Andrej Karpathy, a well-known AI researcher who joined Anthropic in May 2026, wrote on X that much of the world has little understanding of what advanced artificial intelligence systems can do.
Karpathy's estimates
- He estimates roughly 6 billion people have little or no experience using a large language model (LLM).
- About 1–2 billion people use chatbots occasionally.
- Approximately 20 million people use the most advanced models professionally, for tasks such as mathematics and programming.
- Only around 5,000 people have access to lab-grade frontier systems.
Karpathy used these figures to illustrate a sharp divide in how the technology is perceived and accessed: casual users and developers see very different aspects of the same underlying systems, and the capabilities of frontier models are known and reachable by only a small number of people.
Responses and interpretations
Aaron Levie, CEO of Box, which offers AI-based services, framed Karpathy’s observations as an opportunity. In his reply he argued that the knowledge gap highlights how early stage we are in the broader adoption of AI.
Nature of current use and technological constraints
Karpathy observed that many still think of AI mainly as chatbots or next-generation search engines and are not exploiting higher-level capabilities. He expects the future to involve systems that distribute complex tasks among parallel agents, but identified several current obstacles:
- Model capabilities vary across task types.
- Large-scale operation is expensive.
- Tools to manage multiple agents are still under development.
- Most people do not have access even to the most powerful models.
In an earlier April post, Karpathy contrasted OpenAI’s free voice mode— which he said struggles even with the simplest questions you might see on an Instagram Reel— with the company’s high-end Codex model, which he described as capable of working for an hour or more to transform an entire codebase.
Rapid progress
Karpathy emphasized the speed of advancement: while many solutions were largely hand-coded at the beginning of the year, by April agents had emerged that can solve programming tasks in moments that would previously have taken days or weeks.
When AI delivers the most value
He argued that AI’s real advantages appear when agents are entrusted with longer, iterative work: given hours to test and refine outputs, agents can improve results to reach the best possible outcome. Karpathy sees a problem in that people often do not trust systems with such substantive tasks; asking a question that can be answered in a single paragraph barely exercises an AI’s true capabilities.
Conditions for broader deployment
Karpathy concluded that companies still need to figure out how to make these systems affordable, safe, and reliable enough for widespread use. Until then, disparities in knowledge and access will significantly shape who experiences and benefits from the most advanced capabilities of artificial intelligence.
