Hand-picked ImageNet photos break InceptionV1 neuron 4e:55's 'pointy vs round' rule
livgorton · x · 2026-10-22
nickcammarata published a long essay arguing interpretability may be the most important scientific problem right now: after OpenAI's early mechanistic work (only 2% of neurons legible with primitive tools), modern agentic tools reveal most neurons — up to trillion-parameter models — as legible.
@MiniUlisse stress-tested the essay's flagship claim that InceptionV1 neuron 4e:55 fires on pointy things and not round ones. On the essay's own photos the rule holds, but across other ImageNet photos all four combinations occur: round fluffy faces fire strongly (Persian cat 81%, Pomeranian 65%) while pointy German shepherds stay silent (3%). Takeaway: tidy neuron-level rules may not survive beyond curated examples — a caution for interpretability narratives.
More from Models
- Carbon checkpoint on LMArena updated, Gemini web and app change in sync — lyraxana · 2026-10-22
- KRAFTON's Raon-1-mini-preview Trends on Hugging Face — KRAFTON · 2026-10-22
- Gemini 4 Carbon Checkpoint Wows at 3D Generation: Opus 5.5 Level? — WorldofAI · 2026-10-22
- GDPval benchmark under fire: first seed artifact in public split looks off — vincent_koc · 2026-10-22
- Rumor: Anthropic's next model could land today or tomorrow per release cadence — himanshustwts · 2026-10-22
- Lighthouse Benchmark: Gemini Argon Runs 8x Cheaper Than Claude Opus — cgarciae88 · 2026-10-22