The short answer
AI capability is not one story but four, sorted by depth of contact: most people have never used an LLM, many use free chatbots casually, a small professional group runs frontier agents daily, and a few thousand hold inside access at the labs. Tier counts are Andrej Karpathy's back-of-envelope estimates, checked against OpenAI's weekly-user figure.
Two people on the same street can hold opposite views of AI and both be telling the truth. One calls it a toy that still fails simple tests. The other watches it close a multi-day engineering task overnight. We have written before about these two groups talking past each other. The picture has sharpened since. It is no longer two crowds. It is a funnel.
The metaphor comes from a thread Andrej Karpathy posted in late 2025, which he suggested dusting off a year later. His point was not that one side is dumb. His point was that experience of AI capability now stacks into four tiers, each one a different size, each one seeing a different product.
TL;DR
- Karpathy's back-of-envelope funnel sorts the three quarters of the world who barely touch an LLM, a casual free-chat tier he put at one to two billion, a professional agent tier in the tens of millions, and an inside-the-labs tier he sized in the low thousands.
- His rough counts lag the verified 2026 data, and that is the interesting part: ChatGPT reported crossing 1 billion weekly active users (OpenAI, confirmed 08/09/2026), yet a viral February 2026 adoption chart still found about 84% of people had never used AI. The funnel moved up; its shape held.
- Three forces carve the funnel: how big your ambition is, which field you work in (machines grade themselves in code and math, not in prose), and which access tier you pay for. A business owner outside tech can still climb one rung with a single measured test.
Tier one: six billion people who have barely touched an LLM (Karpathy's estimate)
Karpathy's first back-of-envelope cut said roughly six billion humans, about three in four of the world, had almost no contact with a large language model (LLM, the software behind chatbots and coding agents). The estimate is his, not a measurement. Treat it as a shape, not a census.
The shape has held up against newer data. Global population sits near 8.1 billion as of February 2026 (UN figures), and the viral adoption chart that circulated on LinkedIn in the same month put the never-used-AI group at about 84% of the world. That count is stricter than Karpathy's sketch, but it points the same direction: the majority of humanity has no personal evidence about AI at all. Their entire opinion comes from headlines, memes, and what relatives forward in group chats.
For a business reader, this tier matters for a quiet reason. Most of your customers, and several of your suppliers, sit here. When they hear "AI will replace us" or "AI is a scam," both sentences land with equal force, because neither can be checked against experience.
Tier two: the casual free-chat crowd, judging by the demo
The next band is people who have opened a chatbot. Karpathy put this casual, occasional tier at one to two billion, again a back-of-envelope number. The 2026 adoption chart found about 16% of the world had tried AI somewhere along the line. Much of that trial happens on free tiers, and free tiers do not show you the frontier. They show you last year's model, at yesterday's price of nothing.
Karpathy's line, paraphrased here: a lot of people tried free ChatGPT once, and let that single meeting settle their view of AI for good. In this tier the technology reads like a slightly better search box. It writes a decent email, then confidently invents a fact, and the invention becomes the story you tell at dinner.
Plenty of professionals live here too, including executives. Their days are full of meetings and people, not repositories and pipelines, so their contact with AI stays indirect and abstract. They can explain the trend charts perfectly and still have never watched an agent run a real task start to finish.
Tier three: the professionals who see the wall go vertical
Below the casual band, the funnel narrows hard. This tier uses frontier-grade LLMs (the strongest models labs currently ship) professionally, mostly in math and software. Karpathy's rough count was around twenty million people. In this tier, projects that once took a team weeks get closed by an agent (software that plans and executes multi-step work on its own, not just a chat reply) from one well-written prompt (a task instruction). Building an app, porting one, translating a codebase, reversing a compiled binary.
He added a personal note, and it is worth keeping as his observation rather than a universal law: not a year ago, he said, he was still typing code character by character. The February 2026 chart agrees on the shape of this tier from the other side. It found only about 0.04% of the world using AI coding tools. A thin slice, but the slice where the capability curve goes vertical.
Karpathy's summary of the whole funnel, restated: the degree to which someone is amazed by AI tracks, almost perfectly, the degree to which they use AI to write software.
Tier four: a few thousand people inside the labs
At the point of the funnel, the crowd goes tiny. Karpathy estimated on the order of five thousand people hold internal access to frontier systems before release, the researchers and engineers inside a handful of labs. Everyone else is watching a preview of a preview.
What is reported to be happening inside that tier: swarms of cooperating agents running for weeks on software projects far too large for any single person to review. Machines hunting and patching security holes at computer speed. Claims of new mathematics and scientific leads. If any of that is real at scale, the bottleneck is no longer the model. It is human comprehension.
That is what makes this tier different in kind, not degree. Recent cyber incidents are still being reverse-engineered months later. AI-generated results are still awaiting human verification, quietly piling up faster than anyone can read them. A world where the frontier outruns our ability to audit it is not a faster version of today. It is a different problem, and almost nobody outside the labs has the vantage to see it coming. We would not over-read the estimates here; nobody outside can check them.
What carves the funnel: ambition, domain, access
Why four tiers and not a smooth slope? Three forces do the shaping, and each one maps to a decision you can actually make.
1. Ambition sets the width of the funnel's mouth. The depth of what you see scales with the size of the problems you bring. Someone with a folder full of hard, genuinely interesting problems pulls the model into its deepest behavior. Someone asking it to summarize a weather report never finds out what it can do. The funnel's consumption side is, first, a question of what you dare to ask.
2. Uneven capability picks the field. AI did not get smarter everywhere at once. Capability peaked in domains that are digital, checkable, and economically valuable: math and code above all. The reason is mechanical. Labs train models with reinforcement learning (learning by trial and reward) against verifiable rewards (a score a program can compute without a human judge). The test passes or it does not. The proof compiles or it does not. Practice meets an honest referee, millions of rounds a day, and skill cliffs upward. Prose has no referee, so meetings, emails, and management judgment stay in tier two even for people paying top prices.
3. Access decides which door you can even walk through. Free, paid, and internal. Your judgment of AI is downstream of the tier you hold, and tiers change hands for roughly the price of a dinner out each month. ChatGPT Plus now stands at around 50 million paying subscribers, according to reporting summarized by SearchEngineLand, so the paid door is far more crowded than Karpathy's original sketch assumed.
The strange part: all four tiers at once
Hold the whole picture for a second, because this is the part no one outside the funnel believes. The vast majority have barely interacted with LLMs. The majority of professionals see modest improvements at best. And a narrow slice is living through a capability curve bending vertical, on the same calendar week, in the same countries, often inside the same companies.
That is the oddity Karpathy flagged, restated in our own words: one technology, four realities, zero shared evidence. Debates about AI are mostly not about AI. They are people in different tiers comparing anecdotes and assuming they describe the same object.
Here is the 5ac take. The funnel is not a fairness problem to solve. It is a map to read before you buy anything, hire anyone, or sign off on a strategy. Position in the funnel predicts confidence better than intelligence does, and it predicts purchase decisions better than either. A vendor selling you "AI" is selling one tier while quoting another tier's results. The single question that deflates most pitches: which tier, which model, which task?
What a business owner outside tech can do this month
So: which tier are you actually in? If your AI experience is a free chat window used a few times a month, that is tier two, and it is a fine place to be. Just do not budget from it. You cannot forecast what a paid agent will do for your shop, clinic, or agency from what a free chatbot did for your curiosity.
Climbing one rung does not require learning to code. It requires one honest experiment on a task your business already pays for in hours. Pick something with a checkable outcome, so your own referee does the grading: reconciling a stack of supplier data, drafting a first-pass document against a rubric you already use, turning a month of customer messages into a tagged list. Run it on the exact paid product you would buy, with your real files, and count saved hours in your ledger. Not in a demo. Your tier matters because your evidence matters.
When the one-task test works, the natural next rung is orchestration. Our guides on controlling agent costs with token budgets and agentic CRM for customer service walk through running agents with approval gates and cost logs on real business operations.
Takeaway
The capability funnel reframes the loudest argument in tech. Nobody in it is stupid; people are just standing at different depths of the same shape, and depth comes from problems you carry, fields you work in, and doors you pay through. Find your tier this week. Run one real task, with a checkable result, on the tier you would actually buy. One rung up, the conversation changes entirely.
Frequently asked questions
What are the four tiers of the AI capability funnel?
Tier one is everyone with barely any contact with an LLM, the widest band of the funnel in Karpathy's back-of-envelope sketch. Tier two is casual users of free chat products. Tier three is professionals running frontier-grade agents on real math and engineering work. Tier four is the small community of people with internal access to unreleased systems inside the labs. Only the tier boundaries are estimates; the ordering is the point.
Why do AI gains show up first in coding and math?
Because machines can grade themselves there. A unit test passes or fails; a proof compiles or does not. Labs train models with reinforcement learning against those verifiable rewards, so capability climbs fast in digital, checkable, economically valuable fields. Prose and judgment work have no automatic referee, so they improve as a slow hum rather than a cliff.
How many people actually use AI agents at work?
Measured data is still thin. The viral February 2026 adoption chart put paying users of AI tools below 1% of the world, and AI coding-tool users far below that. ChatGPT reported crossing one billion weekly active users in September 2026, mostly on free tiers. Wide contact and deep professional use are different things, and the funnel is what separates them.
What should a non-tech business owner do with this funnel?
Locate your tier honestly, then run one real task on the paid tier you would actually buy, with a checkable outcome and your own files. If saved hours land in your ledger, you have moved up on evidence. If they do not, you have learned that too, which beats renting someone else's opinion.
Sources
- Andrej Karpathy's thread on the widening gap in understanding of AI capability (posted around October 2025, viewed October 2026) — source of the four-tier framing and the tier-size estimates; all figures attributed to it are his back-of-envelope numbers, not measurements.
- SearchEngineLand, ChatGPT user-growth reporting (accessed October 12, 2026) — 900 million weekly active users in February 2026, the ~50 million paying-subscriber figure, and the roughly 18-month path from millions of users to a billion-scale audience.
- OpenAI's public confirmation of the ChatGPT weekly-user milestone (reported from a mid-August 2026 snapshot, confirmed 08/09/2026; summarized by CryptoBriefing) — the headline adoption number in the stat block.
- United Nations World Population Prospects (accessed October 12, 2026) — global population near 8.1 billion as of February 2026, the denominator behind every percentage in this piece.
- Viral AI adoption chart, February 2026 (circulated on LinkedIn; a four-band split of the world's AI contact, from never-used to coding tools) — treated here as a widely shared estimate, consistent in shape with Karpathy's funnel.
- 5ac.vn editorial: The AI Capability Gap — our earlier piece on two user groups talking past each other; the funnel in this article is the sharpened version of that picture.
Move one rung up the funnel this month. Tell us the slowest task in your operation, and we will run it as a measured pilot on a real agentic stack, with approval gates, audit trails and a cost log, so your ledger decides what AI can do for your business.