ICA Lens lands on Neuronpedia: decompose tokens in your own prompts
CatAstro_Piyush · x · 2026-11-11
ICA Lens is now live on Neuronpedia, letting anyone decompose tokens in their own prompts and inspect which features activate and how they push logits up or down.
For example, on Qwen3.6-27B's 24-ICALENS-RES layer, gratitude-related tokens activate a "thanks" feature (positive logits include Thank, 致谢, etc.). Neuronpedia now hosts 8.6M neurons across 9 models with 1.3M community explanations, plus tools like persona-neurons, ModelMapper, and Neuron Lens — a major public resource for interpretability research.
More from Research
- Demo: training models to read K/V vector breadcrumbs left by other instances — Sauers_ · 2026-11-11
- EEG can't match implanted BCIs: physics caps scalp signals at ~70Hz — EricMTrautmann · 2026-11-11
- 2B model takes 1.7 days and 22 steps to land one FunSearch hit — cephaloform · 2026-11-11
- Baseten's All-Gather research conference unveils agenda with Noam Brown, Oriol Vinyals, Mark Chen — vjkaruna · 2026-11-11
- The 1960 Doomsday Equation That Predicts Infinite Population on Nov 13, 2026 — csvoss · 2026-11-11
- The Key Assumption: Productivity a = a0·N^(1/k) in the Doomsday Paper — csvoss · 2026-11-11