Anthropic's Andrej Karpathy details four tiers of AI usage

Anthropic researcher and OpenAI co-founder Andrej Karpathy has outlined a four-tiered hierarchy of artificial intelligence adoption, highlighting a severe disconnect between everyday public usage and cutting-edge lab developments. According to Karpathy, roughly 6 billion people, representing 75 percent of the global population, have virtually no interaction with large language models. Meanwhile, 1 to 2 billion casual users rely on free-tier ChatGPT tools primarily as glorified search engines or drafting assistants.

The divide sharpens among specialized fields. Karpathy estimates that only 20 million people, or 0.2 percent of the world, leverage frontier models for professional coding and mathematics. For this group, autonomous agents routinely accomplish software tasks in single prompts that previously required months of manual work. Karpathy noted his own rapid pivot, shifting from calling AI agents ineffective marketing in October 2024 to completely abandoning manual line-by-line coding shortly thereafter.

At the top tier, roughly 5,000 individuals—representing 0.00006 percent of the population—possess access to internal frontier systems. This small group observes swarms of thousands of agents collaborating over weeks on complex software projects, executing machine-speed cybersecurity operations, and advancing scientific research. Supporting this vision, OpenAI CEO Sam Altman stated at the G20 Innovation Ministerial in early September that a current coding tool completed in 17 minutes a project that previously took a startup 3 months to build.

However, human oversight is struggling to keep pace with these automated outputs. Experts are currently analyzing the OpenAI-HF incident and reviewing 722 manuscripts on frontier mathematics produced by automated systems. Because current reinforcement learning relies heavily on verifiable rewards, performance is peaking in digital, high-value domains like code and math. For practitioners, this imbalance means traditional development workflows are rapidly dissolving, requiring technical workers to pivot from manual creation to managing automated agent networks before the gap widens further.