Good morning everyone. This morning we're going to talk about a new partnership in the aerospace sector, a partnership set to reshape the industry: Airbus and Mistral AI. An alliance announced at the end of May, precisely on May 28th, and already being called historic. I don't think I'm wrong in saying that. On one side, we have the European aircraft manufacturer Airbus, which you all obviously know—the world's leading aerospace group, a symbol of European industrial excellence in civil aviation, defense, and space. On the other side, a company you already know very well: Mistral AI, the French champion of artificial intelligence, which in about three years has become the indispensable player in this technology in France and Europe. Their ambition—I won't state it myself, of course, they will go into detail—but we can well imagine it's about revolutionizing how all Airbus products are designed, produced, and flown through sovereign and frugal AI. I summarize in broad strokes, and with Antoine Defond obviously the desire to keep Europe at the forefront in terms of industry and artificial intelligence in an increasingly complex geopolitical environment. Rest assured, I'll stop talking now. To discuss this exceptional, promise-filled collaboration, we have today the two main parties involved: Guillaume Faury, Executive President of Airbus, and Arthur Mensch, Co-founder and CEO of Mistral. Please applaud them. [Applause]
Very well, I'll start. Can you hear me? Yes, it's good. So, generative artificial intelligence—what we are building—what we see is that it's starting to have a very significant impact across the entire industry, in the French sense of the term, which the English call manufacturing. It has an impact because it changes the way work is done in operations: deployment, industrialization, and customer operations. That's the first thing. I'd say it's the easiest and quickest to implement. It has an impact on the entire R&D and design aspect: conception of new systems, improvement in understanding the physics that applies to these systems, and the use of language models to reason and find new innovation opportunities, etc. That's a second category which is, for us, certainly one of the most exciting because it accelerates technological progress, allows for better systems, and it's particularly true for aircraft. Then the third category—and I'll come back to the second—it requires not only language models but also models that understand physics and have a quantitative representation of systems, and the interaction between the two is what allows progress to be accelerated. And finally, the third category is the deployment of artificial intelligence within the systems themselves, and particularly in aircraft, to improve the situational awareness of pilots, to enhance safety especially on the ground, and to generally improve safety everywhere as well as efficiency by aiding flight controls, etc. So, it's really about deployment in systems, embedded AI, which also has implications in defense. That's the third category: operations, design, and deployment in systems.
I wanted to allow—I'll give you the floor right away—but to bounce off a phrase you said recently: 'AI shifts the intelligence of work towards capital and machines.' I think that's a key phrase. What did you mean by that precisely, and what will be the consequences of this observation?
So, generally speaking, there is a part of knowledge work that silicon is today capable of doing. In fact, I think that's a new reality that's been here for nearly three to four years, which means that when you're a software engineer today, a large part of software development can be delegated to artificial intelligence systems. And so the developer, the human, becomes more and more a delegator to systems that are silicon-based, which are systems that are expensive, require capital expenditure, and need to run by transforming electrons into tokens, which are the economic unit of this domain. And so what this means is that a part of value creation is shifting from the human brain to data centers. So that's a reality—it allows for more growth by innovating more, finding better systems, discovering new technologies, etc. It allows for greater efficiency in companies. It has macroeconomic effects we can return to, but simply, it means that employees retrain to use the technology and become better delegators using new tools. It also means that this is a technology with significant impact and of great scale—we can expect several trillions in the next five years. And that also means it's a technology that will certainly fragment across different economic entities, because just as energy is strategically important for countries and societies, intelligence will need to be secured in terms of supply, must be cheap because for it to have value, it must also be cheap—cases where we choose to use artificial intelligence instead of humans, it should be cheaper than humans. And it should also be sustainable because it fundamentally transforms electricity into intelligence, so the carbon cost of that electricity matters a lot.
You've opened up many doors. We'll delve into these different points during the roundtable. Guillaume Faury, same question, still broad for now: What will AI change in industry, in aerospace?
So, first of all, hello everyone. It's a great pleasure to be here and to be here with Arthur. I think what Arthur just explained shows that AI brings about a great number of upheavals and creates many opportunities in numerous fields that are our fields—whether it's employee efficiency in their current missions, processes where we have the opportunity to review, change, and greatly improve many processes, including production processes, and obviously the products themselves. So today, we're not changing sectors; we want to remain in aerospace, space, and defense, which are our sectors of excellence. But we believe that to remain excellent, to become excellent in certain parts of the game in these domains, it is absolutely essential to seize the different opportunities, and the one that appears most important right now is artificial intelligence. We've been doing AI for a long time, actually—more traditional AI with neural networks historically. I might say a lot of nonsense since it's not my area of expertise, but after machine learning and deep learning, we've done things on products, but it's true that generative AI has created an enormous field of opportunity that we're trying to seize. So to remain the best possible in our domains of excellence, there are already many technologies we need to master or work with partners who master these technologies. And here, we see that AI permeates, much like I'd say computing completely changed everything about twenty or thirty years ago. The arrival of AI is changing everything. To be able to do AI, it seems obvious, but you need data. And it's true that at Airbus, we have products that are older for a certain number of them—the A320, which is the most successful product in commercial aviation history. On February 22nd next year, we'll celebrate the 40th anniversary of its first flight. So we have to combine platforms that were designed a very long time ago, production tools built and designed a very long time ago, with the speed at which AI allows things to be improved. And so we started about ten years ago a digitalization of all our processes, products, production systems, and we have other partnerships to carry out this digitalization, as well as partners with which we also do AI. What's very special about the partnership we've established with Mistral and with Arthur's teams is several things. First, it's a great team, just to say it. Second, what you're trying to do aligns exactly with what we are and where we want to go—industrial AI, I'm not sure how to summarize it all, but a bit like you explained, the opportunity to go into products, into our core processes, and into production. And these are very sovereign topics. So we were looking for a partner capable of meeting these different criteria of sovereignty, of helping us with industrial AI—not just office AI or better equipping our employees themselves to be much more efficient in daily tasks, but changing our processes, using all the data we generate in a much more significant way than through humans, and preparing for the entry of AI into civil and military products. So this partnership is very important for us. It's a partnership we've finally decided to establish over 5 years, and 5 years in AI isn't eternity, but not far. We're trusting more than just one or two projects with the Mistral partnership; we're trusting the team, and also the desire you have to work with us to achieve successes in certain areas. We might not succeed in all projects we've chosen to work on, but I'm convinced there are some that will really make a difference. And so here, we'll be in processes and in products. And once again, the sovereign dimension is very important. We wanted people very close to us, very European, with whom we could work in complete trust.
So, I don't think I'm revealing a secret by saying that you've been working together for about at least a year now. We can start going into the details of this agreement, having concrete examples, use cases that have been tested. Are there already good feedback from the field?
Yes, we're working together and we've known each other now for—I don't know if it's been a year or perhaps even a bit more. Now, we have a joint work plan, and you're putting a lot of human and financial resources. We're also providing people with a certain number of objectives. Well, I don't want to reveal all the ideas we have because it has a bit of a commercial, confidential dimension—I don't know how you say that in French. Simply, there are topics that are fairly obvious. And typically, when we develop an aircraft and put it into service, we need to generate from design data, production data, and an enormous amount of documents, whether digital or paper—it's documentation. We need to do this for documentation and data, maintenance manuals. So, production, maintenance—we have a lot of work to do in the certification domain, and actually we reinvent everything each time, even though the process is relatively similar from one to another, but the data we use are different. So, there's a need for learning like humans do. And if we can make this learning happen from digital data and artificial intelligence tools, we can go much faster, be much more efficient, and also make far fewer errors because when we do all this with humans today, there are many errors, with cycles of rework, verification, correction, and it requires enormous time and money. We're trying to go much faster. So typically, these are activities where we must be able to, with experts—because in the end, humans will validate the documentation, ensure that what's in it works. We'll run tests to ensure it works, whether in production, maintenance, or certification—we have a certifier, but we must be able to go much faster. So that's typically what we want to do. There are others in products. Everything in products, we'll rather keep for ourselves. Some will work, others won't. And we've seen, in the year or two we've worked together, that you've already advanced strongly with certain ideas. We had the chance to spend a bit of time together recently, and there's already much potential on product ideas.
Arthur Mensch, how do you organize very concretely to work with a group like Airbus in terms of team and model?
It's not too hard, right? No, it's fine. So, first of all, thank you Guillaume for inviting me here as well, and thank you to CMA Méen for inviting me, and thank you for your trust in this partnership. I'd say the way we organize is that we are a technology company. So, we have a platform that is portable, meaning it can be deployed on our clients' infrastructure so that critical data and critical intellectual property remain confined in the client environment. This is what we can broadly call a sovereignty issue, which is a sovereignty issue for states, but is also a sovereignty issue for all companies that have a lot of intellectual property and knowledge accumulated over decades. So, our technology is deployable on client infrastructure. And what does it do? On one hand, it allows for generating tokens—so it allows systems to generate text and actions that will understand multimodal inputs. It can be text, plans, logs, or what we call the bill of materials. And from that, the models are connected to different data sources from our clients, particularly Airbus. And from there, we can give them tasks to do, and they will generate documentation, for example. They'll generate documentation, present it to humans, and in interaction loops, ensure that this kind of documentation is delivered faster. So, we provide all the infrastructure to do that: models, data connection, interfaces, hosting of interfaces—all in a portable platform. That's the technological part. What we've observed in the three years since we've operated is that what worked very well with our clients was also deploying teams that spend 100% of their time with experts in the subjects we're discussing. We're not experts in aerospace at all, but we know what artificial intelligence agents can and cannot do. And we know when we reach the limit of what they can do, how to do software engineering to go beyond that limit. And so, we spend time with our deployed teams choosing a certain number of topics with Airbus, and over several quarters, we work hand-in-hand with the business teams to deliver value and create use cases that go from the operational part to the really business part, where we build models, train models that understand physics, understand plans, and are really specific to the clients we work with. So, that's how we work, and it's indeed long-term partnerships because what we see with artificial intelligence is that it's possible to have very quick gains, especially on operational topics, in one or two quarters. However, on design topics, for example, there are additional efforts to make because you need to review all processes and change a lot the way you work. I was told at Airbus that they call them 'pizza teams.' Catherine Gestin, Digital Director, told me: three Mistral experts working with three Airbus experts and three business experts. Catherine Gestin also gave some very concrete examples. So, I think I'm not revealing anything confidential, but it's important to understand concretely where AI can improve performance.
So, I'll say quickly: support for technical documentation, assistance to helicopter pilots, assistance in communication between the pilot and the control tower—which is a constant concern—product development obviously, predictive maintenance, more automated factories. We'll get to that, I think it's also an important point. What is the AI... can we already say in terms of productivity gain, in terms of product performance gain, have we evaluated its contribution? Do we have that perspective already, or is it still complicated to know? Who wants to take the question?
I'll maybe answer. I think the honest answer is that today, we don't know, and right now, we're spending money. We don't have productivity gains yet; we're investing in a new technology for which we see many possible use cases. We've amused ourselves, so to speak, since the arrival of ChatGPT with generative AI. We tried many use cases and did them, and I think we put in more money than we gained. And we got quite dispersed. So now we're trying to organize to put energy, time, and money on topics with the biggest potential—return a bit to fundamentals.
But in terms of performance and productivity gains, do you have that perspective?
I think the productivity potentials—I don't even know if we can call it productivity—the efficiency or gain potentials are absolutely enormous. And we're at the very beginning of an adventure. So, I think we don't know how far this will go. What we see is that we have use cases that seem clear, where if we can crack the use of AI on those cases, we'll save a lot of time and money. But that's also the first or second step before things that will be much bigger, much more disruptive. We see physical AI is just starting. There are people who think it won't be through LLMs that we'll get there. So there are even new technologies within AI that are developing. I think today everyone is very humble about what it will really be able to do. Even the experts you are, I believe, are sometimes very surprised by the potential of AI and what it manages to do. So, we who are far from AI professions and use AI, we're also very cautious. So, there you go, I think today we don't know. There have been technologies that fundamentally changed our way of working and our efficiency and speed. This has the potential to be even more revolutionary than others, provided we do all the rest well. It doesn't replace the rest; it adds to it.
Yes, absolutely. So, I'd say there are domains where we see gains and returns right now, and especially on customer operation topics, customer support topics, maintenance of deployed machines, for example. I think we've been working with ASML for a bit longer than with Airbus, and we see on the deployment of their machines that we can improve uptime and similar things. And so on customer operations, since it's easier because the topic is really about being able to look at the logs that machines produce and use them to make decisions—we see gains there. We also see gains in banking, in financial services because it's more virtual, so you have fewer barriers. The physical aspect is always a bit of a barrier because it makes interaction loops longer. And so in knowledge domains that are completely virtualized, like banking, we see gains, especially on customer operations as well as efficiency on markets, things like that. Then, when we go to design problems and embedded deployment problems, there we're on more innovative topics where the opportunity is the same as we see today in software development. In software development, it's very clear—there are enormous productivity gains, and that comes from the fact that it's completely virtual, and you have an interaction loop between the developer and the tester that in fact happens in 5 to 10 minutes because you have continuous integration telling the agent it works or doesn't work, and the agent can make decisions accordingly. When you do aerospace design, for example, you need to test your ideas, and there it's rather simulation that becomes much longer than code tests and can become the bottleneck. And so, to really have gains, our bet is to have language models but also models that understand physics, which will allow for much faster simulation.
What we call physical AI, is that it?
Physical AI can mean many things because you have AI in robots, you have AI in aircraft, and you have AI that understands physics. So you have embedded AI and artificial intelligence that understands physics. In general, when you embed artificial intelligence, you like it to understand physics because it will be connected to sensors that are themselves embedded. But so, our bet, and we've invested a lot in this in recent months, is on smaller models that are quantitative, that we can train with our clients because generally you need access to historical data, and they will model physics and interact with language models, which rather mimic human reasoning, the reasoning of engineers, in fact.
Gentlemen, it's impossible to ignore and not talk about defense. The subject is sensitive obviously, but we cannot avoid discussing it. Arthur Mensch, you're already working, you've signed a partnership with Lamiade, the ministerial agency for defense artificial intelligence. What are you doing with them? What can be said about this partnership?
So, we can say—I can say few things, as often in defense. Simply, we work with operational forces in much the same way as when we work with our clients in the civilian sector. You have to start from use cases and the concrete reality of processes where generative artificial intelligence can help. And what we see is there are really two categories: on one hand, embedding artificial intelligence in systems to give them more autonomy and especially to face systems that are themselves more autonomous, adversarial systems that are more autonomous. And the other topic is the topic of processing heterogeneous data coming from different locations in the field of...
bataille et qui doivent être agrégés de sorte à fournir de l'aide opérationnelle à la décision. Donc ça, c'est des sujets qu'on explore, ça s'explique pas de manière abstraite. C'est-à-dire que la meilleure manière, c'est d'aller travailler avec les forces armées, comprendre comment elles s'engagent, comprendre comment elles travaillent et faire des cas d'utilisation et extraire de ces cas d'utilisation la plateforme qu'il faut pour pouvoir la répliquer à différents endroits.
Et vous dans le domaine militaire, vous avez un rôle de conseil, disons ça. Ce que vous disiez, je crois lors de votre audition à l'Assemblée nationale, mais pas de droit de regard pour résumer simplement.
