Karpathy Estimates Only About 5,000 People Have Access to the Most Advanced AI

Andrej Karpathy, a co-founder of OpenAI and former director of AI at Tesla, estimated in a post how many people on the planet actually have access to the most advanced internal AI systems: roughly 5000. For comparison, hundreds of millions of people use ChatGPT and its counterparts every day. By this estimate, the gap between what the public sees and what employees at leading labs work with is enormous.

The estimate is Karpathy's own. There is no counting methodology, and the labs have not confirmed it. The figure works more as an order-of-magnitude guide. But it drew attention because Karpathy rarely talks about such things at random: he has seen the inner workings of both OpenAI and Tesla.

What is happening inside the labs

According to Karpathy, inside frontier labs it is no longer a single assistant helping a single developer. Thousands of agents assemble into cooperative clusters and work without a break for weeks. This is a different mode from the familiar chat, where a person asks a question and waits seconds for an answer.

The list of tasks he names looks like this:

• large software projects that agents carry out over weeks;
• hunting for zero-day vulnerabilities, meaning holes the developers don't know about;
• participation in cyberattacks and cyber defense at machine speed;
• contributions to scientific discoveries;
• work on frontier mathematics problems.

The timescales deserve a closer look. Continuous work over weeks means the agents hold the project's context, split subtasks among themselves, fix their own mistakes, and don't need a human at every step. Public agent products so far more often operate on the scale of hours and in limited scenarios.

Why the 5000 figure matters

If the estimate is close to reality, a narrow circle of people shapes everyone else's picture of AI. That circle includes lab employees, a few outside researchers with special access, and some people in government and security agencies. Everyone else judges progress by public models, which arrive with a delay and usually with reduced capabilities.

Several consequences follow. Debates about risks and benefits take place at different levels of information: some people argue about chatbots, while others work with clusters of agents. Regulators have a harder time assessing what they cannot see. And companies building products on top of public APIs struggle to plan a roadmap when the real capability ceiling is hidden.

Cybersecurity stands apart. If agents really do find zero-days and operate in both defense and offense at machine speed, the reaction time of human analysts stops being the deciding factor. For defenders it is a chance to close vulnerabilities faster; for attackers, too. Which effect will outweigh the other cannot be determined from a single post.

How to use this information

It is useful to treat such statements as a signal, not as a fact. Practical steps for teams and developers:

1. Assume in your plans that public agents will grow into today's internal capabilities with a delay, and design products so they can be switched to a stronger model without a rewrite. 2. Start building infrastructure for long-running agents now: action logging, permission limits, rollback of changes. In a year or two this will become mandatory. 3. If you are responsible for security, revisit your threat model. Assume that vulnerability discovery is being automated and that the "found a hole — wrote an exploit" cycle is getting shorter. 4. Follow primary sources: statements from the labs, model system cards, and reports from independent evaluators. The words of individual experts, including Karpathy, remain opinion until data confirms them.

The main question for the near term: how quickly will closed capabilities become public, and who will manage to prepare for them.