Karpathy: Public Vastly Underestimates AI Capabilities

Andrej Karpathy left OpenAI to join Anthropic earlier this year, and he's using his new perch to make a pointed argument: almost nobody outside a small circle of researchers actually knows what AI systems can do right now. In a recent post, Karpathy lays out a tiered breakdown of the roughly eight billion people on the planet and how each group interacts with large language models. The numbers are stark.

Six billion people, by his estimate, have barely touched an LLM. Another one to two billion use free chatbot tools the way they'd use a search engine or a grammar checker, treating them as a convenience rather than anything resembling real intelligence.

Karpathy's point isn't just about adoption curves. It's about the distance between what the public experiences and what happens inside labs like his own, where frontier systems are reportedly handling problems that would take trained specialists years to work through. He zeroes in on a narrow slice of professionals, maybe 20 million people in math and programming, who actually use these tools at the frontier of their capability. Fewer still, by his count, get to see the systems before they're released at all.