AI researcher Andrej Karpathy argues that public understanding of AI is splitting into sharply different levels because access to the most capable systems is concentrated among a small group. Most people see a search or writing assistant, while some developers use agents for projects that previously took weeks.
Karpathy estimates that roughly six billion people have barely used a large language model, another one to two billion are casual users and about 20 million professionals use frontier models seriously for coding or mathematics. He places access to internal systems capable of coordinating large agent swarms at around 5,000 people worldwide.
He says those internal systems work over long periods on software, cybersecurity, science and mathematics, creating outputs faster than human reviewers can fully understand. Coding and math are advancing first because they are digital, economically valuable and offer results that training systems can verify automatically.
These numbers and capability descriptions are Karpathy’s own estimates, not an independently measured census or benchmark. They also combine very different products and levels of access. The useful point is narrower: judging advanced agents only by a free chatbot may understate what selected laboratories can test, while judging society by laboratory demonstrations may overstate what ordinary users can reliably deploy today.