Six billion, one to two billion, twenty million, five thousand. Those are the head counts in a funnel that Andrej Karpathy, a founding member of OpenAI and former head of AI at Tesla, sketched on X in October 2026. His subject was who has actually seen what large language models, the technology behind today’s chatbots and coding agents, can do. Around six billion people, he estimated, “have barely come in contact with LLMs at all.” One to two billion are “casual and infrequent users” of free-tier chatbots, who treat them “a bit like a better Google search, a writing aid.” About twenty million professionals in software and mathematics “see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt.” At the narrow end sit roughly five thousand people inside the labs, working with systems the rest of us have only glimpsed.11 Andrej Karpathy, post on X, October 11, 2026, x.com/karpathy/status/2109361546505966046. The post quotes and updates his own post of April 9, 2026.
He calls the figures napkin math, and they should be read that way. The sentence in the post that matters most for this magazine’s readers is about where executives sit. Karpathy puts most of them in the second group, next to his friends outside tech: “execs and many other functions 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.”
That matches most of the chief executives we meet, and the placement has consequences. In a company of 150 people, the chief executive’s read on AI decides how much time and money the company puts into it, and that read is mostly assembled from other people’s reports. Our argument here is that problem size is the lever an executive actually controls, and that the shape of the funnel deserves a place in quarterly planning next to the price and capability trends that already get one.
Four rooms
The October post was an update. In April, Karpathy had described two groups “speaking past each other.” One had tried a free chatbot and remembered mostly the mistakes. The other paid for the best agentic tools and used them for programming, math and research, where the improvements had been “nothing short of staggering.” Six months later he thought the picture had gotten worse. The gap was “widening,” and was now “less ‘two groups of people speaking past each other’ and more a sharp funnel.”
He names three forces that shape it. The first is ambition. “The impact scales with ambition, problem size, and horizon,” he wrote. “A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about.” The second is what he calls jaggedness: “Capability peaks in domains that are digital, verifiable and economically valuable.” The models are trained hardest where answers can be checked automatically and where customers will pay. The third is the most mundane: “Access. Free-tier, paid-tier, internal.”
The result, as he puts it, is that the public mostly sees “a derpy chatbot,” most professionals “still see only a modest uplift,” and a small group is “experiencing the vertigo of the curve going vertical.” Then the line that gives the post its force: “And it is all happening at the same time.”
That last point is what makes the funnel a management problem. If the four rooms were four stages, with everyone passing through them in turn, an executive could wait for the technology to arrive in the second room and adopt it then. Karpathy is describing four simultaneous experiences of the same technology, each one accurate for the person having it. A company contains several of them at once.
Reports from below
Start with the executive’s own seat. The job is mostly conversations: with customers, lenders, the board, the plant manager, the person who wants a raise. Karpathy’s description of that schedule is a description of how most executives learn anything. They hear it from someone who saw it. For most business questions that works well enough, because the people doing the reporting are looking at roughly the same thing. A sales forecast from three regional managers can be averaged.
Reports about AI can’t, because the people sending them are standing in different rooms. Karpathy notes that “even many professionals outside of math&code” are in the second tier. Most of a mid-market company is outside math and code. But most companies also have a few people who work inside it, and they are having a different year. The head of data at a large veterinary group, describing how his team took to coding agents, told us that nobody on the team objected that the machine was better at coding than they were. The reaction was closer to disbelief: “You’re telling me I can do this whole backlog of stuff in an hour?” Another executive at the same company had framed the problem a year earlier: “You could get it with the 10 that are keen on it. And then how do you get adoption from 50, 100, 200.”
The spread shows up inside small teams too. On our own podcast, a colleague of Sam Gaddis, one of Runpoint’s two founders, said teams should expect “big disparities,” with some people who “all of a sudden become 10 times more productive” than others. Then he made it personal. “Sam is occasionally 10 times more productive than I am because he’s adopted some new thing that I haven’t had. And it pisses me off and then it makes me want to catch up.” That’s a firm of a handful of people who all work with these tools every day. In a company of 150 the distance between the most and least fluent person is wider, and the executive is usually somewhere in the middle of it.
Picture what lands on a chief executive’s desk in a typical month. The controller tried a free chatbot on a reconciliation and it invented a number. The marketing lead uses one every day for drafts and calls it “fine.” An engineer, or the one analyst who taught herself to write scripts, says something hard to believe about a project that would have taken a quarter. All three reports are true. Averaged together they produce a picture none of them saw, a technology that is moderately useful at everything. Executives who form their view that way end up with a policy to match: a license for everyone and a modest set of expectations.
The fix is a habit of questioning. When someone reports on AI, ask two things before weighing the report. What problem did you give it, and how big? Which tool, at which tier? Most of the disagreement inside a company dissolves under those two questions, because the people disagreeing ran different experiments. It also gives the executive a rough map of where each person sits, which is worth knowing for reasons “Arm the Few,” earlier in this issue, takes up at length.
Small trials, small verdicts
Karpathy’s first force deserves the most attention from executives, because it is the one they can change on their own. The uplift a person sees, he says, depends on the problem they bring. A short question gets a short, competent answer, and a short, competent answer looks like a slightly better search engine.
Think about how a typical executive first tries the technology. Rewrite this email. Summarize this article. Draft a note about the holiday schedule. The result is fine, and the verdict that follows is fine too: useful, modest, a little overhyped. Karpathy’s April post explains why. “Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability.” The usual executive trial uses a small problem, in the flattest part of the capability curve, often on a free tier. The executive goes back to a calendar full of meetings having learned something true about a narrow case and something false about the technology.
The executives we’ve watched move up a tier mostly did it by changing the size of the problem. This month the founder of a consultancy that implements financial-planning software described to us how she had prepared for a project to rethink her company’s delivery work. Her team had collected what people told them about how projects actually get delivered. She took all of it, along with the firm’s older methodology document, and gave it to a model. What came back, as she described it, laid out “every step and where there’s gaps, what could be automated in a delivery process or not.” That’s a problem with a horizon of months and a few dozen people inside it, and it produced something she could argue with. On the same call she put her motive plainly: “The way we’re built is built for where we were yesterday. I’m trying to get us to where we’re going tomorrow.”
The other move that works is changing the tool along with the problem. In September a portfolio-company CEO told Sam about moving from the friendly desktop version of an agent to the command-line version that engineers use. The learning curve had been “virtually none,” and he was getting about ten times as much done, “despite me not seeing or writing a single line of code.” A tool built for long, hard programming problems served his long, hard problems, which weren’t programming, better than one built for chat.
What would a reservoir of big problems look like in a 150-person company? It would include the questions an executive has carried for years without the staff to answer them: why two plants report different margins on the same product, what five years of lost quotes have in common, how an order actually moves from the first email to the invoice. These are the problems the twenty million in Karpathy’s third tier bring to the tools every day, and an executive who never brings one has no firsthand basis for judging what the tools can do.
Where the peaks are
The second force, jaggedness, is a planning fact, and it should be treated like one. Capability, Karpathy explains, “emerges from reinforcement learning on verifiable rewards,” and the labs spend the most effort where the revenue is. Code improved fastest because code can be checked: “unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge.” The outside evidence for this unevenness goes back a few years. A 2023 field experiment with consultants at Boston Consulting Group found that AI helped a great deal on some tasks and hurt on one that looked similar, which the authors named a “jagged technological frontier.”22 Fabrizio Dell’Acqua et al., “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality,” Harvard Business School Working Paper 24-013, 2023. “Arm the Few,” earlier in this issue, covers the findings.
For an executive, Karpathy’s three conditions translate into questions you can ask about each kind of work in the company. Does the work already live as files and data, or does it live in conversations and on the shop floor? Can someone check an answer quickly and cheaply, the way a test checks code? And is it the kind of work thousands of other companies also pay to have done, so that the labs have a commercial reason to get good at it? Work that passes all three will move first. In most mid-market companies that means software, data cleanup, reconciliations, pulling fields out of documents, and the reporting that sits on top of all of it. Work that fails all three moves last: deciding which customer to let go, reading the mood of a plant, negotiating a supplier contract with someone you’ll see at church.
The middle is where the planning gets interesting. A good deal of work is digital and valuable but hard to check, which keeps it off the peaks. Some of it can be made checkable. A company that has years of past quotes, past credit decisions or past audits has answer keys, and can test a model against cases where it already knows what happened. Tests built from your own past work move a piece of the business a little closer to the peaks, for you specifically, since the labs are unlikely ever to train on your answer keys.
Getting it firsthand
If the problem is that an executive’s read is secondhand, the remedy is some amount of firsthand time. That is easy to say to someone whose calendar is already the reason for the problem. We’d suggest three habits, none of which requires becoming an engineer.
The first is to spend a few hours a week on one real problem of your own, at full size. Choose something you would otherwise give a senior analyst for a month, and use the best tool you can get. Keep at it for weeks, since a single sitting mostly measures how good you are at the first sitting. “Arm the Few” argues for teaching people one or two at a time on projects that matter to them, and we’d put the chief executive first in that line. The point is to have something to compare the reports against.
The second is to borrow eyes from the higher tiers. Every company has a few people, often junior, who are already in Karpathy’s third room or close to it. Executives tend to ask them for a demo, which is the wrong request, because a demo shows their problem. Ask them to sit with you and work on yours. An hour of watching a fluent person take apart a question you care about teaches more than a quarter of summaries.
The third is to test against your own finished work. Pick something your company already did and got right: last year’s pricing review, a diligence memo, an analysis of why a product line lost money. Give the same inputs to a model and compare what comes back with what you know happened. It’s the cheapest way we know to get a firsthand verdict about your own business. Write it down with a date on it. Next quarter, run the same test again.
The case against
There is a fair objection. Karpathy’s narrow end tells a 150-person distributor almost nothing about its next quarter, and most of a mid-market company’s value sits in the jagged valleys: relationships, judgment about people, knowledge of a local market, the trust of customers who’ve known the owner for twenty years. On that view, an executive who spends the day talking to people is doing the part of the job the technology reaches last, and a secondhand read of the rest is a sensible division of labor.
We agree about the valleys. Where we part ways is on delegation. The decisions that set how fast a company moves on AI, how much to spend, which work to redesign and whose judgment to trust, are made by the executive. A decision-maker who knows the technology only from reports will weigh those reports badly, and a few hours a week of firsthand use, on a problem worth the time, is a small price for a better read.
A gap with a half-life
The funnel has its own half-life question. Over the next four quarters, does the gap close or widen?
Karpathy’s own evidence points one way. Between his April post and his October update, by his account, the gap widened. Access tends to diffuse downward over time: features reach paid tiers and then free ones, and the price of a fixed level of capability has been falling fast, as the lead essay in this issue describes. If access were the whole story, the funnel would flatten over a year or two.
The other two forces don’t diffuse on their own. Jaggedness follows where the training effort goes, and nothing guarantees that the work a chief executive does all day becomes a peak on any particular schedule. Ambition doesn’t diffuse at all. A tool that can take on a month-long problem gives nothing extra to someone who brings it a paragraph. Our guess, and it is a guess, is that the access gap narrows over the next four quarters while the other two stay as wide as they are or grow. If that’s right, the tier an executive ends up in will depend less on the subscription and more on the size of the problems they bring.
That gives the funnel a place in a quarterly review. Once a quarter, ask where the leadership team sits, where the company’s most and least fluent people sit, and which parts of the work have moved closer to the peaks since the last time anyone checked. The answers will decay, which is the reason to write them down.
The founder of the planning-software consultancy has her page of gaps, built from her own people’s accounts of how the work really gets done. It’s a large problem, the kind Karpathy says the tools reward, and she brought it herself. What she does with it over the next two quarters will tell her more about AI than anything she reads, this essay included. The useful question for any executive reading this is what their own version of that page would contain, and whether they have tried to build it.