ANDREJ KARPATHY: Firsthand exposure to frontier AI capability narrows sharply from billions of people to millions to thousands

The Gist

Karpathy sorts the world into four groups: most people have hardly used AI, one to two billion use free chatbots casually, tens of millions use top coding agents at work, and a few thousand insiders use unreleased systems. Each smaller group sees much more of what AI can do. Public user counts roughly fit the middle groups, but the top and bottom numbers are rough. This is a steelman: a charitable reconstruction of Andrej Karpathy's posts in their strongest form, not an endorsement of their claims.

Conclusion

Firsthand experience of frontier AI capability is distributed as a steep funnel, with each narrower tier seeing materially more capability, so the earlier two-group description understates how concentrated that experience is.

Premises

1. By rough estimate, most of the world's population, on the order of six billion people, has had little or no deliberate contact with LLMs.
2. On the order of one to two billion people use free consumer chat products occasionally, mostly as a better search engine or writing aid, and many professionals whose work is mainly conversation sit in this tier.
3. On the order of tens of millions of people use frontier-grade agents professionally in programming and mathematics.
4. On the order of thousands of people inside frontier developers have access to unreleased internal systems.
5. Each tier is roughly two or more orders of magnitude smaller than the tier before it, and each sees a materially different level of capability.

Assumptions

• The tier sizes are order-of-magnitude estimates, not measurements.
• Little or no deliberate contact excludes incidental exposure: search-page AI summaries reach more than 2.5 billion monthly users, so passive contact is broader than the bottom tier suggests.
• Active-use figures fit the middle tier: ChatGPT reported 1.2 billion weekly users in September 2026, the Gemini app passed 1 billion monthly users, and Microsoft estimated 18.8% of the working-age population used generative AI in June 2026, with overlap among these counts.
• No public figure isolates professional frontier coding-agent users; the nearest disclosure, 35 million weekly users of ChatGPT Work and Codex combined, includes non-coding work, so the tens-of-millions tier is plausible but unverified.
• Combined frontier-lab headcounts exceed five thousand, and vetted outside partners use some gated models, so the top tier is better read as thousands to low tens of thousands with varying access.

Analysis

Argument type: Inductive.

Potential Fallacies

• Spurious precision / internal numerical inconsistency (Premise 5 versus Premises 1–3) — P5 says each tier is two or more orders of magnitude smaller than the previous one, but the stated figures give a step of only about 3–6x from the first tier to the second, and about 1.5–2 orders from the second to the third. The funnel is steep overall and steepest at the top, but not uniformly so. Rough estimates read better as ranges than as a clean geometric ladder.
• Conflating usage or access with capability exposure (Premise 5 and the conclusion) — The evidence is about how many people use, or can access, each kind of product. The conclusion is about how much capability each group has seen. Skill, task difficulty, model choice, and domain all shape what a user actually encounters, so a free-tier power user on hard problems may see more than a routine professional user. The step from 'narrower tier' to 'materially more capability' needs its own support.
• Irrelevant corroboration (Premise 5 supporting evidence) — The confirmed figure on growth of about one percentage point per quarter is attached to P5, but adoption growth says nothing about the ratios between tiers or the size of the capability gap. Verified numbers for the middle tier can lend credibility to claims they do not actually support.
• Appeal to unverifiable insider access (Premise 4 and the conclusion) — Outsiders cannot check what unreleased internal systems can do, so the claim that the top tier sees materially more capability rests on the source's standing. This is a reasonable reason to take the claim seriously but not independent evidence for it. The same source also has some interest in a narrative of rapid progress.
• Definitional framing (partly circular) (Premise 5 and the conclusion) — Tiers are defined by access and usage type and then described as tiers of capability. Concentration therefore follows partly by definition rather than being discovered. Likewise, the 'understates' comparison with an earlier two-group framing is nearly guaranteed, since a finer partition always shows more structure.

Counterarguments

These challenges are not ranked.

• Premise 5 and Conclusion (High impact) — Capability seen depends on user skill, task difficulty, tooling, and budget as much as on access tier. A skilled paid user in a non-coding field may encounter more capability than a routine programmer. The funnel therefore tracks access, not experienced capability, and may be better described as a skill-and-effort gradient than an access gradient.
• Premise 5 and Premise 4 (High impact) — The gap between public and internal systems may be small or temporary. Frontier models are often released within months, internal systems may be earlier checkpoints, and the main differences may lie in scaffolding and compute rather than raw capability. If so, the top tiers see little that the next tier does not, and the 'materially more capability' claim weakens substantially.
• Premise 5 (Medium impact) — On the argument's own numbers, 6B to 1–2B is under one order of magnitude, and 1–2B to tens of millions is about 1.5–2. The 'two or more orders of magnitude per tier' generalization is therefore false as stated, though the funnel remains steep overall.
• Premise 1 (Medium impact) — Passive exposure through search summaries reaches more than 2.5B monthly users, and AI-mediated services affect many who never chat with an LLM. Restricting the bottom tier to 'deliberate contact' makes it hard to falsify and may overstate how cut off the base of the funnel is from AI.
• Conclusion (Medium impact) — The 'understates concentration' claim is close to trivial, since any finer partition reveals more structure than a two-group split. The argument also never states the earlier description, so it cannot show that description was wrong rather than merely coarse.
• Premise 3 (Medium impact) — Treating programming and mathematics as the canonical professional frontier reflects the author's own community. Frontier-level use may also occur in science, security, medicine, and law, so the tiers may not form a single ladder and professional use is not confined to coding.

Suggested Improvements

• Ratio claim — Replace 'two or more orders of magnitude per tier' with a claim that matches the numbers, such as 'each tier is substantially smaller than the last, with the narrowing steepest at the top'. Present sizes as ranges. This keeps the funnel conclusion and removes the one claim that the argument's own figures contradict.
• Capability gradient evidence — Provide independent evidence that each tier sees different capability, such as benchmark comparisons of the best model available at each tier, documented release lags between internal and public models, or accounts from users at each level. The conclusion's force rests on capability differences, but the evidence offered only establishes headcounts.
• Definition of frontier and tiers — Define 'frontier capability' independently of tier membership, specify whether paid reasoning tiers count, and state whether tiers are nested or overlapping. This avoids defining capability by the tiers and the tiers by capability, and clarifies what a 'funnel' means given acknowledged overlap.
• Earlier framing — State the earlier two-group description explicitly and show what it claimed and how the four-tier model differs. Without it, readers cannot evaluate the 'understates' claim, and the comparison looks like a restatement of finer granularity.
• Dynamics and domain — Acknowledge that tier boundaries shift quickly (about a point per quarter) and that capability exposure varies by domain and skill. Consider presenting the funnel as dated and multi-dimensional. A static single-axis ladder will go stale quickly and hides cross-cutting variation.
• Professional tier measurement — Seek deduplicated telemetry or survey data isolating professional agentic coding and math users, and distinguish agentic use from autocomplete-style use. The middle of the funnel has the weakest direct evidence, so a better figure would anchor the shape.

Scenario Tests

• A representative survey finds that a few percent of adults use paid reasoning models daily, a group much larger than tier 3 and distinct from the casual tier. (Challenges) — The funnel would show a smoother continuum in the middle rather than discrete tiers, and the ratios between tiers would shrink. The concentration claim would still hold, but the four-bucket model would look artificial.
• Independent evaluators with access to both public and gated pre-release models show a large, consistent performance gap on hard tasks. (Supports) — This would supply the missing evidence for the capability gradient and make the top tier's 'materially more capability' claim credible.
• Frontier models are routinely released within weeks of internal availability, and benchmarks show public models matching internal ones on most tasks. (Challenges) — The top tiers would see little more than the tier below, so the funnel would describe headcount rather than capability exposure, and the conclusion would largely collapse.
• The professional agent tier turns out to be about 5 million or about 100 million people. (Neutral) — The funnel's overall shape survives either way, but the ratios between adjacent tiers shift, and a uniform 'two orders of magnitude' claim fails further. The qualitative conclusion is robust to a several-fold error in this tier.
• Agentic tools are bundled into free consumer products and adoption continues to rise by about a point per quarter. (Challenges) — People would move up tiers quickly and the funnel would flatten. The snapshot would be a transient feature of 2026 rather than a stable structure, though the top tier's information lead might persist.

Coherence & Relevance

The argument is coherent as a descriptive sketch: four tier-size premises feed a synthesizing premise that supports the funnel conclusion, and the stated assumptions candidly flag the soft spots (estimate quality, passive exposure, overlap, an unverified middle tier, a fuzzy top tier). The qualitative claim of strong, top-heavy concentration is well supported. The main weaknesses are the numerical overstatement in P5, the unsupported leap from access and headcount to capability seen, and the missing statement of the earlier two-group view. Taken as an inductive case for a concentrated, tiered distribution with fuzzy boundaries, it is moderately strong.

• By rough estimate, most of the world's population, on the order of six billion people, has had little or no deliberate contact with LLMs. (Strong) — Establishes the wide base of the funnel but is largely a residual of the other tiers. It says nothing about capability and depends on a narrow definition of 'deliberate contact'.
• On the order of one to two billion people use free consumer chat products occasionally, mostly as a better search engine or writing aid, and many professionals whose work is mainly conversation sit in this tier. (Strong) — Supports the tier size well. Free-tier users may run weaker models, which links the tier to lower capability exposure, but that link is not shown, and overlapping counts blur the boundary with other tiers.
• On the order of tens of millions of people use frontier-grade agents professionally in programming and mathematics. (Strong) — Needed for the funnel's middle step, but unverified. The link between professional agent use and seeing more capability is plausible but assumed. Frontier-level use outside coding and math is not captured.
• On the order of thousands of people inside frontier developers have access to unreleased internal systems. (Strong) — Defines the apex of the funnel. Whether internal systems are materially more capable than public ones is not evidenced, and access levels within labs and partners vary.
• Each tier is roughly two or more orders of magnitude smaller than the tier before it, and each sees a materially different level of capability. (Moderate) — This premise carries most of the conclusion but is partly redundant with it. Its ratio claim conflicts with P1–P2 and its capability claim lacks evidence. The attached growth statistic is irrelevant to it.

View this argument on LogicFirst.ai