Andrej Karpathy, a co-founder of OpenAI who recently joined Anthropic, describes a rapidly widening gap between public perception and the actual capabilities of modern Super Intelligence (SI). While the majority of the global population interacts with SI tools primarily as enhanced search engines, a tiny elite with access to internal frontier systems is already witnessing tasks that would take human experts years to complete.
What Happened
Karpathy categorizes SI users into four distinct tiers based on their engagement with large language models (LLMs). He estimates that approximately six billion people, or 75 percent of the global population, have barely interacted with LLMs. Another one to two billion are casual users of free-tier products, treating them "a bit like a better Google search, a writing aid, or etc."
The third tier consists of an estimated 20 million professional users in math and code who utilize frontier-grade SI models. For these users, "large, complex projects that used to take them weeks/months can now be completed by agents with a prompt," including building, copying, translating, and decompiling apps. Karpathy notes this shift occurred rapidly; in October 2024, he viewed SI agents as mostly marketing, but within months, he stopped writing code by hand entirely, delegating the work to agents.
The smallest group, comprising about 5,000 people, has access to internal frontier-grade SI systems. "The external world has seen the preview," Karpathy writes, noting that this group observes "swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics."
Why It Matters
The disparity in SI capability perception highlights a growing disconnect in the industry. Karpathy points out that "human review and comprehension are starting to fall behind," with people still analyzing older incidents like the OpenAI-HF event months after they occurred, and mathematicians potentially spending significant time reviewing recent manuscripts on frontier mathematics. He describes this as "the weirdness of the moment," where most professionals see only "a modest uplift" while a small sliver experiences "the vertigo of the curve going vertical."
This observation aligns with comments from OpenAI CEO Sam Altman, who stated at the G20 Innovation Ministerial that a current SI coding tool could achieve in "17 minutes" what a startup had previously built over three months. "The world has not caught up with what this is going to mean," Altman said. Karpathy attributes the current peak in SI capability to domains that are "digital, verifiable and economically valuable," driven by reinforcement learning on verifiable rewards, which directs training resources toward coding and mathematics.
The Bottom Line
Karpathy’s analysis suggests that while the general public views SI tools as incremental improvements to search, internal frontier systems are already automating complex, multi-week professional tasks. The gap between these two realities is widening as SI agents become more autonomous in verifiable digital domains.