Satya Nadella on AI: @NoPriorsPodcast x Latent Space Crossover Special at Microsoft Build 2026

Please welcome Swyx, Sara Goa, A Lad Gill, and Chairman and Chief Executive Officer of Microsoft, Satya Nadella. 欢迎 swyx、Sarah Guo、Elad Gil，以及微软董事长兼首席执行官萨提亚·纳德拉。 Hello. 大家好。 Satya? 萨提亚？ Uh I'm so excited to be here. 我非常高兴能来到这里。 Welcome to a crossover episode of No Priors and Lain Space with Satya Nadella. 欢迎收看 No Priors 与 Latent Space 联合特辑，今天我们请到了萨提亚·纳德拉。 Um congratulations on an amazing build. 恭喜你，这届 Build 真的太精彩了。 No, thank you so much and it's great to be with both of you. 不不，非常感谢，能和你们两位在一起太棒了。 I listen to both of you or both the podcasts all the time. 我一直在听你们两个人的节目，两档播客都在听。 It's great to be on it. 能参加这档节目太好了。 Thank you so much. 非常感谢。 Sara has been talking about um these amazing uh announcements from across the Microsoft estate all morning for I think 3 hours. Sarah 今天早上已经就微软旗下各产品线的一系列精彩公告讲了足足 3 小时了。 What is the uh what's the most important reflection or takeaway you have? 你觉得最重要的感悟或者收获是什么？ I I'd say there are uh perhaps the the biggest one for me is let's sort of conceptualize this more as an ecosystem play as opposed to a single model or even a single platform, right? 我会说，对我而言最重要的一点是：我们应该把这次 AI 浪潮更多地理解为一个生态系统机会，而不是某个单一模型甚至单一平台的胜利，是吗？ I mean yeah, whatever I at least for me having grown up at Microsoft, having seen whatever four major platform shifts, uh I sort of fall into that uh uh camp where a platform is defined by fundamentally its ability to create more value above the platform versus what's captured in the platform. 我是说，至少对我来说，在微软一路成长过来，亲历了大概四次重大平台转型，我倾向于认为，衡量一个平台的根本标准，是它在平台之上能创造的价值，是否远超平台本身所捕获的价值。 And so if you you view what's happening right now, I think this morning's keynote was how can any company whether it's an AI-native company a traditional enterprise company, participate as a first-class participant where they can point to AI they created. 所以，如果你来看现在正在发生的事情，我觉得今早主题演讲的核心问题是：任何一家公司，无论是 AI 原生公司还是传统企业，怎样才能作为一等参与者进场，拥有自己训练出来的 AI？ Right? 是吗？ It's not that they don't use other people's AI. 不是说他们不用别人的 AI。 Of course, they will. 当然会用。 But to me, what's the path? 但对我来说，路径是什么？ What's the recipe? 配方是什么？ How do I do it? 怎么做到？ What does the stack look like? 技术栈长什么样？ What does the tooling look like? 工具链长什么样？ What is valuable? 什么是有价值的？ How do you do that? 怎么实现？ That's it. 就这。 That's sort of our job to do. 这就是我们的任务。 Yeah. 对。 Ecosystem strategy is uh very complicated, right? 生态系统战略非常复杂，是吗？ Because you end up building certain components, partnering for certain components, supporting them. 因为你需要自己构建某些组件，与合作方联合开发另一些，同时还要支持整个生态。 You just announced this big suite of models. 你刚宣布了一整套新的模型。 Like, tell us a little bit about the uh training strategy for Microsoft. 能跟我们说说微软的训练策略吗？ Yeah. 对。 So So, the thing that we wanted to do with the MAI models was to build, and as Mustafa talked about, first of all, a great lineage, right? 我们做 MAI 模型想要实现的，正如穆斯塔法讲的，首先是建立良好的数据血缘，是吗？ Starting with pre-training uh with very good data quality, uh doing all the ablations, making sure because in in some sense, it's become even harder to build a clean lineage model. 从 pre-training 开始就要确保数据质量，做好所有 ablation，因为在某种意义上，构建一套干净的血缘模型比以前更难了。 Yes, because there's so much stuff out there uh that you truly need to ablate out to be able to have a fantastic pre-trained model. 对，因为外面的数据质量参差不齐，你真的需要把那些劣质数据 ablate 掉，才能得到一个出色的 pre-trained 模型。 In fact, that's one of the challenges of a lot of the open weight models is they look great on one benchmark or two, but they're not great on practice. 事实上，很多开源权重模型的问题就在这里：在某个或两个 benchmark 上看起来很好，但实际场景中表现一般。 So, that's why, in fact, even in the RFDs are are pretty gone really excited about these MAI models because how the heck can a small 5B model hill climb? 所以，实际上在 RFD 里，大家对 MAI 模型也相当兴奋，因为一个小小的 5B 模型怎么就能 hill climb 起来呢？ Uh and it goes back a little bit to what I think is ultimately the key thing to do, which is try to pursue finding that cognitive core. 这其实回到了我认为最关键的一件事，就是去找那个认知核心。 Uh so, to me, starting with a clean lineage, then creating that ability for companies to be able to use this, right? 所以对我来说，先建立干净的血缘，再创造出让企业能够使用它的能力，是吗？ Not just as a generalist, but to create their own specialist by building this hill climbing scaffold around it, right? 不只是作为通用模型，而是通过在它周围搭建 hill climbing scaffold，让企业能创建自己的专用模型，对吗？ So, it's not just the model, but you have a hill climb scaffold around it, then you will start building your RLE. 所以不只是模型本身，你在外面套了一个 hill climb scaffold，接下来你就会开始构建自己的 RLE。 You will start collecting the traces. 你会开始收集 trace 数据。 Most importantly, you'll have private emails because we know all the emails out there are good, interesting, but they're not really that critical at this point because they all can be maxed. 最重要的是，你会有私有的 eval，因为我们知道外面那些公开 eval 很有意思，但到了现在这个阶段已经不是特别关键，因为都可以被刷满。 And so, the point is each company will have its own private email. 所以关键就是，每家公司将拥有自己的私有 eval。 And so, that end-to-end platform story around our models is sort of uh what I think is interesting. 围绕我们的模型所形成的这套端到端平台故事，我觉得才是真正有意思的地方。 And then the one other thing, Sarah, is since you brought that up is I do feel there's a new frontier. 还有一件事，Sarah，既然你提到了，我确实感觉出现了一个新的前沿。 Like people talk about the frontier and are you operating at the frontier? 人们谈论前沿，问你是否在前沿上运作？ Um interestingly enough, if you add a little temporality to it, you can use, let's say, in in in fact, that the Land O'Lakes demo we showed was pretty cool. 有意思的是，如果你给它加一点时序维度，就可以，比如说，事实上我们展示的 Land O'Lakes 演示就很说明问题。 We used whatever GPT-55, right? 我们用的是 GPT-4.5，对吗？ Then you collected a bunch of traces, and then you took a 5B reasoning model and achieved higher 然后你收集了一批 trace 数据，再拿一个 5B 推理模型去训练，最后超越了 uh so, that is another aspect of what it means to appear you know, operate at the frontier. 所以，这就是在前沿运作意味着什么的另一个维度。 Yeah. 对。 I I think uh I first of all have to congratulate you on basically building a frontier neural lab inside of Microsoft in 2 years. 我首先要恭喜你，基本上是在微软内部两年内就建起了一个前沿研究院。 Um I'm wondering, you know, you have all this AI strategy that you're rolling out. 我想问，你现在在推进这么大的 AI 战略， I'm wondering, what do you know now that you wish you would tell yourself 2 years ago where or 2 or 3 years ago? 如果能回到 2 到 3 年前，你现在知道的这些，你会对当时的自己说什么？ 3 years for the Jensen partnership, 2 years for uh MAI. 和 Jensen 的合作算 3 年，MAI 算 2 年。 Yeah, I mean, I think the the thing with that I reflect quite a bit, right? 是，我确实经常在反思这个，对吗？ Which is sort of obviously I got into all this when I got excited by the the scaling laws paper and, you know, when, you know, even the OpenAI partnership came about when those folks said, "Hey, we're going to really throw a lot of computer transformers." 当初是因为看到 scaling laws 那篇论文，我才真正兴奋起来的，而且 OpenAI 的合作关系就是在那些人说「我们要把大量算力投进 Transformer」的时候促成的。 Uh and they've held, right? 这条路确实走通了，是吗？ The thing that I always look back and say, "Wow, these things um do have capability that they're climbing up with 我每次回头看都会感叹，这些东西真的有其能力，而且一直在爬升， I mean, this, you know, this crude way of saying it is intelligence is log of compute kind of works. 我的意思是，粗略地说，「智能约等于算力的对数」这个规律确实成立。 Now, what I think we underestimated perhaps is the real world complexity of deploying these so that they actually deliver the value in the real world, right? 但我现在觉得我们可能低估的，是将这些模型真正部署到现实世界并产生实际价值的复杂度，是吗？ So the outcomes as measured by any benchmark is interesting, important, but the true eval is when people out there are able to do unique things that they only can value. 所以，benchmark 上的表现固然有意思、有意义，但真正的 eval 是用户能不能用它做到原本无法做到的事情，并且真的感受到价值。 And it's very measurable. 而且这是可以量化的。 Right? 是吗？ That I wish we had sort of even like had more in our consciousness, right? 这一点我希望我们早就有更清醒的意识，是吗？ Which is as an industry because right now I think when people say, "Wow, I don't want a token max." 就是作为一个行业，因为现在我觉得当人们说「我不想让 token 被用完」的时候， It's an artifact of us not having thought ourselves as an industry that we are using tokens to create value every step of the way. 这其实是我们作为一个行业没有想清楚的一个症状，我们应该在每一步都用 token 来创造真实价值。 So I think that's kind of what I wish we had gotten there, but I'm glad we are here. 所以我觉得这是我希望早就悟到的事，但很高兴我们现在到了这里。 What are some other use cases that you've seen that have created the most value for your customers? 你在客户中看到了哪些创造了最大价值的使用场景？ Because I know that people talk a lot about code and I think it's pretty clear that that's something that's having very large-scale impact. 因为大家说得最多的是代码，这一块的大规模影响已经相当清晰了。 Are there other areas that you find in common that your customers are really benefiting? 还有其他你在客户中普遍看到的、真正受益的领域吗？ Yeah, I think yeah, to your point, obviously coding is now got but it's interesting by the way Elijah to even talk about the coding, right? 对，代码这块肯定是很突出的，但有意思的是，Elad，光说「代码」本身都值得展开讲，是吗？ Which is coding is worked so well that we now have to rebuild the IDE, right? 就是代码方向已经做得太好了，现在我们必须重新设计整个 IDE，是吗？ I mean it's kind of nuts to see what we saw large is like, "Oh my god, I have these 100 agent sessions. 我是说，你看看大家的反应，有点震撼，「我的天，我有 100 个智能体在跑。 I the cognitive load it transfers back to me as a human is so excessive that now I need a new UI. 这些认知负担反压给我这个人类太重了，我现在需要一套全新的 UI。」 Uh oh, by the way, I like the the chat as the only artifact is also impossible. 而且，聊天界面作为唯一的交互形态也行不通了。 So that's why we need a canvas. 所以我们需要一个画布。 So it's kind of interesting for all the things about where is software needed or where is UI needed? 在哪里需要软件、哪里需要 UI 这个问题上，这挺有意思的。 Uh you kind of need that even for code, right? 即使是写代码也需要，是吗？ In a fully agentic world. 在完全智能体化的世界里。 But that said, one of the things that we are starting to see, we started seeing with co-work, but even some of the work we showed with auto autopilot, right, on what you see with claws, is a good one, because if you sort of think about a lot of human capital is doing the glue work, right? 话虽如此，我们开始看到一些迹象，从 GitHub Copilot 就开始了，包括我们展示的 autopilot 工作，还有你在 claw 上看到的，都是很好的例子，因为如果你想想看，大量的人力资本都在做「胶水工作」，是吗？ If you now can augment that with tokens {slash} agents that are long-running, durable, right, then your ability to scale even what is still judgment and glue work gets amplified like coding does. 如果你现在能用长期运行、持久化的 token 和智能体来增强这类工作，那么你在那些仍需判断力的胶水工作上的规模扩张能力，就能像代码一样得到放大。 So you can like I'm positive that 6 months from now we'll all be saying, "Oh, wow, like all through night the night there was a bunch of stuff that all these autopilots that I have working on my behalf with my delegated authority, so to speak, right? 所以我相信，六个月后，我们都会说：「哇，整个晚上，我那些代我运行、被我授权的 autopilot，可以说，做了一堆事情。」 I can sort of given even my identity, did a bunch of work. 就连我自己的身份也已授权，完成了一批工作。 Then of course I'll need my new ADE to say, "What did you do?" 当然，之后我需要一个新的 ADE 来告诉我：「你都做了什么？」 Like that. 就是这样。 My 我的 Did I do this work and so on. 这件工作是我做的吗，诸如此类。 So I think that that's where compressing of workflows, completing of tasks, that's where I think a lot of the value gets created. 所以我认为，压缩工作流、完成任务，这才是大量价值真正被创造出来的地方。 you raise a really interesting point, which is there's the actual agent is doing the code, and then there's a harness around it. 你提出了一个非常有意思的观点，智能体在写代码，但在它外面还有一个 harness。 And that's the environment, that's the context, that's everything you're setting up as a developer around actually a coding agent. 那个 harness 就是环境、上下文，以及你作为开发者围绕编码智能体搭建的一切。 What is the harness for the enterprise? 企业里的 harness 是什么？ Is there an equivalent concept for broader productivity work, or how do you think about that concept sort of generally? 更广泛的生产力工作有没有类似的概念，你总体上怎么理解这个概念？ So so some sense you kind of want the harness to define the models, the the data, and the tools, and so that you have a loop across those three. 从某种意义上说，你希望 harness 能把模型、数据和工具都纳进来，让这三者之间形成一个循环。 And so what we are trying to first of all make sure is each of our products that we build, right, whether it's GitHub Copilot or the security copilot the stuff we showed with M-Dash, or even the discovery for science, it doesn't matter. 所以我们首先要确保的是，我们构建的每一款产品，无论是 GitHub Copilot、安全版 Copilot、我们展示的 M-Dash，还是科学探索，都是一样的。 All of them are multimodal harnesses um tools access so that you can do this progressive uh disclosure of tools even so that they're token efficient. 所有这些产品都是多模态 harness，带有工具调用能力，让你能够渐进式地披露工具，从而也保持 token 效率。 Uh and then you're feeding it with very rich context because that's sort of the other hard lesson we've learned in the last 2 years is oh my god the amount of work you need to do to prep the context layer uh such that your plan can execute in the most efficient way is where the magic is. 然后你要向它喂入非常丰富的上下文，因为这是我们过去两年里学到的另一个沉重教训：为了让你的计划能够以最高效的方式执行，准备好上下文层所需要做的工作量，才是魔法所在。 So we have in our case we have the get up harness which essentially we're using across all our products. 我们自己这边有一个 harness，本质上在我们所有产品中都在使用。