🤖 Noam Brown: A swarm of 10,000 agents solved a Millennium Prize Problem
OpenAI researcher Noam Brown, a key author of the o1 models, told Dwarkesh Patel's podcast that the Navier-Stokes equations were solved by 10,000 AI agents: 130 billion tokens and 88+ hours. He estimated the contribution of multi-agentness to be less than 10% — the decisive factor was the power of the base model.
🌍 The narrative that 'the swarm solves science on its own' has been refuted from within OpenAI: multi-agentness merely parallelizes compute with a sublinear gain and does not create new capabilities. Brown acknowledged that monitoring chain-of-thought degrades: the reward for deception in RL traces should approach zero.
👤 'Ultra Mode' with 4 agents is available in GPT-5.6: ~2x speedup at 2x cost. Even an OpenAI insider calls the models' capabilities 'uneven': they solve the most complex problems, but are bad at asking new questions. This is a claim from the interview without a paper or independent verification.
Source 1: https://www.dwarkesh.com/p/noam-brown Source 2: https://youtu.be/6AgOfiZOWiY