments that connect to related endings and eventually to the name Miyazaki. A surface inspection could classify its activations as unrelated. Following the layers reveals a structured sequence of increasingly specific recognition.

The central reversal is therefore historical as well as technical. The examples that pushed researchers away from neurons might now support direct neuron study. Better observation changes the interpretation without changing the trained network.

Nick Cammarata's superposition challenge will stand or fall partly on these reanalyses. If other researchers reproduce them, the field must reconsider what its canonical examples actually demonstrated. If they fail, the original caution about neurons will remain justified.

Nick Cammarata's Superposition Challenge Targets Sparse Autoencoders

The primary contest is between studying a model's native mechanisms and decomposing its activations into a new feature dictionary.

Superposition research gained influence because individual neurons often seemed unreliable as units of explanation. Researchers responded with sparse autoencoders, or SAEs. These auxiliary models reconstruct neural activity using a larger set of mostly inactive features.

The approach aims to turn a dense activation into a more interpretable list of components. It is closely associated with toy-model evidence showing that networks can represent more sparse features than available dimensions. Those experiments offer a clear mechanism for superposition under controlled conditions.

Cammarata does not argue that toy models are useless. His objection concerns the transfer from simplified settings to trained networks. A mechanism that appears in a designed environment does not necessarily dominate a modern language model.

He also questions whether an extracted feature dictionary preserves the model's native organization. Every reconstruction introduces choices about sparsity, feature count, training data, and acceptable error. The resulting features can be understandable while still departing from the original computation.

That concern becomes more serious when researchers use the reconstruction as their main object of study. An SAE can reveal useful patterns yet miss information needed for exact behavior. It can also split one native mechanism across several learned features.

Anthropic has used sparse representations to map concepts inside language models. Its 2024 feature-mapping research identified patterns involving places, people, code, bias, and safety-relevant behavior. Later circuit work connected extracted features across model computations.

Cammarata argues that comparable circuits can sometimes be recovered from native Qwen3-32B neurons. His examples trace information between groups of neurons and measure changes after removing them. He presents this as evidence that decomposition is not always a necessary first step.

The disagreement is not simply neurons versus features. Both sides use activation examples, causal interventions, automated labeling, and circuit analysis. The real question concerns which representation receives priority.

A decomposition-first researcher treats native activations as tangled and searches for a cleaner basis. Cammarata treats the native model as the main scientific object. He would use additional representations only after direct investigation identifies a specific need.

This distinction affects error handling. A reconstructed circuit can require error terms for behavior that its feature set misses. A native-neuron circuit avoids reconstruction loss, although its human explanation can still be incomplete or wrong.

It also affects scalability. Mapping millions of neurons appears daunting, even with automation. Sparse dictionaries can contain even more features, however, and each feature requires labeling, evaluation, and connection to downstream behavior.

Cammarata sees AI research agents as the factor that changes this equation. Agents can propose neuron labels, generate tests, inspect attribution graphs, and revise failed explanations. Human researchers can then focus on unusual mechanisms and contested cases.

OpenAI explored a related direction in 2023 by asking GPT-4 to explain GPT-2 neurons. Its automated neuron study generated and scored explanations across the model. OpenAI reported that most explanations scored poorly, especially in larger models and later layers.

Cammarata interprets that history differently from a simple story of messy neurons. He says earlier methods often asked a simulator to predict precise activations. Such scoring can punish an explanation for missing weak details that matter little to a neuron's primary role.

Attention creates another complication. A neuron associated with a subject can fire near punctuation because an attention head placed contextual information there. A simulator might judge the explanation inaccurate even when the neuron has a coherent circuit function.

Cammarata therefore favors several forms of evidence. These include high-activation examples, generated counterexamples, output weights, connected neurons, training checkpoints, and interventions. No single metric decides whether an explanation is correct.

This broader process resembles experimental biology more than dictionary construction. Researchers form a hypothesis about a component, perturb it, inspect connected structures, and refine the account. Cammarata argues that coherence emerges as the surrounding map fills in.

His preferred analogy is a city rather than a knot. A city can look confusing without a map, yet its districts and roads serve related purposes. Cutting it into abstract pieces might hide the organization that makes it understandable.

Sparse autoencoders will not disappear if his account gains support. They can remain useful for searching large activation spaces or isolating a desired pattern. The pressure falls on claims that researchers must untangle every network before studying its computation.

The stronger version of Nick Cammarata's superposition challenge says the network has already performed much of the organization. Interpretability then becomes a mapping problem rather than a recovery problem. That would redirect substantial research effort toward native circuits.

An Independent Retest Puts the Headline Number Under Pressure

The first public retest supports some neuron-level structure, but it does not reproduce Goodfire's 90% figure under stricter definitions.

The independent research group IRSI examined several claims from the essay soon after publication. Its preregistered retest evaluated Goodfire's summaries, sampled Qwen3-32B neurons, random directions, and deliberately mixed units.

Three AI classifiers reviewed Goodfire's written descriptions for 500 Qwen3-32B neurons. After balancing results across eight layer groups, their majority vote classified 78.3% of described poles as having one main job. The reported 95% interval ran from 75.5% to 81.1%.

That result is below Goodfire's roughly 90% estimate, but it is not a collapse to near zero. It suggests that semantic coherence appears frequently under an outside classification process. It also shows that the result moves substantially when evaluators use different criteria.

IRSI's own dashboard-based labeler produced a stricter outcome. It marked 57% to 65% of randomly sampled neurons as one-job units. The same system classified about 70% of neurons from Goodfire's curated gallery that way.

The investigators warned that their instrument probably uses a stricter standard than Goodfire. Their procedure also misclassified roughly one-fifth of planted two-job mixtures as one-job units. Those errors prevent a clean verdict about which estimate is closer to reality.

The comparison with random directions produced mixed evidence. In the larger, post hoc sample, 65% of random neurons received one-job labels, compared with 31% of random directions. That 34-point difference cleared the researchers' planned threshold.

However, the preregistered sample contained only 48 neurons. Its 26-point gap missed the 30-point threshold. The larger sample offers suggestive support, but it does not count as a preregistered confirmation.

Layer position complicated the result further. In early layers, random neurons and random directions received similar one-job rates in one post hoc analysis. The distinction grew clearer elsewhere in the model.

These findings expose the difficulty behind any clean interpretability percentage. Researchers must decide which activations count, whether related subfunctions belong together, and whether negative and positive poles deserve separate labels. Each choice affects the result.

Automated graders add another uncertainty. Large language models can produce fluent explanations for noisy patterns. They might also reject valid structures that do not fit familiar categories.

Blinding was incomplete in parts of the retest. One classifier could distinguish neuron dashboards from random-direction dashboards with an area-under-curve score near 0.85. Some AI labelers had also seen the preregistration, although others had not.

Goodfire's released gallery creates a separate selection question. The essay describes 1,000 randomly chosen investigations across two models. Yet outside evaluators reading the summaries do not independently recreate the entire measurement pipeline.

The audit also measures neuron poles, not complete model understanding. A model can contain individually coherent units while relying on interactions that remain hard to explain. Attention, residual streams, nonlinear gates, and distributed circuits can create complexity above the neuron level.

A readable component does not guarantee a safe system. Engineers still need to identify which circuits matter during a specific decision. They must determine how those circuits change across prompts, languages, fine-tuning, and deployment conditions.

The 90% neuron insight also does not prove that deception or unsafe planning would appear under an obvious label. A dangerous behavior might emerge through ordinary components arranged in an unfamiliar circuit. Clean parts can produce an opaque whole.

Cammarata anticipates part of this objection. His essay argues that neurons become easier to interpret when researchers study their surrounding ecosystem. Connections, attention patterns, and causal interventions constrain plausible explanations.

That is a research program, not a completed safety method. It needs standardized evaluations that reward causal accuracy rather than persuasive descriptions. It also needs tests across multiple model families, architectures, and training regimes.

The technical companion paper will be important because the current essay spans several levels of argument. Its vision results, language examples, and theory of learning support one another. They do not yet form a single benchmark that outside teams can reproduce directly.

The skeptical position should therefore remain precise. The evidence does not justify dismissing the essay as anecdotal. It also does not justify announcing that AI interpretability has been solved.

A reasonable reading sits between those poles. Native neurons appear more structured than random directions in several tests. The exact share of one-job neurons remains sensitive to definitions, sampling, and evaluation methods.

Recursive Generalization Offers a Different Theory of Neural Learning

The deeper proposal replaces arbitrary feature packing with a developmental account of how related computational roles grow together.

Superposition frames capacity mainly as a storage problem. A network has more useful features than available dimensions, so it overlaps sparse representations. Low collision rates allow the model to preserve more information than a simple one-feature-per-neuron system.

Recursive generalization frames the problem around learning dynamics. Neurons develop through gradients generated by related examples and nearby circuits. New roles form where existing machinery already provides a useful starting point.

Under this view, a broad car neuron can divide into specialized car mechanisms as capacity grows. A spelling-pattern neuron can acquire branches for names or words that share the same fragments. These roles have a traceable lineage rather than an arbitrary pairing.

Cammarata argues that weak, unrelated features face a difficult path. They must survive interference from dense weak activations while contributing little to the training loss. Related extensions can instead reuse established circuits and receive more consistent gradient support.

He describes this process using ideas such as circuit racing and gradient starvation. Competing mechanisms race to reduce remaining error. Once one route learns a function, it can deprive less efficient alternatives of useful gradient.

That theory produces testable predictions. Rare facts should travel through circuits related to their subject, not through unrelated dimensions chosen only because collisions are unlikely. Training checkpoints should reveal branches growing from semantically or functionally connected mechanisms.

The proposal also predicts that larger, better-trained models can become easier to map in some respects. More capacity permits narrower neurons, while surrounding circuits supply additional interpretive context. Scale increases the number of parts but can reduce ambiguity within each part.

That conclusion runs against a common intuition. Larger models seem harder to understand because they contain more parameters. Cammarata separates the amount of work from the nature of the work.

A large city requires a bigger map than a village. Its specialized districts can still be easier to identify than a cramped settlement where every building serves several purposes. More components do not necessarily mean more conceptual entanglement.

Modern architectures might contribute to that organization. Cammarata highlights improvements that help deeper layers learn and prevent late-layer overflow. Those changes were adopted for model performance, not interpretability, but they can produce cleaner internal specialization.

The argument remains incomplete around attention. Language models contain far fewer attention heads than multilayer perceptron neurons. Each head can participate in broad routing patterns, making a one-head, one-job description less plausible.

Residual streams present another challenge because they combine contributions from many components. A coherent neuron can write into a shared space where several signals interact. Understanding native neurons does not eliminate the need to study those combined representations.

Recursive generalization must also distinguish meaningful lineage from retrospective storytelling. Researchers can often invent a connection between two concepts after seeing them together. Strong evidence requires predictions, held-out tests, and causal interventions.

Training checkpoints provide one promising route. If two functions genuinely share a developmental history, researchers should observe the branch forming over time. The sequence should repeat across training runs more often than an arbitrary explanation predicts.

Cross-model comparison offers another test. Similar tasks trained under different seeds should produce related computational motifs, even when individual neurons do not align exactly. Recurrent structures would strengthen the claim that learning organizes mechanisms predictably.

The theory also needs adversarial cases. Researchers should search specifically for neurons whose strongest functions remain unrelated after careful investigation. Those failures would reveal where recursive generalization ends and true superposition begins.

Cammarata sets a demanding challenge for the opposing view: find a convincing superposition circuit inside a real network. Such a circuit should pack unrelated functions and show downstream machinery reading the combined representation. Correlation alone would not be enough.

The same standard should apply to the clean-network thesis. A convincing explanation must predict activations and interventions beyond examples used to create its label. It should identify relevant connections and survive tests designed to falsify it.

If recursive generalization holds, researchers gain more than clearer diagrams. They gain a theory connecting interpretability to training. That connection might allow engineers to monitor how unwanted mechanisms form rather than only inspecting a finished model.

It might also enable narrower interventions. Teams could identify an environment, dataset pattern, or reinforcement signal that feeds a harmful circuit. They could then adjust that source without retraining every interpretability tool.

Those possibilities remain downstream implications. The current evidence supports targeted experiments, not deployment assurances. Enterprise buyers and developers should not treat a readable-neuron claim as a safety certificate.

For knowledge workers, the immediate relevance is indirect but real. Model providers increasingly use interpretability claims to support trust, auditing, and governance. Buyers need to know whether those claims describe correlations, causal circuits, or comprehensive behavioral coverage.

Nick Cammarata's superposition challenge sharpens that distinction. It suggests the model's native organization deserves another look. It does not remove the obligation to test what those mechanisms actually do.

Three Signals Will Decide Whether the Challenge Holds

The next phase should turn an ambitious working hypothesis into reproducible measurements, competing explanations, and model-level safety tests.

The first signal is the promised technical paper on language-model neurons. It should document sampling, thresholds, automated-agent prompts, intervention methods, and failure cases. Releasing code and activation data would let independent teams rerun the Qwen3-32B analysis.

A result near 90% under several prespecified definitions would strengthen the central claim. A wide spread across reasonable definitions would shift attention from the headline percentage toward the narrower finding that native neurons outperform random directions.

The paper must also clarify how Goodfire selected and evaluated the 1,000 neuron investigations. Random sampling should be reproducible across layers and activation densities. Negative and positive poles need consistent treatment.

The second signal is independent replication across architectures. Researchers should test dense and mixture-of-experts language models, including models trained by different organizations. Vision results alone cannot settle how transformer circuits represent language and reasoning.

Replication should include both older and newer models. If architecture improvements explain cleaner neurons, the difference should appear systematically. Researchers should be able to connect specific design changes to measurable interpretability gains.

Cross-model experiments can also test scale. Cammarata reports similar patterns in models reaching trillions of parameters, including Kimi K3, with details deferred to later work. Outside confirmation would show whether organization persists beyond one Qwen model.

The third signal is a direct contest between native-neuron and decomposition-first explanations. Both approaches should analyze the same behaviors using the same held-out prompts. Evaluators should compare causal fidelity, reconstruction error, coverage, and research cost.

A useful benchmark would include ordinary knowledge, multilingual processing, long reasoning, refusal behavior, and adversarial prompts. It should reward explanations that predict targeted interventions. Natural-language plausibility should not decide the outcome alone.

Native neurons would gain credibility if they explain circuits with less error and fewer auxiliary assumptions. Sparse representations would retain the advantage if they consistently reveal causal structure that neuron-level analysis misses.

Researchers should also inspect hybrid methods. Native neurons might define the primary map, while sparse features help locate unusual activity within residual streams. The evidence may support a division of labor rather than one universal representation.

Safety evaluation remains the hardest test. A method must detect mechanisms that matter during rare or adversarial behavior, not only frequent patterns in a corpus. Clean neuron labels mean little if critical circuits remain invisible under deployment conditions.

The public discussion should avoid two premature conclusions. One says the 90% estimate has solved the black-box problem. The other says a lower independent result makes the entire proposal irrelevant.

The more consequential possibility lies between them. Interpretability researchers may have underestimated the value of a model's own neurons while overestimating the need for universal decomposition. Correcting that bias would change experiments, tools, and funding priorities.

Cammarata's personal reversal gives the debate momentum, but authority cannot settle it. His earlier role in circuit research makes the critique historically important. Reproducible causal evidence must determine whether the new account wins.

Developers should watch the companion paper, cross-model replications, and head-to-head circuit benchmarks in that order. Together, those signals can separate a compelling research essay from a durable shift in AI interpretability.

The practical question is now clear: can independent teams use native neurons to predict and control model behavior better than existing feature decompositions? If they can, Nick Cammarata's superposition challenge will become an engineering program. If they cannot, the 90% neuron insight will remain a valuable warning about measurement, not a new foundation for model safety.