Karpathy: shared understanding of LLM capability is collapsing into a steep funnel

AI researcher Andrej Karpathy says the public's shared picture of what large language models can do is coming apart as adoption grows. In a post on X that reached Hacker News's front page, Karpathy — resurfacing remarks he first published in April — argued the situation has moved beyond two groups simply failing to communicate and now resembles a steep funnel, where each level has a radically different experience of the same technology.

A funnel of four experiences

Karpathy's figures are explicitly back-of-the-envelope, but they sketch four tiers. Around 6 billion people, roughly 75 percent of the world, have barely encountered LLMs at all. Another 1 to 2 billion, around 20 percent, are casual users of free chatbot products; to them the technology reads as a somewhat better search engine or a writing assistant, and he notes that even many professionals outside math and programming sit in this tier, since their sense of the systems remains second-hand and abstract.

The third tier is where his account shifts. He estimates about 20 million people, some 0.2 percent of the population, work professionally with frontier-grade models in math and code. This group is watching agents complete large projects — building, cloning, porting and even decompiling applications — starting from a prompt, work that used to take weeks or months. Karpathy adds that less than a year ago he was still typing code by hand, character by character, with the occasional autocomplete.

At the narrowest point, roughly 5,000 people — about 0.00006 percent — have internal access to frontier systems. What the outside world has seen, he writes, is only a preview: thousands of agents collaborating for weeks on software mega-projects, producing zero-day exploits, running cyber offense and defense at automated speed, and generating new science and mathematics. As evidence that human comprehension is falling behind, he points to researchers still unpicking the OpenAI-HF incident from months ago, and to mathematicians facing a pile of 722 frontier-mathematics manuscripts.

Three forces driving the split

Karpathy attributes the funnel to three factors. The first is that 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 genuinely cares about. This, he says, separates casual users from professionals. The second is the jaggedness of these systems: capability concentrates in domains that are digital, verifiable and economically valuable, a shape he traces to reinforcement learning on verifiable rewards over a curated mixture of environments selected for revenue potential. This separates math and code from other professions. The third is access itself — free tier, paid tier, internal — which separates insiders from everyone else.

Why it matters

If Karpathy's framing is even roughly right, public debate, corporate planning and regulation are all being calibrated against wildly different baselines. Most of humanity's lived experience of AI is a chatbot that helps with a paragraph; a thin professional slice is watching project timelines collapse; and a tiny group of insiders is reportedly working with systems whose output humans are struggling to review. His examples of review lagging — incident archaeology and a backlog of machine-generated mathematics — point at a governance problem that grows with capability. The numbers are one researcher's estimates rather than survey data, and he presents them as such. But the post's traction on Hacker News suggests the core observation lands: the oddity of this moment, in Karpathy's telling, is that all of these experiences are unfolding at the same time.