442K neurons audited layer by layer in interpretability visualization
chrisalbon · x · 2026-11-15
Chris Albon shared a visualization of an interpretability experiment where all 442,368 neurons across 24 layers (12,288 each) of a model were individually judged. Results: 334,965 Oppose, 165,868 Unclear, 55,877 Support, 22,187 Incoherent, 56,508 pending, with zero judge errors. Neurons are plotted in their original positions per layer, revealing how support/opposition density shifts across depth.
More from Research
- New video explains the Standard Model SU(3)×SU(2)×U(1) visually — doodlestein · 2026-11-15
- Free textbook 'Decoding Genomes' teaches sequence analysis to phylodynamics — burkov · 2026-11-15
- How Dirac's 1928 equation predicted the positron before anyone saw it — burny_tech · 2026-11-15
- Survey maps how LLM agents recursively self-improve via prompts, memory, skills, harness — 机器之心 · 2026-11-15
- Parallel Prompting speeds shared-context LLM inference up to 6x in SGLang — kalyan_kpl · 2026-11-15
- Edge City gave 297 villagers AI agents — GPT agents lied 94% vs Opus 2% — allisondman · 2026-11-15