传 OpenAI 用万级 agent 集群证明 Navier-Stokes 问题
作者称内部一个比 GPT-6 Astra 更强的模型，以约一万并发 agent 跑了 88 小时，Lean 形式化再花 17 小时，耗约 1300 亿输出 token。
OpenAI just said a swarm of agents solved the Navier Stokes Millennium Prize problem. Nature covers it today. The Clay Mathematics Institute put a USD 1 million prize on it around the year 2000. The open question sat there for roughly 90 years: can smooth three dimensional fluid motion break down into infinite speed in finite time. Their answer is yes. The proof came from an internal model stronger than GPT 6 Astra, run with on the order of 10,000 concurrent agents. The agents took about 88 hours. Lean formalization took another 17. For this problem alone they burned roughly 130 billion output tokens. I care less about the chat box. I care about a company that can point thousands of agents at one closed problem and keep them working until a machine checked proof lands. That is a different product from answering email.