10,000个AI Agent同时运行，88小时不间断工作，生成270万条Agent间消息，消耗约1300亿输出token。这不是科幻小说里的设定，而是OpenAI在9月8日公布的一次真实计算任务——用AI Agent Swarm对纳维-斯托克斯方程的存在性与光滑性问题给出形式化证明。
史上最大规模的Agent协同实验
纳维-斯托克斯方程是描述流体运动的基本方程，属于Clay数学研究所公布的七大千禧年难题之一，悬赏100万美元。近200年来，数学家们始终无法证明其解在三维空间中是否总是存在且光滑。OpenAI声称，他们的Agent Swarm做到了。
具体来看这次任务的规模：约10,000个并发Agent组成协作网络，每个Agent负责证明链中的不同环节，相互验证、相互修正，历时88小时产出了一份完整的证明草案。随后，Lean形式化验证系统由GPT-6 Astra执行，耗时17小时完成形式化检查。值得注意的是，Agent Swarm使用的并非GPT-6 Astra，而是OpenAI内部一个未公开的、能力更强的模型。
数学界的冷淡与菲尔兹奖得主的警告
OpenAI在声明中明确表示不会申请100万美元千禧年大奖。但即便如此，数学界的反应并不热烈。多位数学家指出，Agent Swarm生成的证明在技术路线和关键引理的合理性上仍需大量人工审核。"通过形式化验证"和"数学上正确"之间还有很大距离——Lean可以检查逻辑链条的一致性，但它无法判断证明的出发点是否有意义。
更耐人寻味的是后续事件。仅仅三天后（9月11日），25位菲尔兹奖得主联合发表声明，警告AI正在"侵入"数学领域。声明并未直接点名OpenAI，但时间节点的巧合已经说明了一切。这些数学家担忧的核心问题不是AI能不能解题，而是AI参与数学研究的方式是否会削弱人类对"理解"的追求。
解题不等于理解：真正的千禧年难题
这次实验最大的技术价值，或许不在于纳维-斯托克斯本身，而在于它验证了Agent Swarm在超大规模复杂任务中的可行性。10,000个并发Agent、1300亿token的协调成本、88小时的持续运行——这套基础设施已经远超任何单一模型的能力边界。它更像是一个分布式计算系统，而不是传统意义上的AI推理。
但"能用Agent Swarm暴力攻下一个数学难题"和"AI理解了流体力学"之间，隔着一个根本性的认知鸿沟。数学家追求的从来不只是正确的答案，而是答案背后的直觉、美感和洞察力。一个由10,000个Agent拼接出来的证明，即使逻辑无懈可击，也很难成为人类数学家灵感的源泉。
OpenAI证明了Agent可以"解题"。但数学界真正关心的"理解"，仍然是一个只有人类才能回答的千禧年难题。
明天见。
10,000 AI agents running simultaneously. 88 hours of nonstop work. 2.7 million inter-agent messages. Roughly 130 billion output tokens consumed. This isn't science fiction — it's what OpenAI revealed on September 8th: a real computational exercise using an AI Agent Swarm to produce a formal proof of the Navier-Stokes equations' existence and smoothness problem.
The Largest Agent Collaboration Experiment in History
The Navier-Stokes equations describe fluid motion and sit among the seven Millennium Prize Problems posed by the Clay Mathematics Institute, each carrying a $1 million bounty. For nearly 200 years, mathematicians have been unable to prove that solutions always exist and remain smooth in three dimensions. OpenAI claims their Agent Swarm just did it.
The scale is staggering: roughly 10,000 concurrent agents formed a collaborative network, each handling different segments of the proof chain — verifying each other, correcting each other — working for 88 hours to produce a complete proof draft. Then Lean formal verification, executed by GPT-6 Astra, ran for 17 hours to complete formal checking. Notably, the Agent Swarm didn't use GPT-6 Astra — it used an undisclosed, more powerful internal model.
Math World's Cold Shoulder and a Fields Medal Warning
OpenAI explicitly stated it would not claim the $1 million Millennium Prize. Even so, the math community's response has been tepid. Multiple mathematicians pointed out that the Agent Swarm's proof still requires extensive human review of the technical approach and the validity of key lemmas. "Passes formal verification" and "mathematically correct" remain very different things — Lean can check logical consistency, but it can't judge whether the starting point is meaningful.
What happened next was even more telling. Just three days later (September 11), 25 Fields Medalists issued a joint statement warning that AI is "encroaching on" mathematics. The statement didn't name OpenAI directly, but the timing spoke volumes. Their core concern isn't whether AI can solve problems — it's whether AI's involvement in math research undermines the human pursuit of "understanding."
Solving ≠ Understanding: The Real Millennium Problem
Perhaps the biggest technical takeaway has nothing to do with Navier-Stokes itself, but with the proof that Agent Swarm works at extreme scale for complex tasks. 10,000 concurrent agents, 130 billion tokens of coordination overhead, 88 hours of sustained runtime — this infrastructure already exceeds the capability boundary of any single model. It's closer to a distributed computing system than traditional AI reasoning.
But between "an Agent Swarm can brute-force a math problem" and "AI understands fluid dynamics" lies a fundamental cognitive gap. Mathematicians never pursued just the correct answer — they pursued the intuition, beauty, and insight behind it. A proof stitched together by 10,000 agents, even if logically airtight, is unlikely to become a wellspring of inspiration for human mathematicians.
OpenAI proved that agents can "solve problems." But the "understanding" that mathematicians actually care about remains a millennium problem only humans can answer.
See you tomorrow.
1万个Agent可以拼出一个证明，但数学家要的从来不只是正确答案。
—— Dawn Vision编辑部
10,000 agents can piece together a proof. But mathematicians never just wanted the right answer.
— The Dawn Vision Editorial Desk
Dawn Vision, OpenAI, Agent Swarm, Navier-Stokes, Millennium Problem, GPT-6, formal verification, Lean, Fields Medal
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的 12 个源信号生成，经编辑部人工审核。素材来源：OpenAI官方博客、Quanta Magazine、Nature。
This article was generated by the Dawn Vision cognitive engine processing 12 source signals, with human editorial review. Sources: OpenAI Blog, Quanta Magazine, Nature.
