Summary
David Shapiro covers Goodfire’s new essay, “Neural networks are already clean,” published October 15, 2026 by Nick Camarata, and the heated reaction it sparked. Camarata argues that ambitious mechanistic interpretability, meaning understanding all of a model, may be solved by the time it is needed most.
The video explains the background: neurons that seem to do several unrelated jobs, known as polysemanticity, and the superposition theory that led researchers to use sparse autoencoders as a workaround. Camarata revisits the famous Inception V1 neuron that fires on both cat faces and car fronts and finds it is simply a V-shaped detector. He reports that about 99% of last-layer neurons in a modern network of the same size are easily interpretable, and introduces the idea of recursive generalization to explain the remaining complexity.
Goodfire’s preliminary audits of Qwen3 32B estimate roughly 90% of neuron poles have one main job, about 7% have two related jobs, and only 2 to 4% remain unexplained. An AI agent testing the explanations produced strongly firing examples 76% of the time, versus about 1% for unrelated topics.
Shapiro also covers the caveats, including Camarata’s roughly 85% confidence, skepticism from researchers such as Neel Nanda, and Eliezer Yudkowsky’s sharp response.
📺 Source: David Shapiro · Published October 18, 2026
🏷️ Format: News Analysis
