On Sunday night Andrej Karpathy put a log-scale sketch on X: roughly 6 billion people have barely touched an LLM, about a billion and a half use a free chatbot casually, about 20 million use frontier models professionally for math and code, and roughly 5,000 have internal access to frontier-grade systems inside the labs. He called it a sharp funnel, and said the gap in shared understanding of LLM capability is widening. By Monday morning the Hacker News thread sat at 303 points and 486 comments, most of them litigating the shape of that funnel — access, price, secrecy.

They are arguing about the wrong axis. The funnel isn’t steep because frontier capability is hoarded. It’s steep because there is no scoreboard outside of math and code.

The Funnel Is a Grading Function, Not a Distribution Channel

Karpathy made this point himself at Sequoia’s Ascent event this year: LLMs automate what you can verify. Frontier labs train them in giant reinforcement-learning environments with verification rewards, which is why they arrive as jagged entities — peaking in math, code, and adjacent domains, and staying rough everywhere the ground truth runs out. Capability isn’t a fluid that diffuses through a pipe. It’s a rating function that climbs wherever somebody can check the answer.

Which is why the access debate is close to moot. Anthropic’s own model table puts the free tier on Sonnet 5 (launched June 30) and Haiku, with Opus and the Fable-branded models behind the paywall. ChatGPT’s free tier now runs with advertising in the US and Europe. Gemini AI Plus costs $4.99 a month. If the constraint were price, we’d have solved it by now.

Handing everyone an Opus-class model would not change the fact that a claims adjuster, a hospital scheduler, a city planner, and a mid-market HR lead have no ground truth to grade against. They get a slightly better draft. Nobody can score it, so nobody can compound it.

The most honest number in the industry right now isn’t 5,000. It’s 15. Per a pricing tracker re-verified on October 5, OpenAI’s GPT-6 Pro caps out at 15 messages a month on Business Standard — against 50 a week at the $100 tier and 200 a week at $200. Somebody ran the unit economics and discovered that, outside a handful of departments, heavy usage is a cost with no measurable return. That is what a missing scoreboard looks like in a spreadsheet.

The 5,000 Is the Least Interesting Number on the Chart

Lab insiders have always had more than everyone else. The figure that should keep people up is the 20 million: the math-and-code professionals whose output is now legible to a grader. Those are the workers whose productivity can be measured, compared, benchmarked, and eventually paid for accordingly — and the first cohort where employers can actually tell who is worth a multiplier and who isn’t.

That is where the political fight is heading, and it will not be phrased as “democratize frontier access.” It will be phrased as: why does the $400,000 engineer get a compounding multiplier while the $70,000 operations analyst gets marginally faster email? A procurement manager at a 4,000-person logistics company, on a Tuesday renewal call with the seat-count tab open, put it to me more bluntly than any think-tank paper has: “We can tell you exactly what the pricing team got out of it. We cannot tell you what HR got out of it. So next year HR gets five seats.”

Note the logic. This isn’t a firm being stingy. It’s a firm being rational about the only thing it can measure.

Cheap Tokens Can’t Buy a Scoreboard

The reflex on the right is that cheap inference will diffuse capability downward, the way cheap bandwidth did. The reflex on the left is that the state should buy everyone a seat. Both assume the binding constraint is price. It isn’t. $4.99 Gemini and ad-supported ChatGPT already exist, and the free tier of Claude is competent enough for a great many jobs.

The binding constraint is that most work has no reward signal. A better model produces a better artifact with no way to establish that it is better. Organizations that cannot score a gain cannot budget for it, promote on it, or defend it in a board meeting. Capability piles up in the two domains with automatic graders and leaks out everywhere else as vibes.

This is the part my own side should find uncomfortable. If capability compounds only where verification exists, then the sectors conservatives keep promising AI will finally discipline — government administration, health, education, insurance back offices — are precisely the sectors with the worst scoreboards. The fiscal dividend is real, but it is going to arrive late, unevenly, and smaller than the press releases. The places where AI will visibly cut cost per unit over the next three years are the places we already measure in tests and deploys, which is to say the places that were already expensive and already staffed by people with options.

So when the next earnest essay arrives about closing the AI understanding gap — more demos, more explainers, more free tiers — remember what the funnel is actually shaped by. Karpathy’s chart isn’t a map of who’s been let in. It’s a map of where the world keeps score. The 5,000 at the narrow end are not privileged so much as surrounded by instruments. Everyone else is being asked to get better at a game nobody has agreed how to referee.

That is a harder problem than access, and no amount of subsidized tokens will solve it.

Sources