Mistral：欧洲 AI 冠军，追求战略自主

你于 2023 年 4 月创立了 Mistral。到现在还不到三年。在这段时间里，你建立了一家我认为现在有 800 名员工的公司，大约 800 人。所以，在有些人可能刚读完博士的时间里，你创立了 Mistral，目前拥有 800 名员工。三位联合创始人。你本人曾在巴黎的 Google DeepMind 工作过。Guillaume Lample 在 Facebook 工作过，他在 Facebook 读的博士。Timothée Lacroix 也是 Facebook 的博士生。所以，Google DeepMind、Facebook。你们三位联合创始人从这些经历中为 Mistral 带来了什么？你们想从那些公司借鉴什么，又想做出什么差异化？

You founded Mistral in April 2023. So, it's less than 3 years old. In that time, you have built a company with I think 800 employees now, but about 800 employees. So, in the time that some people might do a PhD, you founded Mistral with a workforce of currently 800 people. Three co-founders. You yourself spent time with Google DeepMind in Paris. Guillaume Lample with Facebook. He did his PhD with Facebook. And Timothée Lacroix also a PhD student with Facebook. So, Google DeepMind, Facebook. What did you three as co-founders bring from your experiences there to Mistral? What did you want to take from those companies and what did you want to differentiate?

嗯，我在谷歌待了两年多一点。在此之前，我其实在法国学术界。所以，我可以说大部分科学知识我都是在法国学术界学到的。但在 DeepMind 学到的，以及 Guillaume 和 Timothée 在 1 公里外的 Meta 办公室学到的，实际上是如何大规模训练 AI 模型。AI 是一项资本密集型的研发工作，因为你需要大量算力来训练强大的模型，训练那些大型模型，让它们同时在许多 GPU 上运行才能变得更好。嗯，你需要获得 GPU。事实证明，谷歌当时已经有很多资金可以投入，并提供了足够的基础设施来在该领域进行有趣的研究。所以，这就是我学到的。Guillaume 和 Timothée 也学到了这些。我们在 2023 年想做的，实际上是创建一个欧洲在生成式 AI 领域的冠军，因为我们看到了颠覆性技术的浪潮即将到来，它将对数字主权至关重要，对业务连续性至关重要，并将创造大量价值。我们从内部看到的是，当时没有其他选择。如果我们不采取任何行动来推广开源技术，让欧洲在地图上占有一席之地，那么欧洲就不会出现在地图上。这就是我们在 2023 年起步的方式。我们最初是一个研究实验室，然后发展成了公司，并拓展了业务。

Well, I spent a little more than 2 years at Google. Before that I was actually in French academia. So, I would say most of the scientific things I learned were actually in French academia. But what I learned at DeepMind and what Guillaume and Timothée were learning 1 km away at the Meta office was really to train AI models at scale. AI is a capital-intensive R&D endeavor in that you need a lot of computation to actually train strong models and to train models that are big and running on a lot of GPUs at the same time to make them good. Well, you needed to have access to the GPUs. And so, it turns out that Google had a lot of cash to invest already at that time and was providing enough infrastructure to do interesting research in that domain. And so, that's what I learned. That's what Guillaume and Timothée learned as well. What we wanted to do in 2023 was to actually create a European champion in the domain of generative AI because we saw the wave coming of disruptive technology that would become critical for digital sovereignty, that would be critical for business continuity, and that would create a lot of value. And the way we saw it from the inside was that there was no alternative at the time. If we weren't to do anything to promote open source technology and to put Europe on the map, there wouldn't be any Europe on the map. So that's how we got started in 2023. Then we started as a research lab and then grew the company and grew the business.

是的，你提到了主权这个词。这当然非常重要，而你们的 B2B 模式也确实让你们与众不同。你今天早上还与卢森堡首相吕克·弗里登会面，进行了关于主权的会谈。这些对话进行得怎么样？

Yeah, you've mentioned the word sovereignty there. It's very important of course and something that does differentiate you is your B2B model and you did have a meeting this morning with Luke Frieden, Prime Minister of Luxembourg, sovereignty meetings with people like that. How did those conversations go?

嗯，我们其实不太喜欢谈论主权，因为它常常被用来描述欧洲的低端技术。但另一方面，我们谈论的是战略自主。我们所做的，我们的技术之所以与众不同，不是因为它是主权的，而是因为我们与企业、与国家深度合作。我们在多个领域拥有最先进的模型，我们的团队能够快速调整它们，并在其上创建商业应用。这才是真正的差异化，是全球性的差异化。我认为这是我们在 Mistral 学到的第一课：如果你想成为该领域的冠军并保持相关性，你需要出口技术。你不能只专注于为欧洲构建国内技术。你需要成为一个出口商。这是第一点。现在，关于你提到的作为政治话题和术语的主权，人工智能对主权很重要，我认为主要有三个原因。第一个是经济主权。今天白领所做的许多任务将越来越多地委托给 AI 系统，因此这固有地改变了价值在行业和经济中的积累方式。这意味着，如果欧洲仍然 80%依赖美国供应商，就像今天这样，那么数字服务方面的经济平衡将继续发生变化。我们流向美国的价值将被再投资于研发，这将扩大他们与我们之间的差距。所以这是一个大问题，最终会成为一个战略问题，因为它造成了依赖，因为他们有开关按钮等等。所以存在经济主权问题，也存在业务连续性问题。AI 关乎运行流程，它将运行我们经济中的关键流程，无论是公用事业公司、工业公司、服务业还是国防系统。如果你无法保证业务连续性，那么你基本上就会依赖提供技术的国家，成为其附庸。这是我们需要避免的。这尤其适用于国防系统、公共部门，但也适用于国防系统。然后是第三个支柱，即经济主权、战略主权、业务连续性。第三个支柱是文化。这些系统正在构建或生成内容；它们是交互机器。它们以某种方式行为，具有某些文化偏见。如果你有集中式系统，那么你基本上就信任中央提供商，无论他想推动什么想法或观点。我们认为这实际上与民主不相容。因此，为每个国家定制系统，将语言引入稀有语言，特别是将欧洲语言引入我们的模型，这一点非常重要，我认为它将在即将到来的选举中发挥关键作用，因为这项技术是一个重要的影响力来源。我们是一家 B2B 公司，所以我们努力确保所有公司都能使用我们的技术来服务最终用户，但我们非常担心面向消费者的 AI 集中化，这是我们真正应该认识到并视为风险的事情。

Well, we don't really like to speak about sovereignty because it's often a term to describe lower technology in Europe. But on the other hand, we speak about strategic autonomy. What we do, our technology is not differentiated because it's sovereign; it's differentiated because we engage deeply with enterprises, with states. We have state-of-the-art models on multiple domains and our team is able to adapt them quickly and to create business applications on top. So that's the true differentiation, which is a global differentiation. I think that's the first learning we had at Mistral: if you want to be a champion and relevant in the space, you need to be exporting technology. You can't focus on just building domestic technology for Europe. You need to be an exporter. So that's the first thing. Now, to your point on sovereignty as a political topic and a political term, artificial intelligence is important for sovereignty, I'd say for three main reasons. The first one is economic sovereignty. A lot of the tasks that white collars are doing today are going to be more and more delegated to AI systems, and so inherently this is creating a shift in where the value accrues across the industry and across the economy. What that means is if Europe remains fully dependent on US providers to an 80% extent, which it is today, the economic balance is going to continue shifting in terms of digital services. The value that we send to the US is going to be reinvested in R&D, and that's going to increase the gap between what they do and what we do. So that's a big problem, and that's eventually going to become a strategic problem because it creates dependency, because they have an on/off button, etc. So there's a question of economic sovereignty, there's a question of business continuity. AI is about running processes, and it's going to run critical processes in our economy, whether with utility companies, industry companies, services, defense systems. If you can't afford the business continuity, well, you're going to depend on basically becoming a client state to the one that is providing you with the technology. So that's something we need to avoid. And it applies specifically to defense systems, to public sector, but also to defense systems. And then the third pillar, so economic sovereignty, strategic sovereignty, business continuity. The third pillar is cultural. These are systems that are building or generating content; they are interaction machines. They behave in a certain way, they have certain cultural biases. If you have centralized systems, then you're basically trusting the central provider with whatever ideas or opinions he wants to push. And we don't think that this is actually compatible with democracy. And so that aspect of customizing systems for every state, bringing languages to rare languages, European languages to our models specifically is super important, and I think it's going to play a critical role in upcoming elections because this technology is a major source of influence. We're a B2B company, so we try to make sure that all the companies can take our technology and serve end users, but we're quite afraid of the concentration of AI for consumers, and that's something we should really realize and consider as a risk.

谢谢。嗯，我会回到语言这个话题。我很高兴你提到了这一点，因为我们身处欧洲，我甚至不知道这个房间里有多少种语言。以你们的速度扩张，以及人才问题。资金投资，我们在过去一天半里一直在谈论这个，以及以合适的价格找到合适的人才。首先，你说你在学术界学到了你需要的一切。你在法国读博士。太好了。我们在欧洲各地都有杰出的研究人员。你们在哪里找到最优秀的人才？

Thank you. Well, I'm going to come back to languages. I'm very glad that you mentioned that because we're sitting here in Europe and I don't even know how many languages are in this room. Scaling up at the rate that you have and talent. So money investment, we've been speaking about that for the last day and a half and finding the right talent at the right price. So firstly, you said that you have learned everything you needed to learn in academia. That's where you were studying in France for your PhD. Wonderful. We have brilliant researchers across Europe. Where are you finding your best talent?

我们在欧洲各地找到他们。实际上，从历史上看，我们的大型总部在巴黎，那里有优秀的工程学院和博士项目，但我们也在卢森堡找到他们。我们在华沙找到他们。

We find them across Europe. Actually, we find them historically our big headquarters is in Paris, which has great engineering schools, great PhD programs, but we also find them here in Luxembourg. We find them in Warsaw.

我们在德国、希腊招人。英国虽然已不是欧盟成员，但我们在那里也有一个相当大的团队。你们的第二大办公室在卢森堡。但总的来说，我们的策略是在欧洲开设本地办公室，寻找区域性人才库。欧洲有很多区域性人才库。如果你给研究人员或工程师机会留在自己的家乡，提供合适的薪酬和项目，他们更愿意那样做，而不是飞到美国。所以我们的策略就是设立这些本地办公室，招聘初级人才，因为欧洲很擅长培养初级人才。然后有时我们会把一些在欧洲出生的美国员工调回欧洲，他们带来了某种资深经验。

We find them in Germany, in Greece. It's no longer part of the European Union, but we have a fairly big team in the UK as well. Your second biggest office is in Luxembourg. But overall our strategy has been to open local offices and go for regional talent pools in Europe. Europe has a lot of regional talent pools. If you give a researcher or an engineer the opportunity to stay in their hometown with the right compensation and the right project, they would rather do that than fly to the US. So our strategy has been to have these local offices and hire junior people because Europe is great at producing junior people. Then sometimes we relocate some of our fellow European citizens from the US to Europe, and they bring a certain seniority.

你在欧洲招聘初级人才这点很有意思。我想稍微深入一下，因为薪资问题——我知道这个是因为我在伦敦谷歌 DeepMind 有朋友——他们做的工作能拿到七位数。那么，为什么有人会选择为 Mistral 工作，而不是去美国拿那七位数呢？

That's an interesting point that you're hiring junior people in Europe. I'm going to lean on that slightly because the salary tag—I know this because I've got friends at Google DeepMind in London—they can get seven figures for the work they do. So why would somebody choose to work with Mistral as opposed to getting those seven figures in the US?

首先，他们在我们这里也能拿到类似的薪水。尤其是他们看到，我们通过期权和股票给予他们很大的激励。我们的股权轨迹意味着，从纯粹的经济角度来看，加入 Mistral 实际上比加入谷歌更有吸引力。这就是我们的做法。实际上，股票期权制度在欧洲有点噩梦，因为有 27 种不同的制度。统一起来会很好，尽管这很难。

Well, first of all, they get similar salaries with us. Especially what they see in coming with us is that we heavily incentivize them with options and shares. The equity trajectory we've had means that from a pure economic perspective, joining Mistral has been actually more interesting than joining Google. So that's how we work. It's actually something—the stock option regimes is a bit of a nightmare honestly in Europe because you have 27 of them. Unification of that would be great, although it's very hard.

哦，你提到了今天的话题。昨天默克集团的 CFO 海伦·冯·罗登也提到了类似的事情，虽然方式不同，但类似——欧盟的立法国与国之间还有其他的调整。你谈到了股票，但我想招聘和人力资源流程可能也一样。那么，对于像你这样快速增长的公司，你能提出一些建议吗？什么能让你的生活更简单、更高效，让公司以你这样的速度增长？

Oh, you've touched on the topic of the day. That wasn't the first time something like that was brought to our attention yesterday with Helen von Rhoden, CFO of Merck Group. She was saying that in a different way but similar—with some of the EU legislation country to country, there are other tweaks put on top. So you've talked about shares, but I imagine it might be the same with the hiring, the HR process. So if you could put out a few recommendations for a company that's growing at the rate you're growing, what would make your life simpler and more efficient for companies to grow at the rate you are growing?

我认为欧洲最大的问题是通知期。你不能直接从其他公司挖人，因为他们需要在那里待满 3 个月。所以招聘的粘性比美国高得多。最大的问题——每个国家情况不同，有些更糟——是员工想离开公司必须提前 3 个月通知。这完全是一场灾难。我们应该给员工更多权利，确保如果他们想离开公司，可以在一周内离职。这是最大的障碍。在欧洲，有很多事情比在美国更难做，但招聘的粘性是其中之一。如果只能改进一件事，那就是这个。

Well, I think the biggest problem in Europe are the notice periods. The fact that you can't just get some people from companies because they will need to stay there for 3 months. So the viscosity of hiring is much higher than in the US. The biggest problem—and it applies to every country in different ways, some are worse than others—is the fact that an employee who wants to leave his company has to give a 3-month notice period. That's a full catastrophe. We should give more rights to employees. Make sure that if they want to leave their company, they can leave in like a week. That's the biggest source of... There are many things that are harder to build here in Europe than in the US, but the viscosity of hiring is one of them. If there was one thing to improve, it's really this.

第二件事是什么？

What would the second thing be?

我认为有些事情欧洲本身无法改变太多，因为这关乎时间和系统的深度，以及已经规模化公司的积累。但我们发现的最大障碍是招聘高级人才、高管、有规模化市场推广团队经验的人、有规模化营销团队经验的人。所以人才短缺并不在你预期的地方。当然，欧洲很擅长培养工程师，我们也找到了如何留住这些工程师、确保他们不飞往美国的方法。但因为生态系统年轻，没有那么多公司已经完成了从起步到 IPO 的规模化历程，所以你找不到……如果我在硅谷找 CMO，一周内可以面试 10 个，下周就能雇到一个。在这里，基本上没有 CMO 能真正做我们需要的事情。所以这里存在严重短缺，我们对此无能为力，因为这完全关乎经验，以及是否有已经培养出这些人的公司，这些高管会加入更年轻的公司。

Well, I think there are things that inherently Europe cannot do much about because it's about time and depth of system and accumulation of different companies that have already scaled. But the biggest hurdle we find ourselves in is to hire senior people, executives, people that have scaled go-to-market teams, people that have scaled marketing teams. So the talent shortage is not where you would expect it. Of course, Europe is great at training engineers, and we have figured out how to retain those engineers and make sure they're not flying away to the US. But because the ecosystem is young, because there aren't that many companies that have already done the scaling journey all the way to IPO, etc., you can't find... If I'm looking for a CMO, if I was in Silicon Valley, I could interview 10 of them in a week and hire a CMO the week after. Here, there's basically zero CMO that can actually do what we need to do in Europe. So there is a strong shortage there, and nothing we can do about it because it's all about experience and having already the companies that have trained those people, those executives that will join younger companies.

你提到了工程师。我知道你当然需要工程师。Anthropic 的联合创始人兼总裁丹妮拉·阿莫迪最近在媒体上因说人文学科变得比以往更重要而闻名。你怎么看？你们雇了多少人文学科背景的人？

You mentioned engineers. I know that you need engineers, of course. Daniela Amodei, co-founder and president of Anthropic, very recently is famed in the media for saying humanities subjects are becoming more important than ever. What do you think about that? How many humanities degrees are you hiring?

我们雇了很多……我的意思是，我们倾向于招工程师，包括销售团队，但我们的研究团队实际上是一群博士，其中一些人文学科背景。我们有记者在调整模型，确保它们以特定方式表现，确保它们没有偏见。本质上我们做内容生成，内容生成就是媒体。所以你需要知道媒体如何运作、写作如何运作的人。我们的模型实际上在创意写作方面很棒，因此被选中。所以我们有创意人士使用我们的模型来做这件事。我们与一个非营利组织合作，他们正在创作一部新的莫里哀戏剧。他们是一个完整的人文学科部门，是莫里哀专家，他们用我们的模型生成了一部新的莫里哀风格戏剧，但使用了人工智能。所以有很多创意工作可做。还有很多关于如何处理历史话题等主题。所以我们在公司内部有一个完整的团队叫“模型行为”，他们大多是人文学科背景。

We hire a lot of... I mean, we have a bias about engineers including in sales team, but our research team is actually a team of PhDs and some of them are actually humanity trained. So we have journalists that are tweaking the models and making sure that they behave in certain ways, making sure that they are unbiased. Inherently we do content generation, and content generation is media. So you need to have people that know how media is done, how writing is done. Our models are actually great at creative writing and they get picked up for that. So we get creative people that use our models to do that. Working with a non-profit organization that is creating a new play of Molière. They are a full humanity department, specialists of Molière, and they used our models to generate a novel Molière play in the style of Molière but using artificial intelligence. So there are many creative things to do. There's also a lot of topics around how do you treat history topics, etc. So we have a full team called model behavior in the company, and those are mostly humanity trained.

我只想指出，并非所有记者都有人文学位。我还想指出，丹妮拉的兄弟达里乌斯学的是物理学和生物物理学。所以我认为他们两人之间有一个很好的平衡。很遗憾，我和你的时间不多了。所以我将简短地谈一下欧洲的语言和大语言模型。你们是怎么处理的？因为这是个难题。

I just want to point out that not all journalists have humanities degrees. I also want to point out that Daniela's brother Darius studied physics and biophysics. So I think they have a brilliant balance there between the two of them. I don't have much time sadly with you. So I'm going to just jump on a bit very briefly on languages in Europe, large language models. How are you approaching that? Because it's a tough question.

归根结底，如果你想让模型在某种语言上表现强劲，你需要有该语言的内容。然后在训练时，你需要选择是否让模型在英语上稍微变差，但在德语或法语上稍微变好。

Well, at the end of the day, what matters if you want to make a model strong in a certain language is that you need to have access to content in that language. Then you need to pick at the time of training whether you're going to make the model slightly worse in English but slightly better in German or in French.

因此我们做出的选择实际上在语言之间更加均衡，因为我们的市场固有地有 60%在欧洲。所以我们服务这些市场。这对文本很重要，对音频也很重要。我们发布了转录音频转录机器，它们非常擅长处理所有欧洲语言。我们的做法是故意让我们的模型在英语上稍差一些，但在其他语言上稍好一些。我们让它变得更好的方法是选择放入模型的语言混合比例。如果你多放一点法语的比例，那么模型在法语上会好很多。但这意味着你需要少放一点英语的比例。所以最终，多样性……这是一个权衡。你的容量有限，所以如果你想让模型具有多元文化，那么它在谈论美国文化话题等方面就会稍逊一筹。这其实是件好事，因为美国模型的问题在于它们非常以美国为中心。所以我们在这方面下了很大功夫。公共部门可以提供帮助的方式是提供这些内容，并创建我们可以训练的自由版权内容目录。这就是我们与多个国家合作的方式。我们与希腊合作，与摩洛哥合作，我们应该与欧洲每个国家都这样做。

And so we make choices that are actually much more even across the languages because inherently our market is 60% in Europe. So we serve those. It matters for text, it matters for audio as well. We have released transcription audio transcription machines that are great at dealing with all the European languages. And the way we do it is really to deliberately choose to make our model slightly worse in English but slightly better in other languages. And the way we can make it better is to choose the mix of languages that you put into the model. If you put a little more percentage of French, then the model is going to be much better at French. But then that means you need to put a little less percentage of English. So at the end, the diversity... It's a trade-off. You have a limited capacity, so if you want to make the model multicultural, then it's going to be a slightly less proficient at speaking about US cultural topics and these kinds of things. And it's actually a good thing because the problem with models from the US is that they're very US-centered. So we work heavily on that. And the way the public sector can help is really by providing that content and by creating catalogs of free-of-rights content that we can train on. And that's what we do with multiple states. What we do with Greece, that's what we do with Morocco, that's what we should do with every state in Europe.

而且你们也是开源的，这很棒。我时间不多了，我们面前还有 Crystal。用 30 秒回答，欧洲 AI 的未来是什么？

And you're open source as well, which is a great thing. I have almost no time left and we have Crystal in front of us. In 30 seconds, what's the future of AI in Europe?

我认为人工智能是技术的一个转折点，我们拥有的机会是创建大规模 AI 云服务提供商，并减少对我们本不该接受的数字服务的依赖。但之前的依赖是关于数据存储的，主要是关于隐私。如果这样的领军者不出现，我们可能创造的新依赖是流程依赖，以及业务连续性问题、业务连续性风险。所以我们需要这样的参与者出现。我们正在提出一个垂直整合的 AI 云，我们将继续扩大规模，而实现这一点的唯一途径是通过欧洲的团结、通过需求、通过与欧洲领军者合作。这就是我们今天实际上能够实现的。

I think artificial intelligence is such an inflection point in technology that the opportunity we have is to create large-scale AI cloud service providers and reduce the dependency on digital services that we have accepted that we should never have accepted. But the previous dependency was about data storage. It was mostly about privacy. The new dependency that we may create if such champions do not appear is a process dependency and a business continuity problem and a business continuity risk. So we need such actors to emerge. We are proposing a vertically integrated AI cloud and we will continue to scale, and the only way we can do it is through European solidarity, through demand, through working with European champions. And that's what we are actually able to achieve today.

Arthur，非常感谢您来到 EIB 全球论坛。我们很高兴有您，祝您和您的联合创始人为欧洲所做的一切好运。谢谢。

Arthur, thank you so much for coming here to the EIB Global Forum. We're very happy to have you and good luck with everything that you and your co-founders are doing for Europe. Thank you.

谢谢。

Thank you.

我有一个迫切的问题。大家都知道下周我们庆祝国际妇女节。所以我敢问您，您的团队中有多少女性，以及您如何防止模型中的性别偏见？

I have a burning question. You all know that next week we celebrate International Women's Day. So I dare asking you how many women you have in your teams and how do you prevent gender biases in your models?

哦，好的。谢谢，Julie。我之所以能回答你的问题，只是因为你是 Julie Becker。这是个好问题。让我把数字说准确。我想我们的研究团队中大约有三分之一是女性。我的领导团队中超过一半是女性，工程师中有 25%是女性。所以不是 50%，但如果你看看我们在欧洲所做的培训，我们倾向于在女性从事研究或科学主题时很早就让她们退出，我认为这是一场灾难。所以我们尽力通过招聘更多女性来稍微纠正这一点，因为她们通常不会申请，我们在招聘过程中能够做到完全公平。但如果你看看，甚至在大学退出时也存在一些偏见。女性申请高科技公司的比例低于男性。

Oh, well. Thank you, Julie. I'm only allowed to have your question because you are Julie Becker. It's a good question. Let me get my number right. I think we have around a third of our research team are women. More than half of my leadership team are women, and 25% of our engineers are women. So it's not 50%, but if you look at the training we do in Europe, we have a tendency to actually exit the women very early when it comes to working on research or scientific topics, which I think is a disaster. So we try our best to redress that a little bit by hiring more women, because often times they would not apply, and we are able to do that with full equity at the hiring process level. But if you look at there are some biases that exist even at the exit of universities. Women would apply less to high-tech companies than men.

我想你是在谈论申请系统中的偏见。我认为她也是指 AI 中的普遍偏见。

I think you're talking about bias in the application system. I think she also meant bias within AI in general.

是的，当然。所以我们对此有评估。管理偏见是一个棘手的话题，因为模型可以生成很多东西，而且偏见应该在哪里是一个分形边界，但我们大力致力于拥有正确的评估，以便当我们生产模型时，它们能按照我们的要求去做。

Yeah, absolutely. So we have evaluations for that. It's a hard topic on managing biases because models can generate many things and it's a kind of fractal frontier of where the biases should be, but we work heavily on having the right evaluations so that when we produce the models, they are doing what we ask them to do.

谢谢你的精彩问题。谢谢 Arthur 回答这个问题。

Thank you for the wonderful question. Thank you Arthur for answering the question.