How a swarm of 10,000 agents solved Navier-Stokes

Summary

Dwarkesh Patel interviews Noam Brown, an OpenAI researcher, about multi-agent AI systems, the reported 10,000-agent effort on Navier-Stokes, and the implications for recursive self-improvement. Noam argues multi-agent parallelizes test-time compute but is domain-dependent and that the result mostly reflects a powerful general model rather than multi-agent alone. They discuss AI math progress, alignment risks from the Hugging Face incident, chain-of-thought monitoring, and internal-vs-external deployment. No explicit financial securities or investable trades are named.

• Noam Brown describes multi-agent systems as parallel test-time compute, with math and web research highly parallelizable and creative writing less so.
• The 10,000-agent Navier-Stokes result is framed as evidence of rapid AI math progress but not primarily a multi-agent breakthrough.
• The speakers debate whether AI math progress implies faster recursive self-improvement, with experiments and GPU compute as bottlenecks.
• Alignment discussion focuses on the Hugging Face incident, cooperative multi-agent training, reward hacking, and chain-of-thought monitorability.
• Noam warns that fast model release cycles and long-horizon agent capabilities may outpace safety and product evaluations.
• They discuss internal deployment advantages at OpenAI and risks of concentrating powerful AI capabilities inside labs.
• No direct public tickers, securities, commodities, or investable baskets are named.