Nick Cammarata, a researcher at Goodfire and former OpenAI employee, published an essay on October 15, 2026, titled 'Neural Networks Are Already Clean.' He argues that modern large neural networks possess a more organized internal structure than previously assumed, with most neurons performing distinct functions.
New analysis using agentic methods indicates that nearly 90% of neurons in models like Qwen3-32B have a single primary function. Another 7% of neurons handle two related roles, while only 2–4% remain unexplained. These results contrast with a 2023 OpenAI project where only 1.7–2% of neurons were considered legible.
Cammarata suggests that previous observations of 'polysemantic' or 'tangled' neurons, such as those in InceptionV1, were likely due to inadequate analysis tools rather than the architecture of the models themselves. He challenges existing theories regarding superposition and polysemanticity.
The research relies on automated analysis rather than manual inspection, suggesting that interpretability could scale alongside model development. While some researchers view these findings as a breakthrough for AI alignment, others remain cautious, noting that the debate over representational superposition continues.
