Karpathy: the gap between casual and heavy AI users is turning into a funnel

In an X post on 11 October 2026, Andrej Karpathy revisited an April post and said the gap in how people understand AI is widening. His napkin maths: about 6 billion people have barely used AI chatbots; 1 to 2 billion use free-tier tools now and then, like a better search or a writing aid, including many professionals outside maths and code; about 20 million use frontier models for maths and code and see projects that took weeks or months done by agents; and about 5,000 people inside labs see internal systems far ahead of that. He gives three reasons: the payoff grows with the size and ambition of the problem you bring; AI is strongest in work that is digital and easy to check, like code and maths; and access differs (free, paid, internal). These are his estimates, not survey data. Hacker News thread (331 comments): No consensus. Sceptics say his claim that big projects can now be done "with a prompt" does not match their experience, since real work still needs specs and careful review, and some read the internal-systems tier as "trust us, give us money". Others say tools like Claude Code and Codex clearly work and that giving models a terminal is what changed things; one argued free models already cover what most people need. A point raised on both sides: checking AI output is becoming the real bottleneck. Sorted take: If your view of AI comes from a free chatbot, test a paid frontier model on one big, checkable job you care about before you decide. Either way, keep the review step, because verifying is now the slow part.