Karpathy Puts Internal Frontier AI Access at About 5,000

Andrej Karpathy wrote on X on Sunday, 11 October, that the gap in shared understanding of LLM capability is widening, and that it now looks like “a sharp funnel” rather than two groups talking past each other. By his own napkin math, about 6 billion people have barely touched LLMs, while around 5,000 people have access to frontier-grade systems inside the labs.

Karpathy was re-sharing a post he wrote in April. He said the four tiers of the funnel are shaped by the size of the problems people bring, by how unevenly capability is spread across domains, and by access: “Free-tier, paid-tier, internal.”

The Four Tiers, in His Numbers

Karpathy labelled his figures as rough estimates. “Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all,” he wrote.

The next tier is casual use. “Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products,” he wrote, a group that “sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc.”

The third tier is professional use in technical work. “Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt,” Karpathy wrote.

The last tier is inside the labs: “the ~5,000 people (~0.00006%) with access to frontier-grade systems internally.” The chart below sets out his four estimates on a log scale.

What Was Posted on X

Andrej Karpathy’s post of 11 October 2026, re-sharing his April post on the gap in how people experience LLMs. It is the primary source for this piece: the four tiers (about 6 billion, 1 to 2 billion, about 20 million and about 5,000 people) and the three drivers are his own napkin estimates and reasoning, not measured data.

What the Wide Top of the Funnel Sees

Karpathy did not dismiss the casual tier. “I have non-tech friends who (reasonably, imo) say they have not much use for it at all,” he wrote.

He put many professionals in the same tier. Executives and other functions, he wrote, “spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.”

Math, Code and the 5,000

For the professional tier he described a fast change in his own work: “less than 1 year ago, I was writing code manually by hand,” he wrote.

Of the internal tier, he wrote: “The external world has seen the preview.” He described what it looks like as “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.”

He added a caution about people keeping up. “Meanwhile, human review and comprehension are starting to fall behind,” Karpathy wrote. As examples he cited people still working on the “archeology” of “the OpenAI-HF incident from many months ago,” and wrote that “Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.”

Why He Says the Funnel Forms

Karpathy gave three drivers. The first is the size of the problem: “The impact scales with ambition, problem size, and horizon.” A question with a paragraph answer, he wrote, “barely stresses the system.”

The second is what he calls jaggedness. “Capability peaks in domains that are digital, verifiable and economically valuable,” he wrote, because LLM capability “emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential.”

The third is access, in three tiers: “Free-tier, paid-tier, internal.”

What He Wrote in April

The April post Karpathy re-shared described two groups rather than four. It said many people tried “the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much,” while those using frontier agentic models professionally in technical domains had seen improvements he called “nothing short of staggering.”

That post gave two reasons for the uneven progress: the technical domains “offer explicit reward functions that are verifiable,” and “they are a lot more valuable in b2b settings.” Its summary was that “the people in these two groups are speaking past each other.”

The Structural Read

Karpathy’s post makes three moves. The first is scale: by his napkin math, about 75% of people have barely touched LLMs and about 0.2% use frontier models professionally in math and code.

As we read it, the second move is to explain the gap by where capability peaks. He wrote that capability emerges from reinforcement learning on verifiable rewards, which favours digital, verifiable and economically valuable domains.

The third move is about access. He said the internal tier sees agent swarms working for weeks on large software projects, and that human review and comprehension are starting to fall behind.

Three Implications

PRODUCT TEAMS Karpathy wrote that capability peaks in domains that are digital, verifiable and economically valuable, so in his account the uplift users see depends on the domain a product serves.

EXECUTIVES He placed many executives in the casual tier, writing that their understanding is “second-hand and a bit abstract” because much of their time goes on talking to other people.

REVIEWERS AND RESEARCHERS He wrote that human review and comprehension are starting to fall behind, and that mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Product Overhang Doctrine.

The analysis puts it this way: “Model capability is exponential; product capability is approximately linear.”

As we read it, Karpathy’s funnel describes that gap from the user’s side: what each tier sees depends on the problems, domains and access it brings, not on the frontier alone.

What Is Not Established

Karpathy calls his figures napkin math and gives no method or data behind the 6 billion, 1 to 2 billion, 20 million or 5,000 estimates. He does not say which organisations the internal tier covers, what counts as frontier-grade, or which system the 722 manuscripts came from, and he does not explain the OpenAI-HF incident in the post. The chart’s bar for the 1 to 2 billion tier is drawn at 1.5 billion, which is our choice of midpoint.

The Bottom Line

Karpathy’s post describes a gap in exposure, not a measured survey. By his estimates, most people have barely used LLMs, about 20 million see agents finish weeks of technical work, and about 5,000 inside the labs see more. He said the gap is shaped by problem size, by capability peaking in verifiable and valuable domains, and by access tier.

His own summary of the moment: “So this is the weirdness of the moment.” He added that “a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.”