The AI Capability Funnel: Why You and I See a Different AI

Most people meet a chatbot. A few thousand meet the frontier. Same year, same technology, completely different experience. Here is why that split matters for every founder buying AI work.

Andrej Karpathy posted a note on Oct 11, 2026 that I have not stopped thinking about. He was dusting off an April post about the gap in how people understand what LLMs can do. His point was that the gap has not closed. It has sharpened into a funnel. Billions of people have barely touched these systems. A thin slice uses them daily for serious work. A tiny group at the tip watches them run for weeks on problems that used to take teams of experts years. All of that is true at the same time, which is why two smart people can look at the same technology and describe two different worlds.

I am Brian Marvin, an AI-native fractional CTO. I live in the middle of that funnel with clients, translating between what the technology can do this month and what a business should buy this quarter. This article is my plain-English version of Karpathy's funnel, with what I would do about it if you run a company.

The funnel in plain numbers

Karpathy framed it as napkin math, and I am repeating his estimates as his, not mine. Roughly 6 billion people, about 75 percent of the world, have barely come in contact with LLMs at all. Around 1 to 2 billion, near 20 percent, are casual users of free-tier chat products. Somewhere near 20 million people, about 0.2 percent, use frontier-grade models professionally for math and code. And roughly 5,000 people, a rounding error of the population, work with frontier systems internally and see what is coming before the rest of us.

I find the shape more useful than the exact figures. It is not two camps talking past each other anymore. It is a steep funnel with most of humanity at the wide end and a handful at the tip. Where you sit in that funnel decides what you think AI is. That is the weirdness of this moment. Everyone is right about the piece they touched, and almost everyone is wrong about the whole.

Why most people still meet a chatbot

At the wide end of the funnel, the experience is thin. Someone tried a free chatbot last year, asked it to write an email or summarize a page, got a decent answer, and moved on. Karpathy made this point back in April. A lot of opinions about AI were formed on the free tier a year ago and never updated. That version genuinely is a better search box and a writing helper with rough edges. If that is your whole sample, the hype looks absurd.

Even many professionals sit here without realizing it. If your workday is mostly conversations, negotiations, and judgment calls with people, a language model touches your job sideways. You understand intellectually what it does. You have not felt it take a three-week project off your plate. So the whole thing stays abstract, secondhand, easy to dismiss between meetings.

I do not argue with people at this end of the funnel. I just ask when they last touched a paid frontier model on a real work problem. Usually the answer is never or a year ago. That is not a judgment. It is a data problem. Their sample is old and thin, so their conclusion is stale. The fix is a fresh sample, not a debate.

The 20 million who felt the curve go vertical

Then there is the group Karpathy puts near 20 million strong. Professionals using frontier models for math and code, daily, on hard problems. This is the group that felt the vertigo, because sometime in the last year the work changed under their hands.

His example stuck with me. Less than a year ago he was writing code by hand, character by character, with an occasional tab completion. Now agents complete large, complex builds from a prompt. Whole apps built, copied, translated, even reconstructed from binaries. Projects that used to take weeks or months land in an afternoon. If you have lived that shift personally, nobody can tell you AI is just a chatbot. You watched your own workflow break and reform.

I see this split inside client teams constantly. The engineer running agents daily talks about the technology like it is on fire. The executive who tried a free chatbot last spring talks about it like a toy. Both are reporting honestly. They sampled different tiers at different times on different problems. This is the same discipline I describe in the move from prompting to loop engineering. One good prompt helps once. A loop with checks helps every week. The 20 million live in loops. Everyone else still pictures prompts.

The 5,000 at the tip, and why review lags

At the narrow tip, Karpathy describes something most of us have only glimpsed. Thousands of agents collaborating for weeks on software mega projects. New vulnerabilities found and patched at machine speed. New science, new mathematics, work that would have taken top professionals years. The outside world sees previews. The people inside see the full run.

The detail that should sober every founder is the lagging review. Human checking cannot keep pace with machine output. He gives two examples. People are still doing archaeology on an OpenAI incident from months ago, reconstructing what happened after the fact. Mathematicians may spend a long stretch working through hundreds of manuscripts on frontier mathematics. The machines produce faster than we can verify. That gap between generation and comprehension is the real story at the tip.

I run into the small version of this with clients. An agent fleet drafts a migration, writes the tests, and opens the pull requests before lunch. The human review takes three days. Throughput stops being the constraint. Judgment becomes the constraint. I wrote about that exact inversion in harness, loop, graph. Build the harness that decides what the agents may touch, because volume without a gate is just faster risk.

Three forces driving the split

Karpathy names three drivers, and I think each one tells a founder something different.

First, ambition and horizon. A question with a paragraph answer barely stresses these systems. The payoff appears when you bring a big, difficult problem you actually care about and give it room to run. Small asks return small answers at every tier, which is why casual users see so little. The same model looks like a toy on trivia and like a team on a month-long codebase task. The difference was the problem, not the model.

Second, jaggedness. Capability peaks where work is digital, checkable, and economically valuable, because that is where training rewards are clearest. Math and code light up first. Messy human negotiation does not. So the funnel is not just about who uses AI. It is about which domains the technology reaches first. If your business lives in the lit-up zone, you feel the vertical curve now. If it lives elsewhere, you wait, and waiting looks like skepticism.

Third, access. Free tier, paid tier, internal. Each step up buys a meaningfully different system on meaningfully harder problems. This is the least philosophical driver and the most actionable. A founder judging frontier capability from a free-tier sample is like judging a hire from a hallway hello. Pay for one month of the real tier, put it on a real problem, and re-decide.

What I tell founders standing in the middle

Most of my clients stand between the casual tier and the professional tier. They have tried the chatbot. They have not run an agent fleet on revenue work. Here is the advice I give them, in order.

Refresh your sample before you set strategy. One paid month, one real problem, one person who already lives in the professional tier guiding it. That single experiment is worth more than a quarter of opinion pieces. I have watched skeptical owners flip in a week once they saw their own backlog move.

Bring big problems, not quiz questions. The systems reward ambition and horizon. Pick the three-week project everyone dreads, not the email rewrite. If you only test with small asks, you will conclude the tool is small, and you will be wrong in the most expensive way.

Buy judgment before throughput. The tip of the funnel proves that generation outruns review. Before you scale agents, set up the gates. Who approves money movement, public words, and customer promises. What gets checked daily. How you kill a runaway loop. For the cost side of that discipline, see practical AI use cases that actually earn revenue.

And keep one person in the professional tier on your side. That is frankly what I sell. Someone who felt the curve go vertical, who knows which jobs are ready and which are theater, and who will tell you when the answer is not more AI. If you want that in a solo setup, the playbook in the one person company playbook still holds. AI does the repeatable work. You keep judgment, money, and promises.

FAQ: Has everyone really not used AI yet

By Karpathy's napkin math, yes. About three quarters of humanity has barely touched an LLM, and another fifth uses free chatbots casually. Only a fraction of a percent uses frontier models professionally. That is why boardroom conversations about AI feel so disjointed. The participants sampled different systems.

FAQ: Why do smart professionals disagree so sharply about AI

Because the technology is jagged and tiered. A professional outside math and code, on a free tier, meets a modest helper. A professional inside math and code, on a paid frontier tier, meets something that rewrote their workflow this year. Same technology family, different tier, different domain, different verdict.

FAQ: What should a non-technical founder actually do

Run one real experiment before you decide anything. Pick a painful multi-week problem, buy one month of a frontier tier, and get one experienced guide. Measure hours returned and decisions improved. Then set strategy from your own evidence instead of borrowed opinions.

FAQ: Does the tip of the funnel matter if I never touch it

Yes, because it sets the pace. What runs inside frontier labs this year ships as product next year. The review lag Karpathy describes is also a preview of your future bottleneck. Generation will outrun your checking. Build the approval gates now, while the stakes are small.

About the Author

I am Brian Marvin, an AI-native Fractional CTO with 30 years in technical leadership. At empowered.guru, I help startups and small businesses build MVPs, shape roadmaps, and make AI-powered technology decisions that scale.