Karpathy: only ~5,000 people see frontier AI, 6 billion have barely touched LLMs

APPSO · wechat · 2026-10-12

Andrej Karpathy argues the perception gap on LLM capabilities is widening into an extreme funnel, with four tiers of exposure:

6 billion people (75%) have barely touched an LLM;
1–2 billion (20%) occasionally use free ChatGPT for email and lookups and find it unimpressive;
20 million (0.2%) deeply use agents for coding and math — reading/writing files, running tests, iterating — completing projects that once took weeks;
only 5,000 people see frontier internal systems: agent teams spending weeks on large software projects, machine-speed cyber attack/defense, and math progress so fast that mathematicians are still digesting 722 frontier math manuscripts.

Why AI feels 'useless' to most: frontier LLMs only shine on large, complex, long-horizon problems; capabilities explode first in verifiable, economically valuable domains (code, math) where RL gets clean reward signals; METR estimates agent capability doubles every 7 months, yet a 2025 randomized study found experienced devs took 19% longer with AI, and 2026 evaluations suffer sample bias as devs refuse no-AI conditions; a16z data shows the top 1% of payers contribute 19.5% of consumer AI revenue, and only 35M of ChatGPT's 810M weekly users pay.

The piece also critiques the industry's launch-cycle culture: OpenAI and Anthropic's relentless release-and-marketing cadence increasingly substitutes narrative for real experience — launches replace experience, hype becomes reality.

Original post →

Karpathy: only ~5,000 people see frontier AI, 6 billion have barely touched LLMs

More from AGI Musings