Karpathy at Anthropic: 5000 People See an AI the Rest of Us Don't

Andrej Karpathy, an OpenAI co-founder who now works at Anthropic, points to a growing gap in knowledge about AI. By his estimate, about 75% of the world's population has barely interacted with large language models. Roughly 5000 people, by contrast, have access to internal frontier systems, where swarms of agents do work that would take the best experts years. The outside world, Karpathy says, has seen only a "preview" of what AI can do. At best, the public sees it as a strange chatbot.

Three layers of perception

Karpathy's picture sorts people into groups by their proximity to the technology. The largest group, three quarters of humanity, has barely encountered LLMs. That is not necessarily a rejection: many people simply have never opened ChatGPT or anything like it. At the top sits a narrow tier of a couple of thousand people, about 5000, who work with systems that are not available from the outside. In between are those who use chatbots and coding assistants and judge the limits of the possible by them. That last layer is my own inference; the summary of Karpathy's remarks does not name it directly.

His main thesis concerns exactly this middle layer. A chatbot in a browser window shows only part of what models can do when they run as a swarm, with tools and over long task horizons. According to Karpathy, at that level agents do in days what would take top specialists years.

What is known and what is not

The source is brief, so it matters to separate claims from verified data. The figures of 75% and 5000 look like Karpathy's own estimates, and the summary gives no methodology. It is not disclosed which companies employ these five thousand people. It is also unknown which tasks he means or how "years of expert work" was measured. None of these claims can currently be verified externally.

Key points:

• about 75% of the world's population has barely used LLMs;
• roughly 5000 people have access to internal frontier systems;
• agent swarms complete tasks that would take top experts years;
• the public image of AI remains that of a "strange chatbot".

Why these words deserve attention

Karpathy was part of OpenAI's core team, led AI development for Autopilot at Tesla, and popularized the term vibe coding in 2025. He speaks as an industry insider, and such testimony is rare. But the insider view has a downside. Lab employees have an interest in making their systems' capabilities look large and the gap between closed and open models feel like a chasm. It helps to combine trust in the speaker's experience with a sober demand for evidence.

There is also an independent reason to take the thesis seriously. If internal agent systems really are ahead of public products, then the assessments of markets, regulators, and the companies themselves rest on an outdated picture. Decisions about jobs, security, and investment are made based on what is visible from the outside.

How to work with this

For developers and teams, the practical takeaway is simple: judge AI not by a conversation in a chat window but by long tasks. Give an agent a real ticket, a repository, or a research question, and see how long it works without intervention and how accurate the result is.

Companies should measure their own internal gap: who in the organization already works with agents on a regular basis, and who sticks to chat. That difference is usually where the speed advantage hides.

Finally, it is worth treating bold claims about closed systems as hypotheses. Until the labs show metrics and reproducible examples, Karpathy's words remain the testimony of one authoritative participant, not a proven fact. Watch the public releases: they will show how far the "preview" lags behind what the five thousand people inside are seeing.