I Can See Everything
Zvi (Don't Worry About the Vase) TheZvi
Nick Cammarata and Goodfire argued that neural network neurons are “clean” and mostly map to single concepts, sparking wide debate about how solvable mechanistic interpretability is and how much it changes alignment expectations. In preliminary audits of Qwen3-32B, they estimated about 90% of neuron “poles” have one main job, with about 7% having two related jobs and 2–4% not yet understood. The consensus shifts toward more optimism about understanding LLM internals with better tools, while also pushing back against overhyped comparisons and claims of near-term interpretability or alignment breakthroughs.
Why it matters
Nick Cammarata of Goodfire makes a bold claim: Neural Networks are Already Clean. This was his follow up after various resulting online discussions. This is clearly good work that tells us new and important things. The question is, how big … Continue reading →
