新智元报道
118511个。
这是清华大学沈阳教授团队用AI,从网上能公开收集到的资料里,整理出的数学猜想总数。
当然,这些文档不是给人看的,是给你的AI看的。
团队还让AI给这些猜想排了座次,评出前100名。在AI眼里,中国人最熟悉的哥德巴赫猜想,只排第49。
今天,新智元深度对话沈阳教授。
作为横跨文、理、工、医的跨界学者,他曾先后在计算机、信息管理、新闻传播、人工智能、临床医学等多个领域任教授或指导博士后,在AI绘画、AI音乐与AI人文社科领域的探索亦引发广泛社会关注。
访谈中,沈阳教授系统抛出四重核心思考:自进化数学(PCU框架)、自进化算法优化(PAE框架)、自进化软件工程(ECX范式),以及AI社会未来结构的两种终局前景。这不仅是一份技术演进路线图,更是一份人类在智能洪流中夺回「掌舵权」的思想备忘录。
近期,OpenAI在数学领域的突破引发热议。舆论都在追问:AI究竟解出了多少道难题?是否已经超越人类数学家?
但在我看来,一个更震撼的事实被忽略了——此次公开的722篇手稿只是冰山一角。只要系统愿意,它随时能抛出7000甚至70000篇手稿,等待人类去验证。
【核心转折】未来,机器生产数学知识的速度,将长期超过人类理解它们的速度。 人类将第一次大规模拥有自己尚未真正理解、却又必须持续验证和使用的数学知识。数学研究的重心,正从「攻克单个难题」转向「创造解决一类问题的新方法」;原本属于数学天才的部分大脑功能,开始迁移到持续运行的智能系统中。
数学由此可能成为最早实现大规模「知识自我繁殖」的学科:旧定理帮助发现新定理,新定理催生新工具,新工具又挖掘出更多数学结构,形成持续扩张的研究闭环。这也倒逼传统数学教育改革——AI对有判断力的资深学者友好,对敢于探索的年轻人友好,唯独削弱了单纯靠资历筑起的学术壁垒。
当AI能够轻易量产数学成果时,未来数学家最重要的工作不再仅仅是证明新定理,而是创造基于AI自我进化的新概念、新表达方式与新理论体系。 这要求我们重新审视一个底层问题:今天使用的数学概念,究竟是数学世界唯一的表达方式,还是人类为理解世界而建立的认知工具?
人类习惯使用符号、公式和图形,是因为它们契合人脑的思考习惯。人类现有的数学体系,或许只是适应人脑的一种「数学界面」,既非表达全部数学真理的唯一方式,更未必是最优方式。 AI不受人脑生理限制,未来很可能发现逻辑完全成立、却难以直接转化为人类熟悉概念的数学结构。到那时,数学家的重要任务,就是为这些陌生结构创造人类能够理解的「解释界面」。
基于这一变化,我提出面向AI时代的数学研究框架——PCU(Production 知识生产、Certification 成果认证、Understanding 认知理解):
这也将彻底改写数学的评价标准。当定理可以低成本量产,单纯增加定理数量便不再珍贵,数学将从「产出稀缺」走向「人类检验稀缺」。 若缺乏筛选、组织与整合能力,海量AI成果只会堆积成难以利用的「知识垃圾」。真正稀缺的,是能被反复调用、能统一大量既有结论、能催生新理论的「枢纽型成果」。
沿着这一思路,我进一步提出「自进化数学」。传统数学主要研究:在已有公理、定义和规则下,我们能证明什么?自进化数学则追问:我们能否创造全新的概念、表达方式与推理结构,让过去无法开展的数学研究变得可能?
它不仅研究数学知识本身,更研究数学如何借力人机协同,将繁复的底层推演自动提炼为新工具,持续生长出新的数学能力。最有价值的突破,未必是解出最难的一道题,而是创造一种新方法,让成千上万道题迎刃而解。
由此,我们还能利用AI探索更深远的「跨智能数学」。人类智能、不同架构的AI乃至未来的其他智能系统,会形成各自的数学概念、表示方式与证明偏好。有些人类觉得极难的问题,机器换种表示就能轻松解决;反之,人类凭直觉就能理解的结构,机器也未必采用相同的思考方式。数学真理未必因智能不同而改变,但通往真理的道路远比过去想象的更加丰富。
回顾数学史,真正改变数学面貌的突破,往往是新概念、新表示方式和新理论体系被创造出来。AI数学革命真正值得期待的,不是机器解决了多少著名猜想,而是建立起一套「自进化数学」体系,让AI与人类持续共创数学的未来。
今年,我把主要精力放在了一件事上:利用AI,让算法不断实现自我进化。
这项研究的起点源于AI绘画。过去三年,我累计使用AI创作了14万幅绘画作品,在全球举办了13次AI绘画展。在海量创作中,我形成了一套探索方法:先确定大致方向,再生成大量不同方案,通过比较、筛选和反复修改,不断逼近更优结果。
这套方法完全可以迁移到算法研究中。两者有一个底层共性:不必一开始就知道最优答案,而是先打开尽可能多的可能性,再通过持续反馈寻找突破。 本质区别在于:绘画追求审美上的新颖独特,而算法必须接受残酷的客观检验——若不能提升性能、降低成本或在真实环境中稳定运行,就谈不上成功。简单说,AI负责创造更多可能,实验负责判断哪些可能真正有价值。
随着实践深入,我开始思考一个更核心的问题:能不能让每一次算法优化的成功经验,都转化为下一次算法优化的能力? 这才是我真正希望实现的「自进化」。
沿着这条路线,我们团队,同时切入了数据压缩、优化求解、核聚变模拟、图像识别、工业声纹检测、CAD几何求解、文本分类等多个跨模态硬核领域。这在前AI时代不可思议,目前已取得七项阶段性成果:
除了这七项公开成果,我们还有近五十项在研算法探索——有的效果显著,有的进展缓慢,也有的最终失败。但失败的算法实验和成功的实验一样重要,因为我们要建立的不是只会展示成功案例的橱窗,而是一个能从成败中持续学习的进化系统。
回看这些跨界案例,我们既没用同一种算法,也没套用固定模型,真正相同的是背后的研究逻辑:每进入一个新领域,都反复追问三件事——当前算法卡在哪里?限制能否被突破?怎样用实验验证突破的有效性? 在此基础上,我提炼出通用的PAE算法自进化框架:
- 透析(Penetration,看清瓶颈):拆解算法运行全过程(数据、结构、训练、推理与资源消耗),找到真正限制性能的根源;不仅研究算法,更研究「研究算法的方法」。
- 增益(Advancement,兑现价值):让优化在客观指标与真实场景中产生可验证的净收益。
- 开域(Expansion,打开新空间):这是我最重视的部分。「向内开域」通过创造新的算法表示与求解机制,让过去难以企及的算法空间变得可探索;「向外开域」通过大幅降低计算成本与运行时间,释放资源,使更高频的实验与更复杂的实时交互成为可能。
透析发现突破口,增益兑现收益,开域创造下一轮可能。为此,我们构建了双层循环机制:内层循环不断优化具体算法(让算法越来越强),外层循环不断优化发现问题、设计实验和选择方案的方法(让创造算法的系统越来越强)。同时,我们按突破难度将优化分为四个层级:基于人类算法的弱优化、对AI优化结果再提升的强优化、挑战强优化边界的极限优化,以及在既有约束下看似无解的不可能优化。
【算力突围定律】算力差距,并不必然等于算法性能差距。 过去大家迷信「大力出奇迹」,但在具体算法问题中,搜索路径、问题分解方式与收敛效率同样关键。如果别人需要盲目搜索一万个候选方案,而我们通过深刻的「透析」与算法构造,只需搜索几个关键方案,那么即便计算资源相差一万倍,我们依然能在特定任务上实现反超。
当拥有强大算力的AI系统把容易摘的果子都摘走后,人类的机会恰恰存在于AI尚未找到的搜索路径中——人类的角色将从「亲自完成每一步计算」,转向发现关键约束、创造新问题表示、判断探索方向以及设计验证机制。
近期,我们已经启动用AI优化量子计算机自进化测控平台的研发,初步看来效果很不错,并在探索「自进化工业」与「中医AI自进化」的早期理论与实践。如果有算力当然更好,实在没有,就在算法效率上找思路。跨领域突破离不开行业专家的深度参与——核聚变专家懂物理意义,工业工程师懂故障逻辑,医学专家懂临床价值。领域专家负责提出真问题、判断专业价值,我们的AI自进化系统负责大规模探索优化空间。如果你所在的行业有长期难以突破的算法与技术瓶颈,欢迎到我的微博(@清新研究)后台留言,我们一起完成具有一定概率的验证与落地。
当AI已经能够生成代码、搭建应用、优化算法,软件工程的变革,难道只是让传统开发流程变得更快吗?
AI虽未消除工程难题,却大幅降低了将想法转化为原型的成本。软件研发的重心正发生根本转移——比起「如何把功能做出来」,如今更重要的是回答三个问题:它是否值得实现?相比其他方案有何优势?上线后为何值得长期保留?
基于这一思考,并在我此前提出的「四阶校验」(AI校验、人校验、社会校验、时间与规律校验)理论基础上,我进一步提出了ECX三维并行研发范式:
- E(Evolution,性能进化):AI自动识别系统瓶颈并生成优化方案,持续提升运行效率、稳定性与资源利用率。
- C(Candidate,功能候选):综合用户反馈、竞品分析与AI构想,批量生成候选方案;通过比较、组合与筛选,决定功能的去留与演进。
- X(eXperience,交互验证):以人为本,通过真实用户的交互反馈与行为数据,识别体验痛点与需求缺口,作为下一轮进化的路标。
这三个维度并非串行推进的流水线,而是并行运行、相互触发的进化闭环。目前,我们正通过「小森邻」与北航何静副教授主持研发的「AI科学家」两个项目,检验核心命题:软件能否在持续使用中,一边优化已有能力,一边生长出全新的能力?
「小森邻」是一次典型的AI原生软件研发实践:仅靠两个人的团队,获客6万,研发两个月即完成天使轮融资。 今年3月,毫无编程基础的主理人青青在社交平台分享软件构想,零粉丝账号首帖即获三千点赞,完成了最初的需求验证。随后,她借助AI「氛围编程(Vibe Coding)」在一个月内做出初代Demo,并引入一名专业程序员,形成了「非技术主理人+专业程序员」的极简双人模式,完整跑通了ECX机制:
- 在E(性能进化)维度:软件的「关系网」功能在渲染60个角色、180条关系时曾出现明显卡顿。借助AI辅助定位与优化路径分析,团队采用批量绘制、跳过无效线段与缓存复用等策略,将渲染耗时从138–182毫秒直接压缩至18–19毫秒,耗时锐减86%–90%。软件可以在六年前的手机流畅运行。
- 在C(功能候选)维度:青青累计整理了30万字的用户留言,利用AI进行需求归类与场景拆解。AI针对同一需求生成多种方案,经原型对比与成本评估后择优上线。近期每次功能发布都能获得过万点赞,最高是一个功能在小红书获得了5.7万个点赞。
- 在X(交互验证)维度:通过高粘性的用户社群持续收集真实操作习惯与审美偏好,让用户成为产品进化的共创者。
如果说「小森邻」探索的是轻量级应用的敏捷演进,那么已拥有近30万用户的科研平台「AI科学家(让科研像聊天一样简单)」面对的则是复杂工具链的演进。在面向大规模科研用户的迭代中,该项目展现出软件进化的另一重特征:新能力不仅来自开发新功能,更来自已有能力的重新组合。 例如近期更新中,系统将原本独立的检索、推演、代码生成与写作模块打通,涌现出了「自主生成算法—验证算法有效性—形成论文手稿」的复合工作流。
通过这两款产品的实践,我将AI时代的软件进化概括为两种基本运动:
- 向内优化:在已有能力边界内持续提质,让执行更快、更稳、成本更低(如小森邻的渲染提速、AI科学家的模型调度优化)。
- 向外生长:不断拓宽解决问题的边界,通过新增模块或重组工具链,生长出过去没有的新能力。
当向内优化与向外生长都能由AI加速完成时,生成越容易,校验就越重要。 因此,在ECX体系中,人类并未退出研发,而是从重复编码转向了目标设定、价值判断与风险控制。
拆解AI原生软件的研发过程,实际上存在三个核心角色:意图提出者(I,Intent)、原型开发者(P,Prototype)和工程落地者(L,Landing)。在「小森邻」中,青青一人兼任了I和P,专业程序员担纲L;而在「AI科学家」项目中,何静副教授凭一己之力包揽了I、P、L全部三个角色。这也从底层解释了为什么OPC(一人公司)模式在当下能够真正成立——这是AI技术赋能与新型组织架构共同催生的必然产物。
未来的软件竞争,本质上是进化机制的竞争。软件工程正在从「交付一款静态产品」,走向「建立一套能够持续生成、验证和进化软件能力的生命系统」。
跳出具体的技术与工程,我们必须直面一个更宏大的命题。过去几百年支撑现代社会运转的两条底层铁律——「资本赚钱离不开普通人的体力与脑力」以及「逃避现实的代价极其高昂」——正在被「人工智能+清洁能源+机器人」的闭环彻底打破。在这场剧变面前,人类文明正站在决定命运的十字路口。
如果我们随波逐流,任由技术垄断与算法惯性失控,社会就会滑向「高技术、低理性、强固化」的三层折叠深渊;但如果我们主动干预,用公共制度驯服技术野心,这恰恰也是人类摆脱繁重生存奔波、迈向「三层协同世界」的历史契机:
在智能社会的基座上,最大的隐患不是物资匮乏,而是「温柔的废掉」。如果任由商业算法唯利是图,毫秒级响应微表情的短视频流与永远顺从的AI伴侣,就会变成24小时投喂爽感的「电子子宫」。长此以往,大众负责逻辑推演与抗压的大脑前额叶就会系统性退化。即便未来发放「全民基本收入(UBI)」,那也绝非乌托邦——底层大众虽然温饱无忧,却在沉迷中彻底丧失了维护自身权益的社会议价权。
技术兜底绝不等于精神缴械。我们必须通过制度与文化建设争取正面前景:
- 建立算法的「防沉迷护栏」:立法限制掠夺注意力的成瘾性算法设计,打破信息茧房,保护公众的神经系统不被商业资本殖民。
- 为「发展赋能」:将基本生活保障与社区公共参与、终身教育、线下文体活动深度绑定,让摆脱生存焦虑的普通人不仅有饭吃,更有丰富的真实社区连接与精神成长空间。
对于夹在中间的专业人士、超级个体与微型团队来说,AI时代绝非天然的黄金时代,而是布满了制度陷阱。最危险的滑坡,是沦为「数字佃农」与「肉身背锅侠」:当标准化脑力劳动被吞噬,中层只能依附顶层巨头的大模型接口(API)讨生活。巨头只需改一行参数、涨一次接口费,或在下一次更新中将某项功能收归官方,就能让成千上万个小团队一夜清零。更残酷的是,当AI连审美与直觉都超越人类时,由于AI没有肉身、不能坐牢,中层人类留存的唯一价值,可能只剩下在医疗事故、金融爆仓与工程坍塌的责任书上签字画押,充当承担无限连带责任的「替罪羊」。
要打破这种异化,中层力量必须主动争取技术生态的博弈权:
- 推动开源生态与反垄断:支持开源模型与分布式算力,打破少数科技巨头对底层接口的生杀大权,把「租地种田」变成「自带农具」。
- 把「背锅的义务」转化为「掌舵的权力」:既然人类必须为最终结果承担法律与道德责任,那么人类专业者就必须拥有对黑箱算法的一票否决权与审查权。只要我们坚持算法透明、严守安全底线,那份只有人类血肉之躯才能承担的责任,就会变成任何超级算力都无法剥夺的最高指挥权。
顺着博弈论往下推演,智能时代最令人警惕的社会死锁,是「顶层算力寡头」与「底层沉迷大众」的跨层合谋。对沉迷虚拟爽感的大众而言,发放补贴、提供免费娱乐的算力巨头是「大恩人」,而中间那群试图关掉短视频、逼大家面对现实的理性清醒者,反而是最讨厌的说教者;对顶层寡头而言,温顺的底层最好管理,要求算法透明、试图拥有独立算力的理性中层才最危险。
一旦顶层垄断资本利用底层的多数选票,合法地收缴中层的技术自主权,人类就会陷入棘手的「洞穴拔电源悖论」:当大多数人自愿留在温暖逼真的虚拟洞穴里时,外面的清醒者该如何说服洞穴里的人不拔掉自己的电源?
解开这个死局,靠居高临下的启蒙说教毫无胜算,唯一的出路是用行动重建「真实连接」:
- 放下精英傲慢,与大众利益绑定:清醒者与创新者不能只顾自己突围,而要用技术去解决普通人最关心的养老、医疗、教育与社区陪伴等真问题,让大众切身感受到「开放理性的技术」比「封闭的电子鸦片」更能带来长久的幸福。
- 降低真实世界的挫败感:人们逃进虚拟洞穴,往往是因为现实社会容错率太低。只有建立更宽容的社会安全网与更温暖的线下社区,让普通人在现实中试错不至于倾家荡产,真实世界才能重新赢回人心。
将这套模型投射到中国的现实土壤中,既有独特的结构性阵痛,也蕴藏着破局的深厚底气。我们必须清醒直面三大现实挑战:
- 「做题家契约」的剧烈贬值:城市家庭耗时二十年、倾注心血培养出的「高准确率拟合标准答案」的应试脑力,在AI面前一夜之间失去市场溢价,引发深层教育焦虑。
- 超级平台的挤压效应:巨头「全栈通吃」的惯性,容易让中小型创新团队寸草不生。
- 精神内卷与低质泄压:若仅靠廉价物资兜底加上高浓度AI微短剧泄压,虽能维持低成本稳定,却会悄然耗尽全社会的创新活力。
但与此同时,中国手中握着一张全球独一无二的王牌——国家统筹的「新基建与全产业链底座」。如果说硅谷的优势在于前沿算法,中国的底气则在于「最充沛的绿色电力与最庞大的实体制造躯干」。只要引导得当,我们完全可以走出一条极具韧性的正面路径:
- 用「便宜的电与算力」赋能百业:中国牢牢掌握着特高压电网、绿色新能源、国家级算力枢纽与全门类制造业供应链。只要严格规范平台垄断,就能把这些超级基建变成千万中小企业都能低成本调用的「公共水电」。
- 推动教育从「刷题」转向「破题」:趁着标准答案贬值的契机,把年轻一代从机械刷题中解放出来,转向培养好奇心、动手能力与解决复杂现实问题的能力。
- 升级「数字以工代赈」:将基层的半公共岗位从单纯的「网格化管理」,升级为养老照护、生态修复、文化传承与社区营造等真正创造社会价值的温暖事业。
只要跨过应试思维的转型阵痛,中国几代人积累下的工程师红利与实干精神,就能借着AI的东风,从「按标准答案解题」真正走向「向未知领域开疆拓土」。
面对庞大的算法帝国,普通人究竟凭什么避免被淘汰?答案就藏在AI系统自身无法克服的死穴里——「模型自噬(Model Autophagy)」。
当全社会都在消费AI生成的内容,并用这些二手合成数据去训练下一代AI时,算法生态就会像近亲繁殖一样迅速陷入逻辑退化与崩溃。这意味着,AI越是强大,就越是极度渴求人类在真实物理世界中探索、试错所产生的「一手真实数据」。 虚拟世界永远无法自我闭环,真实世界才是算法的生命源泉。
而在中国的珠三角、长三角以及广大创新一线,全球最廉价的光伏储能、开源硬件与本地开源大模型,正在赋予普通人前所未有的物理底气:
- 技术自耕农化(把工具握在自己手里):普通极客与中小团队完全可以利用廉价绿电、开源芯片与本地大模型,建立起哪怕断网也能独立运转、动手造物的能力,成为既懂前沿代码、又能拿扳手解决实际问题的「现代墨家」。
- 锻造心智护城河:在算法泛滥的时代,主动锻炼大脑的专注力与逻辑力,是最具回报率的自我投资。未来真正的护城河,不再是你记住了多少免费的知识,而是戒断算法投喂、保持深度阅读、忍受枯燥与挫折、进行长链条逻辑推演的自律能力。
【终局结语】悲观的结局只属于彻底放弃思考的顺从者。只要我们坚持走出温暖的算法洞穴,去和真实的物理世界发生碰撞、去亲手创造并承担后果,人类就永远是这个智能时代真正的造物主与掌舵人。
编辑:所罗门
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.
118511.
This is the total number of mathematical guesses compiled by the team of Professor Shenyang of the University of Qinghua using AI from publicly available information available online.
Of course, they're not for people. They're for your AI.
The team also gave AI seats to these guesses and ranked top 100. In AI's eyes, the Chinese's most familiar Gothbach guess is only 49th.
Today, Professor Shen Yang of the Nixuan In-depth Dialogue.
As a cross-border scholar, he has been a professor or mentor in a number of fields, including computers, information management, news dissemination, artificial intelligence, clinical medicine, etc., and has also generated widespread social interest in the fields of painting, music and humanities.
During the interview, Prof. Shenyang put four core ideas: self-evolution mathematics (PCU framework), self-evolution algorithm optimization (PAE framework), self-evolution software engineering (ECX model), and two final perspectives on the future structure of AI society. This is not just a road map for technological evolution, but also a memo of the idea of humans taking back the "manager" in the smart flood.
Recently, Open AI's breakthrough in mathematics has given rise to enthusiasm. Public opinion is asking: How many problems did AI solve? Has it gone beyond human mathematicians?
But, in my view, a more dramatic fact has been ignored — the 722 manuscripts that were made public this time are only the tip of the iceberg. If the system is willing, it can throw 7,000 or even 70,000 manuscripts at any time, waiting for humans to verify them.
In the future, machines will produce mathematical knowledge at a faster rate than humans can understand. For the first time, humankind will have a large-scale mathematical knowledge that it does not really understand, but that must be continuously validated and used. The focus of mathematical research is shifting from "hunting one problem" to "creating new solutions to one problem"; part of the brain function, which was originally a mathematical genius, began to migrate to a functioning smart system.
Mathematics can thus become the first discipline to achieve large-scale "knowledge self-reproduction": old theorems help to discover new theorems, new theorems generate new tools, and new tools to dig up more mathematical structures and create a continuing expanding research loop. This has also forced traditional mathematics education reform — AI is friendly to judgemental senior scholars and to young people who dare to explore, but only weakens academic barriers based solely on seniority.
When AI can easily produce mathematical results, the most important job for future mathematicians is not just to prove the new theorem, but to create new concepts, new expressions and new theoretical systems based on AI’s own evolution. This requires us to revisit a bottom-up question: is the concept of mathematics used today the only expression of the mathematical world, or is it the cognitive tool built by humans to understand the world?
Humans are accustomed to using symbols, formulas and graphics because they fit the thinking habits of the human brain. The mathematical systems that exist in humans, perhaps just a "mathematical interface" adapted to the human brain, are not the only way to express all the mathematical truth, much less the best. AI is free from the physical limits of the human brain, and it is likely that in the future it will find that logic is fully established but that it will be difficult to translate directly into mathematical structures of human familiarity with concepts. By then, the important task for mathematicians is to create an "interface of interpretation" that humans can understand for these strange structures.
Based on this change, I propose a framework of mathematical research for the age of AI - the PCU (Producing Knowledge Production, Certification Results, Understanding):
This will also result in a thorough rewriting of the evaluation criteria for mathematics. When the theorem can produce low-cost quantities, simply increasing the number of theoretics is no longer precious, and mathematics will move from "result scarcity" to "human testing scarcity". In the absence of screening, organization and integration capacity, the volume of AI results can only accumulate into hard-to-use "knowledge trash". What is really scarce is a "junctive outcome" that can be repeatedly called upon, that unites a large number of established conclusions, and that generates new theories.
In this line of thinking, I further propose "self-evolutionary mathematics". Traditional mathematics is a major study: what can we prove when justice, definitions and rules exist? Self-evolutionary mathematics asks: Can we create completely new concepts, expressions and reasoning structures that make it possible to conduct mathematical research that was not possible in the past?
It researches not only mathematical knowledge per se, but also how mathematics can leverage human power to convert complex bottom-ups into new tools that sustain new mathematical capabilities. The most valuable breakthrough is not necessarily the most difficult issue, but rather the creation of a new approach that will solve thousands of problems.
As a result, we can also use AI to explore more far-reaching "cross-intellectual mathematics." Human intelligence, different structures of AI and other future intelligent systems shape their own mathematical concepts, expressions and probative preferences. There are issues that human beings find extremely difficult, and machine substitution means easy to solve; conversely, structures that humans can understand intuitively are not necessarily the same way of thinking. Mathematic truths do not necessarily change with different intelligence, but the path to truth is far richer than it used to be.
Looking back at the history of mathematics, the breakthrough in truly changing the mathematical landscape is often the creation of new concepts, new expressions and new theoretical systems. What is really to be expected from the AI Mathematical Revolution is not how many famous guesses have been solved by the machine, but the creation of a system of "self-evolutionary mathematics" that allows AI and humans to continue to build the future of mathematics.
This year, I focused on one thing: using AI to keep algorithms evolving.
The starting point for this study is the painting of AI. Over the past three years, I have used AI to create 140,000 paintings and held 13 AI painting exhibitions around the world. In mass creation, I have developed a methodology of exploration: to define the broad direction, then to generate a large number of different programmes, and to move closer and closer to better results through comparison, screening and repeated modifications.
The methodology could well be migrated to algorithm studies. There is a bottom-up commonality between the two: it is not necessary to know the best answers from the outset, but to open up as many possibilities as possible before finding a breakthrough through continuous feedback. The essential difference is that painting seeks the novelty of aesthetic aesthetics, and algorithms must be subjected to a cruel and objective test — failure to improve performance, reduce costs or operate steadily in a real environment. In short, AI is responsible for creating more possibilities, and the experiment is responsible for judging what may really be of value.
As practice went further, I began to reflect on a more central question: can successful experiences of each algorithm optimized be translated into the ability to optimize the next one? That's what I really want to do.
Along this line, our team has also moved into multiple cross-model hard core areas, such as data compression, optimisation, fusion simulation, image identification, industrial acoustic detection, CAD geometry, text classification, etc. This was incredible in the former AI era and seven milestones have been achieved:
In addition to these seven public results, we have nearly 50 other algorithms being explored — with significant results, slow progress and ultimate failure. But failure algorithms are as important as successful experiments, because we want to build not just windows that show success stories, but an evolutionary system that can learn from success or failure.
Looking back at these cross-border cases, we don't use the same algorithm, we don't apply a fixed model. The real same is the logic behind the study: every time we enter a new field, we ask three things over and over again — where are the current algorithm cards? Can the restrictions be broken? How do you test a breakthrough? On this basis, I developed a common PAE framework for self-evolution:
- dialysis, identifying bottlenecks: dismantling algorithms to run the whole process (data, structure, training, reasoning and resource consumption) to find the source of a truly restrictive energy; studying not only algorithms but also “methodology of algorithms”.
- Gains (Advancement, Realizing Value): to allow optimization to produce verifiable net gains in objective indicators and real scenes.
- Expansion, opening up new spaces: this is the most important part of my mind. By creating new algorithms to express and solve the algorithmic spaces that were difficult to access in the past, it is possible to release resources from higher frequency experiments and more complex real-time interactions by significantly reducing costing and running time.
The dialysis found a breakthrough, realized gains and created a potential for the next round. To this end, we have constructed a two-tiered cycle mechanism: the inner cycle constantly optimizes specific algorithms (making algorithms stronger), and the outer cycle constantly optimizes methods of detecting problems, designing experiments and selecting options (making algorithms more powerful). At the same time, we divide the optimization into four levels, depending on the difficulty of the breakthrough: weak optimization based on human algorithms, strong optimization of the optimisation of AI, challenges to the optimization of boundaries at limits, and seemingly incomprehensible optimization under existing constraints.
The computational gap does not necessarily equal the algorithmic performance gap. In the past, we used to be superstitious about "demonstrating miracles," but in specific algorithms, search paths, problem decomposers are as critical as efficiency. If others need to search blindly for 10,000 candidates, and we need to search for a few key options through deep "dialysis" and algorithms, we can still do the opposite on a given mission, even if the computation of resources is 10,000 times different.
When AI, with its powerful calculus, removes all the easy fruits, human opportunity lies in the search path that AI has yet to find. – The role of human beings will move from “self-enumerating” to identifying key constraints, creating new problem indications, judging the direction of exploration, and designing certification mechanisms.
In the recent past, we have initiated the development of a self-evolutional monitoring platform using the AIMS computer, which initially appears to have worked well, and is exploring early theories and practices of the "self-evolution industry" and the "Chinese medical AI" evolution. If it's good to be smart, it's not. Cross-cutting breakthroughs require the deep involvement of industry experts — nuclear fusion experts understand the physical meaning, industrial engineers understand the logic of failure and medical experts understand clinical values. The experts in the field are responsible for raising real questions and judging professional values, and our AIA self-evolution system is responsible for large-scale exploration of the optimal space. If your industry has long and hard-to-reach algorithms and technological bottlenecks, welcome to my tweet backstage, and together we'll complete the test and landing with a certain probability.
When AI has been able to generate codes, build applications, optimize algorithms and change software engineering, is it simply to make traditional development processes faster?
While AI did not resolve the engineering challenge, it significantly reduced the cost of transforming ideas into prototypes. There is a fundamental shift in the focus of software development — more important now than “how to make it work”, are three questions answered: is it worth it? What are the advantages of other programmes? Why should it be worth staying on the line for a long time?
On the basis of this reflection, and on the basis of my earlier theory of "Ai verification, human verification, social verification, time and regularity" I further proposed a three-dimensional parallel development paradigm for ECX:
- E (Evolution, Evolution of Performance): AAIS bottlenecks and the generation of optimisation programmes to continuously improve operational efficiency, stability and resource utilization.
- C (Candidate, functional candidate): Comprehensive user feedback, competitive analysis and AI concept, mass generation of candidate options; determination of function retention and evolution through comparison, grouping and screening.
- X (eXperion, interactive validation): People-centred, through interactive feedback and behavioural data from real users, to identify pain points and needs gaps as road signs for the next round of evolution.
These three dimensions are not streaming water lines that are carried forward in a staggered manner, but are co-operating, mutually-triggered, evolutionary closed loops. At present, we are testing the core theme of the "AI Scientists" project, which is being developed under the auspices of "Son Neighbors" and "Associate Professor Ho Jing Ho of North Navigation": Can software be used on a continuous basis while optimizing the capabilities that are already available while developing new ones?
Sam Neighbors is a typical AI-based software development practice: with a team of two people alone, 60,000 visitors, two months after the development of the Angel Wheel. In March of this year, the primer, who had no basis for programming, shared the software concept on social platforms, with a zero fan account for 3,000 hits, completing the initial needs validation. Subsequently, she made her first-generation Demo within a month with the help of AI 'Vibe Coding' and introduced a professional programmer, creating a very simple two-person model of "untechnical principal plus professional programmer" and running through the ECX mechanism:
- On the E (performance evolution) dimension: The software's "relationship" function appeared in MingZ Q114 Zton while reproducing 60 roles and 180 relationships. Using AI-aided positioning and optimization path analysis, teams use such strategies as batch mapping, skipping invalid lines and reusing cache to compress the rendering time directly from 138-182 ms to 18-19 ms, a sharp reduction of 86-90%. The software was running on mobile phones six years ago.
- In C (functional candidate) dimension: 300 thousand words of user messages were accumulated, and AI was used for needs classification and scenario dismantling. AI produces multiple programmes for the same needs, which are available on the basis of prototyping and cost assessment. publish has recently received 57,000 praises for each function, the highest being REDnote.
- X (interactive validation) dimensions: through the continuous collection of real operating habits and aesthetic preferences by the highly sticky user community, users become co-founders of product evolution.
If "Sam's neighbor" explores a quick evolution of lightweight applications, the scientific platform "AI Scientist" (which makes it as simple as chatting) that already has nearly 300,000 users faces the evolution of complex tool chains. In an iterative approach to large-scale scientific users, the project presents another feature of software evolution: new capabilities come not only from the development of new functions but also from the regrouping of existing capabilities. For example, in a recent update, the system introduced the original stand-alone search, rollover, code generation and writing module, which resulted in a composite workflow of the "Autonomous Generating Algorithms - Validation of the Algorithms - Formation of the dissertation manuscript".
Through the practice of these two products, I have outlined the evolution of the software of the AI era as two basic movements:
- Internal optimization: continuous upgrading within the already capable borders to make implementation faster, more stable and less costly (e.g., speed of retrofitting by the neighbours of Sen, optimization of model movement by AI scientists).
- Outward growth: widening the problem-solving boundary to produce new capabilities that were not previously available through new modules or reorganised tool chains.
The easier the generation, the more important the validation is when both inward optimization and outward growth can be accomplished by AI. Thus, in the ECX system, humans did not withdraw from R & D, but instead moved from recoding to targeting, value judgement and risk control.
There are in fact three core roles in the process of developing and dismantling AI-based software: the intention proponent (I, Intent), the prototype developer (P, Prototype) and the project lander (L, Landing). In the "Sen Neighbor", one of them was also I and P, a professional programmer, L., while in the "AiScientist" project, Deputy Professor Ho was able to take on all three roles, I, P and L. And that explains why the OPC model is actually set up in the present -- This is a corollary of AI technology empowerment and new organizational structures.
Future competition for software is essentially a competition for evolutionary mechanisms. The software project is moving from "delivery of a static product" to "development of a life system capable of generating, validating and evolutionary software on a continuous basis".
Out of specific technologies and engineering, we must confront a bigger proposition. The two bottom iron laws that have underpinned the functioning of modern society over the last few centuries — “capital cannot be earned without the physical and intellectual strength of ordinary people” and “extremely expensive to escape reality” — are being completely broken by the closes of “Agent intelligence — clean energy — robots.” In the face of this upheaval, human civilization stands at a crossroads in determining fate.
If we go with the waves and let technology monopolize and algorithmic inertia get out of control, society slides to the three layers of the abyss of "high technology, irrationality, robustness", but if we intervene to tame technological ambitions with the public system, this is also the historic opportunity for humanity to emerge from its heavy existence and move towards "three layers of synergy":
At the very base of an intelligent society, the greatest danger is not the lack of material, but the "sweet loss". If commercial algorithms are left to profit, the short video stream of milliseconds responding to micro-impressions and an AI partner who will always be submissive will turn into a 24-hour "electronic uterus" feeding a sense of comfort. In the long run, the frontal lobe of the brain, which is responsible for the logic and resistance to pressure, is systematically degraded. Even if UBI were to be released in the future, it would not be utopian — the bottom-class population, while well fed, has lost its social bargaining power to defend itself.
The technology is in no way tantamount to a mental weapon. We must seek positive prospects through institutional and cultural building:
- The creation of an algorithm's "prevent the fascination fence": legislation restricts the design of addictive algorithms that plunder attention, breaks the information cocoon and protects the public's nervous system from commercial capital.
- To "development empowerment": to bind basic livelihood security to community public participation, life-long education, and the depth of below-line activities, so that ordinary people who are free from existential anxiety not only have food, but also have more room for real community connections and spiritual growth.
For professionals, super-individuals and micro-teams caught in the middle, the AI era was not a natural golden age, but a system trap. The most dangerous landslide is that of a "digital tenant" and "manager man": when standardized brain work is swallowed up, the middle layer can only live with the LLM interface (API) of the top giant. One line of argument, an additional interface fee, or the officialization of a function in the next update would allow thousands of small teams to be wiped out overnight. Even more brutal, when AI is superior to human beings in its aesthetics and instincts, because AI has no body and cannot go to prison, the only value that mid-level humans retain may be to sign a pledge on a letter of responsibility for medical accidents, financial implosions, and the collapse of the project, and to act as a “sacrifice” of unlimited joint and several liability.
To break this alienation, mid-level forces must take the initiative to compete for technological ecology:
- Promotion of open-source ecology and anti-monopoly: support for open-source models and distributed algorithms, breaking down the life-killing power of a few technology giants over the bottom interface and transforming the term "rent land farming" into a "handy farming tool".
- Translating the obligation to take the fall to the "right to take the wheel": since human beings must be held legally and morally responsible for the end result, human professionals must have the right of veto and censorship over black box algorithms. As long as we adhere to the transparent and secure threshold of algorithms, the responsibility that can be assumed only by the flesh and blood of man becomes the supreme command that no supercalculations can deny.
Following the game theory, the most alarming social locks of the intelligent age are the oligopoly at the top and the bottom of the crowd. For people who are obsessed with virtual pleasures, the credit giants who provide subsidies and provide free entertainment are the "grateful people," while those in the middle who try to shut down short video, who force you to face reality, to be rational and sober, are the most hateful preachers; for the top oligopoly, the soft bottom is best managed, requiring that algorithms be transparent, and that rational mid-levels that try to be independent.
Once the top-level monopoly of capital takes advantage of the lower-level majority vote and legally seizes the middle-level technological autonomy, humans will be caught in the thorny “cave-to-power paradox”: When most people volunteer to stay in a warm, virtual cave, how do the soberers out there convince people in a cave not to pull out their power?
The only way out of this situation is to re-establish the "real connection" with action:
- Put down the arrogance of the elite and tie it to the public interest that soberers and innovators must not just stop by themselves, but use technology to solve the real problems of old age, medical care, education and community solidarity that are of greatest concern to ordinary people, so that the "open and rational technology" is more likely to bring about lasting happiness than "closed electronic opium."
- Reducing the frustration of the real world: people escape into virtual caves, often because of the low tolerance of the real world. The real world will be able to win back only if more tolerant social safety nets and warmer sub-lined communities are established so that ordinary people are not mistaken in reality to lose their fortune.
To project the model into China's real soil, there are unique structural pains and deep-seated gases. We must confront three real challenges:
- The sharp devaluation of the "problemers contract": the 20-year-old "high-accuracy standard answer" developed by urban households, which lost a market premium overnight in front of AI, gave rise to deep educational anxiety.
- Squeeze effect of super-platforms: the inertia of giants to eat all, which tends to keep small and medium-sized innovative teams alive.
- Psychiatry and low-quality excretion: If low-cost materials are used only to cover the floor with high-intensity AI microclaves, low-cost stability can be maintained, but the innovation of society as a whole can be depleted.
But at the same time, China holds a unique global ace in its hands — the nationally integrated “new infrastructure and the bottom of the entire industrial chain”. If Silicon Valley has the advantage of cutting-edge algorithms, the bottom line in China is "the most abundant green power and the largest physical tortilla." With the right guidance, we can walk a very resilient and positive path:
- Powerful with "cheap power and calculus": China is firmly in control of the ultra-high-voltage grid, green new energy sources, national computing hubs and the entire manufacturing supply chain. Strictly regulated platform monopolies can transform these superstructures into “public utilities” that are available to millions of SMEs at low cost.
- Promoting a shift in education from a "brushing" to a "buttering": freeing the younger generation from a machine-brushing opportunity to take advantage of the devaluation of standard answers to develop curiosity, ability to act and ability to solve complex realities.
- Upgrading of the term "digitals for work": from simple "grid management" of semi-public jobs at the grass-roots level to a warm business of truly creating social values, such as old-age care, ecological restoration, cultural heritage and community-building.
As long as the transition pains of prodigious thinking are crossed, the engineers’ dividends and realism accumulated by generations in China can be used in the East Wind of AI to truly move from a “standard answer” to a “broadening of the land to the unknown.”
How can ordinary people avoid being eliminated in the face of a vast algorithmic empire? The answer lies in a dead hole that the AI system itself cannot overcome: "Model Autophagy."
When society as a whole consumes AI-generated content and uses these second-hand synthetic data to train the next generation of AI, algorithmic ecology falls into logical degradation and collapse as quickly as proximate reproduction. This means that the more powerful AI is, the more eager it is for humans to explore and try the "real data of the first hand" in the real physical world. The virtual world can never close itself, and the real world is the source of algorithms.
In China, the Pearl Triangle, the Long Triangle and the vast array of innovations, the world's cheapest photovoltaic energy, open-source hardware and local open-source LLM are giving ordinary people a physical base gas that is unprecedented:
- Technology farming (to hold tools in their hands): ordinary bingoists and small and medium-sized teams can fully use cheap green power, open-source chips and local LLM to build the capacity to operate independently and make their own work, even without a net, and become “modern inkmen” with both front-line code and wrench solutions to real problems.
- Cultivation of the river of mind: In the age of the proliferation of algorithms, the concentration and logic of the active exercise of the brain is the most rewarding self-investment. The true moat of the future is no longer the amount of free knowledge you remember, but the self-regulating ability to feed, keep reading in depth, endure boredom and frustration, and carry out long chain logic.
The result of the pessimism is only for those who give up their thinking completely. As long as we persist in coming out of the warm algorithm caves, collisions with the real physical world, making them with our own hands and bearing the consequences, humanity will always be the true creator and master of this intelligent age.
Editor: Solomon
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.