Karpathy: The World Has Only Seen a Preview of AI's Capabilities

Andrej Karpathy, an OpenAI co-founder who recently moved to Anthropic, wrote on X that access to AI breaks down into four tiers. About six billion people, three quarters of the world's population, have had almost no contact with large language models, AI systems trained on huge amounts of text. Another one to two billion people use only free tools, like an improved search engine. Twenty million coding and math specialists already trust agents, AI systems that carry out multi-step tasks on their own, with projects that used to take weeks or months. Only about five thousand people have access to Anthropic's internal systems, where swarms of thousands of agents work for weeks on software projects that would take the best professionals in the field years. Karpathy wrote that "the external world has only seen a preview" of what these systems can really do.

According to Karpathy, this jump in capability is concentrated in areas that are digital, verifiable, and economically valuable. Right now that means coding and math, because training AI relies on checking whether results are correct, and companies pour their training resources into the most profitable areas. He also said human review can no longer keep pace with AI, even at the very top tier. Mathematicians are still working through hundreds of AI-generated manuscripts, and some incidents are still being pieced together months after they happened.