Researcher Andrej Karpathy wrote on X on Oct. 11 that the gap in shared understanding of AI capability is widening, describing it as less like two groups talking past each other and more like a sharp funnel.
By his own estimate, about 6 billion people, roughly 75% of the world's population, have barely touched a large language model. Another 1 billion to 2 billion are casual users of free consumer chat products. Only about 20 million people, roughly 0.2% of the population, have seen firsthand that agents can now finish complex projects that used to take weeks or months with a single prompt. Access to frontier, unreleased lab systems is narrower still, Karpathy estimated, at around 5,000 people.
Karpathy argued the gap is about exposure and domain, not just the model. The same system looks like a mediocre chatbot or like finished professional work depending on whether the task is digital, verifiable and economically valuable, the conditions under which reinforcement learning on curated environments produces the largest capability gains, he wrote.
The post drew more than 475 points and 690 comments on Hacker News. Commenters who use agents with code execution and terminal access described weeks-long projects compressed into days, while skeptics countered that the gains are concentrated in automatically verifiable work like coding and remain narrow or invisible everywhere else.
Karpathy co-founded OpenAI and later led Tesla's AI team, so his read on who actually sees the frontier carries weight inside the industry. For founders, the practical takeaway is blunt: if your target users are not among the roughly 20 million people already running agents on hard problems, you cannot assume they share your sense of what AI can do today.