Sam Altman reveals OpenAI barely understood how to scale AI agent swarms even as they tackled some of mathematics' hardest problems. He expects major progress in swarm coordination within 3 months and believes the same approach could automate AI research itself.