Andrej Karpathy says the gap in public understanding of AI is widening

An everyday writing aid and a system that finishes months of programming work can produce very different impressions of AI. In an October 11, 2026 X post, Andrej Karpathy argued that this divide is widening. His account contrasts casual chatbot use with much stronger professional and internal systems, but rests on observations and rough estimates rather than verified measurements.

A funnel, not two opposing camps

Karpathy revisited an argument he made in April, replacing the image of two groups talking past each other with a narrowing funnel. He divided users into four levels, moving from almost no contact with large language models—the systems behind chatbots—to professional use and access inside AI labs.

Karpathy’s rough population estimates

• About 6 billion people have barely encountered large language models at all.
• Another 1–2 billion use free-tier chatbot products casually or infrequently, often for search-like questions or help with writing.
• Around 20 million use leading AI models professionally in mathematics and programming, where he says agents can finish substantial projects.
• Roughly 5,000 have access to internal frontier systems, which he describes as capable of much larger, longer-running efforts.

He calls the population arithmetic “napkin math.” The categories describe his view of the divide, not a measured census. Nor does his account separate users neatly by job title. He places many professionals outside mathematics and coding alongside casual users, including executives whose work centers on conversations with other people.

A question with a paragraph answer barely stresses the system.

Andrej Karpathy, October 11 X post

The task changes the impression

His explanation has three parts: ambition and task size, uneven strengths across fields, and access. On the first, he argues that the impact grows with the difficulty and duration of the work. A short question does not reveal the same capabilities as a large project that someone genuinely needs completed.

For coding, his examples include building, copying and translating applications, as well as reconstructing code from compiled software. He says agents now complete complex projects that once required weeks or months. He also offers a personal marker of the speed of change: less than a year earlier, he was still writing code manually.

Those gains are not evenly distributed, in his account. He says capability peaks where work is digital, its results can be checked, and it has economic value. He attributes that pattern to training that rewards verifiable results, with companies directing the mix of training tasks toward revenue potential. Mathematics and coding are his leading examples.

Inside the narrowest tier

Access is the third dividing line: free products, paid products and internal systems expose people to different levels of capability. His most sweeping description concerns the last group. In the same post, he describes thousands of agents collaborating over weeks on major software projects, rather than one chatbot supplying an answer.

He describes work spanning the discovery of previously unknown software vulnerabilities, automated cyberattacks and defenses, science and mathematics. These are Karpathy’s descriptions of internal capabilities, not independently verified performance results. He characterizes the work as something that would previously have taken top professionals years, while warning that human review and comprehension are beginning to fall behind.