Scaling Test Time Compute to Multi-Agent Civilizations — Noam Brown, OpenAI

In this episode of the Latent Space podcast, Alessio and Swyx interview Noam Brown from OpenAI, focusing on his work on AI, particularly the Cicero diplomacy-playing AI, and the O-series models. They discuss how playing Diplomacy himself has changed since working on Cicero, the challenges of detecting AI in online games, and the safety aspects of AI persuasion. The conversation shifts to O-series models, deep research, and the "thinking fast and slow" analogy in AI. Noam shares his coding stack, emphasizing the use of Windsurf and Codex, and expresses his vision for the future of AI, including multi-agent systems and the importance of test-time compute. The discussion also covers the complexities of self-play in AI, the role of world models, and touches on robotics and the potential of humanoid versus non-humanoid designs.