Karpathy Says Shared Understanding of AI Collapses Into a Steep Funnel

Andrej Karpathy argued on Sunday, 11 October 2026, that the public’s shared picture of what large language models can do is collapsing into a steep funnel rather than a simple gap between skeptics and enthusiasts. In an X post timed at 19:12 UTC and reconstructed the same day by CellCog and Deniz Genc, the former Tesla AI director and OpenAI researcher sketched four experience tiers—from billions of people with almost no LLM contact to a few thousand insiders with internal frontier access—and said all of those realities are unfolding at once.

Four tiers in Karpathy’s napkin math

Karpathy labels his counts as napkin math, and CellCog stresses that the post cites no survey for any figure. The shape, not a census, is the claim. About 6 billion people, roughly 75 percent of the world, have barely come into contact with LLMs. Another 1 to 2 billion, about 20 percent, are casual users of free chatbot products; to them the technology looks like a somewhat better search engine or writing assistant, and Karpathy places many professionals outside math and programming in that tier because their sense of the systems remains second-hand.

The third tier is where his account sharpens. He estimates about 20 million people, roughly 0.2 percent, work professionally with frontier-grade models in math and code. That group, he says, is watching agents complete large projects—building, cloning, porting and even decompiling applications—from a prompt, work that used to take weeks or months. He adds that less than a year ago he was still typing code by hand with occasional autocomplete. At the narrow tip, roughly 5,000 people—about 0.00006 percent—have internal access to frontier systems. What outsiders see, he writes, is only a preview: multi-agent work on mega-projects, automated cyber offense and defense, and new science and mathematics that human review is struggling to keep up with.

CellCog reported the Sunday post had hundreds of thousands of views within hours and noted that Karpathy was resurfacing themes from an April post that framed two groups talking past each other. The October version adds the top insider tier and treats the gap as widening. Deniz Genc’s Hacker News-facing summary likewise emphasizes that each level has a radically different lived experience of the “same” technology.

Why he says the funnel keeps widening

Karpathy attributes the split to three forces. First, impact scales with ambition: a question with a paragraph-long answer barely stresses the system, so real payoff requires a reservoir of big, hard problems the user cares about. That, in his telling, separates casual users from professionals. Second, capability is jagged: it concentrates in domains that are digital, verifiable and economically valuable, a shape he ties to reinforcement learning on verifiable rewards. That separates math and code from many other professions. Third, access itself—free tier, paid tier, internal—separates lab insiders from everyone else.

Those drivers matter for companies buying agents and for policymakers writing rules. If most executives still experience AI as a chatbot that helps with email, while a thin professional slice watches project timelines collapse, then board decks, procurement checklists and draft regulations may be calibrated against incompatible baselines. PromptCrates reporting on enterprise trust in autonomous agents and on AI loss-of-control incidents already sits in that mismatch: public debate often lags the workflows that tool-using agents enable for people with the right problems and the right access.

What the post does and does not prove

CellCog’s same-day analysis is careful about limits. The 6 billion, 1–2 billion, 20 million and 5,000 figures are Karpathy’s estimates; the 5,000 are unnamed; and X marked the post as edited after publication. CellCog also places vendor-published user counts beside his middle tiers without claiming they validate his exact numbers: OpenAI has spoken of collective weekly ChatGPT users on the order of a billion, Google has claimed more than a billion monthly Gemini app users, Microsoft has reported tens of millions of Copilot users, and OpenAI has said more than five million people use Codex weekly. Those metrics measure different things than “understanding,” so they can at best show that a large casual tier and a smaller serious-developer tier are plausible orders of magnitude.

Karpathy points to public glimpses of the top tier that outsiders are still reconstructing, including lingering analysis of an OpenAI–Hugging Face incident and a backlog of frontier mathematics manuscripts. Deniz Genc similarly notes researchers still unpicking incident archaeology and mathematicians facing hundreds of machine-generated papers. Those examples support his claim that human review can lag generation; they do not, by themselves, prove the 5,000 figure.

CellCog’s conflict disclosure—that it builds AI employees for roles outside code—does not change the primary facts of Karpathy’s post, but it does explain why the firm spent the same day testing his tiers against vendor user counts. For news readers, the useful discipline is the same: treat the funnel as a researcher’s map of unequal experience, quote the napkin-math caveat every time the 5,000 figure appears, and avoid converting his X thread into a lab headcount.

For PromptCrates readers comparing this cultural map with product launches such as OpenAI’s ultrafast GPT-6.1 modes or fundraising stories like Typesafe’s Jev raise, the usable takeaway is narrower than any scoreboard. Karpathy is arguing that simultaneous, incompatible experiences of AI are now the default, that the professional math-and-code slice sees agentic project collapse the casual majority does not, and that a tiny internal cohort sees capabilities the public only hears about second-hand. The verifiable news from 11 October 2026 is that framing, attributed to his X post via CellCog and Deniz Genc—not a new model release, not a lab census of the 5,000, and not a survey-backed proof of the napkin math.

• CellCog: Karpathy’s AI funnel, only 5,000 see the frontier
• Deniz Genc: Karpathy on the steep LLM capability funnel
• PromptCrates: VentureBeat on enterprise trust in autonomous agents
• PromptCrates: AI loss-of-control incidents July surge
• PromptCrates: OpenAI ultrafast GPT-6.1 modes
• PromptCrates: Typesafe Jev raise and valuation