Karpathy puts agent-built software projects in front of about 20 million people, and says review is lagging

Karpathy sorts people into a funnel by what he calls napkin math. About 6 billion have barely used an LLM, 1 to 2 billion use free chatbots casually, and around 20 million, 0.2% of the population, use frontier models for math and code. He says that last group sees projects that used to take weeks or months finished by agents from a prompt, and lists building, copying, translating and decompiling apps from binaries. A fourth tier of roughly 5,000 people has internal access to frontier systems. What the outside world has seen of that tier, he writes, looks like "swarms of thousands of agents" working on software projects for weeks, and he adds that "human review and comprehension are starting to fall behind."

He gives three drivers: impact grows with problem size and time horizon, access differs by tier, and capability peaks in domains that are "digital, verifiable and economically valuable," because it comes from reinforcement learning on verifiable rewards. The post had about 871,000 views on the evening of October 11. The Hacker News thread was at 122 points and 157 comments about two hours after it was submitted.

Why it matters: This is one practitioner's estimate, not a measurement. The tier sizes and the internal-access claims cannot be checked from the post. The part a builder can use is the stated reason capability concentrates in code: the work is verifiable, which is also what makes a verification loop possible. His point that review is falling behind comes with no figures and no proposed fix.