Karpathy: Only 5,000 People Have Seen the True Potential of AI

Andrej Karpathy, founding member of OpenAI and recently joined Anthropic, has published an analysis on X that should make any technology leader think. His thesis is blunt: most of the world has no idea what AI can actually do, and that gap keeps growing.

According to his calculations, approximately 6 billion people have barely used a language model; between 1 and 2 billion use chatbots casually; about 20 million use frontier models professionally in mathematics and code; and only around 5,000 people have access to the most advanced systems inside the labs themselves. A distribution that draws an extremely narrow funnel at the top.

Aaron Levie, CEO of Box, responded to Karpathy on the same social network and agreed with the diagnosis: "This is a good way to understand how early we really are in AI diffusion." Levie argues that most people still think of AI as a chatbot or a next-generation search engine, without exploring its more advanced applications.

Agent swarms: the next frontier

Levie bets on what he calls "agent swarms": systems that divide complex work among multiple agents operating in parallel. In his own words, they "represent possibly the definitive form factor of what agents will do within our workflows." However, he acknowledges that "less than a fraction of a fraction of the population uses AI this way."

The obstacles he identifies are several: models are not equally capable in all types of work, running them at scale is expensive, and tools for managing multiple agents are still maturing. Add to that the fact that most people do not even have access to the most powerful models. And beyond technology, companies must connect their data and resolve security and regulatory compliance issues. Levie anticipates "a huge amount of change management" before most sectors can take advantage of these agents.

Where AI really shines

Karpathy provides a technical key worth understanding: "Capability peaks in domains that are digital, verifiable, and economically valuable." Programming and mathematics fit perfectly because results can be tested and improvements measured. Writing or advising are harder to evaluate, and there is more money in improving models for technical work.

He also warns that we do not give these systems enough work: "A question with a one-paragraph answer barely stresses the system." Real gains appear when an agent can spend hours working on a problem, testing changes, and retrying. As an example, his AutoResearch project ran 50 experiments overnight on a single GPU, found a better learning rate, and confirmed the change in Git.

What this means for your small business

The gap described by Karpathy and Levie is not just a curious fact: it is a warning and an opportunity. If your company still uses AI only as a chatbot, you are leaving on the table capabilities that already exist and are especially useful in technical and verifiable environments.

At ForgeNEX we see daily how many small businesses stay on the surface. To start closing that gap, we propose a realistic roadmap:

• Identify digital and verifiable processes where AI can deliver measurable results: code generation, data validation, document automation.
• Start with bounded agents, not complex swarms. A single well-directed agent can free up hours of repetitive work.
• Connect your data and define security policies before scaling. Without traceability, agents become a risk, as we already analyzed in AI Agents Without Traceability: The Identity Hole in Your Small Business.
• Measure the return of each experiment. AI that is not evaluated is not improved.

Levie places the opportunity in "the applied layer of AI," that is, in companies that build specific solutions for each sector on top of frontier models. For a small business, that means it does not need to develop its own model, but rather integrate these capabilities into its processes with the help of technology partners.

The final message is clear: AI is advancing at a pace that most people do not perceive because they only see the tip of the iceberg. The question is not whether your company will adopt advanced agents, but when it will start preparing. And the sooner it does, the less ground it will have to recover.

Source: The New Stack. Analysis and adaptation: ForgeNEX.