AI 模型商品化与行业未来：Mistral CEO 访谈

如果所有领先模型表现都一样，那 AI 业务会是什么样子？它们现在确实有点趋同。我们稍后将从 Mistral CEO 那里了解。欢迎收听《大科技播客》，一档冷静、细致探讨科技世界及更广领域的节目。今天我们有一期精彩节目。我们将讨论 AI 业务和技术竞赛中，一些领先的基础模型开始趋同时会发生什么，以及这如何改变行业权力格局。我们请到了最合适的嘉宾。Arthur Mensch 与我们同在。他是 Mistral 的 CEO 兼联合创始人。Arthur，欢迎你。

What does the AI business look if all the leading models perform the same, which they kind of are, we'll find out with the CEO of Mistral right after this. Welcome to Big Technology Podcast, a show for Coolheaded and Nuance conversation of the tech world and beyond. We have a great show for you today. We're going to talk all about what's happening to the AI business and technology race as some of the leading foundational models start to look the same and how that changes the balance of power in the industry. We're joined by the perfect guest to do it. Arthur Mensch is here with us. He is the CEO and co-founder of Mistral. Arthur, welcome.

我很高兴来到这里。感谢你邀请我们。

I'm happy to be here. And thank you for hosting us.

不，有你在太好了。所以，Mistral 这个名字在 AI 圈内的人非常熟悉，但可能对我们的一些听众和观众来说是新的。那么，对于不熟悉 Mistral 的朋友，我给你们几个数据。Mistral 是一家 AI 模型构建公司，还做一些其他事情，我们稍后会谈到。总部在法国。公司始于 2023 年 4 月，估值 140 亿美元。所以不到三年或两年半就打造了一个 140 亿美元的业务。不错。公司有 500 人。Arthur，你在学术界待过一段时间，又在 DeepMind 待了两年半，现在领导这家公司。

No, it's great to have you. So, Mistral is a name that those who are deep in the AI world know very well, but might be new to some of our listeners and viewers. So, for folks who are new to Mistral, let me give you a couple of stats. It is an AI model builder. Does some other things which we're going to get to. It's based in France. Company is valued at $14 billion after starting in April 2023. So little under three years or two and a half years to make a $14 billion business. Not bad. There's 500 people at the company. And Arthur, you are leading it after spending some time in the academy and two and a half years at DeepMind.

没错。我们的总部在巴黎，但大约 40%的员工实际上在美国，我们的很多活动也在这里，所以我也花很多时间在这里，这也是为什么我们在纽约。

Exactly. We're headquartered in Paris but we have around 40% workforce which is actually in the US and a lot of our activity is actually here so that's why I'm spending a lot of time as well and that's why we are here in New York.

好的。很高兴你能来演播室。我们直接进入我认为当今 AI 最紧迫的问题。有很多讨论说谷歌在 2025 年底开始赶上 OpenAI 的模型，而 OpenAI 的模型与其他模型大致相当。在我看来，我们似乎比我想象的要快得多地进入了基础模型的商品化阶段。我原以为会有一场竞赛，一些公司会遥遥领先，其他公司需要一些时间才能赶上。但现在看来，很多模型构建者的前沿模型表现非常相似，很难说哪个最好。你怎么看？

All right. Well, great to have you in studio. Let's just go right to what I think is the most pressing issue for AI today. There's been so much talk about how Google at the end of 2025 started to equal OpenAI's models and how OpenAI's models were somewhat on par with others. And to me it seems like we're just hitting commoditization of the foundational model much faster than I thought it would be. I thought that there was going to be a race where some companies would leap out further ahead and would take others some time to catch up. But it looks like right now you have lots of model builders with their frontier models exhibiting performance that's so similar it's difficult to tell which is the best. So what do you make of that?

我认为这本质上是一项会被商品化的技术。原因在于它其实并不难构建。世界上大约有 10 个实验室知道如何构建这项技术，它们能获取相似的数据，遵循相同的配方和算法，而这些配方和算法实际上很短。训练一个模型所需的知识相当简短，因此它很容易传播。所以你无法创造知识产权差异化差距，因此很难真正超越竞争对手，因为知识的扩散让每个人都做同样的事情。所以问题在于价值在哪里积累？你应该追求什么样的商业模式才能确保最终盈利？我们看到一些竞争对手面临的挑战是，他们投入数十亿或数千亿美元来创造资产，但这些资产折旧很快，因为它们是商品。所以对我们来说，在 Mistral，这始终是一个问题，也是行业最大的问题之一：你需要投入足够多来为企业创造价值，但你也需要合理投资，以便在模型创建是资本密集型的世界中建立合理的单位经济性，而模型最终只是带来参与商品竞争的资产。

I would say that inherently this is a technology that is going to get commoditized. The reason for that is that it's actually not hard to build. You have around 10 labs in the world that know how to build that technology that get access to similar data, that follows the same recipes and algorithms which are very short actually. The knowledge you need to actually train a model is fairly short, so because it's short it actually circulates. So there's no IP differentiation gap that you can create, so it's very hard to actually leapfrog and to be way ahead of the competition because there's some diffusion of knowledge that is just making everybody do the same things. And so the question there is therefore where is the value accruing? And what kind of business model should you pursue to actually make sure that in the end you're turning profitable? And then the challenge that we see with some of our competitors is that they're investing billions or hundreds of billions into creating assets that are depreciating fairly fast because those are commodities. And so for us it has always been at Mistral it has always been a question and one of the biggest questions of the industry is that you need to invest enough to actually bring value to enterprises but you also need to invest reasonably so that you can build unit economics that makes sense in a world where the creation of a model which is capital intensive is actually just bringing you assets that are just in a commodity competition.

那么我们来谈谈这场构建最佳模型的竞赛。就像你提到的，这非常昂贵。OpenAI 将投入 1.4 万亿美元来建设其模型的基础设施，至少它是这么说的。如果模型实际上不相上下，公司会不会说，‘嘿，等等。也许我们投入所有这些钱来构建下一代更好的模型没有意义，因为别人能赶上。’

So let's talk a little bit then about this race to build the best possible model. I mean like you mentioned it's very expensive. OpenAI is going to put $1.4 trillion into building infrastructure for its models or at least it says so. If the models are effectively at par, are companies going to say, 'Hey, wait a second. Maybe it doesn't make sense for us to invest all this money into building the next evolution of a better model because people can catch up.'

我的意思是，从战略上讲，我认为肯定需要设定一个界限。你投入多少来创建足够有价值的资产，让一家科技公司能够为企业或消费者带来价值？归根结底，所有这些投资都需要由下游产生的自由现金流和价值创造来资助。因此，我们公司关注的重点，也是我认为合理的重点，是更多地放在下游应用上，找出企业遇到的摩擦并努力消除这些摩擦。因为归根结底，我认为当今行业面临的主要挑战之一是，AI 在三、四年前带来了很多承诺。但如果你问一家企业，‘你真的从中赚钱了吗？’他们通常会说不。原因在于他们没有进行足够的定制化，也没有从他们想要解决的问题出发反向思考。所以他们考虑的是解决方案，而不是问题。因此，试图帮助他们找到正确的用例并进行适量的定制化，这样当原本需要 20 人团队操作的供应链工作流，突然可以用两个人来完成。这样的例子很多。但行业面临的挑战是，我们需要让企业快速实现价值，以证明所有集体投资的合理性。

I mean, strategically, I think there's definitely some cursor to be set. How much do you invest in creating assets that are valuable enough for one technology company to bring value to an enterprise or to bring value to a consumer? And at the end of the day all of these investments will need to be funded by the free cash flow and value creation that is being made downstream. And so the focus that we have as a company but that I think is the reasonable focus is to be more on the downstream applications and to figure out what is the friction that enterprises are running into and try to lift these frictions because at the end of the day I think one of the major challenges that the industry is facing today is that AI brought a lot of promises like three or four years ago. But if you ask an enterprise, 'Did you actually make money out of it?' they will in general say no. And the reason for that is that they are not customizing things enough and they are not thinking backward from the problem they want to solve. So they think about the solution but they don't think about the problem. And so trying to help them to actually go for the right use cases and actually do the right amount of customization so that when it was a team of 20 people actually operating some supply chain workflow, suddenly you can actually operate that with two people. And there's a lot of examples like this. But the challenge that the industry will face is that we need to get enterprises to value fast enough to justify all of the investments that is collectively being made.

是的，这非常有趣，因为很长一段时间你会听到这些公司专注于模型、模型，对吧？当你想到 OpenAI 时，下一个 GPT-5 是最大的新闻。现在他们开始更多地讨论如何利用你拥有的智能来构建有效的应用程序。我可以分享一个报道：几周前，我有一篇报道，基本上是在纽约市与 Sam Altman 和一群新闻领袖共进午餐，Altman 告诉他们，他们的首要任务之一是为企业构建应用程序。基本上，这将是 2026 年的一个主要优先事项。这有点从‘我们想构建 AGI’到‘我们想为商业构建应用程序’的修辞转变。

Yeah, it is very interesting because for a long time you would hear these companies focus on model, model, right? The next GPT-5 was the biggest news when you think about OpenAI. Now they're starting to talk more about how do you take the intelligence that you have and build the applications that work. Just one bit of reporting that I can share: a couple weeks ago, I had this story basically inside a lunch with Sam Altman and a bunch of news leaders in New York City, and Altman told them that one of their biggest priorities was building applications for enterprise. Basically, it's going to be a major priority in 2026. And it's a little bit of a shift in rhetoric from 'we want to build AGI' to 'we want to build applications for business.'

那么谈谈为什么会这样？这是商品化问题的衍生吗？

So talk about why is that happening? Is it an offshoot of this commoditization issue?

嗯，我认为问题在于，首先，AGI 是一个非常简单的概念，可能对企业来说太简单了。不存在一个能解决世界上所有问题的系统。最终，它永远不会存在。你有大量的问题，就像没有一个人类能解决世界上所有任务一样。你需要一定程度的专业化来真正解决问题。所以我们从魔法思维转向系统思维。我们需要弄清楚用什么数据来让模型在特定任务上更好，需要设置什么反馈循环，以便从与系统交互的人类那里获取更多信号，最终让应用越来越好。在现实中，企业是复杂的系统，你无法用像 AGI 这样的单一抽象来解决。AGI 在很大程度上是我们无法实现的东西，它基本上是“我要让系统随着时间的推移变得更好”的北极星。但因为很难向投资者解释你正在构建的技术永远不会被竞争对手超越，所以叙事发生了转变。公司不再构建一个解决所有问题的单一北极星系统，而是需要深入企业的细节，解决他们的实际问题。在 Mistral，我们提前思考了这一点。这奠定了我们的故事。我们的故事一直是假设最终 AI 会更加去中心化，需要更多的定制化，因为我们正面临可获取数据量的限制和缩放定律的限制。基于这个前提，我们创建了公司：为企业带来更多的定制能力。

Well, I think the issue is, first of all, AGI is a very simple concept, probably too simple for enterprises. There's no such thing as one system that is going to solve all the problems of the world. At the end of the day, it's never going to exist. You have a wealth of problems, just like you don't have any human that is able to solve every task in the world. You need some amount of specialization to actually solve problems. So we're moving from magical thinking to system thinking. We need to figure out what data will be used to make the model better at a specific task, what feedback loop we need to set so that we accrue more signal from humans interacting with the system, so that eventually the application becomes better and better. In real life, enterprises are complex systems, and you can't solve that with a single abstraction like AGI. AGI, to a large extent, is what we were not able to achieve, and it's basically the north star of 'I'm just going to make the system better over time.' But because it's hard to explain to investors that the technology you're building will never be matched by your competitors, there's a shift in the narrative. Companies are not building a single north star system that solves all problems; they need to go into the weeds of enterprises and solve their actual problems. At Mistral, we've been ahead of time in thinking about this. That set our story. Our story has been to assume that eventually AI will be more decentralized, that more customization would be needed because we were running into the limits of the amount of data we could accrue and the limits of scaling laws. Because of that, we created the company on that premise: to bring more customization ability to enterprises.

是的。我们稍后会谈到 Mistral 的故事，但关于这一点还有一个问题。在我看来，而且我想知道你是否认为这是一个转变。你当然走在了前面。但似乎 AI 行业发生了一个转变：以前的想法是让模型变得更聪明，它们就能自己解决这些问题。例如，让模型更聪明，它就能做低级助理的工作，或者为多个系统做数据录入并提交报告。而现在似乎从那个想法转向了实际构建基础设施：模型只是一个组件，基础设施非常重要，像编排和构建在模型之上的应用程序才是价值所在。这很有趣。

Yeah. And we'll get to the Mistral story in a little bit, but one more question about this. It seemed to me, and I wonder if you think this has been a shift. You were ahead of this for sure. But it seems like there's been a shift in the AI industry where the idea was effectively make the models smarter and they'll be able to figure out these problems on their own. For instance, make the model smarter and it will be able to do a lower-level associate's job or maybe do data entry for multiple systems and file reports. And now it seems like there's been a shift from that to actually build out the infrastructure: the models are just one component, the infrastructure is super important, and things like orchestration and working through the applications built on top of the models is where the value will be found. It's interesting.

是的。我认为从系统角度来看，有两个组成部分，而且我们永远会有这两个部分。第一个部分是静态定义，规定工作流应该如何以及系统应该如何表现。这些静态定义由人类设定，定义系统的行为方式。这对应于你用来定义系统的手动信息。然后有一个动态部分，你将模型连接到工具，给模型指令，模型可以自己调用工具。所以它可以决定执行路径。这部分是动态的。还有一个静态部分，你设置护栏或有时决定决策树。我认为认为没有人类指导的动态系统就能解决一切是有点乌托邦和不现实的。过去三年行业发生的变化是，动态部分增长了，因为模型可以思考更长时间，调用多个工具，编写代码。但静态部分仍然极其重要。即使动态部分增长，静态部分也允许你创建更好、更有趣的系统，解决以前无法解决的问题。所以这些静态系统（你可以称之为编排）和动态系统（你可以称之为智能体）的结合将保持极其重要，因为两者一起提升，这样我们就能解决越来越复杂的问题。

Yes. I think if you look at it from a system perspective, you have two components, and we'll always have these two components. The first component is static definitions of how the workflow should be and how a system should behave. Those static definitions are set by humans defining how the system should behave. This corresponds to the manual information you use to define the system. Then there's a dynamic component where you connect a model to tools, give instructions to the model, and the model can call the tools itself. So it can decide on the graph of execution it will follow. That part is dynamic. And there's a static part where you set up guardrails or decide on a tree of decisions sometimes. I think it's a bit utopian and unrealistic to think you can solve everything with a dynamic system without guidance from humans. What has happened in the industry over the last three years is that the dynamic part has grown because models can think for longer, call multiple tools, and code. But the static part remains extremely important. Even if the dynamic part grows, the static part allows you to create systems that are even better and more interesting, and you can solve problems you couldn't before. So the combination of these static systems, which you can call orchestration, and the dynamic systems, which you can call agents, is going to stay super important because the two things are moving up together so that we can tackle more and more complex problems.

好的。那么现在确定了这一点，我在思考业务会是什么。假设模型已经商品化了。那么 AI 中的业务会是什么？我想象一些形式的消费产品，比如聊天机器人，你可以把 OpenAI 放在那个类别里。会有一种业务是让你现有的产品变得更好，比如与 Microsoft Excel 聊天。这可能是现有公司改进产品的一种方式。但还有另一个大类别，我们已经稍微讨论过，那就是企业方面。那么你会如何对这三个类别的商业机会进行排序？

Okay. And so now with that established, I'm thinking through what the businesses will be. Let's say the model has commoditized. So what are the businesses going to be in AI? I imagine some form of consumer products like chatbots, where you could put OpenAI in that bucket. There will be a business where you can make your existing products better, like chatting with Microsoft Excel. That could be one way current companies can make their products better. But then there is this other big bucket, which we've talked about a little bit already, which is the enterprise side of things. So how would you rank the business opportunity in those three buckets?

嗯，在消费端，因为 AI 开始成为你获取信息的方式，你基本上可以建立一个广告业务。这显然会被建立起来。这不是我们公司的重点。然后如果你看企业端，我们基本上是在重新平台化所有企业软件。在企业中，有人、数据和流程。历史上，工具是碎片化的，用于运行多个流程、多个数据系统、多个记录系统，团队也是碎片化的，无法同时访问所有信息。本质上，AI 在企业中允许你从统一的数据开始，甚至从碎片化的数据源开始，因为 AI 能够导航它们。然后你在上面放一个 AI，构建适量的智能，理解企业中正在发生的事情，AI 系统能够生成对每个人类实际工作有用的界面。

Well, on the consumer side, because AI is starting to become the way you access information, you basically have an ads business to be built. That's pretty clearly going to be built. It's not the focus of our company. Then if you look at the enterprise side, we're basically replatforming all enterprise software. In enterprises, you have people, data, and processes. Historically, there was a fragmentation of tools to run multiple processes, multiple data systems, multiple systems of record, and a fragmentation in teams that were not able to access all information at the same time. Essentially, what AI allows you to do in an enterprise is to start with unified data, or even fragmented data sources because the AI can navigate them. Then you put an AI on top that builds the right amount of intelligence, understanding what's going on in the enterprise, and the AI system is able to generate the interfaces that are useful for every human to actually work.

因此，重新构建整个企业软件栈的那部分，是能够创造大量价值的一个方面。在企业中，拥有上下文引擎——这个持续运行、观察正在发生的事情并为其创建文档的系统——以及拥有前端，而前端越来越多地按需生成。比如说，我是一名律师，我想解决我的一个问题并进行非常具体的审查。我只需带上我的文档，然后系统实际上会演变，向我展示正确的小部件和我需要的信息。所以，在上下文引擎之上的生成式界面，该引擎不断更新其对企业中正在发生的事情的表示，而底层是记录系统，这些系统基本上将只是纯粹的数据库。你不再需要之前堆叠在上面的所有东西。这就是发展方向，而这一重新构建将需要十年时间，因为企业需要一段时间来采用这些东西。但其中蕴含着巨大的价值，因为突然之间，你可以围绕这样一个事实重新组织你的公司：对于许多原本需要大量人力的流程，你现在可以运行得更快。一方面是效率。另一件事，也是最重要的——这是企业中的一种业务模式。企业中的第二个方面是与企业合作，帮助他们利用真正专有的数据，例如在制造业中，机器生产的资产，并将其转化为其他人无法复制的智能。因此，当你与一家制造飞机的公司合作时，或者当我们与 ASML 合作，制作专门擅长操作其机器的模型时，让模型在特定物理领域表现出色。这具有巨大的价值，因为你不仅是在公司内部构建效率，而且实际上是在释放由于缺乏人工智能而被锁定的技术进步。因此，新系统提供的这种解锁带来了巨大的增长。实际上，这更难衡量，因为第一个方面是短期的。你可以看到公司在五年后的样子，因为你已经削减了公司的某些部分。你重新引导其他人去创造增长。你可以为此建立模型。在技术方面，我认为这更难一些，因为我们知道有些东西，比如核聚变或更精细的半导体蚀刻。这些都是我们开始遇到物理限制的领域，而人工智能实际上可以帮助解除这些物理限制。因此，技术进步的加速，我认为将是大部分价值创造所在。这将需要一点时间，并且比人工智能将产生的效率提升更难衡量和预测。但这两件事同样重要。

And so that part of replatforming the entire enterprise software stack is the one thing where a lot of value can be created. In the enterprise, owning the context engine—the system that is constantly running, looking at what's happening, and creating documentation for what's happening—and owning the front end as well, which is more and more getting generated on demand. So let's say I'm a lawyer, I want to fix one of my problems and make a very specific review. I just bring my document and then the system actually evolves, showing me the right widgets and the right information I need. So the generative interfaces on top of a context engine that is constantly updating its representation of what's happening in the enterprise, on top of systems of records that are essentially going to be just pure databases. You don't need everything that was sitting on top before. This is where this is going, and that replatforming is going to take a decade because it takes a while to get enterprises to adopt these things. But there's immense value to be created because suddenly you can reorganize your company around the fact that for many processes where you had a lot of people, you can actually run those much faster. That's on one side, efficiency. And the other thing, which is the most important—that's one of the business modalities in the enterprise. The second one in the enterprise is about working with enterprises to help them take their really proprietary data, the assets being produced by their machines if it's in the manufacturing industry for instance, and turning that into intelligence that nobody else can reproduce. And so making models specifically good at a certain kind of physics when you're working with a company doing planes for instance, or when we're working with ASML making models that are specifically good at operating their machines. That's huge value because suddenly you're not building efficiency within the company but you're effectively unlocking technological progress that was locked by the absence of AI. So that unlock that the new systems are providing is immense growth. It's actually harder to measure because the first one is shorter term. You can look at what the company will look like in five years because you've reduced certain parts of the company. You've reoriented other people to be creating growth. You can create models of that. On the technological side, I think it's a little harder because we know there are things like nuclear fusion or sharper engraving of semiconductors for instance. These are things where we are starting to run into physical constraints, and artificial intelligence can actually help to lift those physical constraints. So the acceleration of technological progress is, I think, where most of the value creation will be. It will take a little bit of time and it will be less measurable and less predictable than the efficiency gains that AI is going to produce. But the two things are as important.

好的，让我试着理清一下。如果这将是人工智能世界价值的关键驱动力，那么有两种方式可以实现。一种是构建一个比所有人都好的模型，并以溢价出售。但我们已经讨论过，这似乎不会永远是一种模式。另一种方式是，模型本身并不是价值，而是专有技术和实施方面。所以你可以将模型开源，但随后为企业提供服务，帮助他们弄清楚如何将该模型付诸行动并实际获得结果。这是两种选择吗？

Okay, so let me see if I can sort of game this out here a little bit. So if that is going to be the key driver of value in the AI world, there's two ways to do it. One is to build a model that's better than everybody else and sell it for a premium. But we've already talked about the fact that that doesn't seem like it's going to be a mode forever. And the other way is, you know, the model is actually not the value, it's the know-how and the implementation side of things. So you can make the model open source but then provide a service to businesses to be able to figure out how to take that model and put it into action and actually get results. Are those the two choices?

是的，这就是我们在行业中看到的分叉。我们的观点一直是选择第二种，真正……

Yeah, that's kind of the fork that we see in the industry. And our view there has been to be on the second one, to really...

开源实现。

The open source implementation.

这带来了定制化，但也带来了去中心化。因为，如果你假设整个经济都将运行在人工智能系统上，那么企业只会想确保没有人能关闭他们的系统。就像如果你有一个工厂，你把它连接到电网，你想确保没有人会因为不喜欢你而关闭电网。如果人工智能实际上变成了一种商品，而事实正是如此，如果你把智能视为电力，那么你只想确保你获取智能的途径不会被限制。所以这也是开源技术能够带来的东西之一。因此，如果你使用开源，你就不必担心偏离轨道，比如我说像 Anthropic 的用户条款，然后暂停你做事的能力。如果你使用开源，你基本上可以按照自己的条件运行它。

Which brings customization, but it also brings decentralization. In that, if you assume that the entire economy is going to run on AI systems, well, enterprises will just want to make sure that nobody can turn off their systems. So the same way if you have a factory, you connect it to the grid, you want to make sure that nobody's going to turn off the grid because they don't like you. If AI effectively becomes a commodity, which is what's happening, and if you treat intelligence as electricity, then you just want to make sure that your access to intelligence cannot be throttled. And so that's also one of the things that open source technology can bring. And so if you're using open source, you don't have to worry about going astray of, I'm just saying like Anthropic's user terms, and then pausing your ability to do what you do. If you use open source, you can basically run it on your own terms.

从而暂停你做事的能力。

And so pausing your ability to do what you do.

是的，你按照自己的条件运行它。你创建你需要的冗余。你可以提供更高质量的服务。你可以确保无论地缘政治局势如何，你仍然可以运行这些系统。所以这实际上是 IT 方面的事情。因此，如果我是一名 CIO，我真的将开源视为创造杠杆和独立性的方式。但在科学方面，这也是你能够创建有效利用员工民间知识的系统的唯一途径。那是你几十年来招募的知识。将其转化为其他人无法访问的资产的唯一方法是基于这些开源模型创建你自己的模型。所以这很难。实际构建这些模型很难，对吧？所以这就是你需要正确工具的地方。你需要正确的专业知识。这就是构建开源模型的补充商业模式。

Yeah, you run it on your own terms. You create the redundancy you need. You can serve with higher quality of service. You can make sure that whatever the geopolitical situation may be, you can still run the systems if you want. And so that's really on the IT side. So if I'm a CIO, I really look at open source as a way to create leverage and independence. But on the scientific side, it's also the only way in which you can create systems that are effectively using the folklore knowledge of your employees. That's the knowledge that you've recruited for decades. The only way to turn it into an asset that nobody else gets access to is to create your own models based on those open source models. And so that's hard. It's hard to actually build those, right? And so that's where you need the right tools. You need the right expertise. And that's the complementary business model to building open source models.

但即使是闭源模型提供商，像 Anthropic 这样的公司也会说他们能够用你的数据定制他们的模型。你不相信吗？

But even the closed source model providers, companies like Anthropic will say they'll be able to customize their models with your data. You don't believe that?

他们会这么说，但随后他们会在上面设置一些护栏。所以你基本上是在信任他们的工程师会给你足够的系统深度访问权限。你能永远信任这一点吗？我不确定。所以这个问题既是控制问题，也是定制问题，对吧？

They will say that, but then they will put some guardrails on top of it. So you're basically trusting that their engineers are going to give you enough access to the depth of the system. And can you trust that for eternity? I'm not sure. So the issue there is as much a question of control as a question of customization, right?

就像供应商会试图锁定你。所以如果你获得访问权限，并且如果你在开源模型（比如我们的开源模型或任何人的）之上构建，你基本上就不太会被供应商锁定。而这是一项非常重要的技术，你不想被单一供应商锁定。所以这也是我们带来的机会。

Like a vendor is going to try to lock you in. So if you get access and if you build on top of open source models like our open source models or anyone, you're basically less locked into the vendor. And this is a technology which is so important that you don't want to be locked into a single vendor. So that's also the opportunity we bring.

你知道吗，让我震惊的是？ChatGPT 已经过去三年了，它让很多人意识到了这一点。虽然大型科技听众可能之前就有所了解，尤其是在 ChatGPT 出现之前我们就采访过那些认为这东西有意识的人，但那是另一个话题了。我们今天要说的，我总结一下你提出的两个主要观点。一是今天的人工智能模型无法独自完成所有事情，它们需要编排。第二个重要观点是，要用当前的智能进行这种编排或实施，你需要像托管服务这样的服务。所以我觉得有趣的是，我们从可能朝着一个全能的神级模型努力，变成了这可能是我们一生中见过的最强大的技术，但当你真正想使用它时，它在某种程度上变成了托管服务。

You know what's stunning to me? We're three years past ChatGPT, which brought this into a lot of people's consciousness. Although big technology listeners would have known about it a bit beforehand, especially since we were interviewing people who thought this stuff was sentient before ChatGPT came out, but that's a conversation for another time. What we're basically saying today, I'm going to sum up two of the main points you've made. One is that today's AI models can't do it all themselves; they need orchestration. And the second big point is that to do that sort of orchestration or implementation with the current intelligence, you need a service like a managed service. So it's interesting to me that we've gone from this perspective of maybe working towards a god model that could do it all to the fact that this may be the most powerful technology we've seen in our lifetimes, but when you actually want to use it, it becomes a managed service in a way.

是的，确实如此。我认为这不是历史上第一次观察到这种现象。这是一项新技术，一个新平台。如何使用它的知识仍然非常稀缺。没有多少人能够构建大规模运行、可靠运行并真正解决实际问题的系统。与企业合作时，你总是需要一些上层服务，因为实施很复杂，即使是像数据库这样相当成熟的技术也是如此。但对于人工智能来说，这更加必要，因为它需要改造企业。你还需要帮助思考团队应该如何围绕系统本身运作。而且它确实需要定制化。所以你需要数据科学家知道如何利用数据并将其转化为智能，而今天这仍然是一种稀缺资源。我确实期望软件在这些部署中的比例会增加。今天通过微调、强化学习等方式进行定制化的方式，将会从企业买家那里抽象出来，因为它太复杂了。他们应该只关心拥有自适应系统，从经验和与人的部署中学习，而不是考虑应该使用微调还是强化学习来将知识注入模型。我们正在做的工作是尝试将数据科学家理解的底层例程抽象为业务所有者可以实际使用的高级系统。所以这将会发生，我们正在努力。但服务部分仍然非常重要。今天，如果你是企业，这两者的结合是获得价值的最快途径。所以我们一直在将两者结合起来。

Yes, this is true. I don't think it's the first time we observe this in history. It's a new technology, a new platform. The knowledge on how to use it is still pretty scarce. There aren't that many people who can build systems that perform at scale, run at scale reliably, and actually solve an actual issue. When working with enterprises, you always need to have some services on top because of the complexity of implementation, even with fairly well-understood technology like databases. But for artificial intelligence, it's even more necessary because it requires transforming businesses. You also need to help in thinking how the team should perform around the system itself. And it does require customizing things. So you need data scientists who know how to leverage data and turn it into intelligence, and today this is still a pretty scarce resource. I do expect the part of the software in those deployments to increase. The way customization occurs today with fine-tuning, reinforcement learning, these kinds of things, this is going to be abstracted away from the enterprise buyer because it's too complex. They should just worry about having adaptive systems that learn from experience and from deployment with people, instead of thinking about whether to use fine-tuning or reinforcement learning to put that knowledge into their models. The work we are doing is to try and abstract away from lower-level routines that data scientists understand to higher-level systems that business owners can actually use. So it's going to occur, and we're working on it. But the service part is still going to be quite important. Today, the combination of the two things is the fastest way to value if you're an enterprise. So we've been combining the two.

我在对话开始时称你为模型构建者，然后我停顿了一下，说我们稍后会讨论。现在我们来了。基本上，我从你那里听到的是，Mistral 显然是一个自豪的模型构建者，但似乎没有服务，没有能够与企业坐下来并教他们如何使用它，这将是一个不完整的拼图。那么你认为自己最重要的工作是构建模型，还是提供服务？你主要是模型构建者还是服务提供商？

I started our conversation by calling you a model builder and I paused on it, saying we'd get into it later. Here we are. Basically, what I'm hearing from you is that Mistral is obviously a proud model builder, but it seems like without the services, without being able to sit with a business and show them how to use it, it would be an incomplete puzzle. So do you consider yourself as the most important thing you do building the models, or is the most important thing you do the service? Are you primarily a model builder or primarily a service provider?

我的意思是，我们是为了帮助客户获得价值。所以是服务。但要获得价值，他们需要有好的模型，而要获得价值，他们需要有正确的工具来训练模型。创建这些工具的最佳方式实际上是训练最好的模型。所以这两件事紧密相连。我们创建非常容易定制的模型。我们创建带有工具的模型，然后将其导出给客户，以便他们可以使用，并且我们帮助客户训练他们自己的模型。如果你不能向世界展示你实际上是开源技术的领导者，你就不能去销售给企业，说你将帮助他们创建非常定制的系统。所以这两部分同等重要。第一个使另一个成为可能，并且实际上有一个飞轮效应，因为我们在模型设计上做出的选择是为了支持我们各种各样的客户。一个例子是我们非常强调拥有在物理方面表现出色的模型，因为我们与遇到物理问题的制造公司合作。这就是我们通过让科学团队和业务团队坐在一起建立的飞轮。

I mean, we are there to help our customers get to value. So service. But to get to value, they need to have great models, and to get to value, they need to have the right tools to train the models. The best way to create those tools is effectively to train the best models. So the two things are extremely linked together. We create models that are very easy to customize. We create models with tools that we then export to our customers so that they can use them, and we help our customers train their own models. You can't go and sell to an enterprise that you're going to help them create very custom systems if you can't show to the world that you're effectively the leader in open source technology. So the two parts are equally important. The first enables the other, and there's effectively a flywheel there because we make our choices when it comes to the model design in a way that enables the various customers we have. One example is that we've put a lot of emphasis on having models that are great at physics because we work with manufacturing companies that run into physical problems. So that's the flywheel we have set up by having the science team and the business team actually sit together.

好的，我们请到了 Arthur Mensch。他是 Mistral 的 CEO，也是联合创始人。休息回来后，我们将讨论开源、开源运动与闭源。还记得 DeepSeek 吗？开源本应超越闭源。那么，它超越了吗？我们还将讨论地缘政治和监管，以及这是否会给这家公司带来优势，然后可能会进入一些更实际的例子，因为我们应该谈谈这项技术是如何在实际中应用的。我们马上回来。

Okay, we're here with Arthur Mensch. He is the CEO of Mistral, also co-founder. When we come back after the break, we are going to talk about open source, the open source movement versus closed source. Remember DeepSeek and open source was supposed to surpass closed source. Well, has it? We'll also talk about the geopolitics and regulation and whether that's going to give this company a leg up, and then maybe get into some more practical examples because we should talk about how the technology is being used on the ground. We'll be back right after this.

我们回到了 Big Technology Podcast，嘉宾是 Arthur Mensch，Mistral 的 CEO。Arthur，我想问你关于过去一年开源进展的问题。我记得一月份读到关于 DeepSeek 的报道，当时的主旋律是，这是开源的一次巨大飞跃，很快像 OpenAI 的 GPT、Anthropic 的 Claude 以及可能 Google 的 Gemini 这样的闭源模型就会被开源超越，因为开源社区在共同努力，相互借鉴创新，而闭源社区则各自为战。我们刚刚在节目开头谈到 Gemini 可能商品化了 OpenAI 的 GPT 模型，但关于开源是否达到了年初的期望，并没有这样的讨论。那么是我错过了什么，还是我理解错了？你怎么看？如果有什么阻碍了开源，那是什么？

And we're back here on Big Technology Podcast with Arthur Mensch. He's the CEO of Mistral. Arthur, I want to ask you about the progression of open source over the past year. I remember reading about DeepSeek, doing reporting on DeepSeek in January, and the overriding theme was that it was such a leap forward for open source that soon the closed models like OpenAI's GPT and Anthropic's Claude and maybe Google's Gemini would be surpassed by open source because the open source community was working together and building on each other's innovations where the closed source community was kind of going at it on their own. We just had this moment we talked in the beginning of the show about how maybe Gemini commoditized OpenAI's GPT models, but that conversation was not being had about open source living up to that expectation from the beginning of the year. So am I missing something, or am I reading it wrong? What do you think? If something has held back open source, what has it been?

嗯，如果你看 2024 年的趋势，我会说可能有一个大约六个月的差距。如果你看 2025 年的趋势，我认为差距大约在三个月左右。

Well, if you look at the trends in 2024, I'd say there might have been like a six-month gap. If you look at the trend in 2025, I think the gap is more around three months.

所以我想，明年的差距会是多少，这只能靠大家猜测了。但实际上，这个差距正在显著缩小。原因在于，当你预训练模型达到约 10^26 FLOPs 时，会出现饱和效应。原因是，预训练模型时，你能找到的可压缩数据是有限的。因此，那些可能起步稍晚的实验室，也积累了足够的算力来训练这种规模的模型，而且效率也提高了。这意味着，如今每个实验室都能在几个月内获得 10^26 FLOPs 的设施。

So I guess it's up to anyone to guess what the gap is going to be next year. But effectively this gap has been shrinking quite significantly. The reason is that you have a saturation effect when you pre-train models around 10^26 FLOPs. The reason is that there's only that much data you can find to compress when you pre-train models. So effectively, labs that maybe started a little behind created enough compute capacity to train models at this kind of scale, and efficiency has also increased. So what it means is that today everybody has access to 10^26 FLOPs facilities over the course of a few months.

那是算力的一个度量。

And that's a measure of compute.

那是算力的度量。嗯，那是算力乘以时间的度量。所以你需要 10^26 FLOPs，任何实验室今天都能在几个月内实现。正因为如此，饱和效应意味着开源模型已经赶上了，因为起步较早的闭源模型撞上了预训练的那堵墙。所以这个差距只会继续缩小。看看我们最新发布的开源模型 Death Desk 2，一个编程模型，它的性能大约相当于 Anthropic 两三个月前的水平。所以我认为差距在缩小。而且，我认为这个问题可能问得不对，因为这两者提供了截然不同的价值主张。一方面，它管理良好，你依赖提供商本身；另一方面，它需要更多努力，因为你需要更多地拥有它，学习如何定制它，使用正确的工具，如果你选择部署在自己的设施上，还需要维护它的部署。但最终，这为你提供了对抗闭源提供商所需的杠杆。所以这两类实际上是不同的，但如果只看纯性能方面，它们肯定在趋同。

That's a measure of compute. And well, that's a measure of compute times time. So you need 10^26 FLOPs, which any lab today can achieve in a couple of months. Because of that, the saturation effect means that open source models have caught up, because closed source models that started ahead kind of run into that wall of pre-training. So this is only going to continue shrinking. If you look at the latest open release we did, which is Death Desk 2, a coding model, it's performing around the performance of Anthropic around two or three months ago. So I think the gap is shrinking. And again, I think the question is probably not posed in the right way because it's also offering two very distinct value propositions. On one side, it's well managed and you depend on the provider itself. On the other side, it takes a little more effort because you need to own it more, learn how to customize it, use the right tools, and maintain its deployment if you choose to deploy it on your own facilities. But at the end, this creates the leverage you need against closed source providers. So the two categories are effectively different, but if you look at the pure performance side, they are definitely converging.

你提到了饱和效应。那么，不深入技术细节，模型是不是已经停止变好了？鉴于它们似乎都达到了饱和，AI 模型还会继续变得更好吗？

You mentioned that there's a saturation effect. So without getting too technical, are the models sort of done with getting better? Are AI models going to continue to get better given the fact that they all seem to be hitting saturation?

它们会在越来越具体的领域变得更好。我认为我们集体上已经让它们变得非常聪明，能够推理长上下文、调用多个工具等。但如果你想在银行或制造公司投入生产，模型需要学习公司内部包含的所有知识。所以对于非常精确的方向，比如我想让我的模型在发现材料或设计飞机方面极其出色，我需要付出一些努力，获得正确的奖励信号，找到正确的专家，让他们让我的模型在非常精确的方向上特别擅长。所以我们绝对没有做完这件事，因为我们都在争夺的是正确环境和针对特定能力的信号提供者。广泛的横向推理能力我们还会改进，但没有人会以创造与竞争对手巨大差距的方式来改进。真正的差距在于与垂直领域专家合作，他们确切知道如何设计飞机，并向模型解释如何做。你可以有很多方向，因为可以在物理、化学、制药、生物学中做。所以对我来说，未来两年最令人兴奋的部分是模型将在非常精确的方向上爆发式提升。对我们来说，机会在于拥有正确的平台来实现这种垂直化，无论是与企业合作，还是与专注于非常垂直化能力的 AI 初创公司合作。我们也乐于帮助他们。所以这就是我对这个领域走向的看法。我们一直在关注横向智能的增长，事情变得越来越聪明。未来两年将是关于让模型在某个技能集上变得极其出色。这实际上更令人兴奋，因为我们到了一个点，你可以选择一个领域并让它变得超人，但我们不会同时让它在所有领域都变得超人。

They will get better in more and more specific domains. I think we've really collectively made them very clever and able to reason about long context and able to call multiple tools, etc. But if you want to put them into production in a bank or a manufacturing company, the models need to learn about all the knowledge contained in the companies themselves. So for very precise directions, let's say I want to make my model extremely good at discovering materials or designing planes, I will need to sweat a little bit, get the right reward signal, get the right experts, and ask them to make my model specifically good in that very precise direction. So we are definitely not done doing that, because what we are all racing for is the right environment and the right signal provider for specific capabilities. The broad horizontal reasoning capabilities we're still going to improve, but nobody is going to improve them in a way that creates a strong gap versus competitors. The strong gap is actually in working with vertical experts that know exactly how they design a plane and explain to the model how to do it. You have a wealth of directions you can take, because you can do it in physics, chemistry, pharmaceutical, biology. So to me, the most exciting part of what's going to happen in the next two years is that explosion of very precise directions in which the models are going to get better. For us, the opportunity is to have the right platform for enabling those kinds of verticalization, whether with enterprises or AI startups working on very verticalized capabilities. We're happy to help them as well. So that's my view of where the field is going. We have been about horizontal intelligence growing and things getting more and more clever. The next two years is going to be about taking a model and making it extremely good at a certain skill set. That's actually more exciting because we're getting to a point where you pick a domain and can make it superhuman, but we're not going to make it superhuman in every domain at the same time.

好的。但关于这一点，在我们之前的对话中，你说你不会有一个能做所有事情的模型。但如果训练在特定垂直领域完成，为什么不行呢？

Okay. But on that note, earlier in our conversation you said that you're not going to have a model that can do everything. But if that training gets done in certain verticals, why not?

嗯，我们也到了一个点，你选择的垂直领域并不能真正迁移到其他领域。所以让一个模型既擅长非常精确的生物学又擅长非常精确的物理学是没有意义的，因为这两者之间的迁移实际上非常不明确。问题是，如果你真的想让你的模型同时解决所有问题，你会让它变得非常大、非常昂贵，服务成本也非常高。所以专用模型就是你要专门为生物做一个，为化学做一个，为这个特定的物理问题做一个。

Well, we are also getting to a point where the verticals you choose do not really transfer to the others. So there's no point in making a model that is good at very precise biology and very precise physics, because the transfer between those things is actually pretty unclear. The problem is that if you actually want your model to be able to solve every problem at the same time, you're making it very big, very expensive, and very costly to serve. So specialized models is really you're going to specialize one for bio, one for chemistry, one for this particular physics problem.

嗯，这实际上更有道理，因为如果你想大规模运行它，想让它后台运行，想让它日夜思考特定问题，你希望它尽可能小，因为模型的成本实际上与其大小成正比。如果你通过让模型在多个模式上都很出色来扩大规模，那么如果你想尽可能多地部署和使用它，效率实际上并不高。所以从经济角度看，在特定方向上做专用模型确实有意义。

Well, it actually makes more sense because if you want to run it at scale, if you want it to run in the background, if you want it to run day and night thinking about specific problems, you want it to be as small as possible because the cost of a model is actually proportional to its size. And if you inflate the size by making the model great at multiple modes, you're actually not very efficient if you want to deploy it and use it as much as possible. So if you look at the economies of it, it does make sense to make specialized models in certain directions.

让我问你一些关于 Mistral 竞争领域的问题。我想我们在美国。我就告诉你美国人在说什么，然后你来回应，因为值得讨论。我认为有一些人，不是所有人，但有一些人觉得 Mistral 在欧洲成立是为了有效利用监管俘获，因为美国公司很难在欧洲竞争，因此 Mistral 会在那里接手所有 AI 业务。

Let me ask you a little bit about the Mistral competitive area. I think that we're here in the US. I'll just tell you what people in the US say and let you address it because it's worth talking about. I think there is a feeling among some, not all, but some that Mistral has been set up in Europe to effectively take advantage of regulatory capture because US companies have a hard time competing in Europe and therefore Mistral will be there to pick up all the AI business.

嗯，我们构建技术是为了服务那些希望拥有足够控制权的公司和国家。人工智能不是一项你想完全委托给供应商的技术，尤其是来自外国实体的供应商。这在数据领域之前就成立，在 AI 领域更是如此，原因有很多。其中之一是，如果你依赖外部供应商，你的贸易平衡实际上会增加，你正在进口服务，如果长期进口过多数字服务，就会成为问题。主权对于国防也非常重要。如果你是一个独立国家，你需要独立的国防系统，而你需要自己独立的人工智能，因为它正在进入国防系统。

Well, we've built our technology so that we could serve companies and states that wanted to have enough control. Artificial intelligence is not a technology that you want to fully delegate to a vendor, especially if it's a vendor from a foreign entity. That was true before for data, and it's going to be all the more true for AI for multiple reasons. One of them is that if you depend on an external vendor, your commercial balance effectively increases and you're importing services, which becomes a problem long term if you're importing too many digital services. Sovereignty is also very important for defense. If you're an independent country, you want independent defense systems, and you will need your own independent AI because it's making its way into defense systems.

所以这个宣传对你来说真的很有效，比如‘我们不是美国公司，我们在欧洲，我们能够帮助你在国家安全如国防方面构建具有重要数据保护的东西’？

So it's really working for you, this pitch of being like 'we are not an American company, we're based in Europe, we'll be able to help you build something with important data protection in national security like defense'?

嗯，这是我们构建的技术差异化。因为我们可以在边缘构建，可以在客户希望的任何地方部署。即使我们倒闭，系统仍然可以运行，这对许多行业都很重要，而且越关键越重要。这也意味着我们可以服务那些希望减少对某些供应商依赖的美国客户。我们可以服务那些需要更多定制、更多控制且受更严格监管的银行。这也意味着我们可以服务欧洲工业，历史上我们就在那里起步——创业时你从邻居开始卖，我们就是这么做的。但我们也服务亚洲国家，他们有类似的问题。他们想要即使我们倒闭也能依赖的技术。他们想要可以根据自身文化需求定制的技术。这确实推动了我们的业务——围绕控制、开源和定制的技术差异化。

Well, it's a technological differentiation we've built. Because we can build on the edge, we can deploy wherever our customers want us to deploy. We effectively can die and the system will still be up, which matters for many industries, and the more critical it gets, the more it matters. That also means we can serve US customers that want to depend less on certain providers. We can serve banks that want more customization, more control, and are more regulated. It also means we can serve the European industry, where historically we were based—you sell next door when you start your company, and that's what we did. But we also serve Asian countries, and they have similar problems. They want technology they can rely on even if we were to die. They want technology they can customize to their own cultural needs. That has been driving our business for sure—that technological differentiation around control, open source, and customization.

有没有欧洲政府来找你，说‘我们就是不信任谷歌或 Anthropic，我们宁愿不基于它们构建’？

Do you have European governments coming to you and being like 'we just don't trust Google or Anthropic and we'd prefer not to build on them'?

嗯，实际上有欧洲政府来找我们，因为他们想构建技术并服务他们的公民。他们想提高公共部门的效率，而我们恰好有一个很好的方案——可以在他们的场所部署，我们可以派部署人员帮助他们实现价值。而且我们也是欧洲公司。所以欧洲国家投资欧洲技术实际上非常好，因为他们为我们创造的收入会再投资到欧洲，我们正在有效地围绕自己创建一个生态系统。从欧洲国家流向欧洲技术提供商的收入流非常有益。说实话，在美国，这已经运作了 80 年，我认为在欧洲我们做得还不够。

Well, we have European governments actually coming to us because they want to build the technology and serve their citizens. They want to increase the efficiency of their public sector, and we happen to have a good proposition for them—deployable on their premises, where we can send forward deployment people to help them get to value. And it turns out we're European as well. So it's actually pretty good for European countries to invest in European technology because the revenue they create for us is reinvested in Europe, and we're effectively creating an ecosystem around us. That flow of revenue from European countries to a European technology provider is very beneficial. To be honest, in the US that has been working for the last 80 years, and I think in Europe we haven't been doing it enough.

说到与地理有联系的开源公司或努力，你如何看待中国的开源努力？显然他们引起了很大关注。看起来那里进展相当顺利。

Speaking of open-source companies or efforts that have some links to geography, what do you think about China's open-source effort? Because obviously they've made a lot of noise. It seems like things are going quite well there.

是的，中国在人工智能方面非常强大。我们实际上是第一个发布开源模型的，他们意识到这是一个好策略。他们证明了自己非常强大。我不确定我们是否在竞争，因为开源的好处在于它并不是真正的竞争——你们互相借鉴。

Yeah, China is very strong on artificial intelligence. We were the first actually to release open-source models, and they realized it was a good strategy. They've proved to be very strong. I'm not sure if we're competing because the good thing about open source is it's not really competition—you build on top of one another.

对吧？你看到他们所有的东西，然后学习什么有效。

Right? You see everything they have out there and you learn what works well.

是的。反过来也一样。我们在 2024 年初发布了第一个稀疏混合专家模型，他们在此基础上构建并发布了 DeepSeek 3。

Yeah. And the same is true in reverse. We released the first sparse mixture of experts back at the beginning of 2024, and they built on top and released DeepSeek 3.

DeepSeek 就是基于那个构建的。

DeepSeek was built on top of that.

嗯，这是相同的架构。我们发布了重建这种架构所需的一切，同样地，投资开源的公司发布的所有东西都会被其他开源公司重用。实际上，这本身就是目的。如果你在不同实验室之间分享发现，研发效率会高得多。在中国，这非常有效——他们在不同实验室之间分享知识。在美国，这效率很低，因为美国公司没有投资开源。我们已经领先成为西方的开源提供商，我认为拥有一个西方的开源提供商是非常必要的。

Well, it's the same architecture. We released everything needed to rebuild this kind of architecture, and the same is true. Everything that companies investing in open source release are things that other open source companies reuse. Actually, it's kind of the purpose. R&D is much more efficient if you share findings across different labs. It's been very effective in China—they share knowledge across different labs. It's been pretty inefficient here in the US because US incorporated companies are not investing in open source. We've taken the lead on being the West's open-source provider, and I think it's going to be very much needed to have a Western open-source provider.

你认为中国的战略是什么？你认为美国有很多关于需要保持领先于中国的讨论吗？你认为如果中国在这方面领先会有风险吗？

What do you think China's strategy is? And do you think that there's this very large conversation in the US about the need to stay ahead of China? Do you think there's a risk if China runs away with this?

嗯，我认为中国非常强大。它是垂直整合的。他们有强大的工程师、算力、能源——他们拥有竞争所需的一切。欧洲也拥有竞争所需的一切。我不认为会出现某个人工智能领先其他所有人的情况。如果你看整个世界，每个足够大的主权实体——一个大经济体——都会希望在其 AI 使用和部署中拥有某种形式的自主权。所以这证明了多个卓越中心的出现是合理的。其中一个在欧洲，由我们领导；另一个在中国；然后还有西海岸的一堆公司。

Well, I think China is very strong. It's vertically integrated. They have strong engineers, compute, energy—they have everything they need to compete. Europe also has everything it needs to compete. I don't think we'll be in a setting where anyone is going to have one AI ahead of the others. If you look at the world in its entirety, every large enough sovereign entity—a big economy—is going to want some form of autonomy in its usage and deployment of AI. So that justifies the emergence of multiple centers of excellence. One of them is in Europe, led by us; another is more in China; and then you have a bunch of companies here on the West Coast.

为什么你认为开发这些开源模型符合中国的战略利益？因为他们没有像你这样的业务，对吧？他们不是在全球范围内成为实施者。

Why do you think it's in China's strategic interest to develop these open-source models? Because they don't have a similar business as you do, right? They're not going out globally and becoming implementers.

他们在中国有很大的业务。

They have a big business in China.

当然。在中国构建开源模型的公司通常是云服务提供商。有很多初创公司，但也有像阿里巴巴这样的云服务商。他们拥有这种垂直整合能力，可以在内部创造价值，无论是在中国还是在他们运营和增长的市场。例如在亚洲，对我们来说，我们往往在那里与他们竞争——不是在中国本土，而是在亚洲其他地区——他们内部竞争是有道理的，而他们进入美国市场的最佳方式就是免费提供这些东西。所以这确实合理。这是一种非常自然的做法：在中国建立一个受保护的业务，然后以零成本出口产品。如果我处在他们的位置，我也会这么做。

For sure. The companies that are building open source models in China are actually cloud providers in general. You have a bunch of startups, but you also have Alibaba, which is a cloud provider. They have this vertical integration that allows them to create value internally, in China and also in the markets where they are operating and growing. In Asia, for instance, which for us is a place where we tend to compete with them—not in China itself, but in the rest of Asia—it makes sense for them to compete internally, and their best way of accessing the US market is by just giving things away for free. So it does make sense. It's a very natural thing to do: build a business in China which is protected, then export the thing for zero. I would do the same if I were in their shoes.

好的。在我们结束之前，我想谈谈你正在构建的这项技术的实际应用。这很有趣。你之前谈到 AI 用于物理学、AI 用于其他研究应用、AI 用于国防。这些听起来都不像聊天机器人。所以请谈谈你正在开发的应用，以及我们是否会看到 AI 超越聊天机器人。

Right. All right. I want to talk a little bit before we leave about the practical applications of this technology that you're building. You know, it's interesting. You were talking a little bit about AI being used for physics, AI being used in other research applications, AI being used for defense. None of this sounds like a chatbot. So talk a little bit about the applications that you are working on and whether we're going to see AI move beyond the chatbot.

我的意思是，聊天机器人通常是界面，因为生成式 AI 允许你以人类的方式与机器交互。所以聊天机器人是人机界面，但仅此而已。现在，如果你看看那些让我们非常兴奋的实际应用，有两类：一类是真正的端到端工作流自动化，它从根本上改变了企业的运营方式。例如货物调度。当我们与航运公司 SEMA 合作时，我们帮助他们在船舶进港时调度所有集装箱。他们需要联系数百人，联系港口，联系监管机构，并激活 20 个不同的软件系统。这需要几百人来做，而通过合作自动化这些流程，突然之间你可以节省 80%。

I mean, the chatbot is often the interface because generative AI allows you to interact with machines in a human way. So the chatbot is a human-machine interface, but it's not the rest—it's only that. Now, if you look at the actual applications that are strongly exciting for us, you have two things: things that are really on the end-to-end workflow automation that effectively changes the way a business is fully run. Examples are like cargo dispatching. When we work with SEMA, which is a shipping company, we help them dispatch all of their containers when the ship comes into the port. They need to contact hundreds of people, contact the harbor, contact the regulators, and activate 20 different software systems. It takes a few hundred people to do it, and by working together on how to automate those things, suddenly you can save 80%.

所以大语言模型在进行这些通信，并且也在做决策——不仅仅是打电话，而是决定谁得到什么。

So the LLM is making those communications and also deciding—not just making the call, but deciding who gets what.

它做决策并连接各个部分，你衡量它是否在做正确的事，如果没有，你就改进系统。

It decides and it wires the things, and you measure whether it's doing the right thing, and if it doesn't, then you improve the system.

效果如何？

How's it doing?

它正在工作。实际上已经在某些机构上线了。所以这对我来说非常令人兴奋，因为它有物理足迹。它以安全的方式做决策，并且有效地为公司带来了巨大的效率提升。另一个例子，更偏向增长方面，是我们与 ASML 合作的项目。我们正在与他们合作开发视觉系统。

It's working. It's live actually in certain agencies. So that's very exciting to me because it has a physical footprint. It takes decisions in a safe way and it's effectively bringing a very large efficiency gain to a company. Now another example, which is more on the growth side, are things we do with ASML. We are working with them on vision systems.

对于那些不了解的人，请谈谈 ASML 是什么。ASML 是一家从事计算光刻和扫描的公司，他们的角色是制造那些大型机器，这些机器实际上在晶圆上雕刻，然后这些晶圆被用于例如英伟达的芯片。

Talk a little bit about what ASML is for those that don't know. ASML is a company that is doing computational lithography and scanning, and their role is to build those big machines that are effectively engraving the wafers that are then used as the chips in Nvidia, for instance.

对，所以他们是半导体制造的关键工业组件。

Right, so they're like a key industrial component of semiconductor manufacturing.

他们为半导体制造提供机器。这么专业的东西，你会想生成式 AI 怎么能帮到他们？

They provide the machines for semiconductor manufacturing. Something so specialized, you would think how's generative AI going to help them?

嗯，生成式 AI 模型通常是预测性 AI 模型。它们的一个优点是能够看到并推理所见之物。ASML 需要推理的一件事是来自扫描仪的图像，这些图像用于验证芯片雕刻中是否存在错误。这实际上相当复杂，因为需要一些逻辑思考。图像和逻辑思考的结合使我们能够更快地自动化这些过程，这意味着晶圆厂的生产线吞吐量将增加。在这种情况下，定制化是关键，因为输入的数据类型在其他地方找不到。ASML 是唯一拥有这些图像的公司。所以我们发现一个物理问题，实际上是制造过程中的瓶颈，然后我们去训练模型来有效解决它。这种情况会在很多不同的地方发生。那里需要生成式 AI，因为你需要一个能够推理图像的模型，推理能力至关重要，但为特定问题和特定输入类型定制这些推理模型才是关键。

Well, generative AI models are generally predictive AI models. One good thing they have is that they can see and reason about what they see. One of the things that ASML needs to reason about are the images coming out of their scanners that are verifying whether there are errors in the engraving of the chips. It's actually fairly complex because there's some logical thinking to be done. The combination of images and logical thinking is what enables us to actually automate those things much faster, which means that the throughput down the line of fabs is going to increase. In that setting, customization is key because the kind of input that is coming in is nowhere to be found elsewhere. ASML is the only one who has access to these images. So we find a physical problem that is effectively a bottleneck in a manufacturing process, and we go and train models that are effectively solving it. This is going to occur in many different places. Generative AI is needed there because you need a model that can reason about images, and the reasoning capabilities are critical, but customizing those reasoning models for a specific problem with a specific kind of input is the unlock.

是的，通用 AI 的工业应用对我来说非常令人惊讶和有趣。已经有技术，比如计算机视觉技术，可以查看一台机器或一个输出，并说“那不好”或“那就是我们需要的”。但还没有一个神经中枢，信息可以被传递到那里，然后做出决策，再传达给现场的人。而这正是这项技术正在实现的：整个技术工作流程开始能够由这项技术完成。

Yeah, the industrial applications of general AI to me have been super surprising and interesting. There has been technology, for instance computer vision technology, that can take a look at a piece of machinery or an output and say 'that's not good' or 'that's what we need.' But there hasn't been this nerve center that that information can be channeled to, and then have a decision made about it, and then communicated to somebody in the field. That's what this stuff is enabling: that full line of technical work is starting to be able to be done by this technology.

是的。基本上你需要的是能够感知多种信息的模型，而在制造业中，信息通常是视觉的。所以拥有非常强大的视觉模型非常有用。然后基于这些视觉模型，你可以做出选择，并依靠大语言模型本身来编排调用智能体、进入工作流的下一步、调用工具或在数据库中写入内容。拥有能够看到工厂中正在发生的事情、看到流程中正在发生的事情并采取下一步行动的动态智能体——无论是自动步骤还是调用智能体来验证决策——是创造大量价值的地方。这将重组制造业。制造业已经多次自我重组：当我们发明蒸汽机时，我们不得不围绕中央蒸汽机重建整个工厂，因为那是能源提供者。所以在未来 10 年，所有制造流程都将围绕大语言模型编排器重建。这非常有趣，因为你要解决物理问题。系统有物理足迹。

Yeah. Basically what you need are models that can perceive multiple kinds of information, and often in manufacturing, information is visual. So having very strong visual models is super useful. Then based on those vision models, you can make choices and rely on the LLMs themselves to orchestrate calling an agent, going into the next step of the workflow, calling a tool, or writing something in the database. Having dynamic agents that are able to see what's happening in a factory, see what's happening in a process, and take the next step—whether it's an automatic step or a call to an agent to validate a decision—is where a lot of the value can be created. This is going to reorganize manufacturing. Manufacturing had to reorganize itself multiple times: when we invented the steam engine, we had to rebuild entire factories around a central steam machine because that was the energy provider. So what's going to happen in the next 10 years is that all of the manufacturing processes will be rebuilt around LLM orchestrators. It's super interesting because you have physical problems to solve. The system has a physical footprint.

所以有一些安全问题需要解决。系统本身的复杂性就很大。这对我们这样的工程师来说是一个迷人的问题。

So there's some safety issue that you need to solve. Just the complexity of the system itself is huge. And so that's a fascinating problem for engineers like us.

让我看看我理解得对不对。我认为我们开始看到这些东西的种子开始真正在商业中产生影响。我们有一期节目采访了一位记者，他报道了一些律师如何利用这个工具更好地筛选文件。完美吗？不。我们在评论里看到了。不完美。但它显示了潜力。工业界也是如此，可能还有你提到的其他领域。但仍然感觉处于萌芽阶段。那么，是什么让它从今天的状态变成一种我们真正能在经济中看到影响的有效方式？仅仅是时间和定制化的耐心，还是模型的改进？

Let me see if I'm getting this right. I think what we're starting to see is the seeds of this stuff starting to be able to really have an impact in business. We had an episode with a reporter who was reporting on how some lawyers are really able to use this to sift through documents better. Is it perfect? No. We heard it in the comments. Not perfect. But it's showing potential. Same thing in industry and maybe also in other areas that you touch on. But still feels nascent. So what's going to get it from where it is today to something that is effective in a way that we really see the impact in the economy? Is it just time and patience on customization, or is it improvement of models?

我认为模型正在变得更好，这有帮助。每当你有一个更强的模型，你可以相信它会推理更长时间，并且失败得更少。

I think models are getting better, which helps. Whenever you have a stronger model, you can trust that it's going to reason for a longer period of time and that it's going to fail less.

嗯。

Mh.

但需要拥抱的是迭代。我们永远无法构建开箱即用、一次成功的系统。我们试图向客户传达的一点是，他们需要构建一个在 80%情况下都能工作的原型。但如何从 80%提升到 99%，从而将产品投入生产？方法是从用户那里获得反馈。如果系统不工作，如果你构建的 AI 软件不工作，这意味着你需要更多的数据和信号。这与我们过去构建软件的方式非常不同。因为以前软件不工作时，你基本上会回去编码并修复问题。但因为我们构建的是有机系统，模仿人类的系统，让它们变得更好的方法是给它们反馈，然后重新训练系统。所以这将使你提到的种子变成真正有价值的东西。这会奏效。

But then the thing that needs to be embraced is iterations. We're never going to be able to build systems that work out of the box in a single shot. The one thing that we try to convey to our customers is that they need to build a prototype that's going to work 80% of the time. But then how do they get from 80% to 99% where they can move the thing to production? The way to get it is to actually get feedback from users. If the system is not working, if the AI software you've built is not working, it means that you need more data and signal. That's something quite different from the way we used to build software. Because when the software was not working before, you basically went back to coding and fixed the problem. But because we're building organic systems, systems that imitate humans, the way to make them better is to give them feedback and then to retrain the system. So that will take the seeds that you mentioned and make them actual valuable things. That's going to work.

你提到了律师。我认为这是一个知识密集型领域，物理足迹很少。所以这是一个自然的领域，是最容易的。这并不容易。还没有完成。还有很多细微之处需要修复，才能使模型擅长法律工作。但如果你进入物理世界，那就更加复杂了。所以我们会看到知识世界的应用比物理世界的应用更快地投入生产。但可以说，物理世界的应用将更具变革性。

And you mentioned lawyers. I think it's one of the areas where it's very knowledge intensive, you have very little physical footprint. So it's a natural area, it's the easiest one. It's the easiest thing to do. It's not easy at all. It's not done yet. There's still a lot of subtleties to fix to make models great at lawyering. But if you go into the physical world, then it gets even more complex. So we'll see applications in the knowledge world go faster into production than the ones in the physical world. But arguably the ones in the physical world would be more transformative.

这让我们想到了机器人技术。我们就此打住吧。人们一直在谈论我们如何因为 LLM 或世界模型的进步而看到机器人技术的爆发。但这似乎还很遥远。我的意思是，他们有一个演示，叫什么来着，Neo，那个 neouoid 机器人，有一个人控制它，远程操作。有点奇怪。它们可能会出现在你的家里。所以我们还没有看到机器人技术的进步像我们在软件方面、大型语言模型方面看到的那样快。那么，如果它真的会来，它什么时候会来？

That brings us to robotics. So let's end here. People have been talking about how we could see an explosion in robotics because of LLMs or the advancements in world models. But it still seems far off. I mean they had the demo, what was it, the Neo, the neouoid robot, where there's like a person controlling it, teleoperating it. Kind of weird. They might be in your house. So we haven't seen progress in robotics start to move as fast as we've seen it in the software side, in the large language model side. So where does that come, when does that come, if it ever does?

我认为在机器人技术中，你需要结合两件事才能成功。硬件平台需要有合适的执行器和触觉信号，以良好的经济性大规模构建。这开始成为现实，行业在这个领域取得了很大进展。另一件事是，你需要有足够智能的控制系统来部署在这些机器人上。这就是我们发挥作用的地方。同样，你需要定制模型，因为模型需要针对平台进行定制——无论是人形机器人、轮式机器人还是飞行无人机——并且需要针对任务进行定制，因为任务会带来不同类型的图像。可以采取的行动类型会因任务而异。也许护栏也不同。这种对世界以及硬件平台带来的丰富数据的适应确实需要正确的平台和正确的训练平台。我们在机器人技术上的赌注，以及我们与多家公司（尤其是国防领域）一直在做的事情，就是构建一个平台，允许训练适合用途的模型，然后可能部署在边缘。因为在机器人技术中，从战略上讲，我相信我们首先会在不想派人的领域看到这类系统的部署——消防是一个很好的例子。当部署系统的风险远低于收益时，制造业也会如此，因为有些地方你只希望工厂是黑暗的。我认为中期来看，大部分价值将在那里创造。然后长期来看，可能会有东西放在你家里，但让一些相当强大的东西放在外面有点危险。就像我们过去 15 年一直在等待自动驾驶汽车一样，我们可能还要等很长时间才能在家里看到人形机器人。在此之前，我们将看到制造业的大规模部署，这需要正确的软件平台，而这就是我们正在构建的软件平台。

I think in robotics you have the combination of two things that need to work. Hardware platforms that need to have the right actuators with the right haptic signals, built at scale with good economics. This is starting to be true, and the industry has made a lot of progress on that domain. The other thing is that you need to have control systems that are sufficiently intelligent to be deployed on those robots. That's where we come in. Again, you need to have custom models, because the model needs to be customized to the platform—whether it's a humanoid robot, something on wheels, or a flying drone—and it needs to be customized to the mission, because the mission is going to bring different kinds of images. The kind of actions that can be taken are going to vary across the mission. Maybe the guardrails are different. That adaptation to the world and to the wealth of data that the hardware platform is bringing does require the right platform and the right training platform. Our bet in robotics, and what we've been doing with multiple companies, in defense in particular, is to build that platform that allows to train models fit to purpose that can then be deployed on the edge potentially. Because in robotics, strategically, I believe we'll see deployment of such systems first in areas where you don't want to send humans—firefighting is a very good example. When the risk of deploying the system is way under the benefit, it's going to be the case in manufacturing as well, because there are places where you just want the factory to be dark. I think that's where most of the value will be created, I would say midterm. And then maybe long term, you have things that are sitting in your house, but it's a bit dangerous to have some pretty strong thing out there. The same way we've been waiting for self-driving cars for the last 15 years, we'll probably be waiting for humanoid robotics in-house for a meaningful time. Before that, what we'll see is at-scale deployment in manufacturing, and that will take the right software platform, and that's the software platform that we're building.

好的。好了。真的是最后一个问题。我们谈了很多关于 AI 在商业中的应用。一些企业从中获益良多，一些则没有。显然有潜力，但也有大量的投资。有泡沫吗？我们现在处于泡沫中吗？

Okay. All right. Really the last one. We've talked a lot about AI in business. Some businesses have gotten a lot out of it, some have not. Clearly potential but also just like a ton of investment. Is there a bubble? Are we in a bubble right now?

嗯，我们处于一个需要大量基础设施的环境中。所以我们需要投资，例如我们在欧洲就是这样做的。但企业采用的粘性很高，因为需要时间来理解如何构建软件。需要一些构建工作。你不能购买现成的解决方案，然后相信你的生产力会取得巨大进步。这是许多企业在过去两年中经历的失望。所以还有一些构建工作要做。

Well, we're in a setting where we need a lot of infrastructure. So we need to invest, and that's what we do in Europe for instance. But then the viscosity of adoption in enterprise is high, in that it takes time to understand how to build the software. It takes some building. You can't buy off-the-shelf solutions and then trust that you're going to make immense progress in your productivity. That has been the disappointment that a lot of enterprises went through in the last two years. So there's some building to be done.

你可能需要购买一些基础组件，购买一定数量的分解函数，但随后你需要将自己的知识融入其中。所以这需要一些时间。你需要学习如何构建，然后还需要学习如何重组，这需要更长时间，因为团队会发生变化。你需要更少的管理，因为你需要更少的信息流通基础设施，因为 AI 能让信息更快地流通。某些功能会消失，某些功能会增长。所以重组工作还有很多要做，这将需要数年时间。问题是，今天所做的基础设施投资，是否会在两年、五年或十年内创造长期价值？这决定了有些人是在亏钱还是赚钱。这就是问题所在，我们并不真正知道。所以也许人们投资过度，也许投资不足。有些人肯定会亏钱，有些人肯定会错失机会。但今天，我的看法是，我们可能有点投资过度，有点过度承诺，但不是特别严重，而其他人则不然，因为我们看到在企业中真正创造价值是多么复杂。但最终我们会达到目标。最终整个经济都将运行在 AI 系统上，这是肯定的，但这可能需要 20 年，因为实际上相当复杂。

You need to maybe buy the primitives, buy a certain number of factorized functions, but then you need to bring your own knowledge onto it. So it takes some time. You need to learn how to build and then you need to learn how to reorganize, and that takes even longer because the teams are going to change. You need less management because you need less infrastructure to circulate information because AI allows information to circulate faster. Certain functions are going to disappear. Certain functions are going to grow. So there's just a lot of work to be done on reorganizing things, and it will take years. The question is the infrastructure investments that are being made today. Are they going to create long-term value in two years, in five years, or in 10 years? And that does define whether some people are losing money or making money. That's the problem, and we don't really know. So maybe people are overinvesting, maybe people are underinvesting. Some people will certainly lose money. Some people will certainly miss opportunities as well. But today, I would say my view is that we're maybe overinvesting a little bit and overcommitting a little bit, not massively, but some others, because we see how complex it is to actually create value in enterprises. But eventually we'll get there. Eventually the entire economy is going to run on AI systems, that's for sure, but it might take 20 years because it's actually fairly complex.

好的，网站是 mistral.ai。我们的嘉宾是 Mistral AI 的 CEO Arthur Mensch。Arthur，非常感谢你来做客。真的很感激你能来。

All right, the website is mistral.ai. Our guest has been Arthur Mensch, the CEO of Mistral AI. Arthur, thank you so much for coming in. Really appreciate being here.

谢谢你邀请我。

Thank you for hosting me.

不客气。好了，各位。感谢你们的收听和观看，我们下次在 Big Technology Podcast 再见。

You bet. All right, everybody. Thank you for listening and watching, and we will see you next time on Big Technology Podcast.