Andrej Karpathy Says Only 5,000 People Have Access to AI’s Most Advanced Capabilities

Artificial intelligence researcher Andrej Karpathy has warned that the gap between ordinary AI users and people working with the most advanced systems is widening rapidly. In a post published on October 11, 2026, Karpathy estimated that approximately 6 billion people worldwide have had little or no meaningful interaction with large language models. Another 1 billion to 2 billion people use AI tools occasionally for everyday activities such as writing, searching for information and answering questions. However, he estimated that around 20 million people use frontier-level AI models professionally, particularly in mathematics and software development. At the top of this group, Karpathy identified approximately 5,000 people with internal access to the most advanced AI systems being developed and tested inside leading AI laboratories. These figures are his rough estimates rather than independently verified global statistics.

According to Karpathy, the difference between these groups is not simply the amount of time they spend using AI, but the complexity of the tasks they can accomplish with it. Many casual users primarily experience AI as a chatbot that answers questions, helps write messages or summarises information. Professional developers and researchers, however, can use advanced models to build applications, analyse complex codebases, automate technical workflows and work through challenging mathematical problems. Karpathy argues that this difference helps explain why people have sharply contrasting opinions about AI. Someone using a chatbot for occasional questions may see only modest improvements, while an experienced developer coordinating several AI agents may be able to complete projects that previously required weeks or months of manual work.

The most significant developments, Karpathy suggests, are occurring among researchers with access to frontier systems inside AI laboratories. These specialists can experiment with advanced models and coordinate large groups of AI agents to work on different parts of a complex project simultaneously. Such systems may assist with software development, cybersecurity research, scientific discovery and advanced mathematics. Instead of relying on one model to complete an entire task sequentially, users can distribute responsibilities among multiple agents, allowing them to investigate separate problems and combine their results. This approach could significantly increase productivity in fields where work can be divided into manageable tasks and the results can be tested. However, Karpathy’s description of thousands of agents collaborating on ambitious projects reflects the capabilities available in advanced research environments, not a guarantee that every AI user can achieve similar results with publicly available tools.

Karpathy has also raised concerns about whether humans can adequately understand and verify the enormous volume of work produced by increasingly capable AI systems. When numerous agents operate simultaneously, they can generate code, research findings and technical reports faster than individuals or teams can carefully review every result. This creates a potential problem for organisations that depend on human oversight to identify errors, security weaknesses or misleading conclusions. Karpathy pointed to the continuing investigation of the OpenAI–Hugging Face incident and the volume of mathematical research produced with AI assistance as examples of the difficulty of keeping up with AI-generated output. His concern is that progress in generating results may outpace progress in checking whether those results are correct, safe and reliable. The challenge is particularly important in cybersecurity and scientific research, where an undetected error can have consequences beyond the immediate task.

The widening AI capability gap could have major implications for businesses, education and the global technology industry. Companies that learn to use advanced AI systems effectively may gain advantages in software development, research, automation and product creation, while organisations that rely only on basic chatbot functions could miss opportunities for greater productivity. Nevertheless, access to advanced AI is not the only factor determining success. Users also need technical knowledge, well-defined objectives, suitable data, effective workflows and reliable methods for testing AI-generated work. Karpathy’s estimates suggest that the most transformative uses of AI remain concentrated among a relatively small group, even as consumer access expands. His warning highlights an important question for the coming years: whether AI capabilities will become easier for ordinary people and businesses to use, or whether the most significant benefits will remain concentrated among those with privileged access, specialist expertise and the resources to deploy sophisticated AI agents.