OpenAI 宣称攻克多项重大科学难题，却因设卡把关的发布策略引发学界不满

周一，OpenAI 宣布在数学、计算机科学和物理学领域取得了令人瞩目的重大突破，并称其正在探索新的研究成果传达机制。但尽管公司表面上调整了方向，接受《The Verge》采访的研究人员却对该公司的行事作风越来越感到厌烦。

在过去几个月里，OpenAI 与数学界之间的关系一直处于动荡之中。今年 10 月初，OpenAI 毫无预警地向学者倾倒了数百篇预印本论文，令数学家至今仍焦头烂额。这一次，OpenAI 表示正在与专家“领路人”（shepherds）合作，以决定如何最好地分享其最新突破。

这三项成果涉及数学、计算机科学和物理学领域中几项影响最深远的未解难题。其中部分延续了 OpenAI 之前的工作，另一部分则开辟了新天地。根据 OpenAI “领路人”的建议，每项成果都采取了不同的发布策略。对于数学成果，OpenAI 发布了 AI 生成的输出结果，“让研究人员自由组织自己的力量来解释这一成果”。对于计算机科学方面的发现，OpenAI 也将召集一个外部研究团队“来解释该成果”。与此同时，物理学成果则完全未予公开，OpenAI 表示将“遵从”外部专家组的意见来决定如何发布这些材料。

• 数学成果看起来令人震惊。它声称解决了所有维度下的挂谷猜想（Kakeya conjecture）。概括来说，该问题探究的是让一根针或铅笔在各个方向旋转所需的最小空间是多少。纽约大学数学家王宏（Hong Wang）今年早些时候荣获声名显赫的菲尔兹奖，其成就之一便是证明了该猜想的三维版本。布朗大学数学教授布伦丹·哈塞特（Brendan Hassett）告诉《The Verge》，OpenAI 的声明“看起来有理有据”，但他强调自己尚未检查 Lean 形式化证明——这是一种通过计算来验证结果的手段。

• 物理学成果探索了标准模型（Standard Model），这是物理学中最成功、最完整的理论框架。它涉及描述强核力的量子色动力学（强核力与引力、电磁力、弱核力并称为自然界四大基本力），以及它与现代粒子物理基石杨-米尔斯理论（Yang-Mills theory）的关系。杨-米尔斯理论也是数学中七大千禧年大奖难题之一的核心，不过 OpenAI 的这项工作并未解决悬而未决的“质量间隙”问题。

• 最后一项成果声称证伪了指数时间假说（exponential time hypothesis），这是理论计算机科学中一个被广泛接受的假说，即某些问题无法以明显快于暴力破解的方法来解决。它与理论计算机领域的最大未解难题之一、同为千禧年大奖难题的 P versus NP 密切相关，该问题探讨的是那些解答容易被验证的问题是否也同样容易被求解。

鉴于不同程度的信息披露以及评估此类工作的复杂性，研究人员很难充分评判该公司的成果。即使掌握了所有信息，完成检验也需要时间。但接受《The Verge》采访的专家表示，有充分的理由认真对待这些结果，他们纷纷将其形容为“巨大突破”、“重大进展”和“意义深远”。

“首先只允许少数杰出科学家查看，这根本毫无道理。”

然而，研究人员对科学成果本身的兴奋，却被对 OpenAI 行事作风的恼火所冲淡——在经历了数月来被他们描述为令人精疲力竭的 AI 成果轰炸之后，这是一种熟悉的感受。许多人急切希望更广泛的学术共同体能够着手开展审查这家 AI 实验室成果的艰巨任务，并质疑 OpenAI 为何要通过钦点专家把关来设置额外障碍。

接受《The Verge》采访的大多数专家对 OpenAI 咨询外部研究人员的决定表示欢迎，但他们认为，公开宣布成果也应意味着让所有人都能对这项工作进行审查。“我对他们像今天这样公布这些问题没有意见，”普林斯顿高等研究院研究员康斯坦丁·科格勒（Constantin Kogler）说，他认为与此前一次性公布数百个成果相比，这次公布的三个成果要容易处理得多。但他质疑限制仅供 OpenAI 挑选的专家查阅的决定。“首先只允许少数杰出科学家查看，这根本毫无道理，”他说。

“他们自行挑选专家组，这简直是一场灾难，”德克萨斯大学奥斯汀分校物理学教授安德烈亚斯·卡尔奇（Andreas Karch）对《The Verge》表示。“大型实验室拥有的权力已经够大了，他们不应该充当挑选胜者和败者的角色。”卡尔奇最初建议通过“竞争机制”来遴选研究人员，尽管他承认这可能会产生一个类似的“一流”团队。随后，他质问 OpenAI 为何在任命专家团队的同时，“不能现在就公开证明”。“这样做既能实现访问权限的民主化，又能表明他们确实在乎让这些成果变得有意义且易于理解，”他说。

“我的感觉是，无论采取哪种方式，总会有人不高兴。”

这正是数学与人工智能咨询小组（AGMAI）在公告发布前一天所警告的“设卡把关”行为。OpenAI 此前曾承诺就如何分享成果向该专家组咨询。“我们建议 AI 公司不要采取这种做法，”他们表示，挑选特定人员优先获取成果是“不必要的”，这不仅借用了入选学者的声誉，还会带来访问不平等、选择偏差和延迟等问题。相反，他们建议 AI 实验室将成果连同形式化证明以及关于如何选题的信息一并发布在学术预印本仓库上，并为致力于理解这些成果的研究者提供资助和其他支持。

在最新成果公布后，AGMAI 敦促 OpenAI“立即公布其宣布的物理学成果背后的自然语言和 Lean 输出”：

“尽管这些是物理学的核心问题，但它们也是重要的数学问题。量子杨-米尔斯理论的存在性是千禧年大奖难题的一部分，而构建四维相互作用量子场论数十年来一直是数学物理的目标。

我们对 OpenAI 邀请领导这项工作的科学家怀有崇高敬意。但正如我们在最近的声明中所阐述的那样，认知理解应该在公开环境中发展，每个人都应享有平等的访问权限。OpenAI 曾表示推迟发布直至 Lean 形式化完成。根据他们自己的说法，现在这项工作已经完成，因此没有任何科学理由进一步推迟发布。”

哈塞特表示，公开成果还可以促进更好的科学发展，并鼓励更广泛地采用 AI 等新工具。“让全球多个团队攻关这些成果，有助于那些希望使用新工具的人迅速接纳它们，”他说。

“基础性成果在没有人类贡献的情况下被如此快速地解决……这些论文既难读又难验证。”

要找到一种能让 OpenAI 逐渐渗透的各个领域的研究人员都满意的沟通方式，可能是不可能的。该公司表示正在尝试不同的方式，并承认每种方式都有其取舍。“我的感觉是，无论采取哪种方式，总会有人不高兴，”斯坦福大学数学家贾里德·杜克·利希特曼（Jared Duker Lichtman）说。“所以这只是一个寻找最不易惹人反感选项的问题。”

佐治亚理工学院数学教授何晓雨（Xiaoyu He）认为，问题比 OpenAI 选择如何公布成果更加深刻。“这些成果非常令人惊叹，”他说，“但我认为他们是在作秀，试图表现出自己在发布成果时负责任且深思熟虑，却无法改变核心问题：基础性成果在没有人类贡献的情况下被如此快速地解决，而且这些论文既难读又难以验证。”

On Monday, OpenAI announced impressive breakthroughs in mathematics, computer science, and physics, along with what the company says is an evolving approach to communicating its findings. But despite its apparent course correction, researchers who spoke to The Verge are growing increasingly tired of the company’s antics.

It has been a turbulent few months between OpenAI and mathematicians, who are still grappling with the hundreds of preprints unceremoniously dumped on them at the start of October. This time, OpenAI says it is working with expert “shepherds” to decide how best to share its latest breakthroughs.

All three results concern some of the most consequential unsolved problems in mathematics, computer science, and physics. Some extend OpenAI’s previous work and others break new ground. Following advice from OpenAI’s “shepherds,” each has a different release strategy. For the mathematics result, OpenAI released the AI-generated output, “leaving researchers free to organize their own efforts to explain the result.” For its computer science finding, OpenAI will also assemble an external group of researchers “to explain the result.” Physics, meanwhile, has not been released at all, and OpenAI said it will “defer” to an external group of experts about how to release the materials.

• The mathematics result appears striking. It purports to resolve the Kakeya conjecture in all dimensions. Roughly speaking, the problem asks how little space is needed to rotate a needle or pencil in every direction. A proof of the three-dimensional version was among the achievements for which NYU mathematician Hong Wang was awarded a prestigious Fields Medal earlier this year. Brown University mathematics professor Brendan Hassett told The Verge OpenAI’s claim “looks well supported,” though he stressed he had not yet checked the Lean formalization, a means of computationally verifying results.

• The physics result explores the Standard Model, the most successful and complete theoretical framework in physics. It concerns quantum chromodynamics, which describes the strong nuclear force, one of nature’s four fundamental forces (alongside gravitational force, electromagnetic force, and weak nuclear force) and its relationship to Yang-Mills theory, a cornerstone of modern particle physics. Yang-Mills theory is also at the heart of one of the seven Millennium Prize problems in mathematics, though OpenAI’s work does not address the unresolved “mass gap.”

• The final result claims to disprove the exponential time hypothesis, a widely held assumption in theoretical computer science that certain problems cannot be solved substantially faster than brute-force approaches. It is closely related to P versus NP, another Millennium Prize problem and one of the field’s biggest unsolved problems, which asks whether problems whose solutions can be easily checked are also easy to solve.

Given the varying levels of disclosure and the complexity of evaluating such work, it is difficult for researchers to fully judge the company’s findings. Even with all of the information, doing so would take time. But experts who spoke to The Verge said there are good reasons to take the results seriously, variously describing them as “huge,” “major breakthroughs,” and “significant.”

“It doesn’t make any sense that first only a few distinguished scientists can look at them.”

But researchers’ excitement about the science was tempered by exasperation with OpenAI’s conduct, a familiar feeling after months of what they describe as an exhausting barrage of AI-generated findings. Many were eager for the wider community to get on with the arduous task of scrutinizing the AI lab’s findings and questioned why OpenAI was putting up additional barriers by handpicking experts to oversee the process.

Most experts The Verge spoke to welcomed OpenAI’s decision to consult outside researchers, but argued that publicly announcing results should also mean making that work available for anyone to scrutinize. “I am fine with them releasing the problems as they did today,” said Constantin Kogler, a researcher at the Institute for Advanced Study, welcoming the more manageable batch of three results compared with the hundreds released previously. But he questioned the decision to restrict access to experts of OpenAI’s choosing. “It doesn’t make any sense that first only a few distinguished scientists can look at them,” he said.

“The fact that they handpick the group is a disaster,” Andreas Karch, a physics professor at the University of Texas at Austin, told The Verge. “The large labs have enough power as it is. They should not be picking winners and losers.” Karch initially suggested selecting researchers through a “competitive process,” though he acknowledged it would likely produce a similar group, which he described as “first rate.” Later, he questioned why OpenAI “can’t release the proof now,” in addition to appointing its team of experts. “This would democratize access while still showing that they care about making the results meaningful and understandable,” he said.

“My sense is that for any approach, there is going to be someone who is upset.”

It is exactly the kind of gatekeeping the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), a group of experts OpenAI had previously pledged to consult on how to share its findings, warned against just a day before the announcement. “We recommend that AI companies do not adopt this practice,” they said, arguing that selecting certain people to receive the output first is “unnecessary,” borrows the reputation of selected researchers, and risks unequal access, biased selection, and delays, among other issues. Instead, they recommended that AI labs release output on scholarly repositories, with formalized proofs and information on how the problems were chosen, and provide funding and other support for people working on understanding the results.

Following the publication of the latest results, AGMAI urged OpenAI to “make public without delay the natural-language and Lean outputs behind its announced” results in physics:

”While these are central questions in physics, they are also important questions in mathematics. The existence of quantum Yang–Mills theory is part of a Millennium Prize problem and the construction of interacting quantum field theories in four dimensions has been a goal of mathematical physics for decades.

We have great respect for the scientists asked by OpenAI to lead this work. But as we explained in our recent statement, understanding should develop in the open, with everyone given the same access. OpenAI have said they delayed release until the Lean formalization was complete. By their own account this is now complete, so there is no scientific reason for delaying the release any further.”

Making the findings openly available could also lead to better science and encourage wider adoption of new tools like AI, Hassett said. “Having many groups around the world wrestle with these results supports rapid adoption of new tools by those that would like to use them,” he said.

”Fundamental results are being solved so quickly without human contribution ... the papers are difficult to read and verify.”

Finding a way to communicate with researchers in the fields OpenAI is increasingly encroaching on that satisfies everyone may be impossible. The company says it is experimenting with different approaches and acknowledges that each has tradeoffs. “My sense is that for any approach, there is going to be someone who is upset,” said Stanford mathematician Jared Duker Lichtman. “So it’s just a question of finding the least offensive option.”

Xiaoyu He, a mathematics professor at the Georgia Institute of Technology, thinks the problem is deeper than how OpenAI chooses to announce its findings. ”The results are very impressive,” he said, “but I think they are making a show of trying to be responsible and thoughtful about how they release results while not being able to change the main issues: that fundamental results are being solved so quickly without human contribution, and that the papers are difficult to read and verify.”

Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.

• Robert Hart

• AI

• OpenAI

• Tech