What happened
In an October 18 essay, Hanh Pham audited four figures attributed to Andrej Karpathy’s informal description of AI adoption. The funnel divides people into roughly 6 billion with little contact with large language models, 1 to 2 billion casual users of free products, about 20 million people in mathematics and code using agents for projects that once took weeks, and about 5,000 people inside AI labs with access to unreleased systems. Pham calls the original framing “napkin math” and treats the task as an audit of what can actually be checked.
The essay says the original X post was not directly available to automated readers, so it relies on three independent transcriptions for the four counts and Karpathy’s accompanying explanation. Pham distinguishes the sources: a newsroom report, a video breakdown, and a vendor write-up. Repetition, the essay argues, should not be mistaken for independent verification. The three drivers attributed to Karpathy are that impact rises with the size and time horizon of a problem; model ability is uneven and strongest in digital, verifiable, economically valuable domains; and access is tiered among free, paid, and internal offerings.
Using a 2025 world-population figure of 8,215,424,893, Pham calculates that 6 billion is 73.0% of the population, rather than the quoted approximate 75%. The smaller tiers align more closely with the stated percentages: 20 million is 0.243%, and 5,000 is 0.000061%. But the essay’s central point is not that the divisions fail. It is that every percentage uses all people on Earth as its denominator, including children, people without internet access, and people who may have no practical route to use a model.
Pham recomputes the same counts against people aged 15 and over, and against people online. Against the online population of about 6.05 billion, the 6-billion “barely in contact” tier becomes 99.2%, while the 1-to-2-billion casual tier becomes 16.5% to 33.1%. The essay argues that this turns the bottom tier into something close to a residual of people who are online but not accounted for as casual users. It also notes one publicly described survey method: Pew Research Center’s February 2026 survey of US adults found 49% had used chatbots such as ChatGPT, Gemini, or Copilot, while 24% reported using them daily. That is evidence about one country and a particular definition of use, not a global validation of the funnel.
The 5,000-person internal-access tier remains different. Pham says it cannot be inspected through public reporting, since labs do not disclose a count or a definition for who has access to frontier internal systems. The audit therefore concludes that three tiers can be bracketed against published data, but the smallest and most dramatic tier cannot be independently checked.
What the archive adds
This is a new analytical fact on a known story, rather than a restatement of Karpathy’s original claim. The earlier note, Andrej Karpathy says only 5,000 people have seen what AI can really do, recorded the four-tier framing. Pham’s contribution is to test what the percentages mean under alternate denominators and to identify the insider estimate as publicly unverifiable.
Trend signal
No supplied pattern directly covers the measurement of AI-adoption narratives.
Why it matters
AI teams regularly encounter market claims expressed as a single adoption percentage: how many people use a model, how many workers use agents, or how far a capability has diffused. This audit shows why the numerator is not enough. “Users” may mean people who tried a chatbot, people who use it daily, paid subscribers, online adults, or workers in an addressable segment. Each produces a different planning conclusion.
For a builder, the practical consequence is not to reject rough estimates automatically. It is to ask what population the estimate describes before using it to set a product roadmap, pricing hypothesis, market-size model, or internal adoption target. The public evidence in the essay also separates observable behavior from inaccessible capability: survey responses can measure some forms of usage, while internal access claims may remain assertions until labs expose a definition and evidence.
Limits and uncertainty
This is an independent essay, not a new dataset or a direct release from Karpathy or an AI lab. Its account of the original post depends on secondary transcriptions because the post was inaccessible to automated retrieval. Its alternate denominators are analytically plausible choices, not proof that one is uniquely correct. The cited Pew survey concerns US adults and self-reported chatbot use, so it cannot establish global adoption or verify the 20-million professional-user tier. The essay also cannot disprove the 5,000 estimate; it establishes that public evidence does not permit an outside reader to inspect it.
Skill to build now
When evaluating an AI-adoption figure this week, write down four fields beside it: the numerator, the denominator, the geography, and the behavior being counted. Then calculate at least one alternative denominator relevant to the decision, such as online adults instead of total population, or weekly active users instead of everyone who has ever tried a product. This small audit makes strategy discussions more explicit about what evidence supports the claim.
Reader takeaway
Karpathy’s funnel may work as a rough narrative about unequal AI access, but its percentages are not self-explanatory facts. The key question is which population each tier is divided by. Public usage data can constrain some claims; the estimate of people with internal frontier access remains unverified.
Open question
Will major AI labs publish a usable definition and evidence for who has access to internal frontier systems, or will that tier remain a rhetorical estimate?