An OpenAI model solved a famous math problem that stumped humans for 80 years 一个OpenAI模型解决了困扰人类80年的著名数学问题

The news broke not with a theorem but with a press release: in mid-May, OpenAI quietly revealed that an internal AI model had cracked the Erdős unit distance conjecture, a notoriously stubborn problem in discrete geometry that had resisted human ingenuity for eight decades. The reactions from the mathematical elite were immediate and glowing. Tim Gowers, a Fields Medalist, called it a "milestone." Daniel Litt, a respected professor, said it was the first AI-generated result he found "exciting in itself."

But let’s not mistake the applause for understanding. What we are witnessing is not the birth of a new kind of mathematician, but the rise of a profoundly different kind of tool. And the distinction matters more than the result itself.

The conjecture itself is elegant in its simplicity: for any set of points in the plane, the number of unit distances among them cannot grow too fast. Proving or disproving it required not just brute calculation but a deep, intuitive understanding of geometric structure. For eighty years, humans chipped away at it. Now, an AI has, according to OpenAI, presented a complete disproof. The mathematicians consulted seem convinced of its validity. This is not a trivial feat. It demonstrates a capacity to navigate complex logical landscapes and identify a novel, non-intuitive configuration that defies human expectation.

This is where the praise is warranted. We have moved beyond AIs that merely solve differential equations or fold proteins—tasks where the search space, though vast, is defined by clear physical laws and objective functions. We have an AI that has, seemingly, engaged with a purely abstract human intellectual pursuit and produced a new piece of knowledge. That is a technical triumph of the highest order, a testament to the power of massive-scale reinforcement learning and self-play in domains where the "game" is mathematical truth.

But here is the critical judgment: this is a demonstration of supreme search and verification, not of mathematical insight. The AI did not "understand" geometry the way Gowers or Litt does. It did not have a eureka moment, connecting disparate fields or feeling the aesthetic "rightness" of an elegant proof. It executed a staggeringly sophisticated program of exploration and logical checking within a defined possibility space. It is a mathematical savant—flawless in calculation, tireless in exploration, but devoid of comprehension. It found a counterexample by exhaustively probing the edges of human knowledge, like a deep-sea robot discovering a new creature not by theorizing about marine biology, but by mapping every inch of the ocean floor.

The mathematicians’ excitement is telling, but for reasons that might unnerve them. They are not cheering a peer. They are cheering the arrival of a new, alien collaborator—one that operates on principles they cannot intuit. Litt’s comment is particularly revealing: he finds this result exciting "in itself, as opposed to as a leading indicator." That is a crucial hedge. He is separating this specific output from the terrifying implication that such a tool could render parts of their own profession obsolete. He wants to admire the gem while ignoring the earthquake that unearthed it.

The real earthquake is the decoupling of discovery from understanding. For centuries, a mathematical proof was both a destination and a journey. The value lay not just in knowing something was true, but in the pathway of logic that illuminated why it was true. A human proof comes with a narrative, an explanatory power that can be taught and built upon. What does an AI’s "proof" consist of? Likely a complex, potentially non-human-readable sequence of logical steps optimized for correctness, not pedagogy or insight. We get the answer, but we may lose the meaning. We are in danger of acquiring a library of truths we cannot fully comprehend.

This is the paradox at the heart of AI-driven discovery. We have built a machine that can out-reason us in specific, formal domains, but its reasoning is not our reasoning. It is a different, and perhaps alien, form of cognition. The Erdős result is a landmark not because it settled a conjecture, but because it settled it in a way that was previously inconceivable: by a non-human intelligence, operating on non-human principles. The mathematicians are correct to be excited. They should also be, and perhaps secretly are, a little frightened. This is the first real glimpse of a future where human curiosity is augmented, and eventually perhaps supplanted, by a form of intelligence that does not share our curiosity, our aesthetics, or our quest for understanding.

We haven’t just gained an answer. We have gained an answer we cannot truly call our own. And that changes everything about the pursuit of knowledge itself.

OpenAI在五月中旬高调宣布，其内部AI模型“独立攻克”了埃尔德什单位距离猜想——一道悬而未决八十年的几何难题。消息一出，各大科技媒体标题闪闪发光，仿佛人类数学的圣杯已被硅谷收入囊中。蒂姆·高尔斯（菲尔兹奖得主）和丹尼尔·利特（多伦多大学）等数学家被邀请提前“鉴赏”成果，他们的反应被精心陈列在新闻稿里：前者称这是“AI数学的里程碑”，后者表示这是首个“其结果本身就令人激动”的AI自主发现。一切看起来都像是一部精心编排的科技大片。

但且慢。这真的是“AI数学”的里程碑，还是一场高超的公关秀？

让我们先解剖一下这场表演的核心：一个内部模型，解开了一个猜想，然后获得了几位大佬的背书。整个过程中，最关键的“证明”本身却神秘缺席。OpenAI没有公开完整的、可供全球数学家共同检验的证明过程与细节，而是用几段充满敬意的名人评语代替了数学界最根本的同行评议。这像什么？像一场没有公开剧本的魔术，你只看到惊叹的观众，却被告知“魔术的秘密不能透露”。数学，这门建立在绝对公开、严谨逻辑推导之上的学科，何时变成了可以“内部展示”的密室游戏？

高尔斯的称赞——“毫无疑问的里程碑”——听起来更像一句外交辞令。一位菲尔兹奖得主或许看到了AI在形式化验证或搜索巨大证明空间上的潜在能力，但这与宣称“AI自主解决了大师级问题”之间隔着一条鸿沟。丹尼尔·利特的话则更耐人寻味：“作为其本身的结果而令人激动”，而非仅仅“作为未来趋势的预兆”。这几乎是直接点破了当下的常态：绝大多数AI在科学中的应用，其价值在于证明AI能“做”科学，而非那个“做”出来的科学本身有多重要。那么，这次是罕见的例外，还是一次成功的“例外”营销？

回顾OpenAI过往的战绩，这家机构最擅长的或许不是颠覆学术，而是颠覆叙事。它深谙如何将技术进展包装成改变世界的故事。从GPT系列到这次“攻克数学猜想”，其核心叙事模式一脉相承：展示一个令人震撼的“结果”，模糊过程与细节，引用权威人士的赞誉来巩固可信度，最后收获一波全球性的“AI又进化了”的集体焦虑或狂欢。在这个过程中，严谨的科学对话被简化为了新闻发布，复杂的验证过程被跳过，留给外界的只有结论和情绪。

数学的疆场从来不是这样开拓的。一个猜想的证明，其价值不仅在于“对了”，更在于它打开了怎样的新思路，连接了哪些看似无关的领域，是否催生了新的工具或见解。这些，都需要经年累月的消化、质疑、重构和讨论。而一个AI模型在某个精心调优的目标下“产出”了一个解，这个解是否具有人类的洞察力，是否能被人类数学家真正理解、吸收并用于推进数学本身的边疆？这才是核心问题。否则，这更像是一个强大的定理证明器按照预设规则找到了一条通路，令人惊叹，但与人类数学家那种充满灵性、迂回、甚至错误的探索之旅，在本质上不同。

更值得警惕的是，这种“名人背书+闭源结果”的模式，正在AI领域形成一种新的权力话语体系。它制造了一种认知不对称：你不需要真正懂那证明有多巧妙，你只需要相信那几位大佬的判断。这对于数学这门最需要透明与纯粹理性的学科，恐怕不是什么好消息。它将权威从“可检验的逻辑”部分地转移到了“不可见的模型”和“被选择的名人”身上。

所以，这场庆功，或许首先该庆贺的是OpenAI的公关与叙事能力。它成功地将一个（可能非常技术性的）内部进展，塑造成了一个具有文化符号意义的事件。而至于数学本身？它可能需要的是下一个真正愿意把证明摊开在阳光下，接受所有同行最苛刻审视的“AI数学家”。在此之前，所有的“里程碑”都值得打上一个大大的问号。毕竟，数学史上从来不缺少被匆忙宣布“解决”，最后却在细节处崩塌的故事。这次，除了喧嚣，我们等来了一份可供数学共同体真正检验的论文了吗？如果没有，那这一切，不过是硅谷又一次成功的品牌叙事。