According to Andrej Karpathy’s tweet, the global population can be divided into four tiers based on exposure to large language models (LLMs). Roughly six billion people, about 75 % of the world, have barely come into contact with LLMs at all. The next tier comprises 1‑2 billion individuals, roughly 20 % of the population, who are casual and infrequent users of free‑tier ChatGPT‑like products. This group tends to view the models as improved search engines or writing aids and, in many cases, reports little practical use for them, including many professionals outside of math and coding who interact with the technology only second‑hand.
A much smaller segment, around 20 million people (0.2 %), uses frontier‑grade LLMs for professional work in mathematics and software development. For these users, tasks that previously required weeks or months—such as building applications, translating code, decompiling binaries, or creating agents—can now be completed rapidly through prompting. Karpathy notes that this shift has occurred in less than a year, marking a stark contrast to his own earlier experience of writing code manually character by character.
At the extreme end are approximately five thousand individuals (0.00006 %) who have internal access to the most advanced frontier systems. In this group, thousands of agents can collaborate over weeks on massive software projects, potentially performing activities such as generating zero‑day exploits, conducting cyber attacks and defenses at machine speed, discovering new scientific results, or advancing mathematical frontiers. Human review and comprehension are beginning to lag behind the pace of these automated efforts, as illustrated by ongoing work to “archeolog” incidents like the OpenAI‑HF case and the examination of hundreds of manuscripts on frontier mathematics.
Karpathy attributes the funnel‑shaped distribution to three factors: the scaling of impact with problem ambition and horizon, the jagged nature of LLM capability that peaks in digitally verifiable, economically valuable domains such as math and code, and tiered access (free, paid, internal). He observes that while the general public mostly encounters derpy chatbots, most professionals see only a modest uplift, and a tiny sliver experiences the vertigo of rapidly advancing capability, all occurring simultaneously.
Key facts
- 6 B people (~75 %) have barely interacted with LLMs
- 1‑2 B people (~20 %) are casual free‑tier users
- ≈20 M people (~0.2 %) use frontier‑grade LLMs for math/code work
- ≈5,000 people (~0.00006 %) have internal access to the most advanced systems