OpenAI squeezed 4,000 years of thinking into 88 hours
Ten thousand AI agents spent 130 billion tokens to crack one of math's hardest problems.
Dwarkesh Podcast · Noam Brown – Agent swarms, alignment, & recursive self-improvementIn an interview, podcast host Dwarkesh Patel calculated that OpenAI had squeezed about 4,000 years of full-time human thinking into 88 hours.
The job was solving what Patel called "one of the Millennium Prize problems," seven hard math questions. The system used 10,000 AI agents working together, spending 130 billion tokens in those 88 hours, Patel said.
Patel said that much back-to-back thinking would stretch from ancient Sumeria to today. He said he was surprised there was not a bigger penalty for splitting the work across so many agents.
Noam Brown, a researcher at OpenAI who helped create its reasoning models, downplayed that part. He said he would give multi-agent work less than 10% of the credit, saying the core reason was a very strong general-purpose model that can reason over long periods. OpenAI has not measured how much 10,000 agents helped versus 1,000, because that test is too expensive.
Questions people ask
What is a token?
A token is a small chunk of text, often part of a word, that an AI model reads or writes. Researchers use total tokens as a rough measure of how much thinking or computation a job took.
What is a Millennium Prize problem?
In 2000, a math institute picked seven extremely hard unsolved problems and offered $1 million for a correct solution to each. Only one had been solved before this announcement, so each new solution is a major event.