Arthur Mensch: Europe needs to own its AI future / H Ευρώπη πρέπει να έχει το δικό της ΑΙ μέλλον.

It is important for states, for Europe, to have its own way of creating models and shaping models with their own values, certainly. The approach we have taken is also quite revolutionary, as we promote a decentralized way of creating technology, which is to bring open-weight models to the world and ensure that everyone can contribute and make them their own. The successful brain of tomorrow will be more creative. We must promote more creativity in all the curricula we create. The biggest impact we can have is to turn businesses and states into leaders in AI adoption and make their businesses competitive.

Well, I guess it probably dates back to when I was six or seven years old. I mean, my mother is a physics professor and my father is a computer scientist. So, I had the opportunity to grow up in a family of scientists. So, I was interested in science from the beginning of my life.

Indeed, I mean, I am much more of a computer scientist and I did much more math than neuroscience in my studies. But I should say that when I had finished the Polytechnic and was looking for a way to develop machine learning, which was the name of artificial intelligence at the time, into specific applications. And one of the applications I was interested in was brain scanning. Using unsupervised techniques to decompose the brain into certain parts. So, you just put people in scanners and keep them still for 15 minutes and they generally tend to fall asleep. And then you look at what their brain does as you do this experiment. And you do it for thousands of people. And thanks to this information, you can actually analyze the different functions of the brain. And then, this can be used as biomarkers for detecting certain diseases and that kind of thing. So, it is one of the areas where you have a very strong application, both from a scientific and a clinical point of view. And where the mathematical problem of scale was very interesting. So, that's how I started. And soon, I moved on to more aspects of natural language processing, etc. So, for me, science is only useful when it has results. And this is also true at Mistral. We create models, but the most important thing is that we turn them into business results. And so I thought about starting my PhD. And I think this is a good way to think about where to start a PhD.

So it is. What is interesting, and I know Demis, he was my employer for a couple of years. And what is interesting is that we are building a different kind of intelligence, in the sense that we haven't the faintest idea how the brain works and how we function. But on the other hand, we have somehow started a parallel intelligence, which is a silicon intelligence, which compresses human knowledge. So we start from human intelligence, because in the end, the data that goes into our models is actually produced by humans. But then we turn it into something that is much better than humans in some areas, but much worse in others. And that is interesting about it, and I would assume that what is interesting is that there is another kind of biological intelligence out there in the world, which is probably true, it is very likely that they thought of their own way of creating another kind of silicon intelligence. And that, yes, from a philosophical point of view, is quite an interesting thought.

It matters in many ways. If you look at the model itself, models are predictors of words. They take context into account and then predict words that you can then turn into organization or images, if you like. As long as they are predictors, they must be trained in a specific way. So, the first thing you do is compress a certain kind of human knowledge. Then you have to choose that human knowledge. Then, the second thing you do is give them some demonstration of humans, real humans performing specific tasks in a specific way. And then, the third thing you do is ask people to compare between multiple outputs of the models. So, at each of these stages, you make conscious choices as an AI scientist about what knowledge to embed in the model, how to guide it towards specific directions, what kind of preference you want to embed in the models. And these are editorial choices. This is not... there is no universal reason why you would prefer one specific way of answering a question over another. There is no reason why, yes, there are multiple questions where you need to make some form of editorial choice. It is not easy, for example, to maintain political neutrality. So, and to think about it, it is almost the job of a journalist to take the model and guide it in specific directions. At Mistral, we call them AI trainers. And they are actually usually journalists. We hire journalists to guide the model in specific directions. And so, of course, the trainers matter a lot. The technology matters. Optimizing the weights of the models to produce the right content matters. But what matters most is the data, and the data is collected from people. So this is very true. And we, just looking at the model itself, it is culturally shaped. So these are content production machines. They also understand the way we access information. Increasingly, we access information through conversation and assistance. And for that, it is important to have a variety of providers, because if you only have a few providers, maybe one or two, suddenly they have great power in shaping the content that everyone will read. So focusing exclusively on the model, and I am not referring to the implementation and adaptation to business needs, etc., there is a strong aspect of sovereignty there. There is a strong aspect of cultural sovereignty. You want... It is important for states, for Europe, to have its own way of creating models and shaping models with their own values, certainly.

Well, it was 2023, we were looking for a name. It is really hard to look for a name. And Timothée came up with Mistral. Timothée lives in the south of France, so that's probably why he had a sensitivity to it. At Mistral, you have two vowels, which are I and A, and in French, artificial intelligence is called Intelligence Artificielle. And that is the I and A of Mistral. That is sort of the internal perspective we give regarding finding the name. Actually, it just sounded right and there were no copyrights for it.

Well, the motivation was at an institutional level, we realized with Guillaume and Timothée that there was a need for a European champion and that this was not a technology that we could... that we had to try it, in the sense that it was a technology we knew how to build. This was a technology that was going to drive the next industrial revolution and, if we calculated the path we took, it seemed that Europe would not have a say in it. So, our view there was that Europe needs a champion. We also need a global champion, who will promote a different way of building frontier intelligence. And that way is through transparency, we have been pioneers in introducing open-source models to the world, so that businesses can adapt them to their needs, so that companies that needed strong control over their data flow etc., could deploy them themselves. So, we had these two really perspectives. First of all, it was the technological one that we firmly believe that AI systems will be built on open-source foundations in the coming years, and we wanted to be pioneers in that. And the second aspect was about the fact that we could find 15 people who were AI scientists in Paris, whom we could take on to bring models to market, quickly and that if we didn't do it, no one would, and then it would be too late. So that's how we thought about the opportunity to create the company. And of course, it is less convenient to build a company than to be an AI scientist at DeepMind. I would say it is not that risky, in the sense that we managed to raise money quickly. We were confident in our ability to succeed. We had no idea what was in store for us, of course. You don't know that when you start a company, suddenly, or in three years, your company will have 500 people and then you have to deal with many new challenges. We had no idea how to build a business. All we knew how to do was build models. But it was a huge learning experience. Certainly, you have to figure things out for yourself. And you don't have many people to rely on. You have investors, etc., but they trust you, so many people trust you at the same time. So, you carry a great weight on your shoulders, but you also have much more freedom of movement and you can learn very quickly. So for me, it was the greatest experience. I wouldn't trade it for any other position in any company.

Yes, similar time projects. They happened simultaneously.

My daughter was born nine months after the creation of Mistral, so it was very much like a similar time project. It takes up the rest of my time, I would say. But it is wonderful to see the two babies grow.

I have a great team. They are there... they are very effective and much greater people than me in many areas of business. So I am grateful to have leverage and to have been able to hire excellent people. That is the key. And then I run a bit to maintain my good health, indeed.

Of course, the revolutionary aspect of it is high. We left the established players to create a company that would compete with them and steal a part of their market. So, inherently, in building Mistral, there was a revolutionary piece. Now, the approach we have taken is also quite revolutionary, as we promote a decentralized way of creating technology, which is to bring open-weight models to the world and ensure that everyone can contribute and make them their own. We believe it is important, because this is a way to prevent one or two companies from having too much influence on content creation and the functioning of the entire economy. It also means we had to build a different business model, because our business model is much more about creating value than centrally controlling the platform itself. So, we build models, but then we help companies deploy these models and we help them adapt them to their needs. And it is largely a way to put pressure on the established players as well. Because their positioning is, of course, to say: we are building a model that no one can compete with us on, because we are building the best models on Earth and we will essentially be the winner-takes-all in this market. And our goal was to show that there will not be a winner-takes-all in this market, because we have managed to always be about three, four months ahead or behind the competition with open-source technology. And so it will become, of course, the story is not fully written yet and it remains to be seen who will be successful or not. We believe there will be centralized model providers and decentralized model providers in the future for different kinds of applications. But I think we support an alternative and we believe we need an alternative particularly in Europe. And everywhere on Earth, we believe that businesses should consider alternatives in terms of the AI digital service provider because in the future they will need leverage. It is not economically viable to run your company with a single AI provider, who will try to extract value, if they can.

Absolutely. Yes, it was a learning journey for us. When we started, all we knew how to do was make models. So, we didn't really know how to change, to turn technology into something that would have results for businesses. And what we did was start with designers and look at their problems and how generative AI could be used to solve them. And we realized that, first of all, they needed more tools. So, they needed tools for inference and tools for training, tools for organizing, connecting AI models with their data and content, and tools to operate, run workflows, connect them with agents and such things. So, we created the organization... the organization and analytics layer on top of the models that we initially knew how to make. And what we also realized is that businesses needed a little push, as the problem of creating new software and new applications with AI models is that it is a different kind of software engineering. You have to think much more about the data. You have to think about iterations. These are organic systems. They behave like humans, more or less, the AI models. Therefore, you have to treat them as such. Therefore, you have to train them. You have to, every month, do an iteration to handle the edge cases of the process that the AI runs and turn the next generation, the next iteration, into something that has fewer edge cases. So, the technique we have created and the product we have created focuses on the return from the results to how the technology will handle it and we found that the combination of products, from the model to the front-end layers to back-end organization layers, with specialists, so, AI for deployment engineers or for deployment scientists who have PhDs and can really think about what kind of new model to create based on the problem, we found that this combination was necessary to get businesses to reach value. Because otherwise many times they would stop just short of reaching production. They would make many prototypes, but there is a growing frustration from business leaders because they have been told for the last three years about generative AI that it is a revolution, but they see nothing on their balance sheet. They don't see anything in revenue. They don't see much in net profit. And we want to help them change that. And we believe that if we don't change that fast enough, there is a risk that suddenly we realize that the investment that has been made is too high. And so we balance the investment as well. So we invest a lot. We just raised 1.7 billion euros with ASML, which is one of our big customers. We invest in improving models, but we also invest a lot in the success of businesses. And this is our way of trying to solve the disconnect that we see in some markets... There is a disconnect in some markets.