陶哲轩隔空对话王虹:数学界该学会“消化”AI了

marsbitОпубліковано о 2026-08-20Востаннє оновлено о 2026-08-20

Анотація

陶哲轩与王虹两位菲尔兹奖得主共同指出,数学界需要学会“消化”AI生成的证明。随着AI快速产出大量证明,数学家面临阅读和理解的压力。陶哲轩亲自实践,花费数天消化了一个由AI辅助完成、解决森多夫猜想的证明,将其改写为更简洁、人类可读的形式,甚至发现了其能覆盖更强猜想的潜力。他强调,数学成果的完整流程包括生成、验证、阐释、发表和融入知识体系,AI目前主要加速前两步,而后期的理解、提炼与整合仍需数学家主导。为此,他呼吁学界改变过分追求“首发权”的习惯,提升“消化整理”工作的地位,并推出了登记库Palomar,旨在系统收集和验证AI证明,促进合作,确保成果的完整性与可用性。

AI出的证明太多,也未必全是好事......数学界已经快读不过来了。

不过好消息是——

面对这场危机,两代菲尔兹奖得主,陶哲轩王虹给出了几乎相同的判断:

消化

北大数院最近释出了一篇王虹的万字专访长文,她给出了相当中肯的建议:

数学界不能忽略AI产生的反例或证明,我们需要去学习使用,也需要理解、消化

陶哲轩则花了好几天时间,把一份由AI参与完成的证明,从头到尾消化了个遍

这份证明解决的,正是悬置67年之久的森多夫猜想。

这听着多少有点反常识。既然AI是来帮数学家解题的,答案都出来了,为什么还要再消化一次?

「吹哨人」陶哲轩给出的答案是:

正确的证明,只是数学成果的第一关,它还要被专业数学家理解吸收,才能彻底可用。

打个比方,传统作者像负责到底的父母,完整陪伴一个证明从诞生到发表,而往后与AI共舞,这段关系可能更像「送养」:

有人借助AI把证明推进到生成或验证阶段,再正式交棒另一组人,进行专业阐释、整理和发表。

数学家依旧大有可为。

消化是解决数学百年危机的下一步

说白了,这也是陶哲轩在回应自己一个月前,在ICM演讲中留下的问题:

如果答案可以被AI批量生产,数学研究究竟还要追求什么?

OpenAI新模型Astra连破10道数学与理论计算机科学难题,横跨高维几何、编码理论、群论等多领域。

一位没有受过高等数学科班训练的神经外科医生,让ChatGPT 5.6自主运行约16小时,给出了数值线性代数中Crouzeix猜想的关键证明。

......

数学这堵高墙,眼看着被AI凿出了一扇后门。

连门外汉都能借它突破前沿难题,数学不会也要喜提“下一个土木”剧本吧?(doge)

非也!事实上,一项完整的数学成果从来不只有「得到证明」这一步。

陶哲轩把它拆成五段:

先生成论证,再核验正确;接着要把思路讲给同行听,让它经过发表和共同体检验;最后,还要把结果重新组织,找到它与旧理论的关系,沉淀成以后能调用的标准知识。

AI目前最擅长加速前两步。越往后,就越需要数学家判断哪里才是核心,结果又意味着什么。

而陶哲轩刚处理的森多夫猜想,几乎是这段流程的1比1实战演练。

该猜想研究多项式零点与临界点的距离,此前较低次数和充分大次数都已经得到解决,中间范围仍然长期空悬。

几天前,数学爱好者Lech Mazur使用AI填补了这一缺口,并给出通过Lean验证的形式化证明。

似乎故事到这里就应该宣布结束了。但陶哲轩注意到,这份原始证明还没有被整理成适合人类阅读和发表的数学文本,也就是说,数学成果还不可用

于是他花了数天,在ChatGPT和纸笔推导的共同辅助下,对成果进行了一次完整消化:

追溯文献来源→提炼真正起作用的恒等式→删掉绕路→把机器发现的论证改写成能看出主线的版本。

这一读,反而读出了些新东西。

整理后的论证不仅解决了Sendov猜想,还能覆盖更强的Phelps–Rodriguez猜想。所需的核心工具也比原始形式化呈现出来的要初等得多,Lean代码从约9万行压缩至1.5万行。

换言之,消化不仅是对AI证明的再验证,也是拓宽数学成果的有效途径。

更重要的是,陶哲轩想借助消化,推翻数学界那套“谁先宣布、谁获得优先权”的旧规则。

数学成果的归属,最早看期刊发表日期,后来改看arXiv时间戳。

如今AI证明来得太快,连arXiv都有人嫌慢,一些结果刚生成就被扔上社交平台。为了抢先,验证阐释统统被跳过,最终给数学界人为制造了很多负担。

因此在ICM演讲中,陶哲轩主张,数学界不要再一味追捧「第一个给出证明的人」,应该提高「消化整理证明」的地位。

解释证明、审稿、把成果整理成经典理论,这些工作也要被重视,不能只有产出新定理才算功劳。

如果论文作者自己都没法做一场内行水平报告把结果讲明白,哪怕AI验证得再正确,这篇成果也不应该发表

人都理解不了的证明,又怎么称得上完整成果。

所以陶哲轩呼吁学界都能冷静地坐下来,就AI的能力底线、数学的价值观来一场深刻的讨论,同时强调数学工作中那些无法被机器替代的部分

One More Thing

与此同时,陶哲轩也实打实公开了一个面向Lean验证结果的登记库——Palomar

Palomar会把题目陈述、证明代码、AI参与方式和版本信息登记清楚,再通过独立内核复核,将散落在GitHub、社媒和新闻里的AI证明集合在一处。

它将作为验证与发表之间的消化中继站,让后来者知道AI和同行做到了哪里,还有哪些工作待完成。

对同一问题感兴趣的团队之间,可以通过Palomar合作研究。

值得注意的是,Palomar不负责宣布某个结果有多重要,也不替代同行评审。

具体结果的归属权,未来将由四个时间节点的先后顺序共同决定:

记录证明如何产生的 生成成果 ,如AI聊天记录;

确认逻辑正确的 验证成果 ,如Lean代码;

把思路讲清楚的 阐释成果 ,如公开演讲;

正式承载结果的 发表成果 ,如论文。

中译中就是,谁先交齐一整套完整成果,谁才获得优先权。

目前,陶哲轩整理后的森多夫猜想,已经成为Palomar首批档案之一。

参考链接:

[1]https://arxiv.org/abs/2608.16753

[2]https://mathstodon.xyz/@tao/117123957813926355

[3]https://www.youtube.com/watch?v=M0--ZH1lOzg

[4]https://mp.weixin.qq.com/s/w-UfVMmwTZqXjNsC81CzYw

本文来自微信公众号“量子位”,作者:关注前沿科技

Пов'язані питання

Q这篇文章中,陶哲轩和王虹对数学界面对AI浪潮的主要建议是什么?

A文章中提到,陶哲轩和王虹对数学界面对AI产生大量证明的相同判断是“消化”。王虹建议数学界需要去“学习使用,也需要理解、消化”AI产生的反例或证明。陶哲轩则身体力行,亲自花费数天时间,把一份由AI参与完成的森多夫猜想证明“从头到尾消化了个遍”。

Q陶哲轩认为,一项完整的数学成果需要经历哪五个阶段?

A陶哲轩把一项完整的数学成果拆解成五个阶段:1. 生成论证。2. 核验正确性。3. 向同行解释,让成果经过发表和共同体检验。4. 把结果重新组织。5. 找到它与旧理论的关系,沉淀成以后能调用的标准知识。他认为AI目前最擅长加速前两步,而越往后,就越需要数学家的判断与消化。

Q陶哲轩在处理AI证明的“消化”过程中,具体做了哪些工作?带来了什么新发现?

A陶哲轩在对森多夫猜想的AI证明进行“消化”时,具体做了以下工作:追溯文献来源、提炼真正起作用的恒等式、删除绕路部分、把机器发现的论证改写成能看出主线的版本。这一消化过程带来了新发现:整理后的论证不仅能解决森多夫猜想,还能覆盖更强的Phelps–Rodriguez猜想,且所需的核心工具比原始证明更初等,Lean代码从约9万行压缩至1.5万行。

Q陶哲轩为什么要呼吁改变数学界“谁先宣布、谁获得优先权”的旧规则?他提出的替代方案是什么?

A陶哲轩呼吁改变旧规则,是因为AI证明生成得太快,导致一些人为了抢先,跳过验证和阐释,直接把未消化的结果扔上社交平台,给数学界制造了额外负担,而且这些未经消化的成果“不可用”。他主张提高“消化整理证明”的地位,并认为如果论文作者自己都无法向同行把结果讲明白,哪怕AI验证正确,成果也不应发表。他提出的替代方案是:一项成果的归属权应由四个时间节点的先后顺序共同决定,即生成成果、验证成果、阐释成果和发表成果。谁先交齐这套完整成果,谁才获得优先权。

Q陶哲轩建立的“Palomar”项目是什么?它的主要功能和目的是什么?

A“Palomar”是一个由陶哲轩公开建立的面向Lean验证结果的登记库。它的主要功能是:将题目陈述、证明代码、AI参与方式和版本信息登记清楚,通过独立内核复核,把散落在各处的AI证明集合在一处。其目的是作为“验证与发表之间的消化中继站”,让数学界能清晰了解AI和同行在哪些问题上已取得进展,还有哪些消化和整理工作待完成,促进合作研究。它不负责宣布结果的重要性,也不替代同行评审。

Пов'язані матеріали

Shanghai's $10 Billion Unicorn Is About to Go Public

Shanghai-based automotive-grade millimeter-wave radar chip unicorn Calterah Microelectronics Technology (Shanghai) Co., Ltd. has filed for an IPO on Shanghai's STAR Market, aiming to raise 3.49 billion yuan. Founded in 2014 by Chen Jiashu, a UC Berkeley PhD graduate, and his professor Ali Niknejad, Calterah pioneered CMOS technology for 77GHz radar chips, breaking the decades-long monopoly of international giants like Texas Instruments. Its low-cost, highly integrated solutions enabled millimeter-wave radar to move from luxury to mass-market vehicles. By 2025, Calterah captured a 31.1% share in China's automotive millimeter-wave radar chip market (second domestically, fourth globally), with cumulative shipments exceeding 30 million units. Its client list includes BYD, Geely, Nio, and Volvo. The company has undergone 11 funding rounds, attracting high-profile investors such as the National Integrated Circuit Industry Investment Fund Phase II, China Capital Management, and GD Capital. Its valuation has reached tens of billions of yuan, with a projected post-IPO valuation of approximately 14 billion yuan. Despite rapid revenue growth—increasing from 206 million yuan in 2023 to 632 million yuan in 2025 with a 75.28% CAGR—Calterah remains unprofitable. It reported net losses of 323 million yuan, 334 million yuan, and 193 million yuan from 2023 to 2025, accumulating over 900 million yuan in losses over three and a half years. These losses are primarily attributed to heavy R&D investment, which totaled over 1.039 billion yuan in the reporting period, often exceeding annual revenue. The company faces significant risks, including high customer concentration (its top five customers accounted for over 99% of revenue from 2023-2025, with BYD alone representing over 50% in 2025) and supply chain concentration. Revenue pressure from key customers and dependencies on overseas suppliers for EDA tools and IP pose challenges to sustainable growth. The IPO is seen as crucial for securing capital to expand production, diversify its customer base, and reduce supply chain dependencies.

marsbit24 хв тому

Shanghai's $10 Billion Unicorn Is About to Go Public

marsbit24 хв тому

Ray Dalio's Latest Macro Analysis Full Text: Buy More Gold, Add Some Bitcoin

In his latest macro analysis, Ray Dalio applies his framework from "How Countries Go Broke: The Big Cycle" to the current global debt environment. He highlights recent events like Japan selling U.S. Treasuries and rising U.S. long-term yields as signs of an unsustainable debt dynamic. Dalio explains that excessive government debt leads to either unacceptably high interest rates, severe economic downturns, or significant currency debasement through central bank money printing. He summarizes the U.S. fiscal situation: with $5.5 trillion in revenue, $7.5 trillion in spending, a $2 trillion deficit, and total debt at six times annual revenue, debt servicing costs are immense. Without correction, U.S. debt could reach $55-$60 trillion in a decade. Dalio proposes a "3% Three-Way" solution: reducing the budget deficit to 3% of GDP through balanced spending cuts, tax increases, and interest rate reductions to avoid a traumatic adjustment. In response to FAQs, he argues that the risk of a U.S. debt crisis is high and could materialize within a few years if the current path continues. He dismisses the notion that the dollar's reserve status makes the U.S. immune, citing historical precedents of reserve currency declines. He is also unconvinced by Japan's high-debt stability, noting poor returns for yen-denominated assets. For investors, Dalio recommends diversifying globally, underweighting bonds, and overweighting assets like gold and a small allocation to Bitcoin (around 10-15% to gold) to hedge against currency debasement and poor debt returns.

marsbit1 год тому

Ray Dalio's Latest Macro Analysis Full Text: Buy More Gold, Add Some Bitcoin

marsbit1 год тому

Stripe’s 16-Year Chronicle: From 7 Lines of Code to a $100 Billion Valuation

Stripe's 16-year journey began with a simple promise: "7 lines of code to accept payments." Founded by Patrick and John Collison, the company started by hiding the complexity of bank integrations and merchant accounts behind a clean API, initially targeting developers at startups. This early focus on user experience and technical simplicity fueled rapid adoption. A key early milestone was establishing vital bank partnerships, a challenge overcome by hiring Billy Alvarado, who brought crucial institutional relationship skills. From this foundation, Stripe systematically expanded its product boundaries. It launched Connect for platform payments, Atlas for company formation, Radar for fraud prevention, and Billing for subscriptions. This transformed Stripe from a payment processor into a broader financial infrastructure suite for internet businesses. The COVID-19 pandemic accelerated growth but also led to over-hiring. A 14% layoff in 2022 marked a period of organizational correction. Subsequently, Stripe shifted its growth strategy towards strategic acquisitions to enter new domains quickly. It acquired Bridge (stablecoin infrastructure), Privy (wallet infrastructure), Metronome (usage-based billing), and agreed to buy OpenRouter (AI model routing). These moves signal Stripe's ambition to build a "programmable money system" for the emerging AI and agent-based economy, managing not just currency flows but also the measurement and pricing of computational resources like AI tokens. Internally, Stripe leverages AI agents (like "Minions") to boost engineering productivity. Despite scaling to nearly 8,000 employees and processing $1.9 trillion in payment volume annually, the company remains private. A recent employee tender offer valued it at $159 billion. The core question for Stripe's future is whether it can successfully integrate its expanding product matrix—spanning payments, crypto, and AI infrastructure—into a cohesive platform, positioning itself as the foundational economic layer for autonomous software agents.

marsbit1 год тому

Stripe’s 16-Year Chronicle: From 7 Lines of Code to a $100 Billion Valuation

marsbit1 год тому

Treasury Secretary's Move to Suppress Treasury Yields Ignites 'Currency Debasement Trade'! Gold Hits Three-Month High, Bitcoin Surges Over 25% in a Single Week

US Treasury Secretary Besant's efforts to lower long-term Treasury yields by announcing expanded buybacks had only a brief market impact. However, this move fueled a "currency devaluation trade," weakening the US dollar while boosting both gold (to a three-month high) and Bitcoin (up over 25% for the week). Analysts attribute this reaction to deepening market concerns over the massive US fiscal deficit and structural pressures keeping long-term rates elevated, including fierce competition for capital from global government borrowing and massive AI sector financing. Despite the Treasury's actions, fundamental forces like growth, inflation, and capital demand are seen as limiting its ability to sustainably suppress yields. Bitcoin's strong positive correlation with gold has reinforced its narrative as a hedge against devaluation. While equity markets have shown resilience, some strategists warn that Treasury yields nearing 5% increase pressure on the dollar and high-leverage assets. Figures like Ray Dalio have advised reducing bond exposure in favor of gold and some Bitcoin, citing US debt risks. Market opinions are divided on the sustainability of the devaluation trade, with some noting the lack of a near-term catalyst for its next leg higher. The underlying tension between the Treasury's desire for lower borrowing costs and the Federal Reserve's focus on inflation and reducing market intervention remains a key theme. Upcoming events like Nvidia's earnings and the Jackson Hole symposium will test whether AI profits can continue supporting stocks and if the Fed aligns more with Washington's preference for easier financial conditions.

华尔街日报3 год тому

Treasury Secretary's Move to Suppress Treasury Yields Ignites 'Currency Debasement Trade'! Gold Hits Three-Month High, Bitcoin Surges Over 25% in a Single Week

华尔街日报3 год тому

Торгівля

Спот
活动图片