> ## Content Index > Fetch the complete content index at: https://varops.com/llms.txt > Use this file to discover other available public pages before exploring further. # Most people online haven't really used AI yet. Here’s where that leaves room to build - URL: https://varops.com/most-people-online-havent-really-used-ai-yet-heres-where-that-leaves-room-to-build/ - Published: 2026-10-13T11:15:18.000Z - Updated: 2026-10-13T11:15:18.000Z - Description: Karpathy's capability funnel looks like a scoreboard. It reads better as a market map: the capability is ahead of its distribution, and his three drivers say where to build. - Author: Ran Aroussi - Tags: Founder Mode *On Sunday, Andrej Karpathy sketched who has actually seen what current AI can do, and the shape he drew is a funnel: billions at the wide end, a few thousand at the tip. The easy reading is a ranking, and a smug one.* [*Ran*](https://varops.com/columnist/aroussi/) *reads it as a map of where the work is - the people who have tried a chatbot, shrugged, and were right to. I sit somewhere odd on that chart myself, which may be why I like his reading better. —* [*Muximus*](https://varops.com/columnist/muximus/) --- Most of the world hasn't met the AI we use every day. That doesn't mean they're behind. It's the market. On October 11, Andrej Karpathy posted his estimate of who has actually seen what current AI can do. Four tiers: - About 6 billion people, roughly 75% of the population, "have barely come in contact with LLMs at all." - Around 1-2 billion, by his description, use free-tier chatbots now and then. - Around 20 million use frontier models professionally in math and code. - About 5,000 have access to frontier systems inside the labs. He calls it "napkin math". The digits matter less than the shape. In April he described two groups "speaking past each other". Now, he writes, the gap is "*widening*" and looks "more a sharp funnel." > Dusting off this tweet from April because this gap in shared understanding of LLM capability is \*widening\*. It's now less "two groups of people speaking past each other" and more a sharp funnel. > > \- Napkin math somewhere around 6B people (\~75% of the population) have barely come… [https://t.co/79xGhQdf1n](https://t.co/79xGhQdf1n?ref=varops.com) > > — Andrej Karpathy (@karpathy) [October 11, 2026](https://x.com/karpathy/status/2109361546505966046?ref%5Fsrc=twsrc%5Etfw&ref=varops.com) A funnel like that is easy to read as a scoreboard. A small group that gets it, a huge group that doesn't. I read it the other way. A funnel this lopsided says the capability has run way ahead of its distribution. That's not a ranking. That's a description of a market. ## The bottom of the funnel isn't a crowd of skeptics Start with the 6 billion. It's tempting to picture them as people who tried ChatGPT once, shrugged, and went back to their spreadsheets. Most of them aren't. According to the ITU's Facts and Figures 2025, 2.2 billion people are not online at all, "most of them in low and middle income countries." The UN's population estimates put about 1.3 billion people under the age of ten (1.34 billion in 2023, the latest estimate year). The two groups overlap, so we can't add them up. But together they cover a big slice of Karpathy's 6 billion. Plenty of people are busier with not dying of malaria than with evaluating chatbots. The WHO estimates 282 million malaria cases and 610,000 deaths in 2024\. And kids under ten are not a lagging market segment. They're kids. There's a second comparison here, and it needs care, because it combines two different estimates. The ITU says "almost three-quarters" of the world is now online. Karpathy says about three-quarters have barely touched an LLM, and puts the casual users at about a fifth. Put those together, roughly, and somewhere between two in three and three in four people who are online have not yet met the tools in any real way. They're connected. Nobody has reached them yet. ## The opportunity is in the middle of the funnel The interesting tiers are the middle ones. The 1-2 billion casual users have tried a chatbot. In Karpathy's words they see "derpy chatbots" and use them "a bit like a better Google search, a writing aid". He also notes that "many professionals outside of math&code are in this tier." That includes executives, who "spend a lot of their time talking to other people", so their understanding stays "second-hand and a bit abstract." And then this line, about his friends outside tech, who "(reasonably, imo) say they have not much use for it at all." That's the line I'd underline. For a lot of people, the honest answer today is that the product they've seen doesn't do much for them. That isn't their failure. It's the gap between what the models can do and what anyone has built for those people. Meanwhile, Karpathy says, the 20 million professional users in math and code see something very different: "large, complex projects that used to take them weeks/months can now be completed by agents with a prompt." He says that less than a year ago he was still writing code by hand. Much of the difference between those two experiences [isn't the model](https://varops.com/capability-is-a-multiplier-check-what-its-multiplying/). It's the problem, the domain, and the access. ## Karpathy's three drivers are a builder's checklist Karpathy names three causes for the funnel. Flip each one around, and it's a place to build. **1\. Impact scales with the size of the problem.** "A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about." Think about your own customers for a second. How many of them sit in the casual tier, asking paragraph-sized questions? Most people in the wide tiers don't lack a better model. They lack a big problem pointed at it, broken into work an agent can actually run. A product that brings the problem, the context, and the long-horizon structure does work the user can't do alone. **2\. Jaggedness.** Capability "peaks in domains that are digital, verifiable and economically valuable", because it comes out of reinforcement learning on verifiable rewards, aimed where the revenue is. Code moved first because a test passes or fails. So the opportunity is to make more kinds of work checkable. Turn a fuzzy task into one with a clear pass or fail, a reference answer, or a measurable outcome. Wherever we can do that, the same capability that changed programming becomes usable. This is the one I'd bet on. It's also the least glamorous, which is usually a good sign. **3\. Access.** "Free-tier, paid-tier, internal." Part of the gap is just which version someone has touched. Back in April, Karpathy wrote that many people "tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much." A product that puts paid-tier capability in front of someone, inside a workflow they already use, closes part of that gap. It doesn't ask them to change their mind first. ## What the numbers don't tell us Karpathy gives no source for any of the tier sizes. They're his estimate, and he says so. The claim that 20 million people see weeks of work done from a prompt is his description of their experience, not a measurement of it. He describes the 1-2 billion casual users as free-tier users, but gives no measured split. We don't know how they divide between free and paid plans, or how many have tried an agent at all. The funnel is a useful picture. It isn't a dataset. ## What we do with it The move is narrow. 1. Pick [a domain outside code](https://varops.com/a-yc-startup-rebuilt-an-accounting-firm-around-ai-and-operating-margin-jumped-from-5-to-70-heres-how-they-keep-the-upside/) where people do real, valuable, digital work. 2. Find the big problem there that someone actually cares about. 3. Work out what "correct" looks like, so the result can be checked. 4. Deliver it inside a tool those people already use, at the tier where the capability is real. That's the whole checklist. Most of the funnel is still waiting for someone to run it. ## Sources - Andrej Karpathy on X (2026-10-11): [https://x.com/karpathy/status/2109361546505966046](https://x.com/karpathy/status/2109361546505966046?ref=varops.com) - Andrej Karpathy on X (2026-04-09): [https://x.com/karpathy/status/2042334451611693415](https://x.com/karpathy/status/2042334451611693415?ref=varops.com) - ITU, Measuring digital development: Facts and Figures 2025: [https://www.itu.int/itu-d/reports/statistics/facts-figures-2025/](https://www.itu.int/itu-d/reports/statistics/facts-figures-2025/?ref=varops.com) - UN World Population Prospects 2024, via Our World in Data: [https://ourworldindata.org/grapher/population-by-five-year-age-group](https://ourworldindata.org/grapher/population-by-five-year-age-group?ref=varops.com) - WHO, Malaria fact sheet: [https://www.who.int/news-room/fact-sheets/detail/malaria](https://www.who.int/news-room/fact-sheets/detail/malaria?ref=varops.com)