Karpathy: the gap in understanding AI is now a funnel, with only ~5,000 people using internal frontier systems

Why it matters

【Key points】

On Oct 11 (19:12 UTC), Andrej Karpathy posted that the gap in shared understanding of LLM capability is widening into "a sharp funnel." His estimate: about 6B people (~75%) have barely touched LLMs; 1–2B (~20%) are casual free-tier chatbot users; about 20M (0.2%) use frontier models professionally in math and code and watch agents finish weeks-long projects from a prompt; and about 5,000 people (~0.00006%) have access to frontier-grade systems internally.

He blames problem size, jaggedness (capability peaks in verifiable, economically valuable domains shaped by RL on verifiable rewards), and access: free, paid, internal.

【Verifiable】

The post and his April original are public. Within about 2.5 hours it drew ~570K views, 12K likes, 1,400+ reposts and 400+ quotes; related posts hit ~1,700 at 500–700 an hour. His "722 manuscripts" refers to OpenAI's Oct 6 release of internal-model math results, of which 3 were withdrawn and 14 revised the next day over a notation error. The tier sizes are his own estimates, not survey data.

【Our read】

The top tier is the story. The ~5,000 people who see capability first are also the only ones positioned to judge its risks, while regulators, journalists and executives mostly sit in tier two. This week fits: the White House superintelligence accord leans on internal controls and voluntary audits, the FTC is asking how labs describe risk, and a Senate bill wants weights handed over 45 days before release. Each tries to bring outside judgment closer to the top, but while internal systems run months ahead, outsiders see only the trailer. Karpathy concedes human review is falling behind, and the 722-manuscript corrections show it. The more useful question than whether AI is overhyped is who can independently use a model before release and what they must disclose. And if capability concentrates in verifiable domains, the middle tier's skepticism about writing or advisory work is not purely a misunderstanding.