Karpathy's four-tier capability funnel is compelling, and none of its numbers are checkable
Andrej Karpathy's four-tier funnel of LLM adoption drew 39,000 likes and a wall of counters, and every figure in it is his own napkin math, not survey data.
Vlad Makarovreviewed and published
On October 11 Andrej Karpathy published an essay-length post on X that resurfaced something he had written six months earlier, and the resurfacing is the message: the gap he described in April has not narrowed. An OpenAI co-founder and former Tesla AI director, he calls it a sharp funnel rather than two groups talking past each other. The post drew 39,008 likes and 5,126 reposts, a Hacker News thread of 671 points and 976 comments, and a crosspost to r/singularity. Almost none of the argument inside it can be checked from outside.
The funnel, in his own numbers
His post is mostly a distribution claim. He sorts the world by contact with large language models into four populations, and every figure in it is an estimate he labels napkin math. Roughly 6 billion people, about 75% of the population, "have barely come in contact with LLMs at all". Another 1 to 2 billion use free-tier products casually, see "derpy chatbots", and treat them as a better search box.
| Tier | His estimate | What they encounter |
|---|---|---|
| Outside | ~6B people (~75%) | "barely come in contact with LLMs at all" |
| Casual | 1-2B (~20%) | free-tier chatbots; "a better Google search" |
| Professional | ~20M (0.2%) | agents finishing week-long projects from a prompt |
| Internal | ~5,000 (0.00006%) | frontier systems; human review "falling behind" |
He notes that even many professionals outside math and code sit in the second tier, and that executives "understand what is happening intellectually" while their knowledge stays second-hand and abstract. The last two tiers are where the argument lives. Tier three is his hinge — some 20 million people who watch agents complete work that used to take weeks or months: "Building apps, copying apps, translating apps, decompiling apps from binaries." He marks the recency himself, writing that "less than 1 year ago, I was writing code manually by hand". Tier four is roughly 5,000 people with internal access to frontier systems, described as swarms of thousands of agents minting zero-days and advancing mathematics while "human review and comprehension are starting to fall behind".
What he says drives the spread
Karpathy offers three drivers rather than one. The first is that impact scales with ambition, problem size and horizon, so "a question with a paragraph answer barely stresses the system". The second is the jaggedness of the models, which he traces to training: capability peaks in domains that are digital, verifiable and economically valuable, because it "emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential". The third is access — free tier, paid tier, internal. Two of those are sociological observations. The second is a claim about how these systems are built, and it is the most interesting thing in the post.
Two claims stacked, and only one is measurable
Read closely, the funnel is two arguments sharing a diagram. The first is that usage is lopsided: that a billion-odd people have touched a chatbot and most of the rest have not. That is checkable in principle, since surveys, handset and broadband data, and the labs' own user figures all bear on it, even if none of them produce his specific counts. The second is that capability is lopsided along verifiable domains. His reinforcement-learning explanation supports that as a hypothesis about training; nothing in the post measures it. A shape that follows from a training recipe is not the same as a shape anybody has counted.
The pushback lands on four different joints
In the Hacker News thread, the objections do not converge on one flaw. User rbehrends goes at the productivity claim, reporting careful experiments with spec-driven design and concluding that agents never autonomously delivered "what I'd consider a finished product" — edges, architectural mistakes and performance problems persist outside throwaway software. tripleee goes at the rhetoric, arguing that Karpathy "is intentionally forgoing all nuance in order to make this sound dramatic", and that the cases that do work are ports and translations, "where all the rules are already documented in the best specification language possible". komali2 goes at the denominator, noting that "a quarter of the world is under 15" and that computing percentages against the entire global population flatters the story. socializer goes at the premise underneath it, arguing something close to peak LLM: most people have no special need for a model, and it is not worth $200 a month to them.
The top of the funnel is the part nobody can inspect
Engagement is evidence that people want a way to think about this, never evidence that the way is correct. The tier that carries the most weight, the 5,000 insiders, is also the least available to inspection: he is describing systems that are not sold, not benchmarked publicly and not documented, and the activities he attributes to them are precisely the ones outside researchers cannot reproduce. The sharper framing sits in the April 9 post he resurfaced — then at 22,507 likes — where he wrote that people who pay for and use frontier agentic models professionally in technical domains are "subject to the highest amount of AI Psychosis". That is a useful warning about self-selection, and it applies to the funnel's top two tiers as much as to anyone else.
What would settle it
Two things, and neither exists. The first is published usage data with a methodology attached — a survey that tests whether the 20-million tier is real at the claimed size rather than inferring it from enthusiasm. The second is independent verification of what frontier access produces, because the 5,000-person tier is defined by a capability the public cannot see. Vendor seat counts will not do the work; a licence sold to an enterprise measures autocomplete and meeting notes, not the week-long agent projects that define his third tier. His closing line is the part that rings true: a small sliver of professionals feel "the vertigo of the curve going vertical" while everyone else sees a derpy chatbot. The sizes are the part nobody can check.
