2024-08-07
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
Editor’s Notehave“European OpenAI”Mistral AI, known as the "Mistral AI", has a valuation of $6 billion just one year after its establishment.
Since its establishment, Mistral AI has frequently made efforts in the open source field, and recently released their new generation flagship model Mistral Large 2, with less than one-third of the parameters, it achieves performance comparable to Llama 3.1 405B.
Recently, Arthur Mensch, co-founder and CEO of Mistral AI, discussed in detail in an interview with Time magazine. How Mistral AI attracts scarce artificial intelligence (AI) talent, how to achieve profitability, and the lack of an AI ecosystem in Europe。
The core ideas are as follows:
Image: Arthur Mensch, co-founder and CEO of Mistral AI
Academic Headlines has made a simple translation without changing the main idea of the original text. The content is as follows:
Over the past year, Paris-based Mistral AI has rapidly emerged as one of Europe’s most influential AI companies. The startup has released six language models that can answer questions, generate code, and perform basic reasoning.
In June, Mistral AI said it had raised $645 million in a funding round that valued it at more than $6 billion. In February, they reached an agreement with Microsoft to provide their models to the latter's customers in exchange for access to Microsoft's computing resources.
Mistral AI co-founder and CEO Arthur Mensch argues during the debate on the landmark AI ActRather than regulating underlying models like Mistral, lawmakers should focus on regulating how others use these models.He also opposes restrictions on AI developers’ freedom to share their creations. “I don’t see any risks in the open source model,” he said. “I only see benefits.”
TIME spoke with Mensch about how to attract scarce AI talent, how Mistral AI can become profitable, and the lack of an AI ecosystem in Europe.
Q: A few months ago, your chief business officer Florian Bressand told CNBC that more than half of the Llama R&D team now works at Mistral. How did you attract so many talented researchers from Meta?
Initially, we recruited our own friends. We were able to do this because we had made some meaningful contributions to the field, so people knew that we were interesting to work with. Then, starting in December, we started hiring people we were not so familiar with. This is due to the strategy we followed, which is to push the field in a more open direction. This is also the mission of many scientists who, for similar reasons to us, prefer the old way of free communication and information flow.
Q: There are very few people in the world who can train an AI system like Mistral. I know that France has a well-developed AI industry, but do you think you have managed to recruit a considerable number (or even all) of the people who understand AI?
Not everyone. We had a lot of friends in the industry who were at Google, OpenAI, and a few people stayed at Meta. But certainly we attracted 15 people who knew how to train these models. It’s hard to estimate the size of the talent pool, but I think it was probably 10% of the people who knew how to work on this stuff at the time.
Q: Mistral AI has been raising funds. What have you spent the money on?
We spent our money mainly on computing. The structure of this industry is different from that of the software industry because you need to invest a lot of money at the beginning to build a scientific research team and build cutting-edge models.
Q: Executives at nearly every other base model company have talked about their expectations of spending $100 billion on computing over the next few years. Do you share similar expectations?
We have burned through around €25 million in the last 12 months to get to where we are today - with a global presence and models that are leaders in performance and efficiency. Our thesis is that we can use capital more efficiently, that the technology we are developing is actually capital intensive, but with good ideas, we can do it with less expenditure than our competitors. We have proven this in 2023-2024 and expect to continue to do so in 2024-2025. Obviously, we will spend more. But we will still spend a fraction of what our competitors spend.
Q: Are you currently profitable?
Not yet. Our investment is quite large and profitability is not something that is expected for a 12-month-old startup.
Q: What is your profit plan? What is your business model?
Our business model is to build cutting-edge models and make them available to developers. We are building a developer platform that enables developers to customize AI models and develop differentiated AI applications - they can deploy the technology where they want to deploy it, so it is possible not to use public cloud services, which allows them to customize the model instead of using the general model behind a closed opaque API as it is now. Finally, we also pay great attention to the efficiency of the model, so we can achieve a certain inference capability to make the model as fast and as cheap as possible.
That's what we're building: a developer platform that we host ourselves and then offer to our customers through APIs and managed services. But we also deploy the platform for customers who want full control of the technology, so we can give them access to the software and full control over the data used in their applications.
Q: Is it fair to say that your plan is to make AI models that are nearly as good as your competitors’ at a lower cost to you and your customers, and to make them more publicly available? Or is it that you want to match your competitors’ most advanced models, or “frontier models,” in terms of absolute capability?
We plan to continue to catch up and eventually be as competitive as other companies. But in reality, our business model is different from other companies. We prefer to share, customize and deploy our technology. We no longer have control over these aspects.
Q: You recently made your most powerful models available as APIs, whereas all your models were open to the public at the beginning. Why did you make this change?
We haven't changed on that. We've always wanted to have the leading model in open source, but also have some advanced features that are only available through monetized services.
A large portion of what we offer is open source, which enables developers to take our technology and build whatever they need with it. Ultimately, when you want to move the workloads they've built into production, or you want to make them better, more efficient, better managed, and cheaper to maintain, these developers will come and use our platform and use our underlying optimization models to improve the performance and speed of inference capabilities.
We're going to keep doing this. Open source is very important to us. We're building a developer platform on top of it, which will obviously be monetized because we do need a business model that works. But we hope to bring additional value to developers who use our open source model.
Q: You often say that Europe cannot rely on American AI companies and needs a homegrown cutting-edge model. Mistral AI is one of the most prominent AI companies in Europe, but it has a partnership with Microsoft to get the computing power it needs. Does Mistral AI's reliance on Microsoft in this regard limit its ability to act as a cutting-edge sovereign AI player?
We have four cloud providers. We are cloud independent by design, and that has been our strategy since day one. Our models are available through Microsoft Azure, but also through Amazon Web Services and Google Cloud Platform. We use all three as cloud providers. We also use different cloud providers, especially CoreWeave, for training. We built our own technology stack and distribution channels to create the independence that we think our customers need.
Q: In addition to establishing AI labs in Europe, should Europe also try to build its own sovereign computing infrastructure?
I think it will be good for the ecosystem. But Europe is not an independent actor that makes decisions and builds something out of thin air. It's an ecosystem question of how to ensure that Europe can effectively provide some computing infrastructure.
This is very important for our customers because some of them are European customers and they do want some form of sovereignty over the cloud infrastructure that they use. In this regard, some of the accessibility of our model, reasoning and platform is actually already deployed in Europe. But there can be some improvements. This is not determined by Europe. It's an ecosystem that needs to recognize that there are some needs that can be solved. We hope to have some European cloud computing partners in the near future.
Q: Cedric O, former French digital affairs minister and one of your co-founders, warned that the AI Act could “kill” Mistral AI. The bill has been passed, but codes of conduct for general-purpose AI models have yet to be developed. What should they look like?
Generally speaking, the AI Act is very doable because the constraints we're subject to are constraints that we already meet. We've already documented how we use models, how we evaluate models, and that's become a requirement for leading-edge models. So it's OK to do that.
We still have some discussions to have about transparency in our training datasets, which is something we very much want to achieve, but it needs to be weighed against trade secrets. A lot of our intellectual property is also in the way we process data and select data. This is also the intellectual property of others. As a small company, we are very careful about our intellectual property because it is the only thing we have. So from that perspective, we are confident that we can find a way that all parties can accept.
We are asked to participate and provide input into the development of technical specifications. We also want Europe to make independent choices that will promote the development of the ecosystem and keep everyone happy.
Q: There’s been a lot of talk from executives at your competitors about how AI will change the world in the next five or ten years, and what they’re worried about, and the kinds of changes that might come with the development of more powerful AI systems. Do you have any predictions about how AI will change the world?
We built a powerful technology, but I think there is a tendency to assume that this powerful technology can solve all problems. At Mistral AI, we are very focused on making sure that our technology can improve productivity, bring reasoning capabilities to certain verticals, certain areas, and thus generate societal benefits.
Everything humans create is a tool, and the new tools we bring bring new abstract capabilities. So in a sense, you can think of it as a more abstract programming language. We've been programming in languages that computers can understand for 50 years. Now we can create systems just by talking to the system in English, French, or whatever language. This brings a new way of abstraction for workers and developers, which obviously changes the way we work in the next 10 years.
I think if we do it right and make sure that everybody has this tool in their hands - and that's really why we created Mistral - we can make sure that it improves the lives of everybody around the world, across socioeconomic groups. And to do that, for us, it starts with differentiated applications in areas like health, education, and so on. It's also very important to make sure that people are trained and have access to technology, and that people have access to that technology - making it available in a more open way than otherwise is a way to accelerate the development of technology. That's not enough, political decision-makers also have to put in place support programs to accelerate internet access in parts of the world that don't yet have internet access. But I think the new tool that we're developing - generative AI - has a positive role to play in helping people use this new tool.
Q: Can you imagine a scenario in the future where you've developed an AI model, or you're developing a model, and you notice something about its capabilities, where you decide it's better not to open source the model, but keep it behind an API, or not even deploy it behind an API?
We won't be in that situation for the foreseeable future. The models we build have predictive capabilities. We've found that the only way to collectively manage software and how it's used is to open source. That's true for cybersecurity. That's true for operating systems. So the most secure technology today is open source technology.
In a sense, AI doesn't change anything about software. It's just a more abstract way of defining software. So I don't see any risk in the open source model. I only see benefits. It's a neutral tool that can be used to do anything. We didn't ban C because you can make malware in C. The models we release are no different. So it's still very important to control the quality of the applications that are put on the market. But the technology used to build these applications is not the only thing that can be regulated.
Original author: Will Henshall
Collect and organize the latest news information and make it available for free for everyone to read
