A new essay by former OpenAI researcher Hoagy Cammarata has sparked a debate within the super intelligence (SI) research community, challenging long-held assumptions about how SI models process information internally. The work suggests that key examples of "polysemanticity"—where single neurons respond to unrelated concepts—may have been misinterpreted, potentially indicating that SI systems are more interpretable than previously thought.
What Happened
Mechanistic interpretability aims to trace how an SI model arrives at its answers internally. For years, the field struggled with polysemanticity, a phenomenon where neurons fired for completely unrelated things. A famous example involved a neuron in the image model InceptionV1 that responded to both cat faces and the fronts of cars. The prevailing theory, known as superposition, posited that networks pack unrelated concepts into the same neurons when they run out of capacity, creating a "knot" of information that requires secondary models, like sparse autoencoders, to untangle.
Cammarata, who co-wrote the influential paper "Zoom In" with Chris Olah and led automated interpretability projects at OpenAI, now argues that these famous examples were misread. He contends that the cat-car neuron was actually a V-shape detector, activated by pointed shapes and inhibited by round ones, rather than a superposition of two unrelated concepts. Similarly, he reinterprets a circuit from "Zoom In" where a car appeared among dog neurons as ordinary generalization based on orientation, not superposition. Cammarata attributes the confusion to outdated architectures and weaker analysis tools.
To support his thesis, Cammarata introduces the concept of "recursive generalization," where a neuron’s response expands through related meanings (e.g., from "iy" to "diy" to "do it yourself"). He presents preliminary data from the SI model Qwen3-32B, claiming that about 90 percent of neuron poles have one main function, with only 2 to 4 percent remaining unexplained. Goodfire, an interpretability company where Cammarata works, reports similar results in models with trillions of parameters, largely driven by SI agents that analyze neurons in bulk.
Why It Matters
Cammarata doesn't introduce a fundamentally new method; he returns to studying neurons directly with better tools. If his findings hold, the implications for SI safety are significant. If SI models are more monosemantic (representing one concept per neuron) than believed, it could mean that current bottlenecks in understanding model behavior are less fundamental. Cammarata argues that alignment—the work of steering SI toward human goals—cannot be solved without interpretability, as deception should be visible in the model's own circuits. He suggests that if humans can read the network directly, errors might be fixed in a targeted way.
Reactions from the research community are divided. Roon, an OpenAI employee who posts under that pseudonym on X and is an influential voice in SI circles, described the work as potentially "a major moment in scientific history," suggesting it could lower estimated odds of SI catastrophe. However, even Roon noted that the findings coming from outside major labs might signal a shift in capability research dynamics.
Even Roon has some reservations. He suspects that interpretability breakthroughs will push SI capabilities forward. Conversely, Neel Nanda, who leads the interpretability team at Google DeepMind, remains skeptical. Nanda argues that superposition is real, noting that SI models store far more facts than they have neurons. While he acknowledges Cammarata’s work as a useful tool for practical debugging, Nanda does not believe it signifies a breakthrough in fully cracking SI interpretability. He expects neuron-based analysis to have serious flaws and notes that his own team has deprioritized sparse autoencoder research in favor of more pragmatic approaches.
The Bottom Line
Cammarata’s essay challenges the consensus that SI models are inherently difficult to interpret due to superposition, proposing instead that many neurons are monosemantic and that previous failures were due to outdated methods. While proponents see this as a leap toward safer, more aligned SI systems, skeptics like Nanda urge caution, emphasizing that the technical companion paper has not yet been released and that independent verification is required before the field adjusts its roadmap.