Karpathy: Only about 5,000 people worldwide have real access to frontier AI

Karpathy mentioned that he recently dug up an old tweet from April this year, when he had already noticed a clear divergence in people's understanding of AI capabilities. In his view, this divergence has now evolved from two groups talking past each other into an increasingly steep capability funnel.

The top of this funnel is extremely narrow: he estimates that globally, the number of people who can truly access the most frontier internal AI systems may be only about 5,000, roughly 0.00006% of the population. The bottom is extremely wide: about 6 billion people, or about 75% of the global population, have barely truly encountered LLMs.

There are two more layers in between. About 1 billion to 2 billion people (about 20% of the population) only occasionally use free ChatGPT-type products; what they see is often still a somewhat clumsy chatbot, and they treat it as a better search or writing assistance tool. Further down, about 20 million people (about 0.2% of the population) use frontier-level LLMs professionally in math and programming, and have already seen with their own eyes that large, complex projects that used to take weeks or even months can now potentially be completed by an Agent with just one prompt.

Karpathy believes this funnel is mainly driven by three factors. First, influence amplifies as ambition, problem scale, and task time horizon increase. If you only ask a question whose final answer is just one paragraph, it is almost impossible to truly squeeze the capabilities of these systems; you need to have a large number of truly important, very large-scale, and highly difficult problems in hand. This mainly determines the gap between ordinary consumers and professional users.

Second, system capabilities have obvious jagged characteristics. The capability peaks of LLMs are concentrated in domains that are digital, verifiable, and economically valuable. He says large model capabilities mainly come from reinforcement learning based on verifiable rewards, and the choice of training environments is itself driven by commercial value. This mainly determines the gap between different professional domains, such as why math and programming stand out in particular.

Third, access permissions are tiered. Free versions, paid versions, and internal systems are not at the same level of capability. Karpathy says this is the strangest part right now: the vast majority of ordinary people have not actually truly encountered these systems, and even if they have, what they see is often just a somewhat dumb chatbot; most professionals currently feel only relatively limited efficiency gains; while a small group of professionals has already begun to experience a kind of dizziness, as the capability curve before their eyes is suddenly becoming almost vertical.

Karpathy says that inside frontier labs, tens of thousands of Agents can form collaborative clusters and run continuously for weeks, advancing large software projects, discovering zero-day vulnerabilities, participating in cyber offense and defense at machine speed, and even driving scientific discovery and frontier mathematics research. He says that if these things were handed to the top professionals in the field, they might have taken years in the past.

These descriptions are his statements about conditions inside frontier labs and cannot be independently verified externally. The source also does not provide the specific scale, operating method, or reproducible details of these Agent clusters.

Karpathy believes that human review and comprehension capacity has already begun to fall behind AI progress. He gave two examples: the OpenAI-HF incident that happened a few months ago, on which people are still doing archaeological analysis; and mathematicians may still need a long time to study those 722 frontier mathematics manuscripts one by one.

The source only mentions both examples and does not provide the specific content of the OpenAI-HF incident, nor does it explain the source and details of the 722 mathematics manuscripts. Therefore, they are better seen as examples he uses to support the judgment that "review cannot keep up," rather than as complete independently verifiable facts.

It needs to be made clear that the numbers 5,000 people, 6 billion people, 1 billion to 2 billion people, 20 million people, and 0.00006% all come from Karpathy's personal rough estimates in the tweet, and are not independent statistics or third-party verification. He himself also expresses the 5,000 figure with "may," and its statistical basis and source are unclear.

The tiered numbers such as 6 billion, 1 billion to 2 billion, and 20 million are likewise vaguely bounded. As a secondhand relay source, Machine Heart does not provide the full first-party original text of the tweet it cites. Therefore, these numbers are better used to understand the tiering framework he proposes, and should not be treated as precise demographic conclusions.

Some netizens associated Karpathy's tweet with a judgment Sam Altman has repeatedly expressed in the past: even if ASI truly arrives, the real world may not completely change overnight. This association is not a statement made by Altman himself this time; it is a netizen's related interpretation.

If viewed along Karpathy's funnel, this association has its own logic: frontier capabilities have already advanced rapidly, but the vast majority of people still have not truly encountered these systems, and have not even formed stable usage habits. Between technology reaching frontier labs and it truly permeating the whole society, there is still a long distance.