Andrej Karpathy revisits his April post about people talking past each other on AI. His new framing is a funnel: people who barely encounter language models, casual chatbot users, professionals using frontier coding and mathematics tools, and a much smaller group with access to internal systems. Their experiences differ not just in degree, but in the kinds of work they ask the systems to do.
He proposes three reasons for the gap:
- The task matters. A short question does not exercise the same capabilities as a difficult project pursued over hours or days. Professional users have a reservoir of consequential problems to try.
- Progress is uneven. Code and some mathematics tasks have checkable outcomes—tests pass or fail, proofs can be examined. Karpathy argues that reinforcement learning and commercial incentives concentrate progress in these areas rather than improving every use case equally.
- Access matters. Free chatbots, paid agents and unreleased internal systems are different products. Experience with one is not a reliable verdict on all three.
The useful distinction is between disagreement about the same tool and disagreement produced by different tools, tasks and access. A chatbot fumbling a simple question does not settle what a coding agent can do; a successful coding project does not establish equivalent competence everywhere else.
His population figures are explicitly “napkin math,” not survey results. The post also describes internal agent swarms, scientific advances and work outpacing human review without supplying evidence that lets readers evaluate those examples. Treat that upper end of the funnel as Karpathy’s account, not a demonstrated capability ceiling. The practical takeaway is narrower: specify the model, access tier, task and verification method before debating whether AI works.