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 clearer internal structure than previously assumed, with most neurons performing distinct roles.
Recent analyses using agentic methods on models like Qwen3-32B indicate that nearly 90% of neurons serve a single primary function. An additional 7% handle two related roles, while only 2–4% of neurons remain unexplained according to these preliminary findings.
These results contrast with Cammarata's 2023 project at OpenAI, where only 1.7–2% of neurons were considered legible using available tools. He suggests that previous observations of 'tangled' or 'polysemantic' neurons, such as those in InceptionV1, were likely due to inadequate analysis methods.
The researcher attributes these improvements to automated, agentic analysis rather than manual inspection. He suggests that this approach allows interpretability to scale alongside model development, potentially aiding AI alignment efforts.
The findings have received mixed reactions from the research community. While some view the work as a potential breakthrough, others caution against overstating the results, noting that concepts like representational superposition remain significant in the field.
