这是「人工智能进入数学」这条线索目前最猛的一次推进,也是最适合用来观察「AI 科研到底怎么算」的一篇标本。它的难点不在语言——长句不多、术语也都不冷——而在它同时写了三件互相拉扯的事:一边报喜(372 道题、88 小时、菲尔兹奖得主都沾边),一边记疑(只有 127 道题带形式化验证代码、推理过程只公布了十道),一边收场(28 位菲尔兹奖得主的公开信、8000 人联名)。
对写作者的启发集中在两处:一是「数据式论证」的密度——全文几乎每一段都由数字驱动(372、700、88、127、10、4000、28、8000、9),这种写法让立场藏在事实后面,比形容词更硬;二是结尾的双关——more existential than numerical 一句话同时回扣标题 Number crunch 与全文真正的忧虑:AI 揭示的问题不再是算术题,而是「数学这门学科还存在吗」。第 ⑧ 段那句引语 the light its proof sheds on other problems 则是本篇最值得整句背下来的表达。
01全文呈现
原文按自然段落编号 ①–⑧,保留原刊用词、拼写与标点(如 OpenAI’s、Fields-medallists 一仍其旧)。本篇为 Science & technology 栏目,无内嵌小标题。
1Tuesday October 6th 2026 may go down as one of the most consequential days in the history of mathematics. That was when OpenAI, an American artificial-intelligence firm, announced that its AI agents, powered by an unreleased model, had solved at least 372 long-standing mathematical problems in a matter of weeks. The findings, detailed in over 700 AI-generated papers, represent research at an unprecedented scale and speed.
2Such a glut has left some queasy. The excitement of potential new mathematical truths is, for many researchers, freighted with concern for the future of their field. Unless more care is taken, they say, the mass production of mathematics will lead to more solutions but less understanding.
3AI’s progress in mathematics has advanced rapidly. Just last month OpenAI claimed that the same internal model had solved the Navier-Stokes problem—one of the landmark Millennium Prize problems—in just 88 hours. Now hundreds of other significant puzzles have been cracked. They cover the entire mathematical corpus, from algebra and geometry to probability and logic, including many which have stumped experts for decades. The average computing power required for the autonomous mathematicians to arrive at a solution was equivalent to only about three hours of reasoning with models that the firm makes publicly available.
4Although the remaining Millennium Prize problems remain unsolved, progress towards one or two does seem to have been made. Most notable among these is the Riemann hypothesis, which concerns the distribution of prime numbers. Also buried in the deluge is a claimed proof of a thorny corollary named the 3D Kakeya maximal conjecture. Hong Wang, a mathematician at NYU’s Courant Institute, was awarded the Fields Medal—the maths equivalent of a Nobel prize—for her work on a closely related problem in July.
5Precisely what breakthroughs have been made, and by what means, will probably take mathematicians years to work out; many of the papers run to over 100 pages. Only 127 of the solved problems are accompanied by a code guaranteeing logical consistency (a norm in the field), and the agents’ internal reasoning has been published for only ten of them. These 372 problems also represent a subset of 4,000 which OpenAI posed to its agents, although the company has not released data about which problems were beyond its capabilities. Such transparency would improve assessments of the technology’s abilities and limitations.
6Some worry that such an approach is not conducive to long-term progress in mathematics. Not long after OpenAI’s solution to the Navier-Stokes problem last month, 28 Fields-medallists published an open letter warning that AI labs’ approach to mathematics was at odds with the field’s goals. In using open maths problems as a benchmark to test new models’ abilities, the letter-writers said, model-makers risked prematurely shutting down lines of inquiry without generating new techniques, questions or, ultimately, understanding. More than 8,000 of their colleagues have since signed on.
7OpenAI claims to have taken note. In working out how to release its new discoveries, the firm says it drew on the recommendations of an independent advisory group, a team of nine esteemed mathematicians formed in the aftermath of the Navier-Stokes announcement. In addition to sharing some details of its models’ internal reasoning, OpenAI also promised to fund a series of conferences to aid understanding of the results.
8All told, the scale of OpenAI’s release means its true impact will be clear only in hindsight. “The real importance of a problem is eventually measured by the new avenues it opens up and the light its proof sheds on other problems,” says Martin Hairer of EPFL, a Swiss university, who is a member of OpenAI’s advisory group as well as being a signatory of the Fields-medallists’ letter. For now, the mathematical problems AI is revealing are more existential than numerical.
02逐段对照精讲 + 全文翻译
① 定调:一个将被载入史册的日子
2026 年 10 月 6 日这个星期二,或许会成为数学史上影响最深远的时刻之一。正是在这一天,美国人工智能公司 OpenAI 宣布:在一个由未发布模型驱动的 AI 智能体的帮助下,它已在数周内攻克了至少 372 道长期悬而未决的数学难题。这些成果写在 700 多篇 AI 生成的论文里,代表着研究在规模与速度上前所未有的突破。
开篇就用《经济学人》最爱的「定日期 + 封历史地位」手法:先给一个具体星期几(Tuesday October 6th 2026,注意 TE 习惯把星期放在最前),再立刻用 may go down as one of the most consequential days in the history of(将作为……史上最……的日子之一载入史册)把一件刚发生的事抬进历史叙事。这里有三个词值得单独记:consequential(影响深远的)比 important 重得多,是社论给事件定级的默认用词;long-standing(长期悬而未决的)与 in a matter of weeks(仅仅数周)形成时间尺度上的反差——难题「长期」,攻克「数周」,张力由此而来。powered by an unreleased model 把「未发布」这个关键信息轻巧地挂在括号式分词里,为后文「透明度」议题埋线。最后 The findings, detailed in…, represent research at an unprecedented scale and speed 是典型的新闻收束句:用过去分词 detailed 把证据(700 篇论文)夹在主语与谓语之间,谓语用 at an unprecedented scale and speed 成对名词收尾,节奏干净。
② 疑虑:更多解答,更少理解
这样一场洪流让一些人感到不安。对许多研究者来说,对新数学真理的兴奋中掺杂着对这个领域未来的忧虑。他们说,如果不加审慎,数学的批量生产将带来更多解答,却留下更少的理解。
全段的难点是两个「反直觉」的措辞选择。Such a glut 里的 glut 本义是「供过于求、泛滥」,作者用它接上段的「700 篇论文、372 道难题」,等于先把喜讯重新定义成「过量供应」——一个词就把情绪翻了面。has left some queasy 的 queasy 本义是「想吐、反胃」,引申为「心里发毛、不安」,比 uneasy 更有身体感,因为这里谈论的正是某种本能的不适。The excitement of A is, for many researchers, freighted with concern for B 是整个第二句的结构骨架:主语 + 系动词 be + 过去分词 freighted(装载、满载)作表语,中间插入 for many researchers 限定范围;freighted with 把抽象情绪写成「装了货的船」,兴奋与忧虑是同一艘船上的两件货。最后 more solutions but less understanding 是全文的 thesis sentence(主旨句):用两个比较级并列、中间以 but 翻转,干净地亮出立场——作者不反对解题,反对的是「解题」取代了「理解」。这种「more A but less B」的对称写法可直接搬进作文。
③ 加速度:从纳维—斯托克斯到数百道
AI 在数学上的进展一直很快。就在上个月,OpenAI 声称同一个内部模型仅用 88 小时就解出了纳维—斯托克斯问题——这是著名的千禧年大奖难题之一。如今又有数百道其他重要谜题被攻克。它们覆盖整个数学领域,从代数、几何到概率论与逻辑学,其中许多已困扰专家数十年。这些「自主数学家」得出一个解答所需的平均算力,只相当于用该公司公开模型进行约三小时的推理。
本段用一系列具体数字把「快」坐实,是全文「数据式论证」的范本。开头 AI’s progress in mathematics has advanced rapidly 句法上略有重复(progress 与 advance 近义),这是 TE 在科技报道里偶尔出现的口语化冗余,读的时候不必当成病句,反而可记下 advance rapidly 这个搭配。Just last month… the same internal model… in just 88 hours 三处时间词(last month / same / just 88 hours)层层收紧,把一个「刚发生」的事件写得像连续剧。Millennium Prize problems 是专有名词(千禧年大奖难题,七道经官方认证的数学难题,每道悬赏百万美元),landmark 此处作形容词「里程碑式的」。cracked(破解)比 solved 口语,带「撬开硬壳」的画面感。stumped experts for decades 的 stump 是「把人难住」,比 puzzle 更强调人的无力。mathematical corpus 的 corpus 原为「语料库、全集」,这里指「全部数学成果的集合」,是个雅致的借词。最后一句是全段唯一的长句:The average computing power required for… to arrive at a solution was equivalent to… 主语被一个 12 词的不定式修饰 required for the autonomous mathematicians to arrive at a solution 拉长,谓语 was equivalent to(相当于)把两种算力画等号——注意 autonomous mathematicians(自主数学家)这个拟人化称谓,把 AI 写成了一个能独立完成研究的主体,是后文「透明度」争议的语言起点。
④ 名单:黎曼假设、Kakeya 与菲尔兹奖
尽管其余千禧年大奖难题仍未解决,但在一两个难题上取得进展似乎确有其事。其中最引人注目的是黎曼假设,它关乎素数的分布。同样被淹没在这场洪流中的,还有一份针对「三维 Kakeya 极大猜想」这一棘手推论的所谓证明。纽约大学柯朗数学研究所的数学家王虹,7 月因在一个密切相关的问题上的工作而获得了菲尔兹奖——数学界的诺贝尔奖。
本段处理「哪些具体难题被攻克」,节奏是「先总退一步,再逐一举例」。Although the remaining Millennium Prize problems remain unsolved, progress towards one or two does seem to have been made 句首让步,does seem to have been made 里的 does 是强调助动词,把「确有其事」的语气顶出来——因为前句刚说过「仍未解决」,这里要防止读者误以为作者改口说「全解开了」。Most notable among these is the Riemann hypothesis 是完全倒装:正常语序 The Riemann hypothesis is most notable among these,倒装让重点(黎曼假设)落回句首,且 which concerns the distribution of prime numbers 这个定语从句紧跟其后补充说明。黎曼假设是数论最核心的未解难题之一,作者点它是在告诉读者「分量够重」。Also buried in the deluge is a claimed proof… 又一句倒装,buried in the deluge(被埋在这场洪流里)延续第 ② 段的 glut 隐喻,把「成果太多反而让人看不清」的意思再推一层;a claimed proof 的 claimed(所谓的、声称的)是全文最省力的一个限定词——作者不否认、也不背书,只说「他们声称」。最后一句 Hong Wang… was awarded the Fields Medal—the maths equivalent of a Nobel prize— 用破折号插入同位语解释菲尔兹奖,a mathematician at NYU’s Courant Institute 作同位语交代身份——注意作者特意强调「与本次被攻克问题『密切相关』的问题」,等于在说:真正被验证的突破,恰恰在王虹已获奖工作的延长线上,这既给了成果可信度,也暗示本次推论的含金量。
⑤ 保留:哪些能验证,哪些说不清
究竟取得了什么突破、又是通过什么手段取得的,数学家们大概需要数年才能弄清;这些论文中有许多超过 100 页。372 道已被解出的问题中,只有 127 道附有保证逻辑一致性的代码(这是该领域的规范),而 AI 智能体的内部推理过程只公布了其中十道。这 372 道题也是 OpenAI 交给智能体的 4000 道题中的一个子集,尽管公司没有公布哪些题目超出了它们的能力范围。这样的透明度本会有助于评估这项技术的能力与局限。
全段是「记疑」的核心,几乎每一句都是「只 / 仅」开头,用数字的落差代替形容词来质疑。Precisely what breakthroughs have been made, and by what means, will probably take mathematicians years to work out 主语是两个并列的名词性从句(what… and by what means),谓语 will take sb years to work out(要花某人数年才弄清)把「现在说不清」写得很克制。many of the papers run to over 100 pages 的 run to 意为「达到(某个数量/篇幅)」,是报纸描写规模的高频小词。Only 127 of the solved problems are accompanied by a code guaranteeing logical consistency:accompanied by(附有)+ 现在分词 guaranteeing 作后置定语,括号里的 a norm in the field(该领域的规范)点明为何这 127 道才「算数」。the agents’ internal reasoning has been published for only ten of them 把 only 提到助动词后,强调「仅十道」——与 372 形成刺眼对比。These 372 problems also represent a subset of 4,000 which OpenAI posed to its agents 的 pose 此处作「提出(问题)」,subset 点出采样偏差:公布的只是被解出的一部分,4000 道里还有「解不出的」没说。beyond its capabilities(超出其能力范围)与前面的 solved 正反对照。Such transparency would improve… 用虚拟语气 would 收尾——作者不直接指责不透明,只说「若透明本会更好」,分寸拿捏得很准。
⑥ 反对:28 位菲尔兹奖得主的公开信
有人担心,这种做法不利于数学的长期进步。上个月 OpenAI 解出纳维—斯托克斯问题后不久,28 位菲尔兹奖得主发表了一封公开信,警告说 AI 实验室对数学的做法与这个领域的目标背道而驰。信的作者们说,模型制造者用公开的数学难题作为测试新模型能力的基准,有可能在不产生新技术、新问题乃至最终的新理解的情况下,过早地关闭研究路径。他们已有 8000 多位同行联名。
全段是反对声音,结构上是「观点 → 权威背书 → 理由 → 声势」四步走。conducive to(有助于、 conducive to sth)是雅思/学术写作高频词,其否定 not conducive to 等于「不利于」。at odds with(与……不一致、背道而驰)比 conflict with 更书面,后接 the field’s goals 把分歧定义为「目标冲突」。28 Fields-medallists 是 Fields medallist 的复数(菲尔兹奖得主,数学最高奖),用权威人数给观点加码。In using open maths problems as a benchmark to test new models’ abilities, the letter-writers said, 先用 In using… as a benchmark 分词短语点出被批评的做法,再插入 the letter-writers said 把下面那句长主张明确归为「信里说的」,避免读者误以为是作者本人观点。model-makers risked prematurely shutting down lines of inquiry:risk doing(冒险去……)、prematurely(过早地)、shut down lines of inquiry(关闭研究路径)三个词串出「因测试而误杀研究方向」的担忧;句尾 without generating new techniques, questions or, ultimately, understanding 用递进收束,ultimately 把最重的词(understanding 理解)压在最后——与第 ② 段 more solutions but less understanding 首尾呼应。have since signed on 的 sign on 意为「签名支持、联署」,since 指「自那封公开信以来」,用 8000 这个压倒性数字收束,让反对声有了分量。
⑦ 回应:九人咨询小组与资助会议
OpenAI 声称已经注意到了这一点。公司表示,在研究如何发布这些新发现时,它参考了一个独立咨询小组的建议——这个小组由九位德高望重的数学家组成,是在纳维—斯托克斯问题公布之后成立的。除了公布其模型内部推理的部分细节,OpenAI 还承诺资助一系列会议,以帮助理解这些成果。
本段是 OpenAI 一方的回应,语气以「宣称 + 举措」为主,暗含作者对「诚意几何」的保留。OpenAI claims to have taken note 的 claim to have done(声称已经做了)是报道他人主张的疏离式起手,比 simply says 多一层「这是他们自己说的」的意味;take note(注意到、放在心上)在此略带反讽——回应一封 8000 人联名的信,用「注意到」显得轻描。In working out how to release…, the firm says it drew on the recommendations of…:in working out(在弄清楚/拟定……的过程中)作状语,drew on(借鉴、利用)是学术与商业写作通用的搭配,宾语是一个带长后置定语的名词短语 a team of nine esteemed mathematicians formed in the aftermath of…;其中 esteemed(受尊敬、德高望重的)是比 respected 更正式的褒义词,in the aftermath of(在……的余波/后果中)是描写事件后续的高频短语。In addition to sharing…, OpenAI also promised to fund a series of conferences to aid understanding of the results:in addition to doing 带两个动作(sharing details / promising to fund),a series of(一系列)与 aid understanding of(帮助理解……)收尾——注意「资助会议」这个动作,作者把它写成「帮助理解成果」的配套,实则是把质疑引向「公开讨论」而非「公开证明」,这是回应的精妙处,也是可被追问的软肋。
⑧ 收束:存在性大于数字性
总而言之,OpenAI 这次发布的规模之大,意味着它真正的影响只有事后才能看清。「一个问题真正的重要性,最终要由它开辟了哪些新路径、它的证明为其他问题投下了什么光来衡量,」洛桑联邦理工学院(EPFL)的马丁·海泽说;他既是 OpenAI 咨询小组的成员,也是那封菲尔兹奖得主公开信的签署人之一。就目前而言,AI 正在揭示的数学问题,更多是存在性的,而非数字性的。
结尾段先收束再点题,是全篇最值得背的两句话。All told,(总而言之、通盘算来)是新闻报道常见的收束副词。its true impact will be clear only in hindsight 的 in hindsight(事后来看、回头审视)是评论高频词,作者用它把「现在下结论还早」写得四平八稳。直接引语 The real importance of a problem is eventually measured by the new avenues it opens up and the light its proof sheds on other problems 是全篇的 golden sentence:系动词 is measured by 后接两个并列名词短语——the new avenues it opens up(它开辟的新路径)与 the light its proof sheds on other problems(它的证明为其他问题投下的光);avenue(路径、门径)与 shed light on(阐明、使……显形)都是学术写作里极漂亮且可直接套用的表达,尤其 shed light on,几乎可以替换任何「解释/说明」。says Martin Hairer of EPFL… who is a member of… as well as being a signatory of… 把引语者的身份做成「双重身份」(既是 OpenAI 顾问,又是反对信签署人),等于提醒读者:连反对者阵营里的人,也不否认 AI 在数学上的冲击力——这是作者给「AI 确实厉害」留的最后一句公道话。最后 For now, the mathematical problems AI is revealing are more existential than numerical 是收束的双关:existential(存在性的、关乎存亡的)与 numerical(数字的)既回扣标题 Number crunch(數字运算),又把全文的忧虑从「算得对不对」抬到「数学这门学科还存在吗」——一个词把报道从科技新闻推成哲学提问。
03核心词汇精讲
标签说明:四级 基础必备 六级 进阶 雅思 学术高频 高分 写作加分
A. 数学、突破与规模
consequential/ˌkɒnsɪˈkwenʃl/六级雅思
释义adj. 影响深远的,重大的(语气强于 important)
派生consequence n. 后果;consequently adv. 因此;inconsequential adj. 无关紧要的
搭配a consequential day 影响深远的一天;a consequential decision 重大决定;of little consequence 无足轻重
unprecedented/ʌnˈpresɪdentɪd/六级雅思
释义adj. 前所未有的,空前的
派生precedent n. 先例;un- 否定前缀;precedented adj. 有先例的
搭配at an unprecedented scale 前所未有的规模;unprecedented speed 空前的速度;an unprecedented move 空前的举动
landmark/ˈlændmɑːk/六级雅思
释义n. 里程碑;地标 adj. 里程碑式的(重大而具标志性的)
派生land(土地)+ mark(标记);landmark case 标志性案例;landmark ruling 里程碑式裁决
搭配a landmark problem 里程碑式难题;a landmark achievement 标志性成就;a landmark in history 历史里程碑
stumped/stʌmpt/高分
释义v. 把……难住,使束手无策(stump 的过去分词/过去式)
派生stump n. 树桩;stumper n. 难题;近义 baffle v. 使困惑、puzzle v. 使迷惑
搭配stumped experts for decades 困扰专家数十年;be stumped by a question 被一个问题难住;a stumper 一道难题
corollary/kəˈrɒləri/高分
释义n. 推论,必然结果(由某定理直接推出的结论)
派生corollaries 复数;近义 implication n. 引申含义、consequence n. 后果
搭配a thorny corollary 一个棘手的推论;a direct corollary 直接推论;as a corollary 作为必然结果
breakthrough/ˈbreɪkθruː/六级
释义n. 突破,重大进展
派生break through 动词短语(突破);近义 advance n. 进展、leap n. 飞跃
搭配a major breakthrough 重大突破;a scientific breakthrough 科学突破;represent a breakthrough 代表一项突破
crack/kræk/四级
释义v. 破解,攻克(难题、密码) n. 裂缝;尝试
派生cracked adj. 破裂的;cracker n. 解密高手;近义 solve v. 解决、decode v. 解码
搭配crack a problem 攻克难题;crack a code 破译密码;crack a case 破案
corpus/ˈkɔːpəs/高分
释义n. 全集,语料库(某一领域全部作品/文本的集合,复数 corpora)
派生corpora 复数;corporate adj. 公司的(同源,本义「身体的」);近义 collection n. 集
搭配the entire mathematical corpus 全部数学成果;a corpus of texts 文本语料库;a reference corpus 参考全集
B. 计算、能力与验证
autonomous/ɔːˈtɒnəməs/六级雅思
释义adj. 自主的,独立的(无需人类干预而运行)
派生autonomy n. 自主、自治;autonomously adv. 自主地;near-synonym independent adj. 独立的
搭配autonomous mathematicians 自主数学家;an autonomous system 自治系统;autonomous decision-making 自主决策
equivalent/ɪˈkwɪvələnt/六级雅思
释义adj. 相等的,相当的 n. 等价物
派生equivalence n. 等价、等同;equivalently adv. 等同地;equal adj. 相等的
搭配be equivalent to 相当于;of equivalent value 价值相当;an exact equivalent 完全对等物
guarantee/ˌɡærənˈtiː/六级
释义v. 保证,确保 n. 保证(书)
派生guaranteed adj. 有保证的;guarantor n. 保证人;近义 ensure v. 确保
搭配guarantee logical consistency 保证逻辑一致性;a guarantee of success 成功的保证;guarantee against 防止……
consistency/kənˈsɪstənsi/六级雅思
释义n. 一致性,连贯性
派生consistent adj. 一致的;consistently adv. 一贯地;inconsistent adj. 不一致的
搭配logical consistency 逻辑一致性;internal consistency 内在连贯;for the sake of consistency 为了前后一致
norm/nɔːm/雅思
释义n. 规范,准则;常态
派生normal adj. 正常的;normative adj. 规范的;near-synonym standard n. 标准
搭配a norm in the field 该领域的规范;social norms 社会规范;against the norm 不合常规
transparency/trænsˈpærənsi/雅思
释义n. 透明度,公开
派生transparent adj. 透明的、公开的;transparently adv. 透明地;反义 opacity n. 不透明
搭配such transparency 这样的透明度;greater transparency 更高的透明度;a lack of transparency 缺乏透明度
assessment/əˈsesmənt/六级
释义n. 评估,评价
派生assess v. 评估;assessable adj. 可评估的;近义 evaluation n. 评价
搭配improve assessments of 改进对……的评估;a fair assessment 公允的评价;risk assessment 风险评估
capability/ˌkeɪpəˈbɪləti/六级
释义n. 能力,才能(常作复数 capabilities 指「本领、机能」)
派生capable adj. 有能力的;capably adv. 干练地;近义 ability n. 能力
搭配beyond its capabilities 超出其能力范围;technical capabilities 技术能力;a capability gap 能力缺口
C. 疑虑、风险与长期主义
glut/ɡlʌt/高分
释义n. 供过于求,泛滥 v. 使充斥
派生glut 作动词:be glutted with 充斥着;近义 surplus n. 过剩、overflow n. 溢出
搭配such a glut 这样一场泛滥;a glut of papers 论文泛滥;a glut of information 信息过载
queasy/ˈkwiːzi/高分
释义adj. 感到不安的,发毛的(本义「反胃、想吐」)
派生queasiness n. 不安、恶心;近义 uneasy adj. 不安的、apprehensive adj. 忧虑的
搭配leave some queasy 让一些人不安;feel queasy about 对……发毛;a queasy feeling 一种不安感
freighted/ˈfreɪtɪd/高分
释义adj. 满载(某种情绪)的(freight 作动词「装载」的过去分词)
派生freight n./v. 货物、装货;freight with 满载着(情绪);近义 laden adj. 负载的
搭配freighted with concern 掺杂着忧虑;be freighted with meaning 饱含深意;emotion-laden 情绪满载的
conducive/kənˈdjuːsɪv/雅思
释义adj. 有助于……的,有益的(常用于 be conducive to)
派生conduce v. 有助于(罕见);反义 detrimental adj. 有害的;近义 favourable adj. 有利的
搭配conducive to progress 有助于进步;not conducive to 不利于;be conducive to learning 有助于学习
at odds with/ət ɒdz wɪð/雅思
释义短语 与……不一致,与……相左
派生odds n. 几率、分歧;settle one's odds 消除分歧;近义 conflict with 与……冲突
搭配be at odds with the goals 与目标背道而驰;at odds with the evidence 与证据矛盾;at odds with each other 彼此不合
benchmark/ˈbentʃmɑːk/六级雅思
释义n. 基准,参照标准 v. 以……为基准衡量
派生bench(长凳)+ mark(标记);benchmarking n. 标杆管理;near-synonym yardstick n. 衡量标准
搭配as a benchmark to test 作为测试……的基准;a benchmark for 衡量……的基准;set a benchmark 设立基准
prematurely/ˈpremətʃəli/雅思
释义adv. 过早地,提前地
派生premature adj. 过早的、未成熟的;prematurely shut down 过早关闭;反义 belatedly adv. 迟来地
搭配prematurely shutting down 过早关闭;a premature conclusion 草率结论;die prematurely 早逝
existential/ˌeɡzɪˈstenʃl/雅思高分
释义adj. 存在性的,关乎存亡的(源自 existence)
派生existentialism n. 存在主义;existential threat 生存性威胁;existence n. 存在
搭配more existential than numerical 更多是存在性的而非数字的;an existential question 存在论问题;existential risk 生存性风险
D. 机构、声誉与判断
esteemed/ɪˈstiːmd/高分
释义adj. 受尊敬的,德高望重的(esteem 的过去分词作形容词)
派生esteem v./n. 尊敬、尊重;self-esteem n. 自尊;近义 respected adj. 受尊敬的
搭配nine esteemed mathematicians 九位德高望重的数学家;an esteemed colleague 受敬重的同事;widely esteemed 广受尊敬
in the aftermath of/ɪn ði ˈɑːftəmæθ əv/高分
释义短语 在……的余波/后果中,在……之后
派生aftermath n. 余波、后果(尤指灾祸后);近义 in the wake of 在……之后
搭配formed in the aftermath of 在……之后成立;in the aftermath of the war 战后;the aftermath of the crisis 危机余波
signatory/ˈsɪɡnətri/高分
释义n. 签署人,缔约国(签署文件/条约的一方)
派生signatory to 的签署方;sign v. 签署;signature n. 签名;近义 signer n. 签名者
搭配a signatory of the letter 公开信的签署人;a signatory to the treaty 条约签署国;co-signatory 联署方
avenue/ˈævənjuː/六级
释义n. 途径,门径(引申义;本义「林荫道、大街」)
派生avenues 复数;近义 path n. 路径、approach n. 途径;a new avenue of research 研究新路径
搭配the new avenues it opens up 它开辟的新路径;explore every avenue 穷尽一切途径;an avenue to success 成功之道
shed light on/ʃed laɪt ɒn/雅思
释义短语 阐明,使……显形,为……提供线索
派生shed(过去式 shed)本义「投下」;light 喻「光、澄清」;近义 illuminate v. 阐明、clarify v. 澄清
搭配the light its proof sheds on other problems 它的证明为其他问题投下的光;shed light on a mystery 揭开谜团;cast light on 同义
in hindsight/ɪn ˈhaɪndsaɪt/雅思
释义短语 事后来看,回头审视
派生hindsight n. 后见之明;with the benefit of hindsight 事后来看;反义 foresight n. 先见之明
搭配clear only in hindsight 只有事后才看清;in hindsight, we should have 事后看来我们本该;20/20 hindsight 事后全看清
take note/teɪk nəʊt/六级
释义短语 注意到,放在心上
派生note n./v. 注意、笔记;take note of 注意到;near-synonym heed v. 留意
搭配claims to have taken note 声称已注意到;take note of the warning 留意这一警告;worthy of note 值得注意
significant/sɪɡˈnɪfɪkənt/四级六级
释义adj. 重大的,显著的
派生significance n. 重要性、意义;significantly adv. 显著地;near-synonym substantial adj. 重大的
搭配significant puzzles 重要谜题;a significant breakthrough 重大突破;statistically significant 统计显著
04长难句拆解(8 句)
1. That was when OpenAI, an American artificial-intelligence firm, announced that its AI agents, powered by an unreleased model, had solved at least 372 long-standing mathematical problems in a matter of weeks.
结构:
That was [when OpenAI, [an American artificial-intelligence firm],
↑ 表语从句(when 引导) ↑ 同位语,交代 OpenAI 是什么
announced [that its AI agents, [powered by an unreleased model],
↑ announced 的宾语从句 ↑ 过去分词作后置定语,修饰 agents
had solved [at least 372 long-standing mathematical problems] [in a matter of weeks]]].
↑ 宾语从句的谓语 had solved ↑ 时间状语:仅仅数周
难点:① 主干是
That was when…,
when 引导表语从句,回指上句的「10 月 6 日」——这是新闻里极常见的「日期 → 事件」收束结构。
② 表语从句内部又套两层:
an American artificial-intelligence firm 是 OpenAI 的同位语,
powered by an unreleased model 是 agents 的过去分词后置定语,两个插入成分把主谓(
OpenAI announced)拉开,读的时候先抓主干,再把这两个修饰放回各自名词旁。
③ 宾语从句谓语用过去完成
had solved,因为它发生在 announced(过去)之前——这是「过去的过去」的标准形态,注意
at least 372… in a matter of weeks 把「数量多」与「时间短」并置,正是全篇数据式论证的起手。
④ 翻译时把两层插入语顺回名词即可:OpenAI(一家美国 AI 公司)宣布,其由未发布模型驱动的 AI 智能体,已在数周内解出至少 372 道难题。
译文:正是在这一天,美国人工智能公司 OpenAI 宣布:在一个由未发布模型驱动的 AI 智能体的帮助下,它已在数周内攻克了至少 372 道长期悬而未决的数学难题。
2. The average computing power required for the autonomous mathematicians to arrive at a solution was equivalent to only about three hours of reasoning with models that the firm makes publicly available.
结构:
[The average computing power [required for the autonomous mathematicians to arrive at a solution]]
↑ 主语 ↑ 不定式短语作后置定语(required for… to…),长达 12 词
was equivalent to [only about three hours of reasoning [with models [that the firm makes publicly available]]].
↑ 系表结构:相当于 ↑ 介词宾语 ↑ 定语从句,修饰 models
难点:① 主语
computing power 与谓语
was 被一个 12 词的不定式定语
required for the autonomous mathematicians to arrive at a solution 隔开——这是全段最易失读的地方,先找到主语,跳过中间修饰,直接接
was equivalent to。
②
required for sb to do sth 是「做……所需要的」,这里的主语是「算力」,动作是「到达一个解」。
③
was equivalent to(相当于)把两种算力画等号,宾语是
three hours of reasoning,后面
with models that the firm makes publicly available 是带定语从句的后置定语,
that 指 models,
makes publicly available 是「使……公开可用」(make + 宾语 + 形容词补语)。
④ 全句的修辞用意:用「平均算力仅约三小时公开推理」来反衬前文的「372 道难题、700 篇论文」——量大而耗算极小,张力由此而生。
译文:这些「自主数学家」得出一个解答所需的平均算力,只相当于用该公司公开模型进行约三小时的推理。
3. Precisely what breakthroughs have been made, and by what means, will probably take mathematicians years to work out; many of the papers run to over 100 pages.
结构:
[[Precisely what breakthroughs have been made], and [by what means]],
↑ 主语从句 1(what 引导) ↑ 主语从句 2(by what means 引导)
will probably take mathematicians years [to work out];
↑ 谓语 take sb years to do ↑ 真正宾语(弄清什么、以何手段)
[many of the papers] run to over 100 pages. ↑ 并列分句:run to 达到
难点:① 主语是两个并列的名词性从句:
what breakthroughs have been made(取得了什么突破)与
by what means(以何种手段),中间用逗号+and 连接,谓语
will take mathematicians years to work out 把「主语(是什么、怎么做的)」整体当作 take 的宾语:
take sb years to do = 要花某人数年才……弄清。
②
Precisely 放在最前,强调「确切地」——因为后文要揭晓的恰恰是「说不清」。
③ 分号后
many of the papers run to over 100 pages 是独立分句,
run to 意为「达到(某篇幅)」,主语 papers 与 run 搭配,是报纸写长度的惯用小词。
④ 整句逻辑:先说「弄清要数年」(时间成本),再补「论文动辄百页」(阅读成本),两层都在铺垫「现在下不了结论」。
译文:究竟取得了什么突破、又是通过什么手段取得的,数学家们大概需要数年才能弄清;这些论文中有许多超过 100 页。
4. Only 127 of the solved problems are accompanied by a code guaranteeing logical consistency (a norm in the field), and the agents’ internal reasoning has been published for only ten of them.
结构:
[Only 127 of the solved problems] are accompanied by [a code [guaranteeing logical consistency (a norm in the field)]],
↑ 主语(Only 强调数量) ↑ 介词宾语:代码 ↑ 现在分词后置定语 + 括号同位语
and [the agents’ internal reasoning] has been published [for only ten of them].
↑ 并列主语 ↑ 谓语(被动) ↑ 范围:仅其中十道
难点:① 两个并列分句都以
only 起手(127 / ten),制造数字落差;第一句
are accompanied by(附有)是被动,主语是 problems,宾语是 code。
②
guaranteeing logical consistency 是现在分词作后置定语修饰 code,括号里的
a norm in the field 是 consistency 的同位语,点明「这在该领域是规范」——作者借此暗示为什么只有 127 道「算数」。
③ 第二句
the agents’ internal reasoning has been published for only ten of them 把 only 放在介词前,强调「仅十道」,与前句 372 形成刺眼对比;
internal reasoning(内部推理过程)是 AI 可解释性领域的核心词。
④ 翻译时保留 two only 的强调语气:只有 127 道「带代码」,仅十道「公布推理」——对比越直白,质疑越有力。
译文:372 道已被解出的问题中,只有 127 道附有保证逻辑一致性的代码(这是该领域的规范),而 AI 智能体的内部推理过程只公布了其中十道。
5. In using open maths problems as a benchmark to test new models’ abilities, the letter-writers said, model-makers risked prematurely shutting down lines of inquiry without generating new techniques, questions or, ultimately, understanding.
结构:
[In using open maths problems as a benchmark to test new models’ abilities],
↑ 分词状语(In using… as a benchmark to…):以……为基准
[the letter-writers said], [model-makers risked prematurely shutting down lines of inquiry
↑ 插入语,标明主张来源 ↑ 主句:risk doing 冒险去做
without generating new techniques, questions or, ultimately, understanding].
↑ without 短语作伴随否定,ultimately 把最重的词压在最后
难点:① 句首
In using A as a benchmark to do B 是分词状语,= By using A as a benchmark to do B,点出被批评的做法(把公开题当测试基准)。
②
the letter-writers said 夹在状语与主句之间作插入语,明确下句长主张是「信里说的」,不是作者观点——这是报道他人立场的标准疏离手法。
③ 主句
model-makers risked prematurely shutting down lines of inquiry:
risk doing(冒险去……)、
prematurely(过早地)、
shut down lines of inquiry(关闭研究路径)三词串出「因测试而误杀方向」。
④ 句尾
without generating new techniques, questions or, ultimately, understanding 用递进收束,
ultimately 把最重的 understanding 压在最后——与第 ② 段
more solutions but less understanding 首尾呼应;
or, ultimately, understanding 里的逗号把 ultimately 夹住,读作「或者,最终,理解」。
译文:信的作者们说,模型制造者用公开的数学难题作为测试新模型能力的基准,有可能在不产生新技术、新问题乃至最终的新理解的情况下,过早地关闭研究路径。
6. In working out how to release its new discoveries, the firm says it drew on the recommendations of an independent advisory group, a team of nine esteemed mathematicians formed in the aftermath of the Navier-Stokes announcement.
结构:
[In working out how to release its new discoveries],
↑ 状语:在拟定如何发布……时(work out 拟定、想清楚)
the firm says [it drew on the recommendations of [an independent advisory group,
↑ 主句 ↑ 宾语从句:drew on 借鉴
[a team of nine esteemed mathematicians [formed in the aftermath of…]]]].
↑ 同位语,修饰 group ↑ 过去分词后置定语
难点:①
In working out how to release… 里
work out 此处不是「解出」而是「拟定、想清楚(办法)」,后接
how to release(如何发布),作时间状语。
② 主句
the firm says it drew on the recommendations of…:
draw on(借鉴、利用)是学术与商业通用搭配,宾语是一个带长后置定语的名词短语。
③
an independent advisory group 后接同位语
a team of nine esteemed mathematicians,再接
formed in the aftermath of the Navier-Stokes announcement 这个过去分词短语作定语;
esteemed(德高望重的)比 respected 正式,
in the aftermath of(在……之后)是描写事件后续的高频短语。
④ 翻译时把两层修饰顺着名词还原即可:一个独立咨询小组,由九位德高望重的数学家组成、在纳维—斯托克斯问题公布后成立。
译文:公司表示,在研究如何发布这些新发现时,它参考了一个独立咨询小组的建议——这个小组由九位德高望重的数学家组成,是在纳维—斯托克斯问题公布之后成立的。
7. “The real importance of a problem is eventually measured by the new avenues it opens up and the light its proof sheds on other problems,” says Martin Hairer of EPFL, a Swiss university, who is a member of OpenAI’s advisory group as well as being a signatory of the Fields-medallists’ letter.
结构:
“[The real importance of a problem] is eventually measured by [the new avenues [it opens up]
↑ 直接引语主语 ↑ 谓语 be measured by ↑ 并列宾语 1(it 指 problem)
and [the light [its proof sheds on other problems]]],” says [Martin Hairer of EPFL, [a Swiss university],
↑ 并列宾语 2(its 指 problem) ↑ 引语动词 ↑ 身份同位语
[who is a member of… as well as being a signatory of…]]. ↑ 双重身份定语从句
难点:① 引语主干
The real importance of a problem is eventually measured by A and B:系动词
is measured by 后接两个并列名词短语作「衡量标准」。
② 并列宾语 1
the new avenues it opens up:
it 指 a problem,
open up avenues 是「开辟路径」;宾语 2
the light its proof sheds on other problems:
its 仍指 problem,
shed light on 是「阐明、使显形」,这里用所有格
its proof sheds 把动作写成「问题的证明投下光」。
③
avenues 与
shed light on 都是学术写作里极漂亮、可直接套用的表达;
measured by 把「重要性」写成可被某种东西度量的量,是隐喻用法。
④ 引语后
says Martin Hairer of EPFL, a Swiss university, who is… as well as being… 把引语者身份做成「双重身份」(OpenAI 顾问 + 反对信签署人),
as well as being 连接两个身份,提醒读者连反对阵营里的人也不否认 AI 的冲击力。
译文:「一个问题真正的重要性,最终要由它开辟了哪些新路径、它的证明为其他问题投下了什么光来衡量,」洛桑联邦理工学院(EPFL)的马丁·海泽说;他既是 OpenAI 咨询小组的成员,也是那封菲尔兹奖得主公开信的签署人之一。
8. For now, the mathematical problems AI is revealing are more existential than numerical.
结构:
[For now], [the mathematical problems [AI is revealing]] are [more existential than numerical].
↑ 时间状语 ↑ 定语从句(AI is revealing 修饰 problems) ↑ 表语:more A than B
难点:① 主干极短:
the mathematical problems are more existential than numerical,
more A than B 是「与其说 B 不如说 A / 更是 A 而非 B」的比较结构,此处强调「存在性」胜过「数字性」。
②
AI is revealing 是省略 that/which 的定语从句,修饰 problems(= problems that AI is revealing),口语化地省掉关系词,读时把 revealing 直接接 problems。
③
For now(就目前而言)收束全篇,把判断限定在当下;
existential(存在性的、关乎存亡的)与
numerical(数字的)构成双关,既回扣标题
Number crunch(数字运算),又把忧虑从「算得对不对」抬到「数学这门学科还存在吗」。
④ 这一句是全篇的 thesis 收束:前面所有数字(372、127、10、4000、28、8000)都指向同一个问题——当 AI 能自己「算」出数学,数学作为人类智识活动的意义是什么。
译文:就目前而言,AI 正在揭示的数学问题,更多是存在性的,而非数字性的。
05写作积累
5.1 六个可直接套用的句型公式
5.2 话题语料(可直接搬进作文)
话题一 · AI 与科学研究
AI for science 科学智能 · autonomous researchers 自主科研智能体 · frontier AI labs 前沿 AI 实验室 · AI-generated papers AI 生成的论文 · formal verification 形式化验证 · interpretability / internal reasoning 可解释性 / 内部推理 · benchmark 基准测试 · unreleased model 未发布模型 · mass production of mathematics 数学的批量生产
话题二 · 数学与千禧年大奖难题
Millennium Prize problems 千禧年大奖难题(七道) · the Navier-Stokes problem 纳维—斯托克斯问题 · the Riemann hypothesis 黎曼假设 · prime numbers 素数 · the 3D Kakeya maximal conjecture 三维 Kakeya 极大猜想 · Fields Medal 菲尔兹奖 · logical consistency 逻辑一致性 · proof 证明 / conjecture 猜想 / corollary 推论
话题三 · 透明度、问责与学术治理
transparency 透明度 · open letter 公开信 · be at odds with the field’s goals 与领域目标背道而驰 · shut down lines of inquiry 关闭研究路径 · independent advisory group 独立咨询小组 · signatory 签署人 · in hindsight 事后来看 · assessment of abilities and limitations 对能力与局限的评估 · beyond one’s capabilities 超出能力范围
5.3 仿写示范 题目:Some argue that using AI to automate scientific discovery will accelerate progress. To what extent do you agree?
Monday’s demo may go down as one of the most consequential moments in the history of research. The excitement of faster discovery is, for many scientists, freighted with concern for the integrity of their field. Unless more care is taken, the mass production of results will lead to more output but less understanding. Only a fraction of the generated studies were accompanied by reproducible code, and the models’ reasoning had been published for only a handful of them. More than 2,000 researchers published an open letter warning that such automation was at odds with the slow, cumulative nature of science. For now, the questions AI is raising are more philosophical than technical.
全段依次用到公式 1(定日期 + may go down as)、公式 2(freighted with concern)、公式 3(more A but less B)、公式 4(Only N… / only a handful)、公式 5(人数 + open letter warning that… at odds with)与公式 6(more A than B 点题)。注意两点:一是所有立场都由数字与动词推出,全程不出现 I think;二是结尾 more philosophical than technical 与原文 more existential than numerical 同构,既收束又把讨论从「工具」抬到「意义」。
06篇章结构、论证手法与文化冷知识
6.1 论证结构(定调 → 疑虑 → 加速 → 举例 → 记疑 → 反对 → 回应 → 收束)
01 定调 · 一个将被载入史册的日子(①)
用「具体星期二 + may go down as 史上最……之一」把刚发生的事抬进历史叙事。三个词定级:consequential(影响深远)、long-standing(长期悬而未决)、in a matter of weeks(仅数周)。最后用 at an unprecedented scale and speed 成对名词收束。
↓
02 疑虑 · 更多解答,更少理解(②)
全段用两个反直觉措辞翻面:glut(把喜讯重定义为过剩)、queasy(本能的不安)。结构骨架 The excitement of A is… freighted with concern for B 把兴奋与忧虑写成同船两货。主旨句 more solutions but less understanding 在此落定。
↓
03 加速度 · 从纳维—斯托克斯到数百道(③)
数据式论证范本:last month / same / just 88 hours 三处时间词层层收紧;cracked、stumped、corpus 三个词点亮难度与广度。唯一长句用 was equivalent to 把「372 道难题」与「仅三小时公开推理」画等号。
↓
04 名单 · 黎曼假设、Kakeya 与菲尔兹奖(④)
先让步(其余未解)再举例。does seem to have been made 的 does 强调「确有其事」;Most notable among these is… 与 Also buried in the deluge is… 两处倒装;claimed proof 的 claimed 不背书不否认;王虹的菲尔兹奖作可信度背书。
↓
05 保留 · 哪些能验证,哪些说不清(⑤)
记疑核心:几乎每句以「只 / 仅」开头。主语从句 what… and by what means + will take years;Only 127… accompanied by a code 与 only ten of them 的数字落差代替形容词质疑;subset of 4,000 点出采样偏差;Such transparency would… 用虚拟收尾。
↓
06 反对 · 28 位菲尔兹奖得主的公开信(⑥)
反对声四步走:观点 → 权威背书 → 理由 → 声势。not conducive to、at odds with 定调;In using… as a benchmark, the letter-writers said, 把主张明确归为「信里说的」;risked prematurely shutting down… + without… ultimately, understanding 与第 ② 段首尾呼应;8000 人联署收势。
↓
07 回应 · 九人咨询小组与资助会议(⑦)
OpenAI 一方回应,以「宣称 + 举措」为主。claims to have taken note 略带疏离;drew on the recommendations of + a team of nine esteemed mathematicians formed in the aftermath of 交代小组来历;In addition to sharing…, promised to fund… 把质疑引向「公开讨论」而非「公开证明」。
↓
08 收束 · 存在性大于数字性(⑧)
先 All told… only in hindsight 收束,再用 golden sentence measured by the new avenues it opens up and the light its proof sheds on other problems 点题。引语者身份做成「双重」(顾问 + 签署人)。结尾 more existential than numerical 双关回扣标题 Number crunch,把忧虑从「算得对不对」抬到「数学还存在吗」。
6.2 七种论证手法(可迁移到自己的议论文)
| 手法 | 原文例子 | 效果 |
| 定日期 + 封历史地位(date-as-history) | Tuesday October 6th 2026 may go down as one of the most consequential days in the history of mathematics. | 新闻开篇的省力写法:先给日历上的一个点,再立刻把它抬进历史叙事。读者不必先读三段背景,就已经知道这篇文章要给一件事定级。consequential 比 important 重,是这类定级句的默认用词。 |
| 情绪装载(freighted with) | The excitement of potential new mathematical truths is, for many researchers, freighted with concern for the future of their field. | 把两种相反情绪写成「同一艘船上的货」:主语 + be freighted with + 宾语,中间插入 for many B 限定范围。比「一方面……另一方面」紧凑,且 freight 的「装货」隐喻让抽象情绪有了重量——可迁移到任何「喜悦中掺杂忧虑」的写法。 |
| 数据式论证(data-driven claim) | Only 127 of the solved problems are accompanied by a code… and the agents’ internal reasoning has been published for only ten of them. | 不靠形容词评价,而靠「只 / 仅 + 数字」的落差来质疑。把关键数字放在句首(Only 127 / only ten),与总数 372 形成刺眼对比,比写「透明度不足」硬得多。全文几乎每段都由数字驱动(372、700、88、127、10、4000、28、8000、9)。 |
| 强调助动词 does(emphatic do) | Although the remaining Millennium Prize problems remain unsolved, progress towards one or two does seem to have been made. | 在助动词前加 does / do / did 把语气顶出来。此处用 does seem 防止读者把「一两个有进展」误读成「全解开了」,因为前句刚说过「仍未解决」。任何需要「纠正潜在误读」的地方都能用。 |
| 归源插入(he-said attribution) | In using open maths problems as a benchmark to test new models’ abilities, the letter-writers said, model-makers risked… | 把一长串主张用插入语 the letter-writers said 明确归为「信里说的」,而非作者观点。这是报道他人立场时保持疏离的标准手法,避免读者把第三方意见误当作报纸立场。 |
| 权威人数加码(authority by numbers) | 28 Fields-medallists published an open letter warning that… More than 8,000 of their colleagues have since signed on. | 用「具体人数 + 权威身份 + 公开信 / 联署」把一种观点写成「业界共识」。N 本身即是论据——28 位诺奖级得主、8000 人联署,数字比形容词更能让人感到「这不是个人意见」。 |
| 结尾双关收题(more A than B coda) | For now, the mathematical problems AI is revealing are more existential than numerical. | 结尾用 more A than B 把全文从具体议题抬到更大的提问,最好让其中一个词回扣标题(numerical 回扣 Number crunch)。For now 把判断限定在当下、留有余地。这是把「科技新闻」收成「哲学提问」的最省力写法。 |
6.3 文化冷知识
千禧年大奖难题(Millennium Prize Problems)是什么
2000 年由克雷数学研究所(Clay Mathematics Institute)设立,共七道,每道悬赏 100 万美元,被公认为当代数学最难的开放问题。包括黎曼假设(Riemann hypothesis,关乎素数分布)、纳维—斯托克斯存在性与光滑性(Navier-Stokes,流体方程)、P vs NP、霍奇猜想、庞加莱猜想(已证)、杨—米尔斯存在性、贝赫和斯维讷通猜想。原文说 OpenAI 解出的纳维—斯托克斯是「其中之一」,黎曼假设则「取得进展」——注意作者措辞的分寸:前者用 solved,后者用 progress… does seem to have been made,并不等同「已证」。
黎曼假设与素数分布
黎曼假设由波恩哈德·黎曼于 1859 年提出,断言黎曼 zeta 函数的非平凡零点都落在复平面的「临界线」上。它之所以被称为「数论皇冠上的明珠」,是因为它与素数(质数)的分布规律深度相关——素数的出现看似随机,却在宏观上遵循由黎曼假设刻画的规律。原文用 which concerns the distribution of prime numbers 一句点明其意义,但并未展开,因为对备考读者而言,记住「它关乎素数分布、是千禧年难题之一」已足够。
菲尔兹奖(Fields Medal)与香港数学家王虹
菲尔兹奖是数学界最高荣誉,每四年颁给 40 岁以下的杰出数学家,常被称作「数学界的诺贝尔奖」。原文提到的王虹(Hong Wang)是纽约大学柯朗数学研究所(Courant Institute)的数学家,2026 年因在三维 Kakeya 问题等相关方向上的工作获奖;而 OpenAI 声称攻克的「三维 Kakeya 极大猜想」(3D Kakeya maximal conjecture)正与王虹获奖工作的方向密切相关——作者特意点出这点,是在暗示这次 AI 的「所谓证明」恰好落在已被人类验证的工作延长线上,既给成果可信度,也暗含保留。
形式化验证与可解释性(formal verification & interpretability)
原文记疑的关键在两点:logical consistency(逻辑一致性)与 internal reasoning(内部推理)。在现代数学与 AI 安全里,形式化验证指用计算机证明助手(如 Lean)把证明写成机器可检查的代码,确保无逻辑漏洞——这正是那 127 道「带代码」题目的含金量。而 interpretability(可解释性)关心的是「模型到底怎么想的」:如果只给结论、不公开推理过程(原文说仅公布十道),外界就无法判断它究竟「证对了」还是「猜对了」。这两条是评估 AI 科研可信度的核心尺度,也是 28 位菲尔兹奖得主公开信的真正关切。
纳维—斯托克斯问题与 AI 的「88 小时」
纳维—斯托克斯方程是描述流体运动的基础方程,其「存在性与光滑性」是千禧年难题之一——通俗地说,就是问「描述流体的数学会不会在某处「爆掉」」。原文称 OpenAI 的同一内部模型「仅用 88 小时」解出它,这是全文最炸裂的单一数据点。需注意两点:一是「解出」在 AI 语境下通常指「生成了一份机器可检查的证明」,最终仍需人类数学家确认;二是这个 88 小时与后文「平均算力仅约三小时公开推理」形成呼应,共同支撑「量大而耗算极小」的叙事。
为什么 28 位菲尔兹奖得主会联名反对
公开信的核心担忧不是「AI 解不出题」,而是「AI 解题的方式会伤害数学这门学科」:当开放题被当成测试模型的基准,模型制造者可能为了刷分而关闭那些「不产生可量化结果」但孕育新思想的研究路径。信里说这「与领域目标背道而驰」(at odds with the field’s goals),并警告会 prematurely shut down lines of inquiry(过早关闭研究路径)。8000 人联署说明这并非个别人怀旧,而是数学共同体对「批量生产数学」的一次集体警觉——这正是原文主旨句 more solutions but less understanding 的现实注脚。