Anthropic's Andrej Karpathy says "the external world has seen the preview" of what AI can actually do

Andrej Karpathy, a co-founder of OpenAI who now works at Anthropic, has published his assessment of how unevenly AI capability is distributed, arguing that the public understanding of large language models badly lags what frontier systems can already do. He frames the user base in four tiers and says the spread between them is widening fast.
The first tier, by his estimate, is about six billion people, roughly 75 percent of the world's population, who have barely come into contact with LLMs at all. The second covers one to two billion casual users of free-tier ChatGPT-like products, who treat them as something like a better search engine or a writing aid. Karpathy notes that many professionals outside math and code sit in this tier too. Executives and similar roles spend most of their time talking to other people, so while they grasp what is happening intellectually, the understanding stays second-hand and a bit abstract.
The third tier is the one he puts the most weight on. He estimates about 20 million people, around 0.2 percent of the world's population, use frontier-grade LLMs professionally in math and code. For this group, he writes, large, complex projects that used to take weeks or months can now be finished by agents from a prompt. His examples include building apps, copying apps, translating apps, and decompiling apps from binaries.
Karpathy says that shift arrived quickly. In October 2024 he described AI agents as mostly marketing that barely worked. Only a few months later he stopped writing code by hand entirely, moving from typing memorized commands character by character in an editor to letting agents handle the work. He then joined Anthropic despite his OpenAI history.
The fourth tier is the one he calls the real gap: roughly 5,000 people worldwide, about 0.00006 percent of the population, with access to internal frontier-grade systems. His description of what they see is vivid: swarms of thousands of agents collaborating over weeks on software mega projects, minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. These are things he says would have taken top professionals in the industry years of work. His summary line is that the external world has seen the preview.
A related problem he raises is comprehension, not just access. Human review and understanding are starting to fall behind, he writes. As evidence he points to people still doing what he calls archeology on the OpenAI-HF incident from months earlier, and he notes that mathematicians may be poring over 722 manuscripts on frontier mathematics for a long while. The implication is that output is arriving faster than expert communities can verify and absorb it, which matters because verification is the bottleneck that keeps results trustworthy.
Karpathy calls the combination the weirdness of the moment. The general public has mostly not interacted with these systems, and when it has, they look like a derpy chatbot. Most professionals see only a modest uplift. A small sliver of professionals experience the vertigo of a curve going vertical. All of it is happening at once.
He is not alone in making this kind of claim. OpenAI CEO Sam Altman said something similar in early September at the G20 Innovation Ministerial, stating that a current AI coding tool could achieve in 17 minutes what a startup had previously built over three months, and adding that the world has not caught up with what this is going to mean.
Karpathy offers a mechanism for why capability clusters where it does. He says LLM performance is currently peaking in domains that are digital, verifiable and economically valuable, because it emerges from reinforcement learning on verifiable rewards, and companies steer training resources toward the domains with the highest revenue potential. Today that means coding and math. This is a useful frame for anyone outside those fields: the gap he describes is not evenly distributed optimism about AI in general, but a specific consequence of where verifiable feedback signals exist and where money is. Tasks with clean, checkable outcomes, software and mathematics being the clearest examples, are the ones that respond to this training approach. Tasks without that property, such as many kinds of judgment, coordination and physical work, do not benefit in the same way yet.
In practice, a developer reading this should take away that the perceived capability of AI depends heavily on which tier you occupy. Someone using a free chatbot at work sees a modest writing and search tool. Someone running agents against a frontier model in a codebase sees multi-week projects compress. The distance between those two experiences is the story. A likely trade-off is that as agent output scales, the cost of review and verification scales with it, and that cost is borne by humans who are still the slower party. The 722 manuscripts reference and the archeology remark both point at that asymmetry.
The specific numbers in the post are estimates from Karpathy rather than audited figures, so treat the tier boundaries as illustrative of relative scale rather than precise counts. The general shape, a very small population with disproportionate access and a very large population with almost none, is the claim being made.
Why it matters: Product teams and developers should treat observed AI capability as a function of access tier, not a global constant, since the tools most people can try differ sharply from internal frontier systems. If that gap is real, roadmaps built on public chatbot experience may underestimate what competitors with frontier access can ship, while review and verification capacity becomes a binding constraint. That reading is inference from the stated estimates, not a claim Karpathy makes.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Andrej Karpathy, co-founder of OpenAI and now at Anthropic, sees a widening knowledge gap in AI. About 75 percent of the world's population has barely touched LLMs. Roughly 5,000 people, meanwhile, have access to internal frontier systems whose agent swarms do work that would take top experts years. The public, at best, sees a weird chatbot. The article Anthropic's Andrej Karpathy says "the external world has seen the preview" of what AI can actually do appeared first on The Decoder .