1956 年，达特茅斯会议为“人工智能”这一研究领域命名。70 年后，当 AI 已经从实验室走向千行百业，一个新的问题摆在研究者面前：如果机器真的能够学习、推理、写代码，甚至参与科学发现，那么未来十年，人工智能还将走向哪里？
在 ICAAS 2026 首场圆桌对话中，诺贝尔化学奖得主、斯坦福大学教授迈克尔·莱维特，2021 年图灵奖得主杰克·唐加拉，以及数学家马骁、中关村学院何纪言，围绕 AI 的技术路线、产业价值、研究方法和自我进化展开了一场跨越计算机科学、数学与自然科学的讨论。
他们的背景不同，对未来的判断也并不完全一致。但这些讨论指向了同一个问题：当 AI 发展进入新的阶段，什么才是推动智能继续向前的关键？
今天的 AI 产业正在向大型 Transformer 架构和相似的训练范式收敛。这种集中让技术进步的速度前所未有，但也带来了一个隐忧：如果整个行业都押注同一条路线，一旦这条路线触及边界，我们还能从哪里寻找新的可能性？
莱维特给出了一个颇具启发性的比喻：面对未知的未来，就应该像一个谨慎的银行家，不要把所有筹码押在同一个地方。杰克·唐加拉也强调，研究者应当探索当前主流路线之外的机会，并接受失败是研究的一部分。
更值得注意的是，圆桌嘉宾认为，AI 研究的多样性本身也在变化：过去更多是算法和任务的多样性，现在是人类解决复杂问题的过程，未来则可能是让 AI 自主学习的多样化环境。
而这种研究范式的变化，也引出了一个更现实的问题：当 AI 从追求模型能力走向自主解决问题，它究竟是在持续突破智能的边界，还是可能因为无法兑现外界期待而遭遇新的瓶颈？这也让圆桌讨论转向了一个更具争议的话题：AI 会不会再次迎来寒冬？
莱维特认为，AI 寒冬可能与过度承诺有关。杰克·唐加拉则认为，AI 已经走过了此前的寒冬阶段，并对未来保持乐观。中关村学院何纪言更进一步提出，智能具有通用性，未来 AI 有望成为像电力一样的社会基础设施。
70 年前，达特茅斯会议提出了关于机器能否学习、推理和自我改进的雄心。70 年后，这些问题并没有消失，只是变得更加具体，也更加复杂。
未来十年，AI 或许会拥有更强的推理能力，能够参与更多科学发现，甚至开始承担训练和改进自身的工作。但这场圆桌最值得关注的启示是：智能的下一片疆域，不只是更大的模型或更高的分数，而是让 AI 能够在真实世界中持续学习、验证、行动，并创造可被证明的价值。
以下为完整对话内容，经由 InfoQ 编辑：
话题一：如果有做「AI 博物馆」，你会捐赠出什么？
刘铁岩：各位早上好。很高兴举行这场圆桌讨论。这是 ICAAS 2026 的首场圆桌，大家已经看到，本次会议的主题正是致敬人工智能七十年历史、展望未来。我们都知道，七十年前，达特茅斯会议为一项非凡的雄心命名——人工智能（artificial intelligence）；那份提案问道：学习、语言、抽象、甚至自我改进，能否被描述到足以让机器执行的程度。今天，我们以提出一些新问题来致敬那份雄心，比如：人工智能七十年的历史，能否告诉我们一些关于未来十年的事。例如，到 2036 年，人工智能应当能为科学做出什么贡献？又将如何改变我们的日常生活？
考虑到这些，今天我们非常荣幸地请到五位杰出嘉宾。他们是：迈克尔·莱维特教授，诺贝尔化学奖得主、生物分子计算方法的先驱；杰克·唐加拉教授，诺贝尔奖得主，他在数值算法与软件方面的工作帮助一代代科学家使用高性能计算机；马骁教授，数学家，曾与邓煜（Yu Deng）合作，通过严格的数学推导把粒子动力学与流体方程联系起来；最后同样重要的，中关村学院何纪言教授，他的团队开发了 ZGCM-1——一个完全开放、完全由 AI 智能体构建的语言模型。欢迎各位来到这场圆桌，非常感谢。
今天这场圆桌，我想做一个不同于传统圆桌的安排：从一座小小的人工智能博物馆的想象开始，以一个面向未来十年的时间胶囊（time capsule）结束。到那时，我们会给每位嘉宾发一张卡片，请你们写下对未来十年的预测。要知道，对人工智能来说，十年也许足够长，但也足够短——短到足以检验你们的预测是否真的成立。那么我们从第一部分开始：想象建造一座很小的人工智能博物馆。我们将请每位嘉宾向这座虚拟博物馆捐赠一件东西、一件展品。它可以是一件物品、一项技术，甚至一次失败。请想一想：你想捐赠什么，为什么？也许我们可以从迈克尔开始？
迈克尔·莱维特（Michael Levitt）：我意识到，我应该在开始之前就把问题先读一遍的。这个问题很难。
刘铁岩：有什么想法吗？我刚才应该找个提前读过问题的人（来回答），这完全是一张白纸³。谁准备好第一个捐赠了？杰克？
杰克·唐加拉（Jack Dongarra）：那我开始吧。我要捐赠的是——大约四十年前，我刚起步做研究时，参与了一个开发数学软件的项目，这套软件的名字叫 LINPACK。LINPACK 是一组用于求解线性方程组的数学软件，这在很多科学领域是非常常见的问题。我们做出的这个软件包是开源的、可自由获取，用来解决线性代数问题。这个项目对整个社区影响很大，因为它可移植、在当时相当高效，而且开源。它后来被社区采纳为标准软件包。
所以我想捐赠的是 LINPACK 用户手册——从这份手册出发，后来又衍生出不少有意思的东西，其中之一叫 LINPACK 基准测试（benchmark）。我们用它衡量高性能计算系统的性能，如今它被用来给超级计算机排名：从最快的机器一直排到榜单末尾，而榜单截断在 500 台计算机。这样的分层排名真正展示了这一领域走过的路。这份榜单我们已经积累了三十、三十五年，呈现出一些有意思的东西，我明天的报告里会展开讲。所以给博物馆的话，我想捐赠这份榜单。
刘铁岩：非常感谢。杰克，我觉得 LINPACK 确实是给这座博物馆的一件好礼物。迈克尔，你准备好了吗？
迈克尔·莱维特（Michael Levitt）：嗯，我的回答恐怕不太原创。但回过头看，对我而言真正重大的跨越之一，大概是在九十年代中期，开始在一台笔记本电脑上用 Linux。突然之间，就像拥有一台大型主机——虽然相当慢，但它就坐在我的膝盖上。我想，计算的可携带性，是我从那以后一直热爱的事情。
刘铁岩：非常感谢。你多捐了一件——AMiner。作为学者，我们每天都在用它。
马骁：我读博士开始做数学到现在八年了，主要研究方向是偏微分方程。所以我的捐赠可能是：我贡献的一些数学想法，和我读过的几篇论文——它们可以成为当前大语言模型的训练数据。目前我感兴趣的，是人工智能与数学的交叉。我注意到，如今 AI 能生成大量错误的事实，能生成错误的证明，能生成与用户意图不符的代码，所以我正在尝试为 AI 研发下一代验证系统，让这些问题在未来能够被解决。
刘铁岩：非常感谢，马骁。我喜欢这个点子——你不仅捐赠了一件展品，还为 AI 模型捐赠了一份训练数据，很好。何纪言呢？
何纪言：如果只选一样，我会选一张训练损失曲线（loss curve）的快照。我们知道，损失曲线就在持续下降、逐步下降，模型性能越来越好。但这条简单的线背后，是我们付出的许多努力：我们建集群、造芯片、搭模型架构、优化算法、设计评测、采集数据、处理数据。这条线非常优雅、非常简单，但在这条线之上，我们与 AI 对话、与 AI 协作——与 AI 相关的一切性能，都来自这条简单的线。
话题二：方法论收敛之辩
刘铁岩：谢谢何纪言。我想，所有这些都是人工智能漫长历史的一个切片，也是你们各自 AI 工作的一个切片。非常感谢大家的捐赠。回望历史，那只是一个起点。接下来，我们把目光投向未来。我想与各位嘉宾讨论的第一个问题，来自方法论视角。我记得，比如三十年前，AI 领域有非常丰富、多样的不同路径，多样性极高。我记得迈克尔你提到过「进化奖励多样性」。但今天，如果你看 AI 的版图，我们几乎都在收敛到某种非常统一的方法——大型 Transformer，甚至统一的训练工具链。这对社区到底是好是坏？这是我想与每一位讨论的问题。也许我们还是从迈克尔开始。回到你那句话——「进化奖励多样性」——那你怎么看今天的收敛？
迈克尔·莱维特（Michael Levitt）：我觉得这是个非常有趣的视角，也挺令人担忧。多样性确实是个好东西，根本原因在于未来是未知的。面对未知的未来，你得像一个好的银行家：这个买一点，那个也买一点，因为我们永远不知道接下来会发生什么。有意思的问题是：什么样的未来会对大语言模型不利？我不知道。在我看来，我们需要找到某种组合的办法。麻烦在于，如果大家都一样，就没有组合可言；而一旦有了两种不同的东西，就有无穷种组合。也许现在正处在摘低垂果实的阶段，我不知道。我的自然反应其实是想掏出手机问问 ChatGPT：「你怎么看计算机领域的多样性？」我不知道（笑）。我确实认为多样性是个好主意。我想，这对架构设计者和训练方案设计者是一个挑战——去尝试做点不一样的东西。
杰克·唐加拉（Jack Dongarra）：这一点我要同意迈克尔。Transformer 提供了一套非常丰富的工具箱，确实让这个领域进展非常快。但正因为所有人都聚焦于它，我们失去了在这个领域内部保持多样性的能力。我们真正需要理解的是，还有哪些别的机会。我希望我们今后不要只盯着它，我们真的需要分头去探索，去理解还有哪些别的能力。做任何研究，你都应当预期会失败。失败是研究的一部分。我们做了件事、失败了，这不是坏事，那只是过程的一部分。我们需要投身于那些看起来非常有前瞻性的方向，这样我们或许才能拥有当前路径之外的替代方案。
刘铁岩：好，接下来我问我们的数学家。据我了解，数学家通常对「用单一方法解决所有问题」持怀疑态度——不同的问题有不同的解法。你怎么看 AI 的这种方法论收敛？
马骁：其实我有一个不同的看法。数学里——也许计算机科学里也是——有一个非常著名而基础的概念，叫图灵完备性（Turing completeness）。从几十年前我们就知道，一台计算机、一台图灵机，可以模拟任何其他计算机所能做的一切。这叫图灵完备性，或者我们称之为通用性（universality）。这是数学中一个非常根本的概念：我们相信存在一种通用的东西，它可以做一切。
那怎么理解当前大语言模型阶段的收敛？你可以说：好，我们已经有了图灵机，它已经什么都能做了，那为什么图灵机还不是通用人工智能（AGI）？原因在于，一台完备的图灵机仍然需要指令。对一台很「笨」的图灵机，执行一个任务需要很长的指令；而对越来越聪明的智能，我们需要的指令越来越短。而要实现更短的指令，就需要来自更聪明的智能的越来越多的数据。这些年我们收集了大量人类数据，AI 在海量数据上训练，于是变得越来越聪明，执行一切所需的指令越来越短。在数学领域，最早的阶段，它不仅需要题目本身，还需要大量的人类指导和指示，需要真正的研究思路；而现在，它只需要问题本身。这就是我对这个问题的理解。
刘铁岩：谢谢马骁。接下来轮到何纪言。我猜这个问题对你来说可能有点难，因为你还非常年轻，按你的年纪，恐怕并不熟悉 AI 的漫长历史——你开始做 AI 的时候，收敛大概已经发生了。那你怎么看？这是你真正想看到的局面，还是你仍想去探索不一样的东西？
何纪言：其实即便在这几十年里，我们也一直有聚焦的技术与聚焦之外的探索。比如现在 AI 产业所有人都聚焦在 Transformer 上，但我们仍然看到了一些局限——我们发现 Transformer 并未完全达到我们的期望⁵。所以在研究中、在学术界，很多人正在走多样化的研究路线。所以我认为 AI 的发展正是依赖这些多样的路线。我们不知道十年之后，Transformer 是不是实现智能系统的唯一途径——我不认为是。
刘铁岩：谢谢何纪言。
话题三：AI 是否经历过寒冬？
刘铁岩：好。这是一个非常好的「利用与探索」（exploitation and exploration）的例子：把主要资源花在主流上，同时也探索未来。从我自己的视角看，AI 是分层的：在最底层，比如基础模型这一层，确实在向大型 Transformer 收敛；但往上看，比如工具层（harness）和各种智能体（agent）层，我们仍然有很多不同的路径和探索。总之，我认为今天我们处在人工智能的黄金年代：人人都在谈 AI，不只是 AI 科学家，还有每天在用 AI 的自然科学家——这确实是黄金年代。但我们不应忘记 AI 漫长七十年历史里那些寒冬：曾有几十年，AI 科学家没有经费、没有资源，不得不离开心爱的研究方向、转向别的领域。
所以下一个问题关于 AI 寒冬。AI 领域变化太快，我们很难预测明年、未来十年、乃至未来三十年会发生什么。为任何可能性做好准备、确保 AI 领域健康地演进到未来，非常重要。第一个问题还是问迈克尔：我不确定在计算生物学领域，是否也有类似「寒冬」的概念——比如没人相信这个领域能健康发展、你拿不到足够经费继续推进。有过这样的时刻吗？
迈克尔·莱维特（Michael Levitt）：我不觉得有过那样的时刻。我认为需要认识到，计算生物学里的很多计算工作是对实验工作的分析——只要实验学家在持续产出数据，就有数据可分析。我想，AI 的那些寒冬，很可能是过度承诺（over promising）的结果。
至于当下，我确实有担忧。比如，我依赖非常、非常便宜的 token。如果 token 价格突然涨一百倍，我会非常难受。就算是现在，等待时间也是问题——用智能体与用一个（搜索）引擎的很大区别在于，你要花大量时间等待。如果你有五个智能体，可以在它们之间切换，但等待仍然很多。如果等待时间变成两倍，我大概会回头自己动手做很多事情。
我不认为寒冬必然会来。但不幸——或者说幸运——的是，监管和治理正在很早的阶段介入 AI。我想这是因为影响太大，比如数据中心能耗之类。当我开始为自己的 token 能耗感到不安时……能耗是一个方面。我要是真聪明的话，就该在笔记本上跑一个开源模型——或者至少在我笔记本旁边再放一台（跑开源模型的）笔记本，这样依赖就会少一些。但这样的东西（能用的本地开源模型）还不存在。
刘铁岩：对，迈克尔提到，有时候寒冬来自过度承诺。我想在另一些情况下，寒冬也可能来自研究方向的选择。我记得理查德·萨顿（Rich Sutton）曾提过「苦涩的教训」（the Bitter Lesson）。杰克，你怎么看计算机科学史上的寒冬——到底是过度承诺所致，还是方向选错了？
杰克·唐加拉（Jack Dongarra）：我要说，AI 已经经历过它的寒冬了，我们早就走过了那个阶段。当时关于 AI 将实现什么，炒作得太厉害，而它没有兑现那些期望，于是经历了一段经费与科研活动都非常低迷的时期。今天则完全相反——正如你所说，这是 AI 活动的黄金年代。未来当然还会有一些让人担忧的阶段，但我认为那些都会被跨越过去。我看到寒冬正在向其他领域走来。在计算这一侧，量子计算是一项我认为很快——甚至已经——要经历寒冬的活动。所以我想说，就寒冬而言，AI 的寒冬已经过去，没什么可担心的；我只看到未来一片光明。量子计算是另一种情况。
刘铁岩：数学会因为 AI 有寒冬吗
马骁：好，这是个有意思的问题。其实这个问题可以拆成两部分：第一，AI 会不会给数学这门学科带来寒冬？第二，AI 会不会给数学家带来寒冬？
对第一个问题，我不这么认为，因为 AI 生成了那么多新的数学结果。更重要的是，AI 不仅在解决数学问题上很强大，它还擅长把其他学科的问题归约（reduce）成数学问题。所以我的预测是：未来也许十年内，AI 会大幅扩展数学的边界，整个数学学科会因为 AI 的能力而繁荣。
另一方面，对数学家来说，AI 确实带来很多挑战。第一，AI 正在取代数学家的部分能力。第二，AI 生成了大量证明，其中一些可能是错的，所以数学家需要更好的组织方式和更好的工具，去检查 AI 生成的证明是否正确。而且不只是证明——对 AI 生成的计算、方案和思路，我们也都需要更稳健的验证。所以基本上我们要做两件事：作为数学家，要快速适应 AI，确保自己不会被 AI 简单地取代；同时也要发展更好的方法，去验证和消化 AI 的产出。
刘铁岩：非常感谢，马骁。最后一位，何纪言。何纪言是我认识挺久的一位年轻学者，我知道他的兴趣非常广泛，不只是 AI，还有 AI for Science、网络安全等许多领域。所以你可以从不同视角回答这个问题——不只是 AI，还可以更宏观。
何纪言：这个问题很有意思。也许我太乐观了，我无法想象 AI 为什么会有寒冬。我认为 AI 提供的是智能，而智能天然是一种通用的东西：它不只对某个特定行业有用，而是对任何行业都有用。举例来说，就算智能不再继续变强，我们依然可以把 AI 用于很多行业——太空探索、芯片设计、能源与电力设计——我们有太多事情可做。我认为 AI 必然会成为社会的基础设施。也许十年后，或更久之后，我们会像今天对待电力那样对待 AI：它无处不在，以至于我们不再专门提及「AI 的存在」。
不过我还有另一个有意思的观点。我和马骁讨论过很多次：我不预期 AI 的寒冬，但这很危险——AI 会给其他行业带来许多个寒冬。
刘铁岩：好。作为 AI 科学家，首先我们当然不想看到 AI 寒冬。本次会议的主题正是 AI 科学家——我们希望把 AI 专家和其他领域的科学家聚在一起，我们当然不希望看到任何我们关心的领域出现寒冬。但要防止寒冬发生，我认为一方面要对 AI 能做什么、不能做什么抱有正确的预期；同时要及时反思当前的技术路线——比如我们刚才讨论的，单一路线的收敛是否健康；还有，我们要真正确保 AI 能创造真实价值。
还有一个有意思的话题——AI for AI，即如何做自我改进，也就是 RSI（递归自我改进）。我知道何纪言今年夏天恰好在这个方向做了一件事：你让七名博士生从零训练了一个模型，并把模型训练的一切都开源了——源代码、配方（recipe）、数据，还有全部日志。能和我们分享一些这背后的故事吗？
话题四： ZGCM-1：七名学生与上百智能体的一个暑假
何纪言：那是一段有意思的时光。我和学生们吃饭，聊当前的趋势、聊我们读的论文。我们很多次在技术报告和论文里看到「我们清洗了数据」，大家都说数据质量是最重要的事。但这到底意味着什么？怎么清洗数据？数据的哪些部分该去掉？用什么指标衡量质量？我们不知道答案。我们还做了很多推理时（inference time）的研究，构建智能体（agent）、构建工具框架（harness），但我们并不知道智能从何而来。所以我们决定去探索。
我们定了一个目标：从零训练一个中等规模的模型。不只是微调一个基座模型，而是自己设计模型架构、采集数据、处理数据、从头完成训练。这就是起点。这个夏天，我们真的做成了这件事：用有限的资源，训练了一个 70 亿参数的大语言模型，跑通了全流程。三个月里我们拼命干，最后把全过程都发布了——包括 GPU 集群管理、基础设施、训练框架、推理框架，包括数据、模型权重，全部开源给社区，收到了很多好的反馈。
刘铁岩：对，我还知道，你们从零开发那个模型时，不只是用人力，还大量借助了智能体。我听说甚至有上百个智能体协同工作，一起搭建框架、做训练。你有什么经验可以分享——如何在困难任务上与智能体协作？
何纪言：是的，我们是一个很小的团队，很难想象怎么去做这么大的项目。像一些大公司从零训练一个模型，通常需要一百人左右；而我们只有几个人。所以我们非常重地依赖 AI——写代码、管理实验、做分析。在这个过程中，我们确实有一些有意思的发现。比如在某些领域，像实现 GPU 算子，AI 真的非常强，我觉得 AI 做得比我好。但在一些系统设计、架构设计上，AI 就没那么好了。我想我们应该给 AI「喂」非常非常多的数据，帮助 AI 建立起对一个庞大系统的全局观。
刘铁岩：何纪言刚刚讲述的是发生在一个小型研究组里的故事：70 亿参数的模型、几名学生、上百个协同工作的智能体。所以何纪言的工作只是一段漫长旅程的起点。要真正实现扩展定律（scaling law），我们需要关注数据的扩展、算力的扩展，以及经验和交互的扩展。我真心希望——我知道何纪言也有很多合作——未来你们能一起在所有这些维度上探索。
何纪言故事的另一个关键词是「开放」。你不仅开放了权重，而是开放了一切：源代码、数据配方，还有日志。我知道杰克，你捐赠给这座小小 AI 博物馆的展品——LINPACK——某种程度上也是开源的东西，而且那是非常非常早期的开源案例，比许多人谈论开源还要早。所以我想听听你对大语言模型领域开放性的看法：开源、开放权重、开放一切，是否有利于社区健康发展？
杰克·唐加拉（Jack Dongarra）：我认为信息的开放获取（open access）至关重要。它能极大促进理解，它展示了可复现性（reproducibility），它让人们能够真正发现现有材料的弱点，而且它能让社区参与进来、共同改进模型的整体状况。
从软件的角度看，这正是我们想看到的：我们希望社区参与进来，帮助构建好得多的东西。事物的开源属性，我认为确实能造福软件、造福模型，并让它们拥有比不开源长得多的生命周期。
话题五：预测与模拟、可验证奖励
刘铁岩：谢谢。从我自己视角看，开放真的很重要。今天，大型 AI 公司拥有训练模型的全部资源和流水线；但在其他领域，比如生物学、化学，那些科学家也许还没有准备好，也没有全部资源和工具链去训练自己的模型。于是很多有意思的实验无法开展，他们只能用 token（调用闭源模型）去做相对高层次的探索，比如 AlphaFold 那类工作。可是，让生物学家从零训练一个生物学模型呢？我认为我们需要某种端到端的开源软件包来实现这一点。
除了大语言模型的开放性，我还想再问迈克尔一个问题，与 AlphaFold 相关——你在报告里也提过。AlphaFold 是一种预测模型：给定氨基酸序列，它会预测蛋白质的三维折叠结构。而你的工作是基于第一性原理的数值模拟——生物分子系统的量子化学模拟。一边是模拟方法，一边是预测方法。对某些问题，猜测预测就够了；但我不确定你是否会全程信任预测，还是在某些情况下，你仍然想依赖模拟？能和我们分享一下吗？
迈克尔·莱维特（Michael Levitt）：好，我想先澄清一点。AlphaFold 不是拿一条序列去预测一个结构——它是拿一千条不同的序列，预测出一个结构。它利用的事实是：这些序列答案相同、但序列本身彼此不同。所以我认为这是一个不同的问题。它不是物理的——我是说，自然界是用物理来做这件事的，而我们（用 AlphaFold 的方式）做不到。所以问题依然存在。
当涉及蛋白质或生物系统如何变化时，塑造其结构的那些力，同样也是驱动它变化的力。不清楚的是，不理解物理是否也能预测这种变化。我认为（AlphaFold 的成功）是一个相当特殊的情形，其中一个特殊之处在于：自然界使用很小的字母表。如果自然界用的是一百万个字符而不是两千个，问题可能就会难得多。而且大多数蛋白质，一旦进化出某种形状，就会反复重用它，而不是重新进化一个新的——所以不同形状的数量远远少于不同序列的数量。从序列空间到结构空间，存在一个非常收敛的映射。我认为这是个独一无二的问题。
回到你刚才提到的「生物专用模型」：我不知道。我实际上认为，现有的语言模型加上文献，已经非常非常强大了。我认为我们真正需要的，是一种完全可以通过计算机访问的实验室。要知道，大多数科学是实验性的：化学是实验性的，物理学大部分是实验性的，生物学也是。但你可以想象：不再是不同的人在各自的实验室里做实验，而是有一个配备机器人做实验的中央设施，你只需要「拨号」接入你的实验——说请用这个、这个、这个和这个做一次实验。如果有了那样的设施，AI 智能体就可以自己跑实验。我认为那会重要得多。所以，人们声称在开发 specialized（专用）生物学模型，我不知道；我想在某些情况下它们可能有用，比如处理大量序列数据。但通用模型实际上出奇地好，诸如此类。
刘铁岩：非常好的洞见。所以在「用专用模型还是通用模型解决问题」之间，确实存在很深入的讨论。迈克尔还提到了 AI 原生的自主实验室（autonomous lab）——如果有那样的设施，AI 智能体就能自己做实验，这对 AI 和生物学两个领域的发展都会非常有趣。
所以我想请马骁进一步展开这个话题：关于不同维度上的可验证奖励——比如数学和其他领域——我们是否真的需要做一些额外的事，来确保有足够的 VR 环境来训练大语言模型，甚至打造新的工具（instrument）来实现它？请马骁，从数学家的视角。
马骁：这是个好问题。大约两年前——或者至少一年前——模型确实还依赖某种形式化验证方法，因为模型本身不具备任何验证能力，所以不得不依赖像 Lean 这样的验证语言。但最近我们意识到，模型本身已经具备了做逻辑验证的能力。我们意识到，你想想 Lean：在 Lean 里，是机械软件帮你检查——比如你的定理（theorem）有前提条件，你应用这个定理时，软件会帮你检查前提是否存在。而今年，随着 AI 越来越强、越来越具有智能体属性（agentic），人们意识到：AI 可以直接用自然语言做这件事。要检查前提条件，AI 自己就能检查，而不需要调用一个工具去查。所以现在，在数学学科内的验证，我认为我们正越来越多地依赖基于自然语言的验证。
不过，对 AI 而言，验证最重要的前提是数据——因为 AI 的每一种能力都不会凭空而来，必须有数据支撑这种验证。逻辑验证这个问题，我认为已经基本解决了——虽然不彻底，但 AI 已经具备用自然语言验证一切的很强能力。
但验证不仅关乎绝对正确性。当你求解一个问题时，你要在一个非常大的空间里搜索，搜索过程中会产生大量候选。在探索的早期阶段就进行验证，能节省大量资源、提升 AI 的能力。而这种面向搜索早期阶段的大规模验证，目前还没有解决。
话题六：数学与自然科学、给年轻人的四条建议
马骁：我有个问题问迈克尔。我想知道你是否注意到了最近 AI for math 的一个突破：OpenAI 最近发布了大约 700 篇论文。其中有一个事实是，这些论文几乎解决了数学里某个学科的全部开放问题。所以你能否预测一个时间线——类似的事情什么时候会发生在生物学？比如，你是否认为明年、或未来某个时间，生物学的某些子学科会被 AI 彻底解决？
迈克尔·莱维特（Michael Levitt）：这是个很好的问题。我认为数学和自然科学之间有一个重大区别：在自然科学里，你很难列出一份「未来二十年都不会被解决的问题」清单，因为我们的语言（本身）不会告诉我们答案——这要靠实验。你可以有一个非常模糊的问题，比如「地球上的生命是如何起源的？」这是个非常难的问题，因为它甚至可能取决于偶然，它依赖于实验证据，而不是一个理论问题。我认为数学重要的不同在于：一切都是由数学框架定义的；而在物理科学、自然科学里，你没有那样的框架。所以我其实认为，我们（生物学）不存在同样的问题。
我在你说话时想到：说到数学，你们应该让 AI 再写下五百个「未来一百年都不会被解决的问题」。在我看来那有点算作弊——因为想出问题、找到问题，本身就是工作的很大一部分。如果我能把一个生物学问题定义得非常非常好，它会变得非常容易解决。问题的定义往往就是关键。但我认为我们没有同样的处境。
还有一点很重要：像我做的这类计算工作，完全依赖实验。我们做的很多工作，其实就是为实验造工具。数学不一样，但我觉得那样也好。
刘铁岩：谢谢迈克尔。最后，我想请杰克给这两位年轻科学家——正处在职业最早期——一些总体建议。
杰克·唐加拉（Jack Dongarra）：好。我通常告诉我的学生四件事。第一，找到你有热情（passion）的东西并投入其中——真正找到让你感受强烈、除了它别的什么都不想的那件事，然后追随这份热情。
第二，既然我们做的是研究——我前面也说过——你应当预期失败。失败是过程的一部分，这不是坏事，它只是整个学习、理解并在其上构建的过程的一部分。
第三，确保不要挑太容易的问题，要挑有挑战性的问题。去找那些你自己觉得有挑战、别人也觉得有挑战的问题，然后去攻克它们。
最后一条，建立人际网络。与人建立网络，这样你们可以交流想法、分享想法、互相碰撞观点，帮助你解决问题；你也以同样的方式帮助别人。我通常告诉我的学生去做的，就是这四件事。
刘铁岩：好，非常感谢杰克。我想这些建议不仅对马骁和何纪言有价值，对在座的每一位都很有价值，谢谢。
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。
In 1956, the Dartmouth Conference named the research field “Agent intelligence”. 70 years later, when AI had moved from a lab to a business, a new problem. Before researchers: If machines can really learn, reason, write codes and even participate in scientific discoveries, where will artificial intelligence go in the next decade?
In the first round-table dialogue ICAAS 2026, the Nobel Prize laureate in chemistry, Professor Michael Levitte, Professor at Stanford University, Jack Tangara, winner of the Turing Award in 2021, and Mathematician Mathematics Mallory, Middle Guan Village College Ho Ho, initiated a discussion on the technical route of AI, industrial values, research methods and self-evolution across computer science, mathematics and natural sciences.
They have different backgrounds, and the judgement about the future is not entirely consistent. But these discussions point to the same question: When AI enters a new phase of development, what is the key to driving intelligence forward?
Today's AI industry is building on a large Transformer architecture and similar training paradigm. This concentration has brought unprecedented speed to technological progress, but it also raises a concern: If the entire industry is on the same route, where can we look for new possibilities once the route touches the border?
Levitte gave an inspiring metaphor: in the face of an unknown future, it should be like a prudent banker, not to place all the chips in the same place. Jack Tangara also stressed that researchers should explore opportunities beyond current mainstream routes and accept that failure was part of the study.
More notably, according to the round-table guests, the diversity of AI research is changing in itself: the diversity of algorithms and tasks in the past has been greater, and it is now the process of human resolution of complex problems, and the future may be a diverse environment for AI to learn on its own.
This paradigm shift also raises a more realistic question: When AI moves from the pursuit of model capabilities to the autonomous solution of problems, is it experiencing a continuous breach of intelligent boundaries or is it likely to encounter new bottlenecks due to the inability to deliver on external expectations? It also turned the round-table discussion to a more controversial topic: Will AI return to the winter?
Levitte believes that the first winter may be associated with excessive commitment. For his part, Jack Tangara believes that AI has gone past the winter phase before and remains optimistic about the future. The Sino-Guang village school doctrine goes on to suggest that intelligence is universal and that AI is expected to become a social infrastructure like electricity in the future.
Seventy years ago, the Dartmouth Conference set out the ambition of machine learning, reasoning and self-improvement. Seventy years later, these issues have not disappeared, but have become more specific and complex.
Over the next decade, AI may have a stronger reasoning capacity to participate in more scientific discoveries and even begin to take on the task of training and improving itself. But the most interesting revelation of this round table is that the next frontier of intelligence is not just a larger model or higher scores, but that it allows AI to learn, validate, act and create proven values in the real world.
Below is the full dialogue, edited by InfoQ:
Topic one: What would you give if you did "AI Museum"?
Good morning, everyone. It was a pleasure to have this round table. This is the first round table of the ICAAS 2026, and as you have seen, the theme of this meeting is precisely to pay tribute to the 70-year history of artificial intelligence and to look to the future. We all know that 70 years ago, the Dartmouth Conference was named for an extraordinary ambition... – Artificial intelligence; the proposal asks whether learning, language, abstraction, or even self-improvement can be described to a level sufficient to allow the machine to operate. Today, we salute that ambition by raising some new questions, such as the 70 years of artificial intelligence, if we could tell us something about the next decade. For example, by 2036, what should artificial intelligence contribute to science? How will we change our daily lives?
With that in mind, it is a great honour for us to have five distinguished guests today. They are: Professor Michael Levitte, Nobel Prize laureate in chemistry, pioneer in biological molecular computation; Professor Jack Tangara, Nobel laureate in numerical algorithms and software, who has helped generations of scientists to use high-performance computers; Professor Machin, mathematician, who has worked with Yu Deng to link particle dynamics to fluid equations through rigorous mathematical guidance; and, last but not least, Professor Ho's words at Central Guanamura College, whose team developed ZGCM-1 — a fully open and fully constructed linguistic model by AI AI agent. Welcome to this round table, and thank you very much.
Today, at this round table, I would like to make a different arrangement from the traditional round table: starting with the imagination of a small artificial intelligence museum and ending with a Time capsule for the next decade. And then we'll send each guest a card, please write a forecast for the next decade. You know, for artificial intelligence, ten years may be long enough, but it is short enough — short enough to test your predictions for real. So let's start with the first part: Imagine building a small museum of artificial intelligence. We will invite each guest to donate to this virtual museum one thing, one exhibit. It can be an object, a technology, or even a failure. Think about it: What do you want to give? Why? Maybe we can start with Michael?
Michael Levitt: I realize that I should have read the question before I started. The problem is difficult.
Liu Iron Rock: Any ideas? I should have found someone who had read the question earlier, and that's a piece of paper. Who's ready for the first donation? Jack?
Jack Dongarra: Then I'll start. What I'm going to donate -- about 40 years ago, when I started doing research, I was involved in a project to develop mathematical software, called LINPACK. LINPACK is a mathematical software for solver linear equations, which is very common in many scientific fields. We made this package, which is open source, freely accessible, to solve linear algebra problems. The project had a significant impact on the community as a whole, because it was portable, efficient and open-source at the time. It was subsequently adopted by the community as a standard software package.
So what I'd like to donate is the LINAPAK User Manual -- starting with this manual and then coming up with a lot of interesting things, one of which is called the LINAPAK Reference Test (Benchmark). We use it to measure the performance of the high performance computing system, which is now used to rank supercomputers: From the fastest machine to the end of the list, the list was cut at 500 computers. This hierarchy really shows the way in this area. This list, which we have accumulated for 30 to 35 years, presents some interesting things that I will address in my report tomorrow. So I'd like to donate this list to the museum.
Liu Iron Rock: Thank you very much. Jack, I think the Linpack is really a good gift to this museum. Michael, are you ready?
Michael Levitt: Well, I'm afraid my answer is not original. But back then, one of the really big leaps for me, probably in the mid-1990s, started using Linux on a laptop. All of a sudden, it's like having a big mainframe -- it's pretty slow, but it's sitting on my knee. I thought, "Calculate portability is something I've loved ever since."
Liu Iron Rock: Thank you very much. You donated one more-- Aminer. As scholars, we use it every day.
MA: I've been doing mathematics for eight years now, and the main course of study is the differential equation. So my donation could be: some of the mathematical ideas I've contributed, and a few of the papers I've read -- they could be training data for the current big language model. Now I'm interested in the intersection of artificial intelligence and mathematics. I've noticed that today AI can generate a lot of wrong facts, create false proofs, and create codes that don't fit the user's intentions, so I'm trying to develop the next generation of authentication systems for AI so that these problems can be solved in the future.
Liu Iron Rock: Thank you very much, Mark. I like this idea -- you donate not only an exhibit, but also a training data for the AI model. What about Ho?
Ho: If it's just the same, I'll pick a snapshot of the training loss curve. We know that the loss curve is declining steadily and gradually, and models are getting better and better. But behind this simple line are many of our efforts: we build clusters, make chips, model structures, optimize algorithms, design assessments, collect data, process data. This line is very elegant and very simple, but above it, we talk with AI, we work with AI -- everything about AI comes from this simple line.
Topic II: Compromise in methodology
Liu Iron Rock: Thank you for your words. I think all of this is a slice of the long history of artificial intelligence and of your respective AI work. Thank you very much for your contributions. Looking back at history, that's just the beginning. Next, we look to the future. The first issue that I would like to discuss with the guests comes from a methodological perspective. I remember, like, 30 years ago, the AI field had very rich and diverse paths, very diverse. I remember Michael when you mentioned "an evolutionary reward for diversity." But today, if you look at the map of AI, we're almost all condensing to a very uniform approach -- large Transformer, even a unified training tool chain. Is this good or bad for the community? That's what I want to discuss with everyone. Maybe we started with Michael. Back to your words, "Evolution rewards diversity," what do you think of today's contraction?
Michael Levitt: I think it's a very interesting perspective and a very worrying one. Diversity is indeed a good thing, essentially because the future is unknown. In the face of the unknown future, you have to be a good banker: this one buys a little, that one buys a little, because we never know what's going to happen next. The interesting question is: What is the future against big language models? I don't know. In my view, we need to find a combination. The trouble is that, if we are all the same, there is no combination; and once there are two different things, there are no pockets. Maybe it's in the middle of picking down the nuts, I don't know. My natural reaction is actually to pull out my cell phone and ask ChatGPT, "What do you think about diversity in the computer sector?" I don't know. I do think diversity is a good idea. I think it is a challenge for architecture designers and designers of training programmes to try to do something different.
Jack Dongarra: I agree with Michael. Transformer has provided a very rich toolkit that has indeed made progress very fast in this area. But it is because all of us are focused on it that we have lost our ability to maintain diversity within this area. What we really need to understand is other opportunities. I hope that in the future we will not just stare at it, and we really need to split up to understand what else we have. Do any research, you should expect failure. Failure is part of the study. We did something and failed. It's not bad, it's just part of the process. We need to engage in those directions that seem very forward-looking so that we may have alternatives beyond the current path.
Liu Iron Rock: Okay, I'm going to ask our mathematician. As I understand it, mathematicians are often skeptical of "one way to solve all problems" — different solutions to different problems. What do you think of AI's method of retreat?
MA: Actually, I have a different opinion. Mathematics -- and perhaps computer science -- have a very well-known and fundamental concept called Turing Completness. We knew decades ago that a computer, a Turing machine, could simulate what any other computer could do. It's called Turing Completeness, or we call Universality. It's a very fundamental concept in mathematics: we believe that there is something common that can do everything.
So how do we understand the contraction of the current big language model phase? You can say, "Well, we have the Turing machine, and it can do anything. So why is the Turing not General Artificial Intelligence (AGI)?" The reason is that a complete Turing machine still requires instructions. For a very "dumb" Turing machine, it takes a long command to carry out a mission; for increasingly intelligent intelligence, we need shorter instructions. Shorter instructions require increasing data from smarter intelligence. We collected a lot of human data over the years, and AI trained on big data, and it became smarter and shorter to execute all the required instructions. In the field of mathematics, at the earliest stage, it needed not only the subject itself, but also a great deal of human guidance and direction and a real approach to research; now, it needed only the problem itself. That is my understanding of the issue.
Liu Iron Rock: Thank you. Then it's Ho. I guess that's a bit difficult for you, because you're very young, and at your age, I'm afraid you're not familiar with the long history of AI -- when you started doing AI, the contraction probably happened. What do you think? Is this what you really want to see, or do you still want to explore something different?
Ho: Even in these decades, we have had focused technologies and unfocused exploration. Now, for example, everyone in AI is focused on Transformer, but we still see some limitations -- we find Transformer does not fully meet our expectations. So in research, in academia, many people are following diverse lines of research. So I think the development of AI depends on these diverse routes. Ten years later, Transformer wasn't the only way to achieve the smart system -- I don't think so.
Liu Iron Rock: Thank you for your words.
Topic III: Has AI experienced winter?
Liu Iron Rock: Good. This is a very good example of "exploitation and exploration": spending major resources on the mainstream while also exploring the future. From my own point of view, AI is a layer: at the bottom, for example, the base model, it's actually leaning towards the large Transformer; but up, for example, the tool layer and the AI agent layer, we still have many different paths and explorations. Anyway, I think we're in the golden age of artificial intelligence today: everyone is talking about AI, not just AI scientists, but natural scientists who use AI every day -- this is really the golden age. But let's not forget the cold winters in AI's 70-year history: for decades, AI scientists had no money, no resources, and had to move away from their beloved research direction to other fields.
So next question about AI winter. AI is too fast to predict what will happen next year, the next decade, and even the next 30 years. It is important to prepare for any possibility and to ensure that the AI field evolves healthyly into the future. The first question is Michael: I'm not sure if there's a concept like "cold winter" in computing biology. – For example, no one believes that this field can develop in good health and that you will not be able to get enough money to move forward. Have you ever had a moment like this?
Michael Levitt: I don't think there was a moment like that. I think it is necessary to recognize that much of the computational work in computing biology is an analysis of the experimental work — data that can be analysed as long as experimental scientists continue to produce data. I think that the colds of AI are probably the result of over-commitment.
As for the moment, I do have concerns. For example, I depend on very, very cheap token. If the price of the token suddenly rises a hundred times, I'll be miserable. Even now, waiting time is a problem — the big difference between using AI agent and using a (search) engine is that you have to spend a lot of time waiting. If you have five AI agents, you can switch between them, but wait a lot. If the waiting time had doubled, I would probably have returned to do a lot of things myself.
I don't think the winter will come. But unfortunately — or fortunately — regulation and governance are intervening at an early stage in AI. I think it's because it's too much, like data centres that run out of energy. When I started to worry about my token being able to use energy, energy consumption was one aspect. If I'm smart, I should run an open source model on the notebook -- or at least put another one next to my laptop, which would be less dependent. But there are no such things (local open source models that can be used).
Liu Iron Rock: Yeah, Michael mentioned that sometimes the winter comes from excessive commitment. In other cases, I think the winter may also come from the choice of the direction of research. I remember Richard Sutton saying "the Bitter Lesson." Jack, what do you think about the cold winters in computer science history -- whether over-commitment or the wrong direction?
Jack Dongarra: I'd say AI has been through its winter, and we've been through that. At that time, when AI was too much to say about what it was going to achieve, it failed to live up to those expectations and went through a period of very low funding and scientific activity. Today it's the opposite -- as you say, it's the golden age of AI activity. There will certainly be some worrying stages in the future, but I think that's going to cross the past. I see the winter coming to other areas. On this side of the calculation, quantum calculations are an activity that I think is fast — even already — going through the winter. So I'd like to say that, in terms of the cold winter, there's nothing to worry about; I just see a bright future. Quantification is another scenario.
Liu Iron Rock: Is it because of AI?
MA: Okay, this is an interesting question. This question can be broken down into two parts: first, would AI bring a cold winter to the subject of mathematics? Second, will AI bring winter to mathematicians?
I don't think so for the first question, because AI produces so many new mathematical results. More importantly, AI is powerful not only in solving math problems, but also in bringing back problems from other disciplines into math. So my prediction is that in the next decade or so, AI will significantly expand the boundaries of mathematics, and the whole math discipline will flourish because of AI's abilities.
On the other hand, for mathematicians, AI does pose many challenges. First, AI is replacing some of the capabilities of mathematicians. Second, AI generates a lot of proof, some of which may be wrong, so mathematicians need better organization and better tools to check if the AI generation is correct. And it's not just proof -- we need a more robust validation of the calculations, scenarios and ideas generated by AI. So basically, we have two things to do: as mathematicians, adapt quickly to AI to ensure that we are not simply replaced by AI; and develop better ways to validate and digest AI outputs.
Liu Iron Rock: Thank you very much, Mark. Last one, Ho. Ho is a young scholar I've known for a long time, and I know that he has a very wide range of interests, not only AI, but also AI for Science and Cybersecurity. So you can answer the question from different angles -- not just AI, but even bigger.
It's an interesting question. Maybe I'm too optimistic, I can't imagine why AI would have winter. I think AI offers intelligence, and it's natural to be universal: it works not for a particular industry, but for any industry. For example, even if intelligence does not continue to grow stronger, we can still use AI in many industries -- space exploration, chip design, energy and power design -- and we have so much to do. I think AI is bound to be the infrastructure of society. Maybe ten years later, or more, we'll treat AI like today's electricity: it's everywhere, so we don't specifically mention the existence of AI.
But I have another interesting point. I've been talking about it many times: I don't expect an AI winter, but it's dangerous -- AI will bring many winters to other industries.
Liu Iron Rock: Good. As an AI scientist, of course we don't want to see an AI winter. The theme of this meeting is precisely AI scientists -- we want to bring together AI experts and scientists from other fields, and we certainly do not want to see any winters in areas of concern to us. But to prevent the winter from happening, I think, on the one hand, have the right expectations about what AI can and cannot do; on the other hand, we need to reflect in a timely manner on the current technological course — for example, whether the single route is healthy or not, as we have just discussed; and, on the other hand, we need to really ensure that AI can create real value.
There is also an interesting topic - AI for AI, how to make self-improvement, RSI. I know that Ho did one thing this summer in this direction: you got seven Ph.D. students to train a model from zero and to open up everything from model training — source code, formulation, data, and all the logs. Can you share with us some stories behind this?
Topic IV: ZGCM-1: Seven students and a summer vacation of hundreds of AI agent
It was an interesting time. I eat with students, talk about current trends, talk about our papers. We've seen "we cleaned the data" many times in technical reports and papers, and everyone says that data quality is the most important thing. But what does that mean? How do you clean the data? Which parts of the data should be removed? What indicators are used to measure quality? We don't know the answer. We've done a lot of reasoning, building AI agent, building tools, but we don't know where intelligence comes from. So we decided to explore.
We set a goal: to train a medium-sized model from zero. It is not just fine-tuning a base model, but designing a model structure, collecting data, processing data and completing training from scratch. That is the starting point. And this summer, we really did this: with limited resources, we trained a large-language model of 7 billion parameters, running through the whole process. In three months, we worked so hard to get the whole process to publish — including GPU cluster management, infrastructure, training frameworks, reasoning frameworks, including data, model weights, all open sources to the community and received a lot of good feedback.
Liu Iron Rock: Yes, I also know that when you developed the model from zero, you used not only manpower, but also AI agent. I hear there are even hundreds of AI agents working together to build frameworks and train. What experience can you share — how to work with AI agent on difficult missions?
Ho: Yeah, we're a small team, and it's hard to imagine how to do this. Like some big companies that train a model from zero, it usually takes about 100 people; we only have a few. So we rely very heavily on AI -- writing codes, managing experiments, doing analyses. In the process, we do have some interesting findings. In some areas, for example, like achieving GPU algorithms, AI is really very strong, and I think AI is doing better than me. But in some systems and architecture, AI is not that good. I think we should give AI a very, very large amount of data to help AI build an overall view of a huge system.
Liu Iron Rock: What Ha Quill has just told is the story that happened in a small research group: a model of 7 billion parameters, several students, and hundreds of co-workers AI agent. So the work of Ho's words is just the beginning of a long journey. To truly achieve the law of expansion, we need to focus on the expansion of data, the expansion of computing, and the expansion of experience and interaction. I sincerely hope — and I know there is much cooperation — that you can explore all these dimensions together in the future.
The other key word for the story is "open." You opened not only the weight, but everything: the source code, the data formula, and the log. I know that Jack, the exhibit you donated to this little Al Museum -- the LINPACK -- is kind of open-source, and that's a very early open-source case, sooner than many people talk about open-source. So I would like to hear your views on the openness of the large-language model: is open source, open weight, open everything, conducive to healthy community development?
Jack Dongarra: I think open access to information is essential. It greatly facilitates understanding by demonstrating replicability, enabling people to truly discover the weaknesses of existing materials, and by involving communities in improving the overall state of the model.
From a software perspective, that's what we want to see: we want community participation to help build something much better. The open-source properties of things, I think, really benefit software, model it, and give them a life cycle much longer than unopenable sources.
Topic five: Forecasts and simulations, verifiable incentives
Thank you. From my own perspective, openness is really important. Today, large AI companies have all the resources for training models and streaming lines; but in other areas, such as biology, chemistry, those scientists may not be ready, and they do not have all the resources and tool chains to train their models. So there's a lot of interesting experiments that can't be done, and they can only use token for relatively high-level exploration, like the AlphaFold kind of work. But what about the biologists training a biological model from zero? I think we need some sort of end-to-end open source package to do that.
In addition to the openness of the large-language model, I would like to ask Michael another question related to AlphaFold, which you mentioned in your report. AlphaFold is a predictive model: a given amino acid sequence that predicts a three-dimensional folding structure for proteins. And your work is based on numerical simulations of the primary principle -- quantum chemistry simulations of biological molecular systems. On the one hand are simulation methods and on the other hand, prediction methods. For some questions, guess predictions are enough; but I am not sure if you will trust projections throughout, or in some cases you still want to rely on simulations? Can we share them?
Michael Levitt: Okay, I want to make one point of clarification. AlphaFold didn't take a sequence to predict a structure... It takes a thousand different sequences, predicts a structure. It uses the fact that these sequence answers are the same, but the sequence itself is different. So I think it's a different question. It's not physical -- I mean, nature does it with physics, and we can't. So the problem remains.
The forces that shape the structure when it comes to proteins or how the biological system changes are the same forces that drive its change. It is not clear whether physics can also predict such changes. I think it's quite a special case, one of which is that nature uses a very small alphabet. If nature uses a million characters instead of 2,000, the problem may be much harder. And most proteins, once they have evolved a shape, are reused over and over again, rather than re-enacting a new one -- so the number of different shapes is much lower than the number of different sequences. From serial to structural space, there is a very strong map. I think this is a unique issue.
Back to the bio-specific model you just mentioned: I don't know. In fact, I believe that the existing language models and documentation are already very, very powerful. I think what we really need is a laboratory that is completely computer-accessible. You know, most science is experimental: chemistry is experimental, physics is mostly experimental and biology is also. But you can imagine: it's no longer different people doing experiments in their own labs, but there's a central facility with robots doing experiments. You just need to dial in your experiments. - Say this, this, this and this. With that facility, AI AI agent can run the experiment on its own. I think that would be much more important. So, people claim to be developing a dedicated biological model, and I don't know; I think they might be useful in some cases, for example, processing a lot of serial data. But generic models are actually amazing, and so on.
Liu Iron Rock: Very good insight. So there's really a lot of discussion between "specialized or generic." Michael also mentioned AI's autonomous lab -- with that kind of facility, AI AI agent can do its own experiments, which is very interesting for both AI and biology.
So I'd like to ask Marion to go further on this topic: Do we really need to do something extra about verifiable incentives in different dimensions — for example, mathematics and other fields — to ensure that there are enough VR environments to train large language models and even to create new tools to achieve it? Please, mathematician.
MA: That's a good question. Some two years ago — or at least one year ago — the model did rely on some formative certification method, since the model itself did not have any certification capacity and had to rely on a certification language such as Lean. But recently we realized that the model itself already had the capacity to do logical validation. We realized, you know, think about it, Lean: in Lean, it's mechanical software to check you out. For example, when you apply theorem, the software will help you check if it exists. And this year, as AI became stronger and more AI agent, it became clear that AI could do this directly in the language of nature. To check the preconditions, AI can check itself without calling a tool to check. So now, in mathematics, I think we're increasingly relying on certification based on natural languages.
For AI, however, the most important prerequisite for validation is data - because none of the capabilities of AI will come empty, and data must support this validation. Logic confirms the problem, and I think it's basically solved. Not completely, but AI already has a strong ability to authenticate everything in natural languages.
But certification is not only about absolute correctness. When you solve a problem, you search in a very big space, and there's a lot of candidates in the search. Validation at an early stage of exploration can save significant resources and enhance AI capabilities. This large-scale validation, which is oriented towards the early stages of search, has not yet been resolved.
Topic six: Mathematics and Natural Sciences, Four Recommendations for Young People
MA: I have a question for Michael. I wonder if you've noticed a recent breakthrough in AI for Math: Open AI recently has about 700 papers in publish. One of the facts is that these papers solve almost all of the openness of a subject in mathematics. So can you predict a timeline -- when something like this will happen to biology? For example, do you think that next year, or some time in the future, certain sub-sciences in biology will be completely solved by AI?
Michael Levitt: That's a good question. I think there is a big difference between mathematics and natural science: in natural science, it is difficult for you to draw up a list of issues that will not be solved in the next 20 years, because our language (in itself) will not tell us the answer — it will depend on experimentation. You can have a very vague question like, "How does life start on Earth?" This is a very difficult question, because it may even depend on coincidence, relying on experimental evidence, rather than a theoretical issue. I think the difference in math is that everything is defined by a mathematical framework, and you don't have that framework in physics or natural sciences. So I actually think we don't have the same problem.
When I spoke to you, I thought that, speaking of math, you should let AI write another 500 "no problem in the next 100 years." It seems to me to be a bit of a cheat — because it is a large part of the job to think of problems and find them. If I could define a biological problem very well, it would be very easy to solve. The definition of the problem is often the key. But I do not think we are in the same situation.
And there's one thing that's important: the kind of computing that I do, it's entirely dependent on experiments. We do a lot of work, actually making tools for experiments. Math is different, but I think that's better.
Thank you, Michael. Finally, I would like to ask Jack to give some general advice to these two young scientists, who are at the earliest stages of their careers.
Jack Dongarra: Okay. I usually tell my students four things. First, find what you have and put into it — really find that thing that makes you feel strong, that you think nothing but it, and follow that passion.
Secondly, since we are doing research — as I said earlier — you should expect failure. Failure is part of the process, not a bad thing, but part of the whole process of learning, understanding and building on it.
Thirdly, ensure that too easy issues are not singled out, but challenging ones. Go to those problems you feel challenging, others feel challenging, then attack them.
Finally, the creation of an interpersonal network. Networking with people so that you can exchange ideas, share ideas, collide with each other and help you solve problems; you help others in the same way. I usually tell my students to do these four things.
Liu Iron Rock: Okay, thank you very much, Jack. I think that these proposals are valuable not only for the horse and the homage, but also for everyone here, thank you.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.