Karpathy: Only ~5,000 People Have Seen Frontier AI's Out-of-Control Capabilities
AI寒武纪 · wechat · 2026-10-12
Andrej Karpathy's new long thread updates his April take, arguing global AI perception is splitting into four torn worlds:
- Tier 1 (6B, 75%): virtually no LLM exposure.
- Tier 2 (1–2B, 20%): free ChatGPT users who see a hallucination-prone chatbot; many execs live here.
- Tier 3 (20M, 0.2%): heavy users of OpenAI Codex and Claude Code in math/coding, paying $200+/month, experiencing "AI vertigo" as weeks of work collapse into one prompt.
- Tier 4 (5,000 people, 0.00006%): those with lab-internal access — agent swarms collaborating for weeks, autonomously finding 0days, attacking and defending networks at machine speed, pushing math frontiers, while human review falls behind.
Why the rift? Three reasons: task scale matters (trivial prompts can't probe capability); capability is unevenly distributed — RL with verifiable rewards favors coding/math, which also carry the commercial value; and access tiers see genuinely different artifacts. His conclusion: "both things are true at once" — free voice mode flubs dumb questions while paid Codex refactors codebases for an hour straight.
Related event: Karpathy: LLM Perception Gap Is Widening Into a Sharp Funnel(26 posts)→
More from AGI Musings
- Agents Could Eat Ecommerce Margins by Shopping on Your Behalf — mobileraj · 2026-10-15
- Founder's take: a personal agent should feel like an elastic on-demand team — andreisavu · 2026-10-15
- ENS physics professor says colleagues all anti-AI, claims serious physics unfunded since the 70s — RachelVT42 · 2026-10-15
- Heartbreaking HN thread: mid-career developers say AI has erased their livelihoods — paraschopra · 2026-10-15
- Dev laments: too many people who don't love The Computer now work on it — mike64_t · 2026-10-15
- Frontier labs' cyber safeguards conveniently also slow Chinese labs' RL training — scaling01 · 2026-10-15