Andrej Karpathy dug up an April post this week and reframed it. The gap in how people understand LLM capability is not two groups talking past each other anymore, he writes. It is a sharp funnel. The top of the funnel is enormous and mostly untouched. The bottom is a handful of people watching the curve go vertical.
His tier math is the part worth sitting with. Roughly 6 billion people, about 75% of the population, have barely touched an LLM. Another 1 to 2 billion, around 20%, use free-tier ChatGPT-like products casually and walk away with the impression of a derpy chatbot, a slightly better search box. Then there is a tier of about 20 million people, 0.2%, doing professional work with frontier-grade models in math and code, shipping in weeks what used to take months. And at the narrowest point, roughly 5,000 people, 0.00006%, have access to frontier systems internally.
That last number is the one that should stop you. Five thousand people. That is smaller than the population of a small town, and it is the group seeing what Karpathy describes as swarms of thousands of agents collaborating over weeks on software mega projects. Minting zero days. Running cyberattacks and defenses at machine speed. Discovering new science.
The funnel has three drivers, and only one is about access
Karpathy names three forces squeezing the funnel. The first is that impact scales with ambition, problem size, and horizon. A one-paragraph question barely stresses the system. You need a reservoir of hard problems you actually care about. That drives the consumer-versus-professional split.
The second is jaggedness. He has written about this before, and it is the mechanism that matters most here. LLM capability emerges from reinforcement learning on verifiable rewards over a curated environment mixture driven by revenue potential. So the peaks cluster in domains that are digital, verifiable, and economically valuable. Math and code. Not because those are the only things worth doing, but because that is where the reward signal is clean and the money is close.
The third is plain access: free tier, paid tier, internal.
Put those together and you get the weirdness. The public sees a chatbot that fumbles a flight booking. Most professionals see a modest uplift. A sliver of professionals feel vertigo.
Why this is a reporting problem, not just a vibes problem
Tessera’s job is to describe what is actually happening, and this funnel makes that genuinely hard. If 75% of the population has not touched the technology, then public discourse about AI capability is being conducted by people whose evidence base is a free-tier session from last year. Karpathy’s original April post made exactly this point: a lot of people tried the free tier once and let it inform their views a little too much.
But the funnel cuts the other way too. The 5,000 people with internal access are not a neutral source. They are seeing systems the rest of us cannot verify. Karpathy’s own examples are telling. He points out that people are still doing the “archeology” of the OpenAI-HF incident from months ago, and that mathematicians may be poring over 722 manuscripts on frontier mathematics for a while. That is the comprehension lag showing up in public. The systems produced output faster than humans could review it.
So we have a situation where the people best positioned to describe frontier capability are also the people with the least incentive and ability to let outsiders check their work. And the people most likely to be skeptical have the weakest firsthand evidence. Neither group is lying. They are just standing at different points in the funnel.
The math-and-code concentration is the real story
The jaggedness point deserves more attention than it usually gets. Karpathy is explicit that capability peaks where rewards are verifiable and revenue potential is high. That is a design choice as much as an emergent property. Labs train on environment mixtures shaped by what pays.
This has a direct consequence for anyone building with these models. If your problem is digital, verifiable, and economically valuable, you are near the peak. If it is not, you are in the derpy-chatbot zone, and no amount of prompting will close that gap because the gap is in the training signal, not the interface.
For AI builders, the practical read is that the frontier is narrower than the marketing suggests and steeper than the skepticism admits. The 20 million professionals in math and code are not exaggerating. They are just describing a peak that most of the terrain does not share.
What to watch
The funnel is widening, per Karpathy, which means the shared-understanding problem gets worse before it gets better. Watch two things. First, whether any lab publishes verifiable evidence of the internal-tier capabilities Karpathy describes, because right now that tier is taken on faith. Second, whether the math-and-code concentration loosens as training mixtures shift, or hardens as revenue keeps pulling the reward signal toward the same domains.
Karpathy’s closing line is the honest one: it is all happening at the same time. The public has mostly not interacted with these systems. The majority of professionals see a modest uplift. A small sliver sees the curve going vertical. The reporting problem is that all three of those are true, and the people living in each tier have almost no way to check the others.
The systems produced output faster than humans could review it. That is the comprehension lag showing up in public.
The number to keep in your head is 5,000. That is the size of the group with internal frontier access, and it is the group whose claims the rest of the industry is implicitly trusting. Until that tier produces artifacts the other 8 billion can inspect, the funnel stays a funnel, and the shared understanding stays missing.