While the essay prompted celebration among AI safety advocates, skeptics noted that the work has not been peer reviewed or independently replicated and that its language model claims rest on preliminary audits rather than a companion technical paper that has yet to be released. Critics also argued that labeling individual neurons does not explain how millions of them interact to drive complex behaviors like deception, and that safety risks remain if optimization pressure is applied to make systems impenetrable to conventional interpretability tools.
In the essay, entitled "Neural Networks Are Already Clean," Cammarata contended that early evidence for superposition was misread—such as a neuron thought to recognize both cats and cars that actually detects V-shapes—and that modern agentic tools show roughly 90% of neuron poles in Qwen3-32B have one main job, ~7% have two related jobs, and 2% to 4% remain unexplained. The findings reverse his 2023 automated interpretability project at OpenAI, which found about 1.7% of neurons legible.