For many people, AI means a chatbot that can help draft text or answer questions. Andrej Karpathy argues that this familiar experience shows only a limited part of what large language models can do, especially when used by specialists with access to more advanced systems.
Four levels of exposure to AI
Karpathy divides LLM users into four broad tiers. At the widest end, about six billion people—roughly 75 percent of the world's population—have barely encountered these tools.
Another one to two billion people use free-tier chatbot products casually. They may treat them as a better search engine or a writing aid, without using them for demanding professional work.
That everyday view can extend into workplaces. Karpathy says many executives and other professionals understand AI in principle, but their experience remains second-hand and abstract when their jobs involve mostly talking with people rather than directly using the systems.
Where AI is changing professional work
A third tier, in Karpathy's estimate, includes about 20 million people, or roughly 0.2 percent of the world's population. These are professional users of frontier-grade LLMs in mathematics and programming, where outputs can be tested and applied to concrete tasks.
For these users, agents can take on complex software projects that previously required weeks or months. Karpathy gives examples such as building or copying apps, translating them, and decompiling apps from binaries. The change is not simply that a chatbot produces a helpful paragraph: users can delegate larger sequences of work.
Karpathy says the transition has happened quickly. In October 2024, he described AI agents as mostly marketing that barely worked. A few months later, he had stopped writing code by hand, moving from entering commands character by character to letting agents handle the work. He later joined Anthropic, despite having co-founded OpenAI.
A much smaller group has access to internal systems
Karpathy estimates that roughly 5,000 people worldwide can use internal frontier-grade systems. He says the external world has seen a preview of their potential, while this small group encounters systems working in parallel across large tasks.
His examples include software projects involving thousands of agents, cyber attacks and defenses operating at machine speeds, scientific discovery, and progress in mathematics. These are tasks he says could otherwise take top professionals years of work.
The gap in access matters because people outside that group may judge AI by public-facing chatbots, while internal users see more powerful capabilities in action. Those experiences can produce sharply different expectations about how far the technology has advanced.
Capability is moving faster than understanding
Karpathy also points to a growing challenge: human review and comprehension may be falling behind what systems produce. He mentions people still examining the OpenAI-HF incident from many months ago and mathematicians who may be working through 722 manuscripts on frontier mathematics.
That leaves several views of AI in place at once. Much of the public has had little contact with LLMs; many users see a chatbot with modest benefits; and a small group of professionals feels the pace of change accelerating. Karpathy calls this coexistence “the weirdness of the moment.”
He says current LLM capabilities are strongest in areas that are digital, verifiable, and economically valuable. Coding and mathematics fit those conditions, and companies direct training resources toward domains with high revenue potential. The implication is that advances may appear uneven: the most striking improvements can emerge first in work that is easier to check and commercially useful.
For readers trying to assess AI's trajectory, the key question is which level of exposure informs their view. A casual interaction may show what a general chatbot can do, but it does not necessarily reveal what specialized users can accomplish with frontier models and agents.