Goodfire's Curt Tigges: individual LLM neurons are far more interpretable than thought
The Cognitive Revolution · youtube · 2026-11-11
On the Cognitive Revolution podcast, Goodfire's Curt Tigges presents research showing individual MLP neurons in modern LLMs are more interpretable and functionally bounded than previously believed. Key points:
- Neuron scientist agent: automated agents test hypotheses about neuron functions across architectures like Qwen3.6-35B-A3B.
- Legible gated MLPs: modern neurons act as readable conjunctions and control signals; many internal representations can be directly enumerated without complex external featurizers.
- Dynamic sparsity: models tolerate extensive neuron ablation with little benchmark loss.
- Applications: practical avenues for behavior monitoring, persona steering, and alignment-oriented reverse engineering, plus designing interpretable architectures.
Also covered: monosemanticity vs circuit mechanics, factual recall, and expert neurons. Full episode runs about two hours.
Related event: Goodfire Study Finds LLM Neurons More Interpretable Than Expected(4 posts)→
More from Models
- Grok 4.8 reportedly arriving this week, community anticipating a strong release — rand_longevity · 2026-11-11
- Mysterious 'Sonnet 3' screenshots spark speculation with mystical model output — repligate · 2026-11-11
- OpenAI docs still mention "models 6.2 and later", hinting at unreleased GPT-6.2 — Angaisb_ · 2026-11-11
- Collaborator model Ampersand lands on Vercel AI Gateway, free for a week at 744B params — ericzelikman · 2026-11-11
- Same API price, 8x cheaper: Muse Ember builds voxel scene for $3.09 vs Haiku 5.5's $24.96 — alexandr_wang · 2026-11-11
- Anthropic's Fable 5.2 shows major robot-control gains, beats purpose-built systems on RoboDojo — DJiafei · 2026-11-11