Andrej Karpathy of Anthropic says the outside world has already seen a preview of what AI can truly do.

Meta description: OpenAI co-founder Andrej Karpathy maps four user levels, from 75 percent barely using LLMs to 5,000 insiders seeing frontier AI agents and capability gaps.

Tags: AI, OpenAI, LLMs, Karpathy, Anthropic

Featured image: abstract technology-style illustration of four concentric user circles and small agent clusters, no people, logos, or text.

Andrej Karpathy, co-founder of OpenAI, describes an AI capability gap between public chatbot use and internal frontier systems. He says about 75 percent of the global population, roughly six billion people, have barely come into contact with large language models, while a much smaller group sees systems already working on tasks that would take top experts years.

Background: Four layers of exposure to large language models

Karpathy groups LLM users into four levels. The largest group is about six billion people, roughly 75 percent of the world. They have barely encountered LLMs at all. The next group is one to two billion casual users of free-tier ChatGPT-like products. In his telling, they treat these systems like a better Google search, a writing aid, or other everyday uses.

The gap is not only among non-technical people. Many professionals outside math and code sit in this tier. Karpathy notes that executives and people in similar roles spend most of their time talking to other people, so they may understand the technology intellectually but experience it as second-hand and abstract.

Development: From marketing hype to agent-written code

The third tier is about 20 million people, or 0.2 percent of the world. These are professional users of frontier-grade LLMs in math and code. They see large, complex projects that used to take weeks or months completed by agents with a prompt. Karpathy lists building apps, copying apps, translating apps, and decompiling apps from binaries as examples.

The change was rapid. In October 2024, Karpathy called AI agents mostly marketing that barely worked. A few months later, he said he stopped writing code by hand entirely. The shift moved him from typing memorized computer code commands into a code editor character by character to letting agents handle the work. He then joined Anthropic, despite being an OpenAI co-founder.

Latest: Internal systems reveal a vertical curve

At the top is a very small group. Karpathy estimates about 5,000 people, roughly 0.00006 percent of the global population, have access to internal frontier-grade systems. The external world, he writes, has seen the preview. This group sees 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, and advancing the frontier of mathematics. These are tasks that would have taken top professionals years.

Human review and comprehension are starting to fall behind. Karpathy points to people still doing archaeology on the OpenAI-HF incident from many months ago, and he notes that mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.

He calls this the weirdness of the moment. The public has mostly not interacted with these systems, and when it has, the experience can look like a derpy chatbot. Most professionals see only a modest uplift. A small sliver is experiencing the vertigo of the curve going vertical. All of this is happening at the same time.

Context: Coding, math, and the next review bottleneck

Karpathy's view overlaps with OpenAI CEO Sam Altman. In early September at the G20 Innovation Ministerial, Altman made a similar claim: a current AI coding tool could achieve in 17 minutes what a startup had previously built over three months. Altman said the world has not caught up with what this is going to mean.

Karpathy says LLM capability is currently peaking in domains that are digital, verifiable, and economically valuable. He ties this to reinforcement learning on verifiable rewards, a training approach that rewards measurable outcomes, and to companies directing training resources toward domains with the highest revenue potential. Right now, that means coding and math. What to watch next is whether human review, comprehension, and public understanding keep pace with systems already operating at machine speed.