How interpretable are neurons in today's large language models?

In the same essay, Cammarata reports a specific, quantified claim beyond the recursive-generalization argument: that neuron-level interpretability in today's large models is already very high. "In preliminary audits of Qwen3-32B, we estimate that roughly 90% of neuron poles have one main job, often with small related offshoots; another ~7% have two related jobs; and only ~2-4% are not yet understood." He adds that "in the sets I've looked at more closely, considering strong activations, it's closer to 97% non-arbitrary mixtures" and that his team "found similar results in models up to trillions of parameters, like Kimi K3," with dense and mixture-of-experts models "roughly similar."

Cammarata attributes the jump from near-zero to near-total interpretability to tools and models improving, not to a different kind of network. He contrasts his figures against two 2023 benchmarks he treats as the field's low-water mark: Anthropic's Towards Monosemanticity, where the median neuron in a studied one-layer model scored 0 out of 14 on its interpretability rubric ("the rater couldn't form any hypothesis"), and OpenAI's automated-interpretability project, which he worked on, which found about 1.7% of neurons interpretable "by our most generous measure." He is explicit that the percentages compress a high-dimensional picture and depend on where the threshold is set for what counts as "one thing," and he calls the headline "cleanliness" number "a broad pointer to the regime a neural network is in, rather than something precise to either optimize directly or hold too tightly."

This figure is the empirical half of the companion note Does recursive generalization explain polysemantic neurons better than superposition?: the mechanism (branch-like growth of related senses) is what Cammarata says produces this measured outcome (near-total single-job neurons). It also bears on Can sparse weight training make neural networks interpretable by design?: if ordinary dense and MoE models are already landing at 90-97% single-job neurons with no sparsity training at all, the capability-for-interpretability trade that sparse-weight training requires may not be necessary to reach useful legibility, at least by Cammarata's own threshold.

The percentages come from Goodfire's own "preliminary audits," are self-reported, not peer-reviewed, and depend on thresholds the author chose for counting a neuron's activations as "one job" versus "related jobs" — a choice he acknowledges "varies" the number. The 97% figure is drawn from "relatively small samples," and no independent party has replicated the measurement. The excerpt does not establish that these rates hold across model families beyond those Cammarata's team has audited; the finding should be read as an optimistic preliminary estimate from one interpretability lab's tooling, not a settled measurement of how interpretable large models are in general.