According to ChainCatcher, AI researcher, OpenAI co-founder, and former Tesla AI director Andrej Karpathy said in a post on X that the shared public understanding of large language model capabilities is widening, shifting from two groups talking past each other into a steeper funnel.
He revisited a related post from April and offered a layered estimate based on levels of exposure.
Karpathy’s breakdown of exposure levels
Karpathy said about 6 billion people, or roughly 75% of the global population, have had almost no exposure to large language models.
Another 1 billion to 2 billion people, around 20%, only use free ChatGPT-like products occasionally. He said many in this group treat them as better search tools or writing aids, and that many non-technical users, as well as professionals outside math and coding, see limited use in them.
He estimated that about 20 million people, or 0.2%, use frontier models professionally in math and coding. According to Karpathy, this group has already seen complex projects that once took weeks or months completed by agents following prompts, including building, copying, and translating applications, as well as decompiling apps from binaries.
Karpathy added that less than a year ago, he was still writing code manually, token by token.
He also said roughly 5,000 people, or about 0.00006%, are able to access frontier-grade systems internally.
What he said about internal previews
Karpathy wrote that what the outside world sees today is only a preview. In his description, thousands of agents can already work together for weeks on large software projects, cyber offense and defense, scientific discovery, and frontier mathematics.
He also said human review is starting to fall behind. As examples, he pointed to OpenAI- and Hugging Face-related incidents that were still being sorted through months later, and said mathematicians may spend a long time studying 722 frontier math papers.
Why he thinks the funnel is getting steeper
Karpathy said the funnel is being driven by several factors: the scale and time horizon of the problems involved, the fact that capabilities stand out more clearly in digitizable and verifiable domains, and the gap between free access, paid access, and internal access.
