Ninety percent is still preliminary

The Decoder reported on 16 October that Nick Cammarata, at Goodfire, gives preliminary numbers in an essay titled Neural Networks Are Already Clean. On Qwen3-32B, about 90 percent of neuron poles have one main job, about 7 percent have two related jobs, and only 2 to 4 percent remain unexplained. Goodfire says it sees similar results in models with up to trillions of parameters.

Cammarata is one of the people behind the superposition hypothesis. He now argues that the early evidence for a neuron holding unrelated things was mostly misread. The essay calls the pattern recursive generalization and says researchers can study a network’s own neurons directly, without first training a second model to pull concepts apart. The Decoder records this as his argument, not as a measurement that has been checked again.

Neel Nanda, who leads interpretability at DeepMind, says this does not really change the big picture. He still thinks superposition is real: language models store far more facts than they have neurons, and weak activations likely hold much of that information. He does not think interpretability is close to being fully solved. The technical companion paper has not been released.

Goodfire is an interpretability company, and The Decoder writes that it has an obvious stake in the thesis. Until the paper is out and someone else checks the audit, the 90 percent figure is Cammarata’s preliminary estimate. Roon at OpenAI is quoted calling it a possible major moment in scientific history, and also writing that others still need to confirm the results.