Karpathy Describes a Gap in Access and Perception of AI

The researcher from Anthropic, Andrej Karpathy, distinguishes four layers of language model usage, including an elite group of a few thousand individuals exposed to very advanced internal systems. While the majority only encounter chatbots, quantitative evidence suggests far superior performance, especially in programming and mathematics.

Concrete Signals of a Growing Gap

Karpathy indicates that around 5,000 individuals worldwide have access to advanced internal systems, which operate like networks of thousands of agents collaborating on vast software projects, capable of generating zero days, executing cyber offensives and defenses at machine speed, and contributing to scientific and mathematical advancements. According to him, human analysis and assimilation are beginning to lag, particularly in the face of the 722 advanced mathematics manuscripts that need to be analyzed. In the same vein, Sam Altman, CEO of OpenAI, stated in early September at the G20 Innovation Ministerial that a current AI coding tool could accomplish in 17 minutes what a startup had built in three months, adding that the world has yet to grasp what this will mean.

Four Layers of Usage, from Billions to Thousands

Karpathy categorizes access to large language models into four groups: approximately 75% of the global population, or nearly six billion people, have little to no contact with these tools. Between one and two billion people use them occasionally through free services, primarily as writing or research aids. A third group, consisting of about 20 million professionals (0.2% of the population), utilizes cutting-edge LLMs for tasks in mathematics and programming, where agents can create, duplicate, translate, or decompile applications from binaries with simple instructions. Finally, around 5,000 individuals (0.00006% of the global population) have access to very advanced internal systems.

Where LLMs Excel and How They Are Perceived

For the majority of the public, language models are primarily seen as tools for writing assistance or research, while the most advanced internal systems are already tackling missions that would take years even for the best specialists. Karpathy emphasizes that many individuals continue to analyze the OpenAI-HF incident that occurred several months earlier and describes a context where there is both limited public interaction, a perception of moderate progress among most professionals, and, for a minority, a pronounced sense of acceleration. He specifies that the current capabilities of LLMs reach their peak in digital, verifiable, and economically valuable domains, which he links to reinforcement learning based on verifiable rewards and the orientation of corporate training resources towards the most lucrative sectors, currently programming and mathematics.