Anthropic researcher Andrej Karpathy described a growing divide in how the public perceives AI capabilities. He noted that about 75% of the global population has had minimal or no interaction with large language models (LLMs).
Karpathy segmented users into four groups. The first group, around six billion people, have barely touched LLMs. The second, one to two billion, use free-tier chatbots like ChatGPT, treating them as writing tools or better search engines.
The third group, estimated at 20 million people, are professional users in math and programming who see firsthand how AI can complete complex projects in a fraction of the time. Karpathy noted that this shift happened rapidly, with him transitioning from writing code manually to letting AI agents handle the work.
A tiny elite, about 5,000 people, have access to internal frontier systems. Karpathy wrote, "The external world has seen the preview." These users witness AI agents collaborating on massive software projects, discovering new science, and advancing mathematics at unprecedented speeds.
Karpathy highlighted that human review is struggling to keep up, with professionals still analyzing old incidents and mathematicians studying recent frontier manuscripts. He called this "the weirdness of the moment," noting that most people still see AI as a basic chatbot.
Karpathy added that AI capabilities are peaking in digital, verifiable, and economically valuable domains, driven by reinforcement learning on verifiable rewards. He did not comment on the timeline for broader adoption, and emphasized the current focus on coding and mathematics.
Source: thedecoder
