Andrej Karpathy Warns the Public Has Only Seen a "Preview" of Frontier AI

A rapidly widening divide is separating public perception of artificial intelligence from the reality of internal frontier systems, according to Andrej Karpathy, co-founder of OpenAI and current researcher at Anthropic. Karpathy warned that while most people perceive large language models (LLMs) as casual search or writing assistants, internal frontier tools are already executing complex projects that previously required years of expert labor.

According to Karpathy, "the external world has seen the preview" of what modern AI can actually accomplish.

The Four Tiers of AI Adoption

To illustrate this disparity, Karpathy broke down the global population into four distinct levels of LLM interaction:

1. The Unreached Majority: Roughly 6 billion people, or about 75 percent of the global population, have barely interacted with LLMs at all.
2. Casual Consumers: Between 1 billion and 2 billion users interact with free-tier, ChatGPT-style tools. These users largely treat AI as "a bit like a better Google search, a writing aid, or etc." Karpathy noted that many professionals outside of coding and mathematics, including corporate executives, fall into this category. Because their daily work centers on interpersonal communication, their grasp of technical AI shifts remains largely abstract and second-hand.
3. Professional Frontier Users: An estimated 20 million people (approximately 0.2 percent of the world) regularly employ frontier-grade models in mathematics and programming. For this group, autonomous agents can now complete large, complex software projects—such as building, duplicating, translating, or decompiling applications from binaries—via prompts instead of weeks or months of manual work.
4. The Frontier Elite: Approximately 5,000 people worldwide—just 0.00006 percent of the population—have access to internal, bleeding-edge frontier systems.

Swarms of Agents and Rapid Evolution

For the small cohort with internal access, Karpathy described a fundamentally different technological reality. These users observe swarms of thousands of agents collaborating over weeks on massive software projects: discovering new scientific principles, advancing frontier mathematics, minting zero-day vulnerabilities, and executing cyber attacks and defenses at machine speeds.

This technological leap occurred remarkably fast. As recently as October 2024, Karpathy publicly characterized AI agents as mostly hype that barely functioned. Only a few months later, he stopped writing computer code by hand entirely, shifting from typing characters into an editor to letting autonomous agents handle the workflow. Karpathy subsequently joined Anthropic.

Similar observations have emerged from other industry leaders. In early September at the G20 Innovation Ministerial, OpenAI CEO Sam Altman stated that a modern AI coding tool could accomplish in "17 minutes" what an entire startup had previously spent three months building, warning that "the world has not caught up with what this is going to mean."

Comprehension Struggles and the Next Frontier

Karpathy highlighted that human review is already struggling to keep pace with model output. He cited ongoing analysis required for incidents from months prior and pointed out that mathematicians "may be poring over the 722 manuscripts on frontier mathematics for a while."

He referred to this divergence as the "weirdness of the moment," where the general public sees a "derpy chatbot," most professionals experience modest productivity gains, and a tiny sliver of technical experts experiences "the vertigo of the curve going vertical."

According to Karpathy, AI capabilities are currently advancing fastest in fields that are "digital, verifiable and economically valuable." Because today's model breakthroughs emerge from reinforcement learning paired with verifiable rewards, major AI labs naturally concentrate compute and training resources into disciplines with the clearest financial return—principally coding and mathematics.