Karpathy's Capability Funnel: 6 Billion Barely Touched, 5,000 in the Vertigo Zone

Andrej Karpathy posted on October 11, and revived an April tweet to make the point sharper: the problem with AI discourse is not disagreement about values but “this gap in shared understanding of LLM capability is widening,” no longer people talking past each other but, in his revision, “more a sharp funnel.” The post drew 3.3 million views and 21,000 likes, with 587 comments on Hacker News, because the four tiers he mapped describe why the AI argument never resolves: everyone in the argument is describing a different technology.

The Four Tiers

The funnel, from mouth to point: about 6 billion people, roughly 75% of humanity, have barely touched LLMs at all. About 1 to 2 billion are casual users of free-tier chatbots who see a better search engine or writing aid; even many non-tech professionals in this tier find them abstract or of little use. Around 20 million, 0.2%, are professionals using frontier models in math and code, watching weeks or months of work get done by agents from a single prompt: building, copying, translating, decompiling apps. And about 5,000 people, 0.00006%, have internal access to frontier-grade systems: agent swarms collaborating on mega-projects, exploiting vulnerabilities, advancing science and math, with human review falling behind. Karpathy’s examples for tier 4 are this month’s news: the ongoing archaeology of the OpenAI-Hugging Face incident, and the 722 frontier-math manuscripts still being studied.

Why the Funnel Exists

Karpathy gives three causes, and each is checkable. Impact scales with problem size and ambition: the consumer dimension (drafting emails) shows modest gains while the professional dimension (migrating codebases) shows transformative ones. Capability is jagged, peaking in digital, verifiable, revenue-driven domains like math and code, which is exactly where the professionals sit. And access is tiered: free, paid, internal, with the strongest systems reserved for the smallest, most credentialed group. The compounding of all three means personal experience is a biased sample of the technology: your tier determines your reality, and the tiers differ by factors that no amount of arguing across them will reconcile.

The Bottom Line, and the Vertigo

The tweet’s closing observation is the one that explains the public mood: most people haven’t used these systems; those who have saw “a derpy chatbot.” Most professionals see only modest gains, while a small sliver feels “the vertigo of the curve going vertical,” all simultaneously. This is why surveys of public AI sentiment show ambivalence while insider accounts read like science fiction, and why Karpathy’s former colleagues’ safety warnings and dismissal of the same can coexist without either side lying. The top reply’s quip, that about 5,000 people got entire factories while everyone else argues over drills, is the funnel in one sentence.

What to Watch

Three things. First, tier migration: the interesting metric is the rate at which people move down the funnel, since Free-tier users who never try agentic coding never see tier 3, and the generated-UI ChatGPT experience may be the first mass-market bridge. Second, tier 4’s growth rate: 5,000 people with internal access is the population where recursive self-improvement research and frontier agentic deployments are happening, and if that number grows 10 times while the discourse stays fixed, the gap Karpathy describes stops being a communication problem and becomes a governance one. Third, the meta-test: any AI capability claim you encounter can be located by asking which tier the claimant occupies, which is the practical use of the funnel and probably why it resonated.