Karpathy says public has seen only AI preview

Karpathy said the global user base could be divided into four groups. He put roughly six billion people in the first group, as people who had barely used LLMs, and another one to two billion in the next group as casual users of free tools.

He said executives and other professionals often understood AI only indirectly. In his account, about 20 million people working in maths and coding were the third tier, because frontier models had started to turn work that once took weeks or months into prompt-driven tasks.

Karpathy also said he had moved from writing code by hand to letting agents do it after saying in October 2024 that agents were mostly marketing. He described around 5,000 people as having access to internal frontier systems.

Those systems, he wrote, could involve thousands of agents working together for weeks on software projects, finding zero-days, and carrying out cyber operations and defences. He said human review was already lagging behind those developments.

He added that model capability was strongest in digital work that could be checked and paid for, because training aimed at verifiable rewards and revenue pushed resources towards coding and maths. Altman had offered a similar warning in September, saying a coding tool had done in 17 minutes work that had earlier taken a startup three months.