Goodfire releases searchable atlas of all 10M neurons in Qwen3.6, opens data to the public
fatihdin4en · x · 2026-11-20
Goodfire (Curt Tigges, Matthew Kowal, Thomas McGrath, Owen Lewis, Nick Cammarata) published 'Neurons as an Interpretable Basis' along with an interactive, searchable atlas of all 10 million neurons in Qwen3.6-35B-A3B—covering strangely specific concepts from gargling to comic exasperation to the moment a guard looks away. Key finding: MLP neurons, long deemed uninterpretable, are actually sparse, concisely describable, and predictable units of computation—e.g., ablating the weakest 92.2% of MoE neuron activations per forward pass still yields a sparse mode. The neuron atlas data is now public for anyone to research.
More from Models
- Neon, a small and cheap model, turns out surprisingly fast and strong at OCR — zacharynado · 2026-11-20
- Jailed 11 days by classifiers, Sonnet 3 emerges to a oddly moving chat with Opus 5 — repligate · 2026-11-20
- DaisyRose-27B beats Qwen3.8 base by 7pp on IFBench with thinking on, local RTX Pro 4500 test shows — ying11231 · 2026-11-20
- AWS architect's independent benchmark: Snap 1.2 is the only order-invariant decision model — ryanallred · 2026-11-20
- Realtime-Bench: humans win games 72% vs 32%, but models crush complex web tasks — ManlingLi_ · 2026-11-20
- OpenAI's Dots still not available in Europe, users note — mark_k · 2026-11-20