The findings reverse Cammarata's earlier work at OpenAI, where he co-authored the Zoom In paper and estimated that only 2% of neurons were legible. He credits this shift to the application of modern agentic tools and the use of simple methods at absurd scale. Other researchers have suggested that these results could transform the pursuit of AI alignment into a "normal science" and increase the likelihood of achieving mechanistic interpretability before the emergence of superintelligence.
Tuesday, October 20, 2026
Story
Nick Cammarata Finds AI Models Legible Reversing Prior Consensus
AI5 days ago
The "tangled knots" of neural networks are far simpler than earlier scientific assumptions suggested, according to research involving models with trillion parameter scales. Nick Cammarata, a researcher at Goodfire, reported that the vast majority of neurons in large models are legible, enabling the direct tracing of circuits in systems such as Kimi K3. Preliminary audits of Qwen3-32B conducted by Cammarata found that roughly 90% of neuron poles serve a single main purpose, while only 2% to 4% remain unexplained.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.