Andrej Karpathy breaks down the capability gap

The piece argues that public understanding of large language model (LLM) capability is diverging into a sharp funnel of experience rather than two groups merely talking past each other. Roughly 6 billion people (~75% of the world) have barely encountered LLMs; about 1-2 billion (~20%) are casual users of free-tier chat interfaces that appear "derpy" and serve as slightly better search or writing aids; around 20 million (0.2%) - mostly professionals in math and code - are using frontier-grade models to compress weeks or months of work into a prompt-driven workflow for building, copying, translating and even decompiling apps; and a tiny cohort of roughly 5,000 individuals (0.00006%) with internal access are running swarms of agents on massive projects.

That funnel is driven by three forces: impact scales with ambition and problem size (small queries don’t stress the system), capability is jagged and peaks in digital/verifiable/economically valuable domains because training favors verifiable rewards, and access tiers (free, paid, internal) concentrate the most powerful tools in few hands. Consequences include rapid, asymmetric productivity gains in specialized domains, emergent high-impact behaviors (zero-days, automated cyber operations, frontier math and science), and a growing gap where human review and broader comprehension are starting to lag behind the systems’ outputs.