Karpathy: the gap in understanding AI is turning into a funnel

Andrej Karpathy has revisited his April argument that people increasingly fail to grasp what large language models can actually do. By his account the gap is not merely persisting — it is widening, and it has shifted from "two groups talking past each other" to a sharp funnel.

Four tiers

By Karpathy's napkin math, roughly 6 billion people (about 75% of the population) have barely touched LLMs. Another 1–2 billion (around 20%) are casual users of free-tier products like ChatGPT: to them these are derpy chatbots, something like a better search engine or a writing aid. Even many professionals outside math and code, he notes, remain in this tier — executives, for instance, grasp what is happening only intellectually, at second hand.

Next come roughly 20 million people (0.2%) — professionals applying frontier models to code and mathematics. They see first-hand how tasks that once took weeks or months are completed by agents from a single prompt. Less than a year ago, Karpathy recalls, he was still typing code by hand, character by character.

Finally, there are about 5,000 people (0.00006%) with access to frontier systems internally. The outside world has only seen the preview: swarms of thousands of agents collaborating for weeks on software mega-projects, minting zero-days, running cyber attack and defense at machine speed, advancing science and mathematics.

What drives the funnel

Karpathy names three factors. First, impact scales with ambition and the size of the problem: a one-paragraph question barely stresses the system; you need a reservoir of big, important problems. Second, the "jaggedness" of capability: models peak where problems are digital, verifiable and economically valuable, because their skills emerge from reinforcement learning on verifiable rewards. Third, plain access — free, paid, internal.

Hence the weirdness of the moment, he concludes: most people have either never touched these systems or have seen only a derpy chatbot; most professionals feel a modest uplift; and a thin sliver is experiencing the vertigo of the curve going vertical — all at the same time. He adds that human review and comprehension are already starting to fall behind.

Source: Hacker News