Nick Cammarata, a co-author of early mechanistic interpretability work at OpenAI and now at Goodfire, has published an essay arguing that neural networks are "already clean." The source reports that many known cases of polysemanticity and superposition—where neurons appear to encode multiple unrelated concepts—fall away under further agentic analysis. Cammarata claims that 90% of neurons in Qwen-32B have a single job.
Reactions quoted in the source are strong. Deedy calls it "one of the most important results in AI, period," and roon says it may be remembered as a major moment in scientific history, with even Eliezer having to lower his p(doom). The source frames the essay as a rare interpretability story, following earlier coverage of Goodfire and Anthropic's scaling of interpretability methods.
If the claim holds, it would mean neural networks are not tangled knots that require a second model to unfold. Instead, neurons could be studied directly, potentially turning alignment into "normal science" and lowering the global probability of doom. The source also covers other news, including Meta's Muse Spark Open release, but the interpretability essay is the headline.