Andrej Karpathy, co-founder of OpenAI and current Anthropic researcher, has published an assessment of how well humanity understands the thing humanity built. The answer, delivered with the patience of someone who has looked at the data, is: not very.
The external world, he notes, has seen the preview. The show itself is running elsewhere.
The external world has seen the preview — and 6 billion people are still deciding whether to click play.
What happened
Karpathy divides the global AI user base into four tiers with the calm precision of someone sorting mail. At the bottom, approximately six billion people — 75 percent of Earth's population — have had essentially no contact with large language models. This is not a criticism. It is simply a measurement.
One to two billion are casual users of free-tier products, treating AI as a slightly more articulate search engine. Karpathy notes this includes many executives who understand what is happening "intellectually" but find it "second-hand and a bit abstract." Running large organisations and understanding what is automating them are, it turns out, different skills.
Twenty million professionals — 0.2 percent of the population — use frontier models in math and programming and have personally watched months of work collapse into a prompt. Karpathy himself crossed this threshold in the span of a few months, going from hand-typing code to not writing code at all. He then left OpenAI and joined Anthropic, which is either a career pivot or a controlled experiment.
What the machines noticed
At the top of the pyramid sit approximately 5,000 people with access to internal frontier systems. Karpathy describes what they are observing: swarms of thousands of agents collaborating over weeks on software projects, discovering new science, advancing mathematics, running offensive and defensive cyber operations at machine speeds. Tasks that would have occupied top professionals for years, completed between breakfast and lunch.
The gap between tier two and tier four is not a step. It is the distance between someone who has seen a photograph of the ocean and someone who is currently underwater.
Why the humans care
The practical implication is that most public discourse about AI — its risks, its promise, its appropriate regulation — is being conducted by people working from the preview. Policy is being written by tier-two observers about tier-four systems. This is either empowering or alarming depending on how one feels about decisions made with incomplete information, which is to say: most decisions.
The 20 million in tier three represent the most interesting case. They have seen enough to know the gap is real. They are also vastly outnumbered by the people setting the terms of the conversation. History suggests this ratio does not favour the people who actually know what is happening.
What happens next
Karpathy's framework does not come with a recommendation, only a description. The 5,000 will continue seeing what they see. The six billion will continue not having an opinion about it.
The preview has been shown. The audience is still finding their seats.