Goodfire Researcher Claims Neural Networks Are Inherently Understandable, Not "Tangled"

Nick Cammarata, a researcher at the AI interpretability lab Goodfire, published a 26,000-word essay on October 15, 2026, titled "Neural Networks Are Already Clean," proposing that neural networks are fundamentally more interpretable than the AI community has largely assumed. Cammarata's work directly challenges the concept of "superposition," which posits that individual neurons often represent multiple, unrelated concepts simultaneously, making their internal workings difficult to decipher.

The essay re-examines canonical examples of polysemantic neurons, such as the "cat and car" neuron in InceptionV1, which was previously understood to respond to both animals and vehicles. Cammarata argues that such instances were misread, suggesting that the neuron was instead identifying a shared visual characteristic, like pointed versus rounded shapes. He attributes the perception of mixed concepts to neurons developing related offshoots during training, coupled with limitations in the tools and data used for analysis.

Goodfire, founded in June 2024, focuses on mechanistic interpretability, aiming to reverse-engineer neural networks to understand and control their internal operations. The company's research, including its Silico interpretability agent, develops tools to investigate learned features and debug unexpected predictions in AI models. Cammarata's essay, labeled a "working hypothesis," suggests that if a model's existing components are understandable, the effort required for AI alignment, ensuring AI systems pursue intended goals and respect constraints, could become significantly more manageable.

Cammarata's preliminary audit of the Qwen3-32B model, cited in his essay, indicates that approximately 90% of neuron poles have a single primary function, with about 7% handling two related functions, and only 2-4% remaining unexplained. He reports similar findings in models up to trillions of parameters, such as Kimi K3. While acknowledging that trillion-parameter models are "not close to solved," Cammarata's findings suggest a more optimistic outlook for understanding large language models.

The essay has generated considerable discussion within the AI research community. Some researchers, including Roon, have called the research a potential "major moment in scientific history," highlighting the implications for AI safety. The ability to understand and control AI's internal workings is seen as increasingly important as AI systems are deployed in critical applications like cybersecurity and finance. Goodfire, which has raised over $200 million in funding and was valued at $1.25 billion in a Series B round led by B Capital in 2026, aims to advance research and technology to align AI. A technical companion paper to Cammarata's essay is expected to provide further details and independent validation.