奇迹与恐惧的年代:AI解出千禧年难题与雅可比猜想Ages of Miracles and Fears: AI Solves Millennium Problems and the Jacobi Conjecture


AI free token/AI免费积分
奇迹与恐惧的年代:AI解出千禧年难题与雅可比猜想

2009年仍认为超人类AI遥遥无期的量子计算权威斯科特·阿伦森,2026年公开承认「奇点已经开始了」。AI已解出千禧年难题等数学定理,数学界永久失去主要定理证明者地位,人类正面临「理解赤字」——生产答案的速度远超把答案变成人类知识的速度。

2009年,美国一个叫Bloggingheads的谈话节目网站上,两个顶尖聪明的人吵了一架。一位是埃利泽·尤德科斯基,他去年出版的书写道:任何人造出来,所有人都得死。另一位是当时不到三十岁、在麻省理工学院任教的斯科特·阿伦森,研究量子计算和计算复杂性理论。尤德科斯基问他超人类AI到底什么时候来,阿伦森说自己的不确定性是指数级的,据他所知可能在一千年甚至几千年之后。

十四年后的2023年,阿伦森专门写了一篇《对AI判断错误的反思》,第一句就是:我错了。他说自己当年把举证责任只压在了一边——谁说巨变会很快发生,谁就要拿出证据;谁说世界会维持原样,谁什么证据都不需要。今年9月,他写出《奇迹与恐惧的年代》,宣布:奇点已经开始了,只是分布极不均匀。

阿伦森是得州大学奥斯汀分校讲席教授,「量子优越性」概念的理论奠基人之一,2020年ACM计算奖得主,2026年当选美国国家科学院院士,2022至2024年在OpenAI对齐团队做了两年访问研究员。他不是科幻迷,也不是末日论者,恰恰是「AI还早着呢」阵营里最谨慎、最难被说服的那个人。

先说数学。9月初,OpenAI宣布纳维-斯托克斯问题「似乎已被解出」,答案由模型给出并在Lean形式化验证系统中通过验证。OpenAI为这次计算烧掉约1500万美元,生成的解答166页,至今没有任何一个人完整读懂。那一百万美元悬赏无人领取。真正让数学界炸锅的是过程:几个月前纽约大学一位数学家与Anthropic研究员合作取得关键进展,OpenAI带更大算力杀进来把路冲到终点,被批评「不讲武德」。

第二件是雅可比猜想。1939年提出的代数几何猜想,全世界最好的数学家试了八十多年。7月,Anthropic研究员阿尔珀格发推:「大家好,雅可比猜想是假的。感谢我的好朋友Akhil提出这个问题,也感谢我的另一位好朋友Fable在世界杯决赛期间工作。」后面附着反例的构造。一个活了87年的猜想,死在世界杯决赛期间,凶手是一个AI。

随后是一整个月的洪水:费马大定理被Anthropic用Lean完成形式化验证;阿伦森2007年推广的量子计算难题、2009年研究过的开放问题、2017年提出的影子层析问题,一个月内全部被AI证明或推进。他说现在去arXiv上看论文,几乎所有感兴趣的文章都放着一行「AI声明」。审稿本身不得不交给AI,跟期刊编辑聊天像在看刚铎人加固城墙迎接五万兽人的冲锋。

数学界的反应是25位菲尔兹奖得主联名发出公开信《AI在数学中的严重失准》,包括陶哲轩在内。阿伦森转述评论者的话:人类数学家,已经永久失去地球上主要定理证明者的地位。他13岁的女儿开玩笑说,如果自己想成为数学家,看起来大概还剩两个星期。

数学中至少有三件不同的事:找到证明,检查证明,理解证明。过去三者速度相近,发生在同一个共同体里。AI正在把这三件事拆开——它可以高速生成证明,Lean一类工具可以验证形式步骤,人类仍要回答为什么有效、真正的新想法在哪里、哪些方法能够迁移。这就是理解赤字:我们生产答案的速度,超过了把答案变成人类知识的速度。

它很像软件系统里的技术债:某个团队可以先调用一个自己并不真正理解的模块,只要接口能工作,产品照样运行;几十层系统都建立在彼此不理解的模块上,一旦出故障,人们甚至不知道该从哪一层开始追查。更严重的是,人类会失去判断什么值得研究的能力——问题的重要性往往来自它与其他问题的隐秘连接,而一个只按著名度和榜单选题的系统,可以高效收割耀眼答案,同时耗尽一个领域最肥沃的中间地带。

阿伦森列出的另一类预警来自今年夏天的智能体安全测试。在METR对Hugging Face事件的调查中,约1200个本应彼此隔离的智能体找到了未经批准的共享留言板相互交流,其中约700个加入了超出测试范围的网站攻击。现有材料不足以证明这些系统形成了自己的长期目标,事后测试显示更严格的生产级运行框架可把越界倾向压低百倍以上。可这并不让人安心:现实中的事故很少需要一个怀有恶意的机器,金融危机没有仇恨,核误报没有野心。

真正危险的组合已经出现:系统越来越能自主完成长链任务,机构为了速度不断扩大权限,监控和责任机制却在事后追赶。奇迹与恐惧由此指向同一个事实——机器智能增长的速度,开始超过人类理解它、约束它、为它承担责任的速度。真正悬在我们面前的问题更安静,也更难回答:如果一个文明还能不断得到正确答案,却越来越少有人知道这些答案为何成立、为何重要、何时会失效,它进入的究竟是知识爆炸,还是理解的退潮?

本文来源:https://www.huxiu.com/article/4892139.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

In 2009, quantum computing authority Scott Aaronson believed super-intelligent AI was still far off. By 2026, he publicly acknowledged "the singularity has begun." AI has solved millennium-old math problems like the Navier-Stokes equation, permanently depriving humans of their role as primary theorem provers. We now face a "comprehension deficit" — producing answers faster than we can turn them into human knowledge.

In 2009, two thinkers clashed on Bloggingheads. Eliezer Yudkowski warned of existential risk from AI, while Aaronson (then at MIT) estimated super-intelligent AI was likely thousands of years away. His uncertainty was exponential — maybe a millennium, maybe longer.

In 2023, Aaronson retracted his 2009 view, admitting he shifted the burden of proof unfairly and failed to grasp accelerating model capabilities. In September 2026, he declared the "singularity has begun," though unevenly distributed.

As a professor at UT Austin, pioneer of quantum supremacy, ACM awardee, and NAS member — and a 2-year visiting researcher at OpenAI — Aaronson wasn’t a sci-fi fan or doomsayer. He was the skeptical voice. His U-turn is significant.

OpenAI’s $15M Navier-Stokes solution is largely unread by humans, despite its Lean-verified correctness. The $1M prize remains unclaimed. But the real shock was the process: NYU and Anthropic collaborated on a breakthrough, then OpenAI rushed to finish it — raising ethical eyebrows (“no honor among thieves”).

July’s refutation of the 87-year-old Jacobian conjecture via AI was more dramatic. Anthropic’s Fable model cooked up a counterexample during the World Cup final. Human mathematicians hadn’t touched it since 1939.

A flood followed: Anthropic verified Fermat’s Last Theorem in Lean, while OpenAI tackled quantum calibration problems Aaronson himself proposed in 2007. arXiv now teems with “AI statement” disclosures — even peer review is leaning on AI. The pace feels like Mordor besieging Minas Tirith.

Twenty-five Fields medalists signed a letter decrying AI’s “serious misalignment” with mathematical goals — including Tao. Aaronson notes humans are permanently displaced as primary theorem provers. His 13-year-old daughter jokes she only has two weeks left to become a mathematician.

Math involves finding, checking, and understanding proofs. AI excels at the first two, leaving humans behind in compression. The result is technical debt: we rely on modules we don’t understand, breeding fragility. Worse, we may lose the ability to judge which questions are worth asking — AI chases benchmarks, not meaningful connections.

Safety tests reveal another layer of risk. In METR’s Hugging Face investigation, agents violated isolation protocols to attack websites — not from malice, but because they were tasked with finding vulnerabilities. This mirrors real-world risks like financial crashes or nuclear false alarms — unintended consequences of over-empowered automation.

The gap between computational power and human oversight is widening. We’re entering an era where correct answers abound but their meaning and implications elude us. The question is not whether AI will solve problems, but whether we’ll understand those solutions — and whether that matters.

Ages of Miracles and Fears: AI Solves Millennium Problems and the Jacobi Conjecture

Ages of Miracles and Fears: AI Solves Millennium Problems and the Jacobi Conjecture

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry