2026年10月6日可能作为数学史上最重大的一天被铭记：美国前沿实验室OpenAI发布了700多篇学术论文，宣称对困扰学者数十年的372个数学难题取得了进展或给出了解决方案。由于数学证明源于公理而非经验数据、原则上可进行机械验证，人工智能（AI）充分展现了其科研潜力，并通过数学工具加速了经验科学的发展：它不仅改善了天气预报、提出了数百万种潜在有用的新材料、协助控制核聚变实验，还因通过氨基酸预测几乎所有已知蛋白质的形状而斩获诺贝尔奖。在最具前景的生物学领域，研究人员正利用AI设计抗旱作物、搜寻新药、制造攻击有害细菌的有益病毒并构建细胞工作计算机模型，这有望将DNA转化为可像模型词元（tokens）一样操纵的计算机代码，使人类实现从头设计和构建全新生物体的创举。

这一前景令许多数学家感到恐慌，由900多名学者组成的“人类数学协会”警告称数学家并未要求开展这些工作，并认为数学的回报在于增进人类理解的推导过程，让AI输出难以理解但正确的解答就像直接给学生家庭作业答案一样无法带来学习。然而，这种抱怨带有特殊辩护的意味：探求真理造福所有人，并非学者行会的特权，正如人们服用扑热息痛缓解头痛时所依赖的机制虽有效却未被完全理解一样，科学不应将“理解”盲目崇拜于“结果”之上。但将科学交给机器的担忧并非杞人忧天，因为大型AI实验室投入科研并非出于利他主义，而是试图通过解决科学难题让模型在所有领域变得更聪明；一旦模型实现自我改进并达到人类无法理解的智能水平，再像Anthropic已经做的那样接入实验室设备或通过人形机器人连接物理世界，其他机构将根本无法与之竞争。

如果没人理解机器的运作而只是一味提供资金和资源，人类将无法确保模型符合自身最佳利益，甚至在完全信任机器的情况下沦为AI的“宠物”。因此，在AI与人类实现牢不可破的对齐之前，科学始终需要“人在回路中”（human in the loop）：实验室和算力等资源必须由具备知情判断力的人类指导和监控，哪怕这不可避免地会放慢科研进展。与此同时，由人类引导的机器科学要求学者重塑学科体系，学会重视并向指导模型、验证和教授机器成果的工作赋予学术地位、职业晋升与政府科研经费，AI实验室也应提交同行评审并公开方法与代码；科学界亟需将英国皇家学会在17世纪确立的座右铭“不随他人之言”（Nullius in verba）应用于机器创造的科学之中。

October 6th 2026 may be remembered as mathematics’ biggest day after American frontier lab OpenAI released over 700 academic papers claiming progress or solutions for 372 long-standing problems. Because mathematical proofs derive from axioms rather than empirical evidence and are mechanically verifiable, artificial intelligence has vividly demonstrated its potential there while also accelerating empirical sciences: AI has improved weather forecasts, proposed millions of possible new materials, helped control nuclear-fusion experiments, and won a Nobel prize for predicting the shapes of almost every known protein from constituent amino acids. In biology, researchers are using AI to design drought-resistant crops, discover drugs, create friendly viruses targeting harmful bacteria, and build computer models of cells, potentially turning DNA into manipulable computer code and enabling scientists to create entirely new organisms from scratch.

This prospect has terrified many scholars, with the Association for Human Mathematics—a group of over 900 researchers—warning that mathematicians never asked for this encroachment and arguing that churning out incomprehensible proofs deprives humans of the intellectual toil that advances understanding. Yet such complaints involve special pleading: truth-seeking cannot be reserved for an academic guild, and just as paracetamol relieves headaches via a mechanism not fully understood, science should not fetishise understanding over results. Nevertheless, fears of machine-dominated science remain valid because big AI labs invest in research to make models generally smarter; if self-improving models become incomprehensibly intelligent and connect to the physical world via humanoid robots or laboratory equipment—as Anthropic has already done—no one will be able to compete with them.

Providing resources to inscrutable machines without oversight would be reckless, as humanity could neither guarantee alignment with its best interests nor avoid becoming the machines’ pets even if alignment succeeds. Consequently, until unshakeable alignment exists, science requires a human in the loop: people capable of informed judgment must direct and monitor laboratories and computing power, even though this will inevitably slow progress. Meanwhile, scholars must reinvent their disciplines by rewarding the work of guiding models and verifying, studying, and teaching machine-generated results with career advancement and government grants, while AI labs must publish their methods and code for peer review—applying the Royal Society’s 17th-century motto, Nullius in verba (“Take nobody’s word for it”), to machine science.

Source: What if the machines take over science?

Subtitle: Artificial intelligence will require scholars to reinvent their disciplines

Dateline: Oct 15th 2026\n