【Cheeky Pint】企业 AI、工作方式与互联网往事 | 纳德拉与 John Collison 对谈 | 中英全文（上）

书童按：AI 进入企业，难处究竟在模型，还是在数据、权限与记忆？这期 Cheeky Pint 里，John Collison 与微软 CEO 萨提亚·纳德拉从这个问题聊起，随后谈到纳德拉如何通过 Teams 了解公司、为什么持续走近开发者，以及人将怎样指挥成群的智能体。最后，两人回到九十年代：微软看到了互联网，却一度押错了实现路径。上篇保留节目开场的精彩片段剪辑，因此会与后文有少量重复。

节目：Cheeky Pint 第 19 期；主持人 John Collison，嘉宾 Satya Nadella。原节目发布于 2025 年 11 月 18 日，本文中的“今天”“现在”均沿用当时语境。

本篇范围：00:00:00—00:27:03。 原节目与音视频 · 英文转录

本系列据完整英文转录逐段翻译，中文在前、英文在后；保留原稿时间戳，长发言按语义分段。Luna 初译，经 DeepSeek-V4.1-Flash 与 Qwen3.6-35B-A3B 从语言和忠实度两个维度交叉校审，再综合定稿。必要的事实背景与转录疑点另作译注；嘉宾的观点和判断保留原意。

系列目录：上篇：企业 AI、工作方式与互联网往事 · 中篇：算力瓶颈、企业主权与智能体商务 · 下篇：模型忠诚、产品捆绑与组织文化

本篇目录

• 开场剪辑
• 企业 AI：数据、记忆与权限
• 在虚拟走廊里管理微软
• 跟着开发者与初创公司走
• 智能体时代的界面：宏观委派、微观引导
• 互联网往事：看对范式，还要找对路径

开场剪辑

Opening highlights

纳德拉（00:00:00） 比尔一直对此着迷。我清楚记得他在九十年代说过这话。他说：“软件只有一个类别，叫信息管理。你得把人、地点和事物都整理成结构化数据，就这么简单。”问题在于，人很复杂。

Satya Nadella (00:00:00) Bill was always obsessed. I remember him distinctly saying this in the nineties. He said, “There’s only one category in software. It’s called information management. You’ve got to schematize people, places and things and that’s it.” The problem is people are messy.

约翰·科里森（00:00:13） 人们会忠于某个模型，还是会忠于某个 AI 品牌？

John Collison (00:00:13) Do people have loyalty to a model or do they have loyalty to an AI brand?

纳德拉（00:00:16） 你需要一组模型，再由智能体在它们之间协调，让这组模型满足你的需求。

Satya Nadella (00:00:16) You want an ensemble of models. You have agents intermediating that ensemble so that it meets your needs.

约翰·科里森（00:00:23） 大家最终的偏好，不会只是更强的智能吗？比如我会打开模型选择器，手动选 o3 来回答“我该去哪儿吃冰淇淋”这种问题。对一些软件公司来说，挑战 Excel 几乎是一种成人礼。它为什么能这么经久不衰？

John Collison (00:00:23) Will everyone’s preference not just be for more intelligence? I’ll go into the picker and manually select o3 for “where should I go get ice cream” query. It’s like a rite of passage for certain software companies to try to take on Excel. Why is it so durable?

纳德拉（00:00:38） 我们多少有点低估它了。就像我可以让它做到——

Satya Nadella (00:00:38) We sort of don’t give it enough credit. It’s like I can make him do—

约翰·科里森（00:00:41） 世界上最容易上手的编程环境。

John Collison (00:00:41) The world’s most approachable programming environment.

纳德拉（00:00:43） 完全没错。这里说的彼得是谁？

Satya Nadella (00:00:43) A hundred percent. And Pieter here is who?

约翰·科里森（00:00:46） 彼得·莱维尔斯。他算是个独立开发者——

John Collison (00:00:46) Pieter Levels. He’s like an indie—

纳德拉（00:00:47） 哦，对。彼得·莱维尔斯，我认识他。

Satya Nadella (00:00:47) Oh yes. Yeah. Pieter Levels. I know him.

约翰·科里森（00:00:50） 当然，你整天泡在线上嘛。看，萨提亚知道彼得·莱维尔斯是谁。这就是微软能成为一家 14 万亿美元公司的原因。数据中心有什么值得一看的吗？还是说，就只是“哇，好多机柜”？¹

John Collison (00:00:50) Of course, you’re so online. See, Satya knows who Pieter Levels is. This is why Microsoft is like a $14 trillion company. Was there anything good to see at the data center? Or is it like, that’s a lot of racks.

纳德拉（00:01:04） 那是最好玩的地方，哥们儿。

Satya Nadella (00:01:04) It’s the most fun place to go, man.

约翰·科里森（00:01:07） 萨提亚·纳德拉于 2014 年接任微软 CEO，但他在公司已经工作了三十多年，见证过许多变化。萨提亚执掌微软期间，公司规模增长了十倍。微软的成功也被归功于他：先是云业务，如今又是 AI 热潮中的表现。

John Collison (00:01:07) Satya Nadella took over as Microsoft CEO in 2014, but he’s been with the company for more than 30 years. And he’s seen a lot. Microsoft has grown by 10x in the time that Satya has been running it and he’s credited with Microsoft’s success—first in cloud and now in the AI boom.

纳德拉（00:01:22） 干杯，约翰。聊得真开心。

Satya Nadella (00:01:22) Cheers, John. It was great.

企业 AI：数据、记忆与权限

Enterprise AI: data, memory and permissions

约翰·科里森（00:01:25） 那大家在 Ignite 大会上可以期待些什么？

John Collison (00:01:25) So what should people be excited about at Ignite?

纳德拉（00:01:28） 对我们来说，Ignite 大会最重要的事，就是确保 AI 能在企业内部普及开来。说到底，关键是：我们不能只羡慕别人的 AI 工厂或 AI 智能体，而得思考怎么建起自己的 AI 工厂。所以，组织好数据层最终可能是最复杂的一环：它必须贯穿整个企业，才能对接智能。我想，我们会花很多精力做这方面的事。

Satya Nadella (00:01:28) The Ignite Conference for us, more than anything else, is about making sure that AI is getting diffused inside of the enterprise, right? I mean, if there is one thing, it’s more about, “Hey, what does it mean not to just admire somebody else’s AI factory or AI agent, but how to build your own AI factory?” So organizing the data layer turns out to be probably the most complicated thing, which spans the enterprise, such that it can meet the intelligence. And so that’s the stuff that I think we’ll probably do a lot of.

约翰·科里森（00:02:00） 在企业环境里，我们其实还没有真正实现“深度研究”。

John Collison (00:02:00) We still don’t really have “deep research” in a corporate context.

纳德拉（00:02:05） 我们已经实现了，这就是 Copilot 要做的事。

Satya Nadella (00:02:05) We do, that’s what Copilot is about.

约翰·科里森（00:02:05） 可大多数人日常并没有这样的能力。那是不是说，他们没有充分用好现有的 AI？

John Collison (00:02:05) But most people day-to-day do not have this. So are they just underusing AI that exists?

纳德拉（00:02:11） 是的。你提到这一点很有意思，因为在我看来，这才是杀手级功能。我们做的一件大事，是利用了支撑我认为每家公司最重要的数据库的那张关系图谱。它连接着你的邮件、文档、Teams 通话等等。顺带一提，人们的工作并非随意、毫无结构地进行；大家做的事总与某个业务事件有关。这种语义上的联系一直存在于人的脑海里，却会流失。现在，我们第一次能更好地把它找回来。

Satya Nadella (00:02:11) Yes. In fact, it’s interesting you brought that up because to me that is the killer feature. So the biggest thing we did was, we took this graph that is underneath what I think is the most important database in any company, which is underneath your email, your documents, your Teams calls, what have you. It’s the relationships that, by the way, people are not working in an ad hoc fashion in an unstructured way, but they’re all doing it in relation of some business event. That semantic connection is in people’s heads and it’s lost and for the first time there’s much better recall of that.

约翰·科里森（00:02:52） 你觉得这在企业里的普及率为什么这么低？我感觉大家用了不少大语言模型工具，也许会上传单个文档，但我不认为多数公司已经把所有公司背景信息都接入日常使用的 AI，让它能面面俱到。

John Collison (00:02:52) Why do you think this is underpenetrated the enterprise? I feel like people are using lots of LLM tools. They are uploading individual documents, maybe, but I don’t think most companies have the all-singing, all-dancing, all of the company’s context is plugged into their everyday AI.

纳德拉（00:03:10） 是的，我会把原因分成两类。第一，这件事才刚刚起步。我常说，至少和我们历史上做过的所有 Office 套件相比，这次推广速度最快，因为这牵涉到变革管理。归根结底，你得把它引进来，让大家用起来。还有，在企业环境里，所有电子取证功能都必须正常运作，数据治理也必须全部到位。我们还得把 Purview 接入 Copilot，这样每当我试图检索机密内容时，系统都能识别它是机密、应用信息权限管理（IRM）等措施。

Satya Nadella (00:03:10) Yeah, in fact, I would say there are two sets of things. One, it’s starting, right? I always say at least compared to anything we have done, in terms of all the Office suites over our history, this is the fastest in that sense, because it’s change management. At the end of the day, you got to get it in, people have to use it. Oh, by the way, in the enterprise setting, it has got to mean all eDiscovery has to work. All of the data governance has to work. We have had to plumb this purview into Copilot such that any time I’m trying to retrieve something that’s confidential, it’s labeled confidential, it’s IRM’d and so on.

纳德拉（00:03:10） 为此我们做了大量工作，我想现在开始看到成效了。另一件事是，让它跨 Microsoft 365 图谱工作是一回事；接下来还得问，你的 ERP 系统怎么办？连接器算是能用，但又不太够用，因为它们就像两根细吸管。我们需要更完善的数据架构，实质上要把所有这些数据在语义层面整合进同一个数据层。

Satya Nadella (00:03:10) So there’s been a significant amount of work and that I think is where we are starting to see the uplift. The other thing I’d say is, it’s one thing to have it work across the Microsoft 365 graph, but then the next thing is, oh, what about your ERP system? The connectors kind of work, but they don’t really, because they’re two thin straws. You just need a much better data architecture where you have to essentially semantically embed all of these into one layer.

约翰·科里森（00:04:15） 好。几十年来，人们一直有个愿景：公司的数据触手可及。我最喜欢的例子是一本讲甲骨文历史的书《Softwar》。书里提到拉里·埃里森做高管简报（EBC），书里说的好像是九十年代在日本的某一场，大概是九十年代末。他向企业高管推销的愿景是：把公司的所有数据放在一个地方。这个推销说法之所以一直不过时，部分原因就在于公司其实并没有把所有数据都掌握在手边。公司总是不肯把数据基础设施这份“蔬菜”吃下去，而向高管兜售的说法始终是：你只要点一下按钮，就能自己找到问题的答案，不用再把请求发给分析师，让对方替你调查。我们这次终于会把数据管道的基础工作做好吗？你也可以反驳这个前提，但这就是我的问题。²

John Collison (00:04:15) Okay. There’s been a vision for decades of your company’s data at your fingertips. My favorite example of this is I really like the book Softwar on the history of Oracle and it talks about Larry Ellison doing EBCs. I think they’re talking about one in Japan in the 1990s, so it’s the late 1990s, and he is pitching executives on all your company’s data in one place. Part of the reason this is an evergreen pitch is because companies don’t actually have all their data at their fingertips. Companies do not eat their data infrastructure vegetables, and the pitch to executives is always you can go answer your questions yourself at the touch of a button as opposed to sending a request to an analyst who goes and does an investigation for you. Will we finally, this time, eat our data plumbing… You can push back on the premise, but that’s my question.

纳德拉（00:05:02） 不。事实上，如果我没记错，“信息触手可及”这个说法是比尔在九十年代一次 COMDEX 演讲上创造的。

Satya Nadella (00:05:02) No. In fact, I think, if I’m not mistaken, Bill coined this term “information at your fingertips” at a COMDEX speech in the nineties.

约翰·科里森（00:05:11） 我觉得是这样。

John Collison (00:05:11) I think that’s right.

纳德拉（00:05:12） 是啊。很长一段时间里，比尔一直对此念念不忘。我清楚记得他在九十年代说过这话；那时我还是个初级员工，在一次评审会上听到了这段话。他说：“软件只有一个类别，叫信息管理。你得把人、地点和事物都整理成结构化数据，就这么简单。除此之外什么都不用做，因为所有软件……”这就是比尔一直以来的梦想。比如说，他讨厌文件系统，因为文件系统没有结构。如果一切都是 SQL 数据库，他就能直接用 SQL 查询所有信息、针对这些信息编程，那他肯定会很喜欢。

Satya Nadella (00:05:12) Yeah, and for the longest time, Bill was always obsessed about, he felt that… In fact, I remember him distinctly saying this in the nineties, which I picked up in one of the reviews I was in as a junior guy sitting around and he said, “There’s only one category in software. It’s called information management. You got to schematize people, places and things and that’s it. You don’t have to do anything more, because all software…” And that was the dream Bill always had because, for example, he hated file systems because they were unstructured. He would’ve loved it if everything was a SQL database and he could just do SQL queries and program against all information.

纳德拉（00:05:12） 在他看来，那是实现“信息触手可及”的优雅方案。问题在于，人很复杂；即使数据有了结构，它也未必真能放进一个索引里，或是让我用一条 SQL 查询全部查出来。所以我认为，这一直是旧世界的根本难题。我以前想不到——我们谁都没想到——AI 竟会以这样的方式出现：深度神经网络只要扩展到一定规模，就能找出模式，而不必依赖经过结构化设计的数据模型。很长一段时间里，我们总在琢磨：“关系要复杂到什么程度，数据模型才能抓住企业的本质？”结果答案竟是：神经网络里有大量参数，再配上强大的算力。

Satya Nadella (00:05:12) That to him was like an elegant solution to information at your fingertips. The problem is people are messy, and even if data is structured, it sort of is not truly available in one index or one SQL query that I can run against all of that. So that has been the fundamental challenge of the old world, I would say. I would’ve not thought, none of us thought that somehow this AI thing and a deep neural network at some scaling will suddenly become the thing that figures out the patterns, not some schematized data model. In fact, one of the longest time, we used to always obsess about, “Oh, how complex do the relationships have to be or the data model needs to capture the essence of an enterprise?” And it turns out it’s lots of parameters in a neural network with a lot of compute power.

约翰·科里森（00:06:41） 德瓦凯什谈过一个非常聪明的远程员工：他五分钟前才入职。他想说明的是，模型可以聪明得超乎想象，可以做检索增强生成（RAG），也可以访问企业里的所有信息，但这和模型本身真正“知道”某件事并不完全一样。因此，除非你在公司内部训练定制模型，否则模型其实不会越来越擅长你们的工作。你问到第一千次时，它并不会比第一次更聪明。你觉得这会怎么发展？

John Collison (00:06:41) Dwarkesh talks about this really smart remote employee who started five minutes ago, getting at the point that the models can be arbitrarily smart and they can do RAG and they can have access to everything in your enterprise, but it’s not quite the same as the model actually knowing something as a model. And so the models, unless you train custom models inside your company, cannot actually get smarter at what it is that you do. And the thousandth query is not any smarter than the first. Where do you think that goes?

纳德拉（00:07:09） 我觉得这里有两件事。如果我理解得没错，他说的都是上下文学习或持续学习。这可以说是终极目标，也和我刚才说的相呼应：如果把模型的认知核心和它掌握的知识分开，就基本上有了持续学习的公式，或者说算法，然后你只要把它放开去学习。至少在我看来，运行时模型之外还有三件事必须解决。第一是记忆，包括各种形式的记忆：短期记忆、长期记忆。

Satya Nadella (00:07:09) I think there are two things there. I mean, if I understand his thing, it’s all about in-context learning or continual learning. That’s sort of the ultimate thing, and it sort of speaks to the thing I was saying, which is if you kind of have the model’s cognitive core separated from its knowledge, then you have essentially the continual learning formula, so to speak, or the algorithm and then you just unleash it. At least there are three things to me that are outside of the model at runtime that I think you kind of have to crack. One is memory and all forms of memory: short, long.

纳德拉（00:07:09） 比如长期归因就是一个大难题，而人类很擅长这件事。所谓长期归因，直觉上说，就是……有人对我说过：“等到 AI 模型既能给予奖励，也能记住如何针对某个错误施以惩罚——因为它具备长期归因的能力——那时你就知道它真的有记忆了。”总之，第一件事是记忆。第二件事是访问权限：模型在运行时必须真正遵守所有权限系统，因为不同角色能访问的内容各不相同。模型必须符合这些权限要求。第三件事是，所有可执行的操作都必须能正常运作。

Satya Nadella (00:07:09) Even these big challenges of humans are great at long-term credit assignment, which is how does, intuitively… Like somebody said to me, “Hey, the day AI models can both reward and remember how to punish for some mistake because they have the ability to do long-term credit assignment, that’s when you’ll know that they have real memory.” But in any case, memory is one. The second one is entitlements, which is they have to really respect all of the permissioning system at runtime because this is where, because there are roles, what access do I have? And so the model needs to meet that, and then the action space all has to work.

纳德拉（00:07:09） 这三件事合在一起，就是模型所处的环境。假如模型拥有操作能力、访问权限和记忆，那么这些东西按定义都在模型之外，却又必须整合进模型之中。比如说，今天的 Copilot 会使用 OpenAI 的模型，也会用 Claude，对吧？我需要整个系统能同时兼容这两类模型。我认为，前沿发展必须朝这个方向推进。

Satya Nadella (00:07:09) So if you bring those three things, because after all, that’s the environment. So if I have actions, entitlements, and memory with these models, and they by definition have to be outside of the model, but be built into the model. So for example, in Copilot today, you use OpenAI models, you even use Claude, right? I need the system to work across both of those and that I think is where the frontier has to move to.

在虚拟走廊里管理微软

Managing Microsoft through virtual corridors

约翰·科里森（00:08:52） 是，是。我还有一百万个 AI 问题想问，不过我想先问问你的工作方式。你平常一天是怎么过的？尤其是，你怎么通过四处走动来了解情况？你会去哪些虚拟走廊逛逛，了解微软在发生什么？你和客户的实际交流是什么样的？就说普通的一天，不算财报会、董事会之类的场合。

John Collison (00:08:52) Yes, yes. I have a million more AI questions, but I want to ask you some questions about your way of working. So what does your day-to-day look like? And in particular, how are you managing by walking around? What virtual corridors are you wandering to just get a sense for what’s going on at Microsoft? What do your customer engagements actually look like? Just for a normal day, not earnings or not a board meeting or something.

纳德拉（00:09:17） 有意思的是，我平常一天的两头都是客户事务。我几乎每天都要和客户交流。很多交流是远程的。我一天大部分时间都在开 Teams 会议，其中至少两三场是和客户开的。这大概是最能让我贴近实际情况的方式，所以我每天至少会安排一两场。然后就是很多会议。做 CEO 之后，我意识到会议大致有两种。

Satya Nadella (00:09:17) Interestingly enough, my normal day, it’s the two ends of it, which is the customer stuff. So there’s not a day that I would say I’m not having… Many of them are remote. I mean, there’s Teams calls for me most of the day, at least two or three of them with some customer. It’s sort of the most helpful way for me to stay most grounded, I would say. So I have at least one or two of those each day. And then I would say there is a lot of meeting time. As a CEO, one of the things I’ve recognized is there are two types of meetings.

纳德拉（00:09:17） 一种会议里，我只需要召集大家，闭嘴听着，因为把人聚到一起才是重点。别表现过头，就坐在那里，因为该做的工作要么已经完成，要么会在会后完成。这是一种。另一种是重要会议，我需要从中学习、做决定，或者传达一些事情。除此之外，我的活动几乎都在 Teams 频道里，对吧？我会一直留意各个 Teams 频道，那里非常有帮助。事实上，我在那里学到的东西最多，也在那里认识最多人。所以说到处走走、在办公室里串门，我真希望能告诉你那就是我的方式。

Satya Nadella (00:09:17) One meeting is where I’m just supposed to convene and keep my mouth shut because convening was the real thing. Don’t overperform and just sit because all the work would’ve either happened, or will happen after. So that’s kind of one. And then the other meetings, which are the important meetings where I do need to learn or I need to make a decision or communicate something. Then I must say it’s kind of like all over, for me, Teams channels, right? I am lingering around Teams channels and they’re most helpful. In fact, if anything I learn the most there. I meet most people there. So wandering the halls, I wish I could tell you that that is the form.

约翰·科里森（00:10:40） 不过，我觉得 Teams 就是新的办公室走廊；你可以在那些频道里四处看看。

John Collison (00:10:40) No, but I think Teams is the new wandering the halls, looking around those channels.

纳德拉（00:10:45） 完全没错。对我来说，最棒的是能……我在那些频道里建立了最多联系。我会发现：“哦，原来他负责 Excel 智能体。哦，他们在看的就是那个评测。”在那里学到的东西，比我做过的其他任何事都多。

Satya Nadella (00:10:45) A hundred percent. And the most beautiful thing is for me to be able to… That’s where I make the most connections. I get to know, “Wow, he’s the person working on Excel Agent. Oh, that’s the eval that they’re looking at.” I learn so much out of it than anything else I’ve done.

约翰·科里森（00:11:00） 所以微软的团队就是各自埋头做产品，然后萨提亚突然冒出来，问他们一个关于产品的问题？

John Collison (00:11:00) So are teams at Microsoft just working away on their product and then Satya pops up and has a question on their product?

纳德拉（00:11:05） 我倒希望是这样。有时我觉得我们的访问权限管得太严了。我有时候真希望能有更多访问权限。事实上，我最大的抱怨就是没法随时加入所有我想去的地方。不过，能直接进去看看确实很有意思，这也让这种互动变得很自然。而且今天的员工也不怕当着你的面表达意见。

Satya Nadella (00:11:05) I wish. Yeah, sometimes I feel like we are way too permissioned. I wish I had more access sometimes. In fact, my biggest complaint is that I can’t drop in everywhere I want to. But yes, it is fun to be able to just go in there and it sort of normalizes it. And then people are also like—today’s workforce is not shy of sharing their opinion with you.

跟着开发者与初创公司走

Following developers and startups

约翰·科里森（00:11:31） 我注意到了，确实。至少在硅谷的一些小圈子里，你很有名。说到你有条理地了解科技行业的动向，我记得你来过 Stripe 在 Mission 区的办公室，还记得吗？那时我们还是个小公司。大概就在你接任 CEO 之后不久吧。我猜是这样。不过那会儿 Stripe 很小，微软很大。

John Collison (00:11:31) I’ve noticed, yeah. You are famous, at least in small corners of the Valley. Speaking of being methodical for staying very connected to what’s going on in tech here, and I remember you came and visited the Stripe office, remember that one on the Mission? Yeah, when we were a pipsqueak company. It was probably right after you took over as CEO, I’m guessing. But Stripe was very small and Microsoft was very big.

纳德拉（00:11:58） 其实还要更早。我第一次去你们办公室时，应该是我刚开始负责 Azure 的时候。

Satya Nadella (00:11:58) Actually before. I think the first time I came to your offices was when I was running Azure first.

约翰·科里森（00:12:02） 好，没错，那还要更早。所以那时 Stripe 才刚起步。你为什么觉得自己比大多数 CEO 更常这样做？其他 CEO 也应该想去见见所有初创公司啊。

John Collison (00:12:02) Okay, yeah. So it was even before that. So that would’ve been very early in Stripe’s journey. Why do you think you do this much more than most other CEOs? Other CEOs should want to meet all the startups too.

纳德拉（00:12:14） 我一直是在这样的环境里成长的；可以说，在微软的成长经历，让开发者关系、技术布道成了我的一种本能。我大概会这样想：“如果你不跟着开发者走……”有两件事一直根植于我心里。第一，如果你不关注开发者的方向，就很难让技术平台跟上需求；第二，想打造技术平台，你确实得了解新的工作负载。至少这两点我一直记着。

Satya Nadella (00:12:14) I’ve always grown up—in some sense, I grew up even at Microsoft, which had those developer relations, evangelism sort of gene in me. I kind of approach, I think a lot of it as, “Hey, if you don’t follow developers…” There are two sorts of things that are ingrained in me. One is if you don’t follow where developers are going, it’s hard to sort of be relevant in terms of tech platforms and then you really need to understand the new workload in order to build a tech platform. Those are the two things that at least I’ve kept.

纳德拉（00:12:14） 所以，只有关注初创公司，你才能知道平台或工作负载会是什么样。于是我一直把精力放在这上面。另一件事是，这能给我很多能量。我总觉得创业者是些神奇的人，能从无到有创造出东西。这简直像变魔术。所以我总会想：他们到底是怎么做到的？

Satya Nadella (00:12:14) And so therefore, the only way… If you’re not following startups, it’s very hard to know what is either the platform or the workload. So that’s a thing that I’ve indexed towards. The other thing is I derive so much energy out of it. I mean I’ve always thought founders are just magical people who create something from nothing. I mean, it just sort of feels like a magic trick. So I’m always like, how the heck does one do that?

约翰·科里森（00:13:08） 你提到关注初创公司的动向，这很有意思。我们一直认为，Stripe 的产品要为初创公司打造。这很重要，因为今天的小型初创公司就是明天的上市公司；在 Stripe 上，这样的例子我们见过很多次。但我们直觉上觉得，初创公司感兴趣的东西，往往能带来更好的产品体验；当时我们还无法证明这一点。如果初创公司想要稳定币、按用量计费之类的东西，我们就应该满足这些需求。这么做不只是为了把初创公司业务做好，企业客户最终也会跟进。我觉得我们花了很多年才验证这个模式，但现在我们确实看到了——

John Collison (00:13:08) Yeah, it is funny you say that about following what the startups are doing. We always conceived of what Stripe was building as it was important to build for startups, both because today’s small startups are tomorrow’s public companies and we’ve seen that again and again on Stripe. But we just felt, at an intuitive level and we felt this before we could prove it, that what the startups were interested in were often better product experiences. And so if the startups want stablecoins or usage-based billing or what have you, we should build for those needs, not just because we’ll have a good startup business, but the enterprises will come around. And it took us, I would say many years to prove out that model, but now we’re really seeing—

纳德拉（00:13:50） 是的。事实上，我觉得你们在这方面算是标杆。我从你们身上学到的一点，是重新发现微软曾经很擅长的事：跟着开发者走，出现在初创公司身边。这也促使我把目光投向 GitHub、Nat，以及相关的其他事情。GitHub 这个资产显然很有价值。我们需要成为开源生态的好管家，但它也是所有初创公司都会用的地方——每家都有的一样东西，就是放在 GitHub 上的代码仓库。³

Satya Nadella (00:13:50) Yeah, in fact, I think you guys are a bit of a gold standard on that. In fact, one of the things that I learned from you guys was rediscovering at some level what Microsoft was very good at, which is following the developer, being where the startups are. And so that’s what sort of led me even to GitHub and NAT and all of the rest, which is to some degree the GitHub asset. Obviously it was a great asset. We needed to be good stewards of an open source ecosystem, but it’s also the place where every startup—the one thing that everybody does have is their repos in GitHub.

纳德拉（00:13:50） 我觉得进入这个圈子对我们很重要，原因不只是“占据一个战略位置很有利”。你们刚才说的很好：我们要去学习，做出更好的产品。有时你会失去对产品体验的审美判断，不知道怎么做才能让使用过程毫无阻碍；而用户的耐心最少，所以从开始使用到获得价值所需的时间必须最大化。⁴

Satya Nadella (00:13:50) And I felt like being in that loop was important for us, not just, “Oh, it’s strategically great to have some position there.” To learn simply and to build better product, I think, is sort of well said. Because you sometimes lose the aesthetic of what is required, what’s that friction-free way to deliver because the least amount of patience is there and the time to value, for example, has to be maximized.

智能体时代的界面：宏观委派、微观引导

Agent interfaces: macro delegation, micro steering

约翰·科里森（00:14:52） 微软有没有在考虑按个人需求生成界面？你想想，软件还困在旧范式里：我们写好一大堆软件，做成最终版，再装在磁盘上发出去。现在软件还是同样那一套，只不过通过云端交付。但实际上，我们也许能实时渲染出你想要的那种界面。你们在朝这个方向发展吗？

John Collison (00:14:52) Is Microsoft thinking about generated UIs that are personalized to… When you think about it, software is stuck in the old paradigm of we write a bunch of software and it goes to Gold master and it goes out on disks, and now that same kind of software it’s delivered in the cloud, but the UI you want is probably, we can render that exact UI in real time. Is that a direction you guys are going?

纳德拉（00:15:14） 肯定在朝这个方向走。某种程度上，现在发生的事是，我们一方面有生成能力。如果你认为所有代码都能生成，那围绕任何东西生成更定制化的用户体验框架也就顺理成章了。尤其是……事实上，很长一段时间以来，微软一直在问：文档、网站和应用程序之间到底有什么区别？所以在某种程度上，没错，你可以随时按自己想要的呈现形式生成其中任何一种。不过有意思的是，尽管大家都在说“这些应用都要消失了”，拿大家熟悉的集成开发环境（IDE）来说吧。⁵

Satya Nadella (00:15:14) I think for sure. At some level what’s happening is on one side our ability to generate. I mean if you sort of say you can generate all code so therefore you can generate some UX scaffolding around anything that’s a lot more custom. So especially… In fact for the longest time, one of the things at Microsoft was what’s the difference between a document, a website and an application, really? And so to some degree, yeah, exactly. So you can generate any one of those at any time depending on what format you want to present it. But at the same, interestingly enough for all the talk of “Hey, all these apps go,” take even our good old IDs.

纳德拉（00:15:14） 从某种意义上说，IDE 正在回归，不管是 Excel 还是 VS Code 都是如此。现实是，AI 会生成输出，而我得弄明白这些输出。事实上，我需要一个出色的编辑器，让我能和 AI 一起查看差异、反复迭代。因此 IDE……最让人兴奋的事情之一，就是出现一批全新的、打磨得很好的 IDE：它们既能和智能层形成遥测反馈回路，又更像抬头显示器。我有成千上万个智能体同时在工作，怎么理解并微调它们的行动？IDE、收件箱和消息工具就会承担这个作用。我以后处理消息或分流任务的方式，不会再和今天一样，而会变得不同。

Satya Nadella (00:15:14) In some sense IDs are back, whether it’s Excel or VS Code because the reality is AI generates output. I need to make sense of that output. In fact, I need a fantastic editor that lets me do diffs and iterations on it with AI. So the ID… one of the most exciting things is new classes of highly refined IDs that have even sort of a telemetry loop with the intelligence layer, but also they kind of act more like heads-up displays. I have thousands of agents going off. How am I going to make sense of the micro steering of thousands of agents? And that is what ID slash inboxes and messaging tools will be, which is I’m not messaging or dealing with triage the way I deal with it today, but it’s going to be different.

约翰·科里森（00:16:53） 好，有意思。所以现在程序员把所有时间都花在 IDE 里，而他们是少数这样工作的职业之一。你的愿景是会出现会计师用的 IDE、律师用的 IDE，还有——

John Collison (00:16:53) Okay, interesting. So you think right now programmers spend all their time in IDE, but they’re one of the few professions that does that. And your vision is the accountant IDE, the lawyer IDE, and—

纳德拉（00:17:04） 我会用什么方式和智能体协作？这有点像大规模的宏观委派：我把一堆指令交给许多智能体，它们就去工作，有时持续几小时、几天——随着模型变得更好，大概会是这样。但它们也会不断向我汇报。因此，这是宏观委派、微观引导。要做到这一点，怎样才能带着上下文进行微观引导？它不能变成下一代通知地狱：系统给我发来一条通知，只有五个词，我却完全不知道真实上下文是什么，诸如此类。

Satya Nadella (00:17:04) What is the metaphor of how I will work with agents? So it’s kind of like massive macro delegation. So there’s lots of agents I go give a bunch of instructions to and they go off and work sometimes for hours, days, let’s say, as the models get better. But they are checking in and so it’s macro delegation, micro steering. So if you take that, how does one do micro steering with context? It can’t be in the next notification hell, which is it sort of notifies me. It has five words. I don’t know exactly what the real context is, or what have you.

纳德拉（00:17:04） 我觉得这正是方向所在，而且它得像跨应用的体验。所以我觉得，所有软件最终长大之后，都会像一个收件箱、一种消息工具，再加上一块带着闪烁光标的画布；只不过这一次，很多工作已经完成了。

Satya Nadella (00:17:04) That I think is where, and that has to be multi app-like. So that’s where I feel like all software finally when it grows up, it looks like an inbox and a messaging tool and a canvas with a blinking screen, except this time around a lot of work happened.

约翰·科里森（00:17:55） 这会是一个应用，还是十个不同的应用？想想当年生产力软件的格局：有三大应用——Word、Excel 和 PowerPoint。有意思的是，数量不是一个，也不是四十个，而是三个。你怎么看？

John Collison (00:17:55) Is that one app? Is that 10 different apps? It’s kind of interesting if you think about the productivity speech that emerged, there were three big apps—Word, Excel and PowerPoint—but it’s interesting that that number was not one and was not 40. It was three. And so how do you think about this?

纳德拉（00:18:16） 我觉得这个判断是对的。在我看来，最后会有少数几个。我这个人有点喜欢把复杂的事情归结到最基本的形式，会想：“天哪，最后还是那些东西，只不过它们承担的工作会变。”拿表格来说，至少对人是这样；当然，也可以讨论智能体之间会用什么工具互相沟通，那是另一回事。现在为了强化学习循环，它们在模拟我们的生产环境，但最终它们会更有效率地创建自己的生产环境，让自己进行强化学习。

Satya Nadella (00:18:16) I think that that’s right. To me it will be a few I think, and in fact the reductionist person in me says, “Man, they’ll be the same things except the job they do is going to be different.” Because I think a table, at least at the human level, because we can all talk about what tools will agents use to communicate with each other? That’s a different thing. Right now for the RL loop, they are simulating our production environment, but they will ultimately be more efficient in creating their own production environments to kind of RL themselves.

纳德拉（00:18:16） 不过先不谈这一点。为了和人沟通，我觉得我们已经发现了一些自己喜欢的形式：电子表格、表格、线性结构的文档，还有收件箱或消息工具。所以这些都是合理的用户界面。只是你刚才问的那个问题在于：当它出现在 IDE 里，带着一组改动时，不能只对我说：“好了，这是一个文件，你去打开它。”你得给我一个能指引我的方案；这方案不只是为了执行任务，也得帮助我完成整个工作流程。

Satya Nadella (00:18:16) But let’s just leave that aside. But in order to communicate with us, I feel like we have discovered some good things that we like. We like spreadsheets and tables and we like documents in linear form. We like inboxes or messaging tools. So these are reasonable UIs, except the question I think you asked is how does this thing have, when it shows up in an IDE with a set of changes, you have to help me more than just say, “Okay, now here is a file, go to that file.” That directed plan, not just to execute, but for me to do my workflow.

纳德拉（00:18:16） 我们正在尝试的一件事是 GitHub Copilot 的 Mission Control。设想是，你启动五六个不同的分支，在每个分支上派出自主智能体。它们各自完成工作、回来之后，你如何对这些拉取请求做分流处理，我认为下一代 IDE 就会从这里诞生。

Satya Nadella (00:18:16) One of the things that we are experimenting with is mission control and GitHub Copilot is the idea is you go have five, six different branches in which you fire off all these autonomous agents, they all do their work, they come back, and then your ability to do PR triage is where I think the next IDE is born.

互联网往事：看对范式，还要找对路径

The early internet: paradigms, products and business models

约翰·科里森（00:19:50） 我一直很惊讶，在科技领域，技术愿景出现得往往远早于技术真正成熟的时候。比如电影《2001 太空漫游》在六十年代就描绘了能听懂语音、还能调用工具的 AI。这个愿景过了五十年才成为现实。再比如，早在八十年代，人们就对能和电脑说话、能做语音转文字和文字转语音感到兴奋；直到现在……我不知道你用不用 SuperWhisper 之类的工具，但现在它终于真的好用了，三年前还不行。比那个愿景晚了四十年。

John Collison (00:19:50) I’m struck by, in technology, how frequently you see the pattern of excitement for and a vision around a technology being so much earlier than the technology actually being ready. Like the movie 2001: A Space Odyssey, which is in the sixties, that was a voice activated AI with tool use capabilities. And it just took 50 years and then people were excited about the idea that you could speak to your computer and text to speech, speech to text. People were excited about that in the eighties and only now… I don’t know if you use SuperWhisper or anything like that, but it’s really, it’s finally really good, but it wasn’t good three years ago. 40 years after the vision.

纳德拉（00:20:33） 你这么一说真是不可思议。我以前住在微软老园区旁边的一套公寓里，当时我在做互动电视。那是 1994 年。

Satya Nadella (00:20:33) Yeah, it’s crazy that you bring that up. In fact, I used to have an apartment right next to the Microsoft campus, that old campus, and I was working on interactive television. This was in ‘94.

约翰·科里森（00:20:44） 信息高速公路。

John Collison (00:20:44) The information superhighway.

纳德拉（00:20:45） 没错。当时有好几件事都很惊人。沿着我的汇报关系往上，依次是里克·拉希德、克雷格·芒迪、内森·迈尔沃尔德，再上面是比尔·盖茨。我当时想：“天哪，这群人的智商加起来可真高。”当然，我们全都错过了互联网，那才是后来真正发生的事。不过我那时把互动电视通过异步传输模式（ATM）交换网络接到了家里，也就是我住的公寓。我记得做过这么一次演示。年轻时我在微软做过一件风险很高的事：演示我们第一个有冗余能力的文件系统，它其实是一个视频服务器。

Satya Nadella (00:20:45) That’s right. In fact, there were multiple things that were stunning. My management chain was Rick Rashid who reported to Craig Mundie, who reported to Nathan Myhrvold, and there was Bill Gates and I was saying, “Man, that’s a lot of IQ.” And of course we all miss the internet. That was the only thing that happened. But I had interactive television, switch ATM, to my home, to my apartment. So I remember doing this demo. One of the high stakes things I did as a young guy at Microsoft was a demo of our first redundant file system, which was a video server.

纳德拉（00:20:45） 约翰·马龙来现场了。比尔当时大概是这样介绍的：“看，这就是互动电视的未来。它还有个好处：即使磁盘出了故障，视频流也能继续播放。”我的任务就是把磁盘拔掉，同时让视频流继续。我们实质上搭建了一个分布式文件系统和流媒体服务器，又通过 ATM 交换网络连到家里。我有五部电影可以看，而且每部都看了好几遍。

Satya Nadella (00:20:45) John Malone was the one who came and Bill was sort of saying, “Hey, here’s the future of interactive television, and guess what? It’s even great because the disc can go haywired and still stream.” And so my job was to remove the disc drive and have the stream continue. But we built essentially a distributed file system and a streaming server and had an ATM switch network to the house, and I had five movies I could watch, and I watched them all multiple times.

约翰·科里森（00:21:49） 我想问问这件事，因为我琢磨了很久，而你是最适合回答的人。微软在九十年代就看到了互联网即将到来的未来，尤其是比尔·盖茨那份著名的《互联网浪潮》备忘录，里面说：“互联网是微软必须专注的一件大事。”并不是说我们没想到互联网，也不是说它在十五个优先事项里排第七；意思是：“听好了，微软唯一该考虑的就是互联网。”但当时大家对互联网的愿景是信息高速公路。它和我们后来所说的互联网有一点微妙的差别，因为当时的想法——而且这个想法非常合理——是：没有人能把互联网接到家里的电脑上。

John Collison (00:21:49) So I want to ask you about this, because I’ve thought a lot about this and you’re the perfect person to ask. Which is, Microsoft saw the internet future that was coming in the nineties, and in particular the famous Bill Gates internet tidal wave memo said, “The internet is the one big thing Microsoft needs to focus on.” It wasn’t like we’re not thinking about the internet. It wasn’t that it was priority number seven of 15. It was like, “Hey guys, listen up. The only thing Microsoft should be thinking about is the internet.” But the vision for the internet at the time was this information superhighway which was subtly different from the internet because the thinking was—and it was very sensible thinking—no one has internet to the computer in their home.

约翰·科里森（00:21:49） 很多人家里甚至没有电脑。大家家里有的是电视，而他们也有有线电视，那能提供高带宽连接。所以我们要在电视上装机顶盒，人们就会通过这种方式使用互联网。大家非常重视这股即将到来的浪潮，解决方案也很合理、考虑得很周全，可它偏偏不是正确的做法。说到这里，显然就会想到规模庞大的 AI……从这件事里，我们应该得到什么启示？

John Collison (00:21:49) A lot of people don’t have a computer in their home. So what people do have is a TV and what they have is cable, which is a high-bandwidth connection. And so we’re going to do these set top boxes on the TV and that is how people will use the internet. Paying a huge amount of attention to this coming wave, pretty sensible, well thought-out solution, and yet not the right approach. So obviously bring that up in the context of the giant AI… What should one take away from that?

纳德拉（00:22:57） 这是个很好的问题。说说我的理解也挺有意思。我没有花太多时间和比尔聊那个时期的事，但即便当时我只是个初级员工，我觉得至少……我对那段历史的理解是，我们好像理解了互联网，又好像没有，因为我们想要交付的是……我觉得我们当时并不相信 TCP/IP 能行。某种程度上，回头看我们当时想做的信息高速公路，核心问题是：“服务质量很重要。”

Satya Nadella (00:22:57) It’s a great one. See, if I look at even my interpretation, it’ll be actually interesting. I’ve not spent as much time talking to Bill about that era, but I felt there were at least, as someone as a sort of an entry level employee at that time, even. My reading of history was that we kind of got the internet, but we didn’t because we wanted to deliver… I don’t think we believed that TCP/IP would work. I mean at some level the information highway, when I look at what we were trying to do was, man, this quality of service is a thing.

纳德拉（00:22:57） “TCP/IP 根本行不通。”所以我们才会和拨号上网的 AOL 竞争。你还记得 MSN 最初是 X.25 网络吧？但后来比尔转了方向。我想是在 1995 年——说起来很好笑，正好在 Windows 95 发布时，他说：“你们知道吗？一切都要变了。”所以我觉得，从 1992 年开始——那时我们大概都看到了第一次演示，对吧？——再到 1993 年 11 月 Mosaic 出现……⁶

Satya Nadella (00:22:57) This TCP/IP is just not going to work. And so therefore we were competing against AOL on dial-up. And even that sort of, you remember MSN was an X.25 network, the first version of it. But that’s when Bill pivoted. So the thing that Bill did was in ‘95, I guess—in fact, it’s funny that right as Windows 95 was launching he says, “You know what? It’s all going to change.” So I feel between ‘92, which is when I think all of us maybe got our first demo, right? November ‘93 is when Mosaic—

约翰·科里森（00:24:10） 对，没错。

John Collison (00:24:10) Yeah, that’s right.

纳德拉（00:24:11） 我记得大概是这样。那时我们都在这件事周围试探。从 1993 年到 1995 年，有两年时间，大家还不清楚互联网会不会采用这种协议和完整技术栈。后来这套技术栈逐渐成形，到 1995 年一切就清楚了，于是我们转了方向。

Satya Nadella (00:24:11) I think something like that. And so we all were kind of dancing around it. So from ‘93 to ‘95, there was that two-year period where it was unclear whether this was going to be the protocol and the full stack. And the stack emerged, and by ‘95 it was clear and then we pivoted.

约翰·科里森（00:24:31） 有意思。所以就在那个时候，其实还看不出开放互联网会胜出。

John Collison (00:24:31) Interesting. So just at that time, it wasn’t actually clear that the open internet would win.

纳德拉（00:24:36） 对。还有一个启示。我一直在琢磨这件事，因为我觉得可以把它延伸到 AI 领域。第一，要找对范式。但即使范式找对了，杀手级应用是什么、商业模式又是什么，也未必清楚。互联网就是如此。谁会想到，开放网络的组织层竟会是一个靠网络效应取胜的搜索引擎？我常说，根本没有所谓的开放网络，只有谷歌网络，因为谷歌主导了它。

Satya Nadella (00:24:36) Yes. And in fact, there’s one more lesson. The interesting thing that I’ve always watched because I think we can parlay this into AI. One is to get the paradigm right. Then it’s not clear. Even if you get the paradigm right, that you may not get what is the killer app or even the business model. That’s always been the case, with the internet, who would’ve thought that for the open web, an organizing layer would be one network effect search engine, right? Because the organizing layer of the web, I always say there’s no such thing as the open web. There’s the Google web, and just because they dominated it.

约翰·科里森（00:25:19） 我们是不是应该想想，当时或许存在着一种一厢情愿的判断：我们自有的专有方案——自由媒体与微软的合资项目——会胜出，而开放网络最终赢了？如果组织要在两种可能性之间作选择，比如我们的专有信息高速公路和开放网络，或许该提醒它们：它们可能会偏向自家的专有方案。

John Collison (00:25:19) Should one reflect on the fact that maybe there was some motivated thinking around our proprietary solution, the Liberty Media-Microsoft joint venture will win. Whereas the open web is what won. And you should maybe caution organizations where, if they’re following two possibilities, our information superhighway proprietary system or the open web, companies will somehow have happy thinking towards the proprietary solution.

纳德拉（00:25:54） 这是个有意思的问题。回头看，我觉得情况很有意思：可以说，AOL 和 MSN 输给了开放网络。但随后，新的 AOL 和 MSN 形式又取代了它们：它们叫搜索引擎，叫应用商店。移动网络的情况其实也很有意思。

Satya Nadella (00:25:54) It’s an interesting one. I think the way, when I look back again, it’s interesting, right? So AOL and MSN kind of lost out, let’s call it, to the open web. Except they were replaced by new forms of AOL and MSN. They’re called search engines. They’re called app stores. The mobile web, in fact, is fascinating.

约翰·科里森（00:26:16） 开放网络只是历史上的一个阶段。

John Collison (00:26:16) The open web was a moment in history.

纳德拉（00:26:17） 历史上的一个阶段。对我来说，其中更深层的启示也许是：即使在开放生态里，组织层也总会出现。许多品类的主导权会转移到那个组织层。只是它究竟会是什么，一直说不准：上一个范式里是搜索引擎，这一次是聊天机器人。这会持续多久？没人知道，但至少今天它很重要。ChatGPT 作为聚合入口的成功毋庸置疑。市场平台和应用商店曾经扮演过这种角色。接下来会是什么？智能体市场或智能体商务会如何改变电子商务？我觉得这些都是值得认真争论和厘清的问题。

Satya Nadella (00:26:17) A moment in history. And so the thing that maybe—the meta thing for me is organizing layers will always emerge even in an open ecosystem. And a lot of the category power moves to that organizing layer, and it’s always unclear, like the last paradigm of this… Last time at a search engine. Today it’s chatbots. How long lasting is that? No one knows, but it’s definitely today. I mean, ChatGPT’s success cannot be denied in terms of what it means as an aggregation point. Marketplaces slash app stores have been a thing. What comes next? What happens to e-commerce in an agentic marketplace or in agentic commerce? I think these are the interesting things that need to be litigated.

继续阅读：中篇：算力瓶颈、企业主权与智能体商务 · 下篇：模型忠诚、产品捆绑与组织文化

译注