Karpathy: Only About 5,000 People See AI’s Fastest-Moving Edge

Andrej Karpathy says the gap between the public perception of artificial intelligence and what advanced models can do in specialized settings has widened dramatically over the past six months.

Different levels of AI experience

In the estimates he presented, nearly 6 billion people have barely used large language models. Another 1–2 billion mainly use free chatbots as improved search engines.

At a more advanced level, Karpathy says roughly 20 million developers and mathematicians are using frontier models to complete in days projects that previously took much longer. These figures are his broad estimates and should not be read as definitive user statistics.

A different world inside AI labs

Karpathy describes a much smaller group of about 5,000 people inside AI laboratories who are seeing another level of capability: thousands of AI agents working for weeks to identify software vulnerabilities or investigate scientific and mathematical problems, potentially producing results faster than people can review and verify them.

He attributes the divide to task complexity, uneven model capabilities and differences in access. Simple everyday questions do not necessarily reveal how a model performs on a large, multi-step project. Models are particularly effective on digital problems with verifiable outcomes, where reinforcement learning can be used to optimize performance.

Two contrasting views at once

The result is that much of the public encounters chatbots that still make obvious mistakes, while a very small group observes a faster and harder-to-comprehend pace of progress. Karpathy’s central point is that both views can be valid at the same time, even as the gap between them continues to grow.