Axiom Math 想阻止 AI 垃圾涌入数学

感觉 AI 每周都在攻克一道新的数学难题。上个月,OpenAI 发布了 719 篇手稿,声称解决了该领域一些最难的难题。但即便是用来确认这些结果的计算机检查,也可能出错。在经历了多年 AI 低质内容泛滥之后,人们有理由怀疑,数学会不会成为下一个。
Axiom Math 押注数学能够避开这种命运——不是靠放慢速度,而是靠站在更坚实的基础上。这家估值 16 亿美元的初创公司希望构建一个经过验证的基础,让 AI 智能体可以依赖它来编写软件,并最终从事科学研究。
周三,该公司宣布已对 Stacks Project 完成形式化验证,这是人类有史以来编写的规模最大的数学成果集合之一。数学家对 Stacks 往往怀有罕见的敬意——哥伦比亚大学数学家 Peter Woit 曾称“没有其他东西真正能与之相提并论”,并将该项目比作“著名的 Bourbaki 丛书”,后者是一套在 20 世纪从基础重建数学的系列著作。
Woit 在哥伦比亚大学的同事 Johan de Jong 于 2005 年发起了 Stacks Project,作为开源参考资料。此后数学家们不断为其贡献内容,从大学水平的代数一路构建到代数、几何与数论领域最前沿的成果。
该公司的 AI 智能体集群将这座文库约 7,654 页的内容转换为约 1,600 万行可由计算机校验的代码。
数学形式化是一种确保其底层结构可靠的方式。数学家把证明改写为一种名为 Lean 的编程语言代码,再由一个称为证明助手的程序逐行检查代码。如果证明通过检查,数学家就能确信它在逻辑上是可靠的。
工程师使用同样的思路——称为形式化验证——来确认芯片设计能否正常工作,而法国的研究人员则用它来捕捉计算家庭福利的软件中的漏洞。
对于像 Stacks Project 这样的库来说,检查每一条定理只是工作的一部分。Stacks 中的定理与定义相互关联,每一个都建立在前面的基础之上。就像家谱一样,Stacks 中的许多结果都可以追溯到之前的成果。
这种相互关联性让 Axiom 起初不确定这个项目能否做成。团队必须在约 1600 万行 Lean 代码中保持每一处连接完好无损。
Axiom CEO Carina Hong 对 TechCrunch 表示:“生活中,宏大的事物与精巧的事物就是无法兼得。[Stacks]是数学成果的集合,是数学的基础。你希望[定理]彼此一致,而不是自相矛盾。你希望这些定义之间不要互相冲突。”
这位 25 岁的 CEO 从斯坦福大学数学博士项目退学,于 2025 年 3 月创办了这家公司。3 月,Axiom 完成了由 Menlo Ventures 领投的 2 亿美元 A 轮融资。
“我们当时不确定 vibe coding 能否保住 Stacks 之美,”Hong 说。
为了完成这个项目,团队给智能体一个宽松的结构,让它们并行运行。Axiom 技术团队成员 Ashvin Swaminathan 表示:“对于 Stacks,我们施加了一种层级结构……[但]它并不是那种规定每个智能体做什么的完整层级。智能体自己在寻找任务,并识别出可以并行推进的工作。”
Axiom 还训练了自家模型,使其擅长数学,Hong 表示,这是大多数其他验证类初创公司没有采取的做法。起初,团队依靠内部一位 Lean 专家来教智能体形成人类式的品味。Hong 说,随着时间推移,这些智能体已经自己培养出了品味。
这一成果填补了形式化数学中最大的空白之一。“当你审视数学中已经被形式化的内容时,会发现存在一个巨大的黑洞。而这个黑洞覆盖了代数几何,”Swaminathan 说。
Swaminathan 说,在此之前,想要用 Lean 检查代数几何领域新成果的研究人员几乎没有什么可以依托的基础。“而现在,他们有了。”
Axiom 认为,每一个经过验证的结果都会成为一个构建模块,因此未来的证明可以依托已经过校验的数学,而不必从零开始。UCLA 数学家、菲尔兹奖得主 Terence Tao 更简单地概括了这一普遍道理:“你拥有的形式化证明越多,你能做出的形式化证明就越多。”
数学家并非 Axiom 唯一的目标受众。开发者长期以来一直难以保证自己的代码没有漏洞和安全隐患。随着 AI 智能体编写世界上越来越多的代码,这一问题愈发突出。
Axiom 押注未来的软件需要建立在经过验证的基础之上,就像如今的 Stacks 一样。
“我们相信,数学是你开始构建这些东西的领域,而且这是一项可迁移的能力,”Swaminathan 说。“智能体可以读取我们的代码、理解它,并在其之上构建新的基础设施。我们的 Lean 形式化如今已成为智能体可以据此构建的底层基础。”
Axiom 的智能体构建 Stacks 的方式——并行工作、自行挑选任务——正是 Hong 预期 AI 未来大部分工作会发生的方式。她认为,AI 智能体需要一个构建良好、可复用的知识体系,才能避免产出垃圾内容。
“你希望[智能体]拥有一个可复用的库,让它们在此基础上继续推进前沿,”Hong 说。“如果你一开始就以糟糕工程选择的方式灌入垃圾内容,那么它们就很难拓展这一边界。”
Hong 对这条路径信心十足,甚至发出了挑战。“我公开挑战所有编程公司,在这一规模上开展形式化工作。”
当你通过我们文章中的链接购买时,我们可能会获得一小笔佣金。这不会影响我们的编辑独立性。
Aditya Mehta 是 TechCrunch 报道 AI 的记者。他获得 Tarbell 人工智能新闻中心的资助,并就读于 UC Berkeley。你可以通过发送邮件至 [email protected] 联系 Aditya,或通过 Signal 上的 adymehta.74 发送加密消息联系他。
正在加载下一篇文章
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Axiom Math wants to keep AI slop out of mathematics)
Axiom Math wants to keep AI slop out of mathematics | TechCrunch
Image Credits:Axiom Math
Axiom Math wants to keep AI slop out of mathematics
8:16 AM PST · November 5, 2026
It feels like AI is cracking a new math problem every week. Last month, OpenAI released 719 manuscriptsclaiming solutions to some of the field’s hardest problems. But even the computer checks meant to confirm those results can slip. After years of AI slop, it’s fair to wonder whether math is next.
Axiom Math is betting math can avoid that fate, not by slowing down, but by standing on firmer ground. The $1.6 billion startup wants to build a verified foundation that AI agents can rely on to write software and, eventually, do science.
On Wednesday, the company announced it had formally verified the Stacks Project, one of the largest bodies of mathematics ever written by humans. Mathematicians often hold Stacks in rare regard — Columbia mathematician Peter Woit has claimed there’s “nothing else really comparable,” comparing the project to “the famous Bourbaki books,” a series that rebuilt math from its foundations in the 20th century.
Woit’s Columbia colleague Johan de Jong started the Stacks Project in 2005 as an open source reference. Mathematicians have contributed to it ever since, building out results from college-level algebra all the way up to the field’s most advanced ideas across algebra, geometry, and number theory.
The company’s fleet of AI agents translates the library’s roughly 7,654 pages into about 16 million lines of computer-checked code.
Formalizing math is a way of making sure the structure underneath it is sound. Mathematicians rewrite the proof as code in a programming language called Lean, and a program called a proof assistant checks the code line by line. If the proof passes, mathematicians can trust that it’s logically sound.
Engineers use the same idea, called formal verification, to confirm their chip designs work, while researchers in France have used it to catch bugs in software that calculates family benefits.
For a library like the Stacks Project, checking each theorem is only part of the job. The theorems and definitions in Stacks are interconnected, and each one builds on the ones before it. Like a family tree, many results in Stacks trace back to the ones that came before it.
That interconnectedness made Axiom unsure the project would work at first. The team had to keep every connection intact across about 16 million lines of Lean code.
“Things that are big and things that are delicate in life just don’t go together,” Axiom CEO Carina Hong told TechCrunch. “[Stacks] is a collection of mathematical results that is the foundation of mathematics. You want [theorems] to be consistent and not, like, paradoxical. You want the definitions to not fight with each other.”
The 25-year-old CEO dropped out of a Stanford math PhD program and founded the company in March 2025. In March, Axiom raised a $200 million Series A led by Menlo Ventures.
“We were not sure if vibe coding would be able to preserve the beauty of Stacks,” Hong said.
To pull off the project, the team gave its agents a loose structure and let them run in parallel. “With Stacks, we imposed a hierarchy … [but] it was not, like, a complete hierarchy of dictating what every single agent was doing,” said Ashvin Swaminathan, a member of Axiom’s technical staff. “The agents were finding tasks for themselves and identified things that they could work on in parallel.”
Axiom also trained its own models to be good at math, an approach Hong said most other verification startups haven’t taken. At first, the team relied on one of its in-house Lean experts to teach the agents a human sense of taste. Over time, Hong said, the agents have picked up taste on their own.
The result fills one of the biggest gaps in formalized math. “When you look at what has been formalized in mathematics, there’s this huge black hole. And that black hole covers algebraic geometry,” Swaminathan said.
Until now, researchers who wanted to check new work in algebraic geometry with Lean had little to build on, Swaminathan said. “Well, now they do.”
Axiom argues that each verified result becomes a building block, so future proofs can lean on checked math instead of starting from scratch. Terence Tao, the UCLA mathematician and Fields Medal winner, has made the general case more simply: “The more formalized proofs you have, the more formalized proofs you can make.”
Mathematicians aren’t Axiom’s only intended audience. Developers have long struggled to guarantee that their code is free of bugs and security holes. That problem grows as AI agents write more of the world’s code.
Axiom is betting that the software of the future will need to rest on verified foundations, the way Stacks now does.
“We believe that mathematics is the domain where you start building these things and that it’s a transferable skill,” Swaminathan said. “Agents can read our code, understand it, and build new infrastructure on top of it. Our Lean formalization is now the substrate that agents can build upon.”
The way Axiom’s agents built Stacks, working in parallel and picking their own tasks, is how Hong expects much of AI’s future work to happen. AI agents, she argues, will need a well-built, reusable body of knowledge to keep from producing slop.
“You want [agents] to have a reusable library for them to build on and keep pushing the frontier,” Hong said. “If you slop drop in a way that’s bad engineering choices in the beginning, then it’s very hard for them to grow that boundary.”
Hong is confident enough in that approach to issue a dare. “I openly challenge all the coding companies to work on formalization at this scale.”
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

Aditya Mehta is a reporter at TechCrunch covering AI. He’s supported by the Tarbell Center for AI Journalism and attended UC Berkeley. You can contact from Aditya by emailing [email protected] or via encrypted message at adymehta.74 on Signal.
Loading the next article
Error loading the next article
原文为英文,正文由自动抓取并保留原文。
原文发布于 TechCrunch AI(Axiom Math wants to keep AI slop out of mathematics),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。