新智元报道
数学界的天，彻底塌了！
统治人类逻辑高地几千年的数学家门，第一次感受到了「灭顶之灾」的寒意。
就在这两天，全球顶尖名校的数学系，直接掀起了一场声势浩大的集体大反抗。
直到目前，全球已有整整1008名数学家重磅联名，加入了一个新成立的组织——「人类数学协会」（AHM）。
光看签署名单，长到一眼划不到底。
他们在此立下「军令状」：从今往后，绝不给商业AI出半点力，也绝不发任何一篇AI写出来的数学论文！
更令人瞠目结舌的是，654名顶尖学者加入了内部的「无AI核心组」，公开向全世界发誓——
以后的日常科研中，不管是找证明思路、挑反例，还是提新课题，100%不用任何AI。
甚至，就连菲尔兹奖得主陶哲轩，都亲自下场转发了声明。
同一周的arXiv上，悬置26年的Barry Simon猜想轰然倒塌。把反例参数推到极限的ChatGPT，连两位人类作者都感到意外。
上下滑动查看
月初，OpenAI一夜倾泻700+篇手稿，成为这一切的导火索。
当时OpenAI测试了约4000个未解问题，每个结果消耗3小时的ChatGPT Pro思考算力，一口气放出719篇文稿。
这批文稿归入372个问题族，约42%附带Lean形式化验证。虽然逻辑上能跑通，但配套的自然语言证明晦涩破碎，还漏掉了大量前人文献。
面对这批半成品，Abhinav Nellore与Johannes Schmitt等四位学者，搭建了名为oai-math-revised的GitHub开源仓库。
10月24日，陶哲轩在博客刊发四位学者的联名文章。文章邀请全球同行提交代码合并请求，替这批AI文稿补齐文献引用并重写推导过程。
仓库提交记录里藏着一个极具戏剧性的细节。早在仓库公开前几天，账号名为teorth的陶哲轩就已经悄悄动手了。
他针对编号003的「Landau-Siegel零点一致排除问题」（关系到素数分布均匀性的数论难题），接连提交了第2号和第3号合并请求。
让人意想不到的是，陶哲轩起草这两次修改，动用的正是Claude Code（Opus 5.5）。
他在核对无误后，把文稿署名栏改成了「OpenAI（含社区贡献）」，并一举补上了Banks、Green及他本人的文献脉络。
截至10月25日，该仓库的13条修订记录里有8条都标注了由Claude Code协助起草。
人类数学家带着一家公司的AI编程工具，去给另一家商业巨头的AI数学论文打补丁。这成了今年秋天数学界充满荒诞感的一幕。
这篇招募帖在陶哲轩博客评论区引发正面冲撞，21条长评迅速分成两派。
不列颠哥伦比亚大学数学教授Lior Silberman态度鲜明。他直言这批文稿的署名作者既然是OpenAI，整理文献、把证明写清楚本来就是作者自己的本职工作。
在他看来，如果商业公司想要学术界提供审稿与润色服务，就该按市场价掏钱。学者没有理由去当免费志愿者。
评论者Skip Tich则点破了背后的训练数据算盘。他指出同行们无偿改出的高质量证明，转头就会被吸进下一代模型训练集。
支持修稿的学者同样给出了沉重的理由。
评论者Nestor打了个比方，跨国巨头在国家公园倾倒了有毒污泥。虽然污染不是志愿者的错，但如果不动手清理，整个学术生态都会受损。
面对质疑，陶哲轩在10月25日清晨亲自写下回复，拿自己补齐Siegel零点文献的提交举例。
他表示这个仓库是为了给学界建一个由社区自己掌控的公共据点。这样才能把分散在各处的纠错、后续研究和批评声音集中到一起。
即便有菲尔兹奖得主亲自出面灭火，关于「人类该不该替商业模型收拾残局」的裂痕依然越撕越大。
正是因为不愿沦为替商业巨头善后的免费劳力，上千名数学家选择了直接筑起高墙。
截至10月25日，新成立的「人类数学协会」（简称AHM）实名成员突破1008人，全员签署了两项核心承诺。
第一是不为任何自动化数学商业实体提供劳动、咨询或宣传。第二是绝不在署名著作中包含任何AI生成的文本。
其中654人进一步加入协会内部的「无AI派」。他们承诺在寻找证明、探索算例乃至提出问题的全流程科研中完全不用AI。
把这上千名数学家逼到公开割席的导火索，正是OpenAI此前高调宣传攻克「Navier-Stokes方程」（千禧年七大数学难题之一）的那场风波。
当时一位数学家把未发表手稿输入ChatGPT润色。OpenAI随后调动约10000个Agent并发运转88小时，生成了一份连专家都难以核验的庞大证明。
这场算力奇观随后被包装成AI突破百万美元悬赏难题的标志性事件，点燃了学界的怒火。
协会在首份声明中把话挑明，直指商业巨头近期的密集造势「展现的是算力霸权，与严谨的学术研究无关」。
声明批评巨头企图用不透明的统计黑箱取代人类理解，把数学家变成替机器输出盖章的「验证技工」。
三天后，协会的期刊工作组又向各大数学期刊喊话。审稿环节要是动用AI必须先过作者这一关，谁也不能背着作者把未发表手稿喂给商业模型。
这份千人割席声明很快遭到AI圈的嘲讽。
技术博主Hesamation拿棋王卡尔森举例，称数学家拒绝AI就像棋手拒绝象棋引擎复盘一样荒唐。
然而，要在今天戒断AI，数学家付出的代价远比放弃一个润色工具沉重得多。
10月22日，博科尼大学的Simon Becker与得克萨斯农工大学的杨梦轩在arXiv上传了一篇论文。
这篇论文否定了数学物理泰斗Barry Simon在2000年提出的「多维L2猜想」。
这道悬置26年的难题源于量子力学中的薛定谔算子。算子的「谱」决定了粒子能向无穷远处自由散射（绝对连续谱），还是被困在局部区域（纯点谱）。
在三维及更高维空间里，粒子散射通道比一维直线宽阔得多。
Barry Simon当年据此猜想，只要高维空间的球面平方可积积分保持有限，正半轴上就一定能保留绝对连续谱。换句话说，粒子不会被完全困死。
二十六年来，无数数学物理学家都卡在了高维空间难以封死所有逃逸角度的几何障碍上。
三维球壳反例构造，每个球壳包含聚焦波包的透镜区（Lens）、窄角通道（Channel）与周期性反射障壁（Reflecting barrier）
这篇论文是怎么做出来的，两位作者专门留出一节，讲清了他们和ChatGPT分三步击穿猜想的全过程。
第一步，两位作者检查了OpenAI那批719篇文稿，发现这道题还没被认领，便与ChatGPT配合设计出同心球壳势函数。
这套构造用微弱角度势能充当「弱透镜」把波包聚焦进窄角通道，再用径向周期势充当「反射障壁」。两人借此成功在有界能量区间内抹掉了绝对连续谱。
第二步，两位学者凭经验判断这套构造最多只能走到这一步，没想到ChatGPT给出了意外推进。
模型在不改动几何结构的前提下微调参数，引入「Combes-Hislop谱平均准则」（消除能级简并、锁定指数衰减的工具），在有界区间内证出了更强的纯点谱。
第三步最让两位学者意外。用他们自己的话说，「令我们大感意外的是，ChatGPT随后把参数逼到了极限，将纯点谱的性质一路推广到了整个正半轴」。
人类学者以为手里的武器只能击穿一面盾牌。机器接过武器调了调准星，一举打穿了整座堡垒。
拿到这条意料之外的参数路线后，两位作者把整篇35页手稿从头到尾重新推演写定，自己为全部证明兜底。
这三场风波撞在一起，是因为延续数百年的「提出问题、斩获荣誉、读懂并验证答案」第一次分家了。
过去的「证明稀缺时代」，人类必须想透沿途的方法才能摸到山顶。率先拿到答案，天然就等于贡献了深刻的理解。
如今到了「证明过剩时代」，算力把这道等号劈开了。
机器靠堆叠Agent就能率先算出结果，把天书翻译成人类知识却要耗费学者漫长的苦工。
要是荣誉还只归第一个算出结果的实体，巨头靠算力就能收割头条，提炼思想的学者反而成了免费劳工。
这也倒逼学术评价与青年培养做出改变，从奖励「谁先算出结果」，转向看重「谁能提出好问题、建立新框架」。
参考资料：
https://x.com/Hesamation/status/2114153199414157752?s=20
https://x.com/Math_files/status/2114406007682326996
编辑：摩西 桃子
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
The sky in mathematics is falling apart!
For the first time, mathematicians, who have ruled the logical heights of mankind for thousands of years, felt the chill of the "end of the plague".
And just these two days, the mathematics department of the world's most famous school has been the scene of a massive collective resistance.
To date, a total of 1,008 mathematicians have joined the AHM, a new organization.
It's a long way to go.
They're here to lay down a "military order": from now on, they will never give half a hand to commercial AI, nor will they send any mathematical paper written by AI!
Even more striking is the fact that 654 top scholars have joined the "no-AI core" inside, publicly swearing to the world that--
In future daily research, 100 per cent of it is done without any AI, whether by looking for proof of ideas, by challenging examples or by proposing new subjects.
Even the Fields Prize winner, Tao Chul-chul, personally transmitted the statement.
On the same week's arXiv, Barry Simon, who had been suspended for 26 years, thought that he would collapse. It's surprising that even two human authors pushed the cross-reference to ChatGPT.
Up-Slide View
At the beginning of the month, Open AI dumped 700+ manuscripts overnight and became the trigger for all this.
At the time, Open AI had tested about 4,000 unresolved problems, each of which consumed three hours of ChatGPT Pro's thinking ability and released 719 scripts in one breath.
The manuscripts were classified as 372 problem groups, with approximately 42 per cent accompanied by a Lean formalization. While logically running, the accompanying natural language has proved to be obscure and has also left out a large amount of previous literature.
In the face of this semi-finished batch, four scholars, Abhinav Nellore and Johannes Schmitt, set up the GitHub open-source warehouse called oai-math-revised.
On October 24, Tau Chul-chul published a joint article by four scholars on his blog. The article invites a global peer to submit a code consolidation request to complete the literature and re-write the extrapolation of the AI script.
The warehouse submitted a very dramatic detail in its submission records. A few days before the warehouse was made public, Tao Cheol-chul, whose account number was Teorth, had been quietly engaged.
He submitted successive requests for consolidation Nos. 2 and 3 in response to the "Landau-Siegel Zero Unanimous Exclusion Problem" (the numerical dilemma related to even distribution of prime numbers).
Surprisingly, Tau Chul-hyun drafted these two changes, using Claude Code (Opus 5.5).
After checking correctly, he changed the title to "Open AI (including community contributions)" and added the links to Banks, Green and his own literature.
As at 25 October, eight of the 13 revised records of the warehouse were marked with the assistance of Claude Code.
Human mathematicians take the AI programming tool of one company to patch the AI mathematical thesis of another commercial giant. It became a scene of a myth in mathematics this fall.
The call for recruitment triggered a positive collision in the comment section of Tao Cheol, with 21 long reviews quickly divided into two groups.
Lior Silberman, professor of mathematics at the University of British Columbia, has a strong attitude. He stated that the author of the manuscript, who was Open AI, had organized the documentation and written the proof that it was the author ' s own job.
In his view, if a commercial company wanted the academic community to provide review and polishing services, it should pay at market prices. Scholars have no reason to volunteer for free.
Commentator Skip Tich points out the underlying training data algorithm. He pointed out that the quality of the proofs that their peers had changed free of charge would be sucked into the next generation model training set.
Scholars who supported the revision also gave heavy reasons.
Commentator Nestor gave an example of transnational giants dumping toxic sludge in national parks. Although pollution was not the fault of volunteers, if it was not cleaned up, the entire academic ecology would be damaged.
In the face of the challenge, Tao Chul-chul responded in person on the morning of October 25th, adding his own examples of the submission of the Siegel zero.
He stated that the warehouse was intended to provide the academic community with a public position under the control of the community itself. This will bring together the voices of error, follow-up research and criticism, which are scattered around.
Even with the Feltz Prize winners themselves putting out the fire, the cracks on whether "humans should clean up the business model" are still getting bigger.
It is because of the unwillingness to become a free labour force to clean up the business giants that thousands of mathematicians have chosen to build the walls directly.
As of October 25, the actual members of the newly established Human Mathematics Association (AHM) had exceeded 1008, with two core commitments signed.
The first is not to provide labour, advice or publicity to any automated mathematical business entity. The second is that there will never be any AI-generated text in the signed book.
Of these, 654 further joined the "no-AI" within the Association. They are committed to not using AI at all in the search for proof, in the search for statistics and even in the process of asking questions.
In the past, OpenAi's high-profile campaign to attack the Navier-Stokes equation (one of the seven major mathematical challenges of the millennium) led to the imposition of thousands of mathematicians on the public cut.
A mathematician then entered unpublished manuscripts into ChatGPT. Open AI then mobilized some 10,000 Agents and operated for 88 hours, producing a huge certificate that could not even be verified by experts.
This event, a landmark event, was subsequently wrapped in the Al-Ai over a million-dollar reward challenge, which set the school community on fire.
In its first statement, the Association singled out the recent intensive campaign of business giants as "a display of arithmetical hegemony, unrelated to rigorous academic research".
The statement criticized the giants for attempting to replace human understanding with non-transparent statistical black boxes, turning mathematicians into "certifying technicians" who had stamped machines.
Three days later, the Association ' s working group on periodicals called out to major mathematical journals. No one can feed unpublished manuscripts to the business model behind the author ' s back if AI is to be used first.
The 1,000-person statement was soon ridiculed by the AI community.
Technical blogger Hesamation takes the example of King Carlson, who says that mathematicians reject AI as absurd as chess players refuse to replay chess engines.
However, to stop AI today, mathematicians pay much more than abandoning a colour tool.
On 22 October, Simon Becker of the University of Boko Haram and Yang Mangwan of the University of Agriculture and Industry of Texas published a paper in arXiv.
This paper refutes the "multi-dimensional L2 Conjecture" that Barry Simon put forward in 2000 in math physics.
This 26-year-old dilemma is due to Scherding's math. The algorithm's spectrometry determines whether particles can disperse freely to endless distances (absolute series) or be trapped in local areas (pure spectra).
In three-dimensional and higher dimensions, particle dispersion channels are much wider than a one-dimensional straight line.
On that basis, Barry Simon assumed that, as long as the ball level in high space remained limited, the absolute continuity spectrum would be preserved on the positive half-axis. In other words, particles will not be completely trapped.
Over the past 26 years, countless mathematic physicists have been stuck in geometrical barriers to the ability to block all escape angles in high-dimensional space.
Three-dimensional shell inverse construction, each containing a focus-packed lens area (Lens), a narrow angle channel (Channel) and a periodic reflection wall
How the paper came out, the authors devoted a section to the whole process of their three-step attack with ChatGPT.
In the first step, the authors examined the 719 articles of Open AI, and found that the subject had not been claimed and that ChatGPT had designed a concentric shell function.
This structure can serve as a "weak lens" to focus the wave pack into narrow angles, and then as a "reflective barrier" with routing cycles. The two men succeeded in removing the absolute continuity spectrum within the energy zone of the boundary.
In the second step, the two scholars judged the architecture to be at best this far from the experience, but the ChatGPT gave an unexpected advance.
The model fine-tunes the parameters without changing geometric structures, introduces the "Combes-Hislop spectrograph averages guidelines" (the tool to eliminate simplicity and lock index decay), and shows a stronger pure point spectra within the boundary.
The third step surprises two scholars. In their own words, "it is a surprise to us that ChatGPT then pushed the parameters to the limit and spread the pure point spectrum to the whole positive axis."
An anthropologist thinks that the weapon in his hand can only pierce one shield. The machine took over the weapon and set the star and hit the whole fort.
Having obtained this unexpected line of parameters, the authors have recast the entire 35-page manuscript from the beginning to the end and have themselves covered all the evidence.
These three waves hit each other because for the first time in hundreds of years, "questions, honours, reading and verifying answers".
In the "proved scarcity" of the past, human beings must find ways to reach the top of the mountain. By taking the lead in getting answers, nature is a contribution to profound understanding.
Now it's "the time to prove excess" and it's the same thing.
Machines can calculate the results first by stacking Agent, and translating the books into human knowledge takes a long and hard time for scholars.
If the honor belongs only to the first entity to calculate the results, the giant can reap the headline by measuring it, and the scholars of thought become free labourers.
This is also forcing academic evaluation and youth development to change from rewarding "who calculates the results first" to focusing on "who can ask good questions and create a new framework".
References:
https://x.com/Hesamation/status/2114153199414157752?s=20
https://x.com/Math_files/status/2114406007682326996
Editor: Moses Peach
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.