WPC 2026 - Workshop 3: AI - Power Struggles and Societal Challenges

Bonjour. Bravo. Coucou toi. Bravo. >> You want to have a minute? Hello. Hello. Good afternoon. Good afternoon. Switch to uh in English. >> So I made a quick and brief introduction. Personally, I said where I'm keen to welcome you here. I have spent 20 years at Capgemini. I was deputy CEO. So I spent a lot of time in that place and I saw it built and so I hope you enjoy it. And I welcome you in here for this workshop. AI power struggles and societal challenges. I proposed this title to the World Policy Conference because every time we have technology disruption, you have the usual dystopian utopian conversation, which I found profoundly useless because at the end of the day you have both. So no need to argue about if it's one or the other. We will end up with both. And I think in the case of AI, that's a little bit where I would like to drive the conversation with your help. We are we have a specific challenges. It goes extremely fast and in an extremely powerful way. There are hundreds of billions deployed at the speed of light and it rock the natural order of the way we tend to address technology disruption. So that's the topic and then being the World Policy Conference, the idea is could we come up with some one or two bright ideas that could help steer the course that we will debate this afternoon. So that's the the scope. To address it, I have a panel of guests highly qualified in in the domain that will make contribution after my introduction. You have Daniel Andler who is professor at the Sorbonne, member of the Academy of Moral and Political He used to be a mathematician turned philosopher. You decide if it's good or bad. I think it's very good for him and for us. Of course about who is the founder and chairman of FDB partners but also a known in France as chairman of ID8 DG World who runs. So I used to say every time at the World Policy Conference I can share everywhere else he's sharing the debates this kind of debate. So I mean everyone is done. We have with us Christine Schwarer, French but working head of data science and artificial intelligence of the Swiss Federal Statistics here. So I'm Swiss so that's the connection. Then we have Toby Simon founder and president of Synapse here. This is a think tank in Bangalore specialized in trilateral conversation and cyber security. We know each other for for a while and met in India many times. Then we have Arthur Snell who used to be a young leader on the panel presenting and now he's in charge of a senior role at Select this and he's in biotechnology and he will highlight the usage of the AI and what you can do with it and the positive less or positive. Then we have Pauline Thompson joining us from Ardian managing director of the infrastructure fund and of data science. Also infrastructure is a separate and big part of the Ardian activities of private equity as probably you heard yesterday from your founder. And then Jay Truesdale coming from Washington DC so with the US perspective here CEO of TD International which is a intelligence entity and they will tell us how this changing the the way it works but also what he can do with the outcome of his work. So, that would be the introduction. So, uh to introduce the topic uh and then give the word to my panelists as I rapidly said in introduction, uh it's not the first technology disruption, but this one uh if we remember ChatGPT, which was the first generative AI, I don't think it needs to be explained anymore, is November 22. And when you look at today, where we are, seriously, in 3 years, uh adoption will take time, but the breakthrough, because we are human beings, but the breakthrough uh in terms of technology and adoption and the the speed at which the ecosystem is built, because it in order to succeed in technology, if you remember Wintel at the time, which was Windows and Intel x86, they build an ecosystem, and then you dominate your area. And I I remember having been at Nvidia uh 10, 12 years ago, and they were already talking to me about we will get ready for when AI comes. So, they are not here by accident. Nvidia started by doing graphic process unit for gaming, then they work a lot with a French company uh which was uh Dassault Systèmes for the digital twins and the visualization for the development, but they were already preparing. And when you see part of the investment made, what he's made, what Jensen Huang, who I not yet met, but I hear he's a very sought-after man now, um he is building the ecosystem. So, he's locking the entire technology agenda uh around the Nvidia processor. And once he has done it, he will make money. So, that's for sure, he's already doing, but you lock it. And that's why you see all of this. And then we will hear later from uh Pauline the impact on energy, which is something we have not seen before in technology. We had If you remember, a lot of debate about uh the network when the internet came. We said, "Oh, there won't be enough bandwidth." At that time, I was at the Redmond at the the headquarters of Microsoft, and they were telling me, "Well, Well, I said, "Why do you do Azure?" They said, "Uh we have 70,000 km of dark fiber, and we don't know what to do with it." So, I say it's twice around the world. So, it's a network. So, then they deployed it to use. And that's how they could enter the market because they could enter the market at marginal cost zero. And it's a good starting point if you want to enter a new technology field. So, that's what we see now. That's the dynamic. And uh if you look at the numbers already what has been achieved so far, I have a few numbers, and colleagues will share others. Uh we since the deployment of AI today, and I will share these numbers after the workshop with everybody's interesting. We we do 35 times faster climate modeling. It's It's amazing. Uh we do 79 times gene editing efficiency. And one example is Demis Hassabis who made his Nobel Prize in chemistry based on the his contribution to not gene editing, but protein discovery. And then, uh there are big debate on uh artificial general intelligence, and probably we'll come back on this during the debate. Now, uh the economic impact, you have different uh you have different views on this. And what we see immediately is the impact on on job, for instance. Uh we estimate 55,000 job losses attributed to AI in 2025 Uh and an additional 78k 78,000 as indirect. So, we see and then there is the trend that you see with physical AI notably but not only towards unqualified job and this create another diversity challenge even if it's not politically correct if politics still exist. Yes, but politically correct with DII women are particularly impacted because they have a lot of unskilled job that are threatened by this. There is a further divide further divides. One is between the haves and the haves not. 20% of the company takes 75% of the profit of the of AI. So, given the exponential nature of AI, this creates a problem. So, the more powerful already will get even more powerful. You see it on the stock market. You see it in the business everywhere. They dominate and there is also of course the battle and we'll discuss it the haves and the haves not in terms of access to the technology. So, the global south is again impacted. You add this to the trade policy and that impacts also the global south and you add the AI. So, it's another accelerator in the divide. So, it's not solving the problem. Related to this, you have different type of answers. So, what we see is that the usual mechanism to control technologies, if you remember, and Arthur can speak better than me about this but when biotechnology emerge, we could establish ethical committees and they have proven to be efficient before we had to do laws. Here we cannot. We run after ourselves or tails. Every initiatives and we can discuss later on the list is not proving successful. You saw there was in France in Paris last year the AI summit with commitment the US and the UK has refused to sign the agreement if you remember well. So it's an open race. So we we struggle to put then there is the regulation as you heard this morning the European response. We regulate before we fight. Then we don't know how to fight but we've regulated. The US has just issued their national policy which is a complete different approach which is a a the federal government will will control. You will tell us Jay what does it mean control? Federal I understand. control and then but for the rest the race is on. China as we know it's a government led. We don't hear much here in Europe but they they are progressing extremely fast and notably they go beyond large language model. If you know that's the big battle we hear now about Autotrope. Gemini enterprise cursor that has just been bought by SpaceX. This is about large language model. So language means this are this is text. This is images. This is video. But then there is what is beyond. We don't learn by reading otherwise we wouldn't be where so we we learn by watching and then you have the large vision model. Yann LeCun is just back from the Silicon Valley in Paris with something that he took from the from there that is in development for 2 3 years in the Silicon Valley called the world model meaning every sensor is a is a source of data and these source of data can be exploited to model new things. So, then you have different type of signals, different type of data that goes beyond the large language model. Uh and you you've seen that a Chinese robot won the world record for semi-marathon. It's It's funny. But it's a big achievement. I mean, to have a machine that run uh you can imagine the impact on the battlefield for instance, all these type of things. That's uh really and it's very complex stuff. It's very because it's a combination not only of software but hardware as well. So, that's a that's the picture we see. So, that what I said at the beginning, utopian dystopian doesn't make really a lot of sense because it is you have both at the same time. I've not mentioned the drones. Perhaps talk about this. This is another application of uh of the of AI as well in the battlefield. So, you have both. You have the development on the protein as I mentioned. The uh the the usual control mechanism don't seem to work right now. And uh I don't see any uh limitation in the deployment of capital because this is geopolitical battle as well uh in the domination through uh technology. So, that's brief introduction. Uh I tried to be a little bit provocative so you can completely disagree with everything I said and that will make the debate interesting. Uh but uh to illustrate and go further and deeper in what I mentioned uh with the panelists, we've agreed to split the intervention uh in two parts. So, uh the first part will be more around what I call the power struggle. So, this race that is going on that we will not stop. Uh for this in the camp of the power struggle, we'll have Bolin who will start on the computer bottleneck and we'll have Christine on data and Dominique on the cybersecurity and Francois on the infrastructure and presenting and highlighting given perspective on what's going on in that field beyond the usual conversation is Europe behind in the race. The answer is yes so we can skip the question and go directly to Colleen. Thank you very much. >> Thank you Patrick and good afternoon everyone. Um so as we were saying artificial intelligence is shaping a new industrial era at an unprecedented unprecedented pace. And if we look at it just as railroads, electricity grids, telecommunication networks that define the previous eras of economic transformation, data centers now are the backbone of this new economy. Uh yet the pace at which we can build them is constrained by several factors which which we will discuss today. So if we start with the basics um when we speak about compute, we need to speak about data centers and energy. Uh so data centers are the physical backbone of the modern digital economy because they house all the servers, the storage systems, the networking equipment that power everything from cloud services to artificial intelligence compute. Uh currently there are over 10,000 data centers globally and yet demand is outstripping supply at an accelerating rate. We can think of it this way, every time we ask ChatGPT a question or stream a video or execute a financial uh transaction, we are relying on a data center somewhere in the world. Uh so these are not abstract digital concepts, they are physical, energy intensive and capital intensive assets. Um on top of that, the data centers themselves are becoming exponentially large facilities. If you think that 5 years ago a 50 megawatt data center was considered a very large site. Now 50 megawatt would be too small for most of the hyperscalers to even consider as a potential data centers. And more and more 1 gigawatt scale campuses are announced throughout the world every day. So why can't we build more and build quicker? Because we face several critical bottlenecks in particular on the energy side. So first grid interconnection queues are longer and longer. New data centers require several several years just to obtain grid connection. In some European markets the wait can exceed 5 years. Second, transformer and equipment supply. High voltage transformers take 2 to 3 years to manufacture and are in global shortage. So this is really a physical supply chain constraint that money alone cannot solve overnight. Third, permitting and planning. There is increasing speculation on what we call now powered land. So which means a real estate site that has power secured for x amount of megawatts. And because of these powered lands have become real an asset by themselves that is traded as very high premiums. And this speculation on powered land is straining planning and regulatory systems which creates even more delays and and and approvals. And fourth of course capital and financing. Data centers are extremely capital capital intensive. And the scale of the new campuses is growing exponentially. So each project is a new capital challenge. Now let's just talk quickly about why this what is driving this demand explosion. Uh, for AI compute, we have two driving forces, training and inference. On the training side, uh, as we were just saying just before, the largest, uh, the large language models now contain over 1 trillion parameters, up from 10 billion just in 2020. That's a 100-time increase in just 5 years. Um, and training on top of that is diversifying from text from purely text uh, data to audio, video, and multimodal systems. So, this is driving, uh, much more compute demand. On the inference side, which is the the running and querying of the models every time we we we query them for our applications, adoption is also scaling rapidly. We are just mentioning that ChatGPT has been the fastest adoption of a new technology in history. Um, this is of importance also when we look at US, China, and Europe because most of the training is now done in the US and China, and Europe is lagging behind in terms of this uh, data center deployments because, apart from Mistral, uh, in fact, all the large language, uh, models are are trained outside of of of of Europe, at least the leading ones. Um, so data center deployments now in Europe account for roughly 15% of worldwide data center capacity installed, um, which is something that I think we will cover but Europe is finally uh, showing the will to catch up, and this is very important as we as we move to the the inference phase. And we if we look at it a bit closer, by 2029, inference workloads are projected to surpass training and will account for 53% of global AI-driven power absorption, reaching 75% by 2031. So, this means that the long-term demand driver is not just building the models, it's also the billions of people and businesses that are using them every day. Um and if we look at the the efficiency gains that for example deep sick models like deep sick uh have have brought, uh it seems that we are in in a textbook case of Jevons paradox because when efficiency reduces cost, the usage increases even faster. Uh and compute demand is growing 4.5 time per year while chip efficiency improves at roughly roughly two times every 18 to 24 months. So, the gap is widening, it's not really uh closing. If we take a look at what the the the hyperscalers are doing, so the hyperscalers are Amazon, Microsoft, Alphabet, Meta, Oracle. They are responding to this dynamic with unprecedented uh capital deployment. Combined CAPEX across these five companies was 127 billion uh in 2021. It's now uh estimated to be at 597 uh billion by the end of the year, uh which will be uh nearly five times the 2021 level. These companies are invested at a pace that uh and and scale that has no precedent in fact in corporate history. Of course, the question on everyone's mind is will AI revenue justify this uh this level of annual infrastructure spending? Um but what's also very interesting to note is that in fact these players have unprecedented levels of cash available, so they can in fact sustain the the race. And maybe at the exception of Oracle, they have so far um been able to finance this CAPEX with very relatively a very um low uh recourse to debt. Um now if we if we look at the the power density that is required for AI compute, we can see that the infrastructure challenge is compounded by the dramatic increase in power density required for the AI chips. Each new generation of Nvidia chips demand significantly more power per rack. The Hopper generation, so the H1 H100 H200, that was the the most common AI chip even in 2022, um required 75 kW per rack. The Blackwell and Ultra now require uh 125 kW per rack. Rubin will exceed 150. And Rubin Ultra, which is expected by 2027, will demand 600 kW per rack. So, that's eight time a multiple of eight in 7 years. And this is not just an incremental change because it requires a complete redesign of the data centers. The chips are so um generate so much heat that we now need to cool them by plunging them in what we call liquid cooling. So, it's a it's a complete again rethinking of the infrastructure. Um This is why all the the energy implications of the of these AI computer are completely uh are staggering. The power availability has now become the number one site selection criteria for data centers, displacing fiber, latency, and cost. Um global data center capacity is expected to still increase by roughly three times uh and and and supported by an estimated 5.3 trillion in cumulative capex, uh and most of it, as we see, is driven by AI. One additional constraint is that data centers tend to deploy by regional hubs. So, many regional grids uh grids were not designed to accommodate such concentrated large-scale loads. And if you look at the key markets, such as Northern Virginia, which is the the largest data center hub in the world, or even in Europe, if you look at Dublin, Amsterdam, uh parts of Germany, uh the grid capacity are already completely uh at at a bottleneck. Uh and that leads to, uh first of course, delays, but also uh in some in some cases, incomplete inability to to to build more. Uh if you look at at Ireland, uh data center energy demand is now 20% of the country's demand. It's It's enormous. Um and on top of that, upgrading and expanding the grid infrastructure, so the substations, the transmission lines, the transformer, requires massive capital expenditure and long lead times. So, you're out you're outpaced by the speed at which data center demand is growing. If we look now uh at a very recent forecast that was provided by the International Energy Agency, uh data center electricity grew 17% in 2025 alone. Um and by 2035, we expect that data center electricity consumption would more than triple uh the the 2020 level. The good news is that many data center operators have committed to sourcing 100% uh renewable energy. That's why the the the share of renewable energy on the graph is is quite large. Um but the intermittent nature of wind and solar generation creates challenges, of course, in matching the the base load demand uh with clean supply. So, data centers still need to rely on fossil fuel backup or grid supply uh power. This is also, of course, a huge opportunity for nuclear, and that's why nuclear has is making a a huge comeback. Uh because nuclear is the only source of power that provides low carbon and base load energy. And this is the reason why countries like France uh can benefit from this massive opportunity as a net exporter of low carbon uh energy. Uh And and this is also why uh if we look here um the the the these constraints uh have pushed more and more operators to look at moving their compute where power is cheap and carbon intensity is lower. And this is why markets like the Nordics in Europe are becoming key data center hubs because you have ample supply of energy, low carbon energy, and and and cheaper energy. If you look at the comparison between, for example, Iceland that is relying mostly on geothermal and and hydroelectricity and the UK, uh Iceland costs 68% less than the UK for 1 MW of of compute. Uh and the carbon footprint uh fact is is different is more dramatic with the a factor of 50 times. So, for non-latency-sensitive applications that do not require to be close to the end user, it can be far more efficient to transport data where clean energy is abundant rather than than than transmitting power over long distances. So, as a conclusion, uh what is clear is that we are at the beginning of a multi-trillion-dollar infrastructure build-out that is reshaping completely the energy system, the real the real estate, and the digital infrastructure markets. And we will need uh still large-scale investment across both digital energy infrastructure uh that will and we need to really, I think, bridge the the the divide between technology, infrastructure, uh power generation, and and and sustainability. Thank you for your attention. >> HELLO. SO, UH I DON'T KNOW IF IT WORKS. SO, THANK YOU FOR PAULINE. Very insightful. We decided to start with foundations and because then the all the rest we talk a lot about the applications or a lot about the impact, but how we make it work and how this will make work is very important. I would for the debate two things. A, capex is massive, but Pauline alluded to the cash flow. Nvidia last year 180 billion of free cash flow. Then you see the numbers and you might worry, but their current business generates a lot and if you make a similarity with the cloud, the successful companies have been the one who could do it on the back of pre-existing investment, be it Google, be it Microsoft or Amazon for building their e-commerce infrastructure. And here we see the same. It just the cash flow is massive, but there will be a need for efficiency. So, whoever wants to develop a new business, they you can't continue at that pace. And this trajectory is unworkable. There are physical limits. So, there will be a need for efficiency and we'll discuss it. So, infrastructure is one part and it consumes data. So, I hand over to Christine who will address the Yeah. data side. Okay. >> Can you hear me? Yes. So, thank you Thank you, everyone. Thanks for joining. Thanks for inviting me. And now for something completely different. So, we are going to talk about data, but not only about data, more specifically about uh data from uh from governments and from administrations. So, um I uh will just give you a a brief um introduction to um why we are doing what we are trying to do. So, I uh as Patrick mentioned, I uh I work for the Swiss Federal Administration. And um 6 years ago, the Federal Council decided they would create a center that does data science and artificial intelligence inside the Federal Administration for the public sector in Switzerland. And so, um there was a mandate that was given for the creation of this center, and the vision was um to use data science broadly speaking, so that includes artificial intelligence, as a way to um develop um innovation, but for public good in Switzerland. So, um we um we really um wanted to help with the first principle of uh the United Nations Statistics Division. So, the first principle says the following, um that statistics, so official statistics, numbers that uh are created by by entities such as such as mine, they are an a crucial element for a a democratic society, because they serve the government, they serve the economy, and they serve the public with data and insights that we can actually trust. And so, I will walk you, so there will be one graph, it's just one visual that I will try to uh to walk you through, and um explain that we are trying to tackle two problems in this uh in that that can um that happen when we collect information and when we distribute data as um as a government. So um when you are um and the the key point being that we want to make sure that trustworthy AI will come from trust trustworthy data. So this is really following the flow of trust from data collection to creating insights. And so when you are um a person, so just like me, a taxpayer or citizen, or when you are a company, you have to give away some of your data to the administration for tax purposes, for um other types of uh of So data is being collected from businesses and from citizens and from taxpayers. So there's one step of data collection. We try not to make it too too much of a burden for uh for individuals and businesses. Then the data is stored within the government. Um and sometimes data partners might contribute further data, so NGOs, other governments, international organizations. And then data is moved within the government, within the infrastructure of the government, national infrastructure, or third-party cloud with which there is an agreement. And then, so that's the part that we see, some information is disclosed. So it can be a GDP, it can be a statistics about demography, it can be very broad on health, energy, transportation. So all the information that we need to actually make evidence-based policy. And there are like really two types of such um information that is released. Data that is open access, so say a GDP, demographics, and so on and so forth. And data that is in restricted access. So, typically, um you're a researcher, you have a research project, you request access to data not normally available, you sign a contract, you get the data, you do your research, and science science moves on, right? And um with my team, we um we're trying to think about ways we can streamline this process for the public good. And I'm going to split this graph in two parts. Like the the first part is when you want to make sure that the data that has been collected is findable. You want to make sure that once it's there, it's once it's available, people can reuse it, people can really leverage it. And so, there's a part which is about facilitating uh discovery. Right? So, a lot of data is produced by governments across the world. And sometimes and you might want to look for information, and you might not get the actual source that comes from a government. Why? Because um because, I would say, search engines or LLMs would might might look for other data sources than the official ones. So, how do we make sure that the data that is produced is AI-ready? So, AI-readiness is a crucial part. And so, it's a broad initiative, not directly my team, but the Federal Statistical Office is trying to uh put together what they call a trusted data observatory. So, meaning that it's uh a catalog of information about the data that is available. So, it's not the data itself. It's what we call metadata, so data about data. How was it produced? By whom was it produced? Can you trust it? Et cetera, et cetera. And we um so it the it's an initiative that is uh hosted in Geneva that has partners such as the World Bank, the OECD, the the um the UNSD. And there's a proof of concept that is expected uh at the end of the year. And what we're trying to achieve here is to make sure that we uh we promote I would say signal over noise in in the way we can promote uh truth. And there's a counterpart, right? The counterpart's the other part of the And I'm not actually moving my slides, which is not very smart. So, it's it was writing It was written here. I apologize. I forgot to click. But there's a a second part, which is the part you have on the on the left-hand side, which is um a bit different. It's what is called sometimes uh dark data. So, this is the data that lives in in in institutions or in companies. It's It's a huge huge huge amount of data. And it's not being used to generate insights. So, it's it stays in silos. It's It's sort of a It's forgotten in data graveyards. And we thought, how can we make sure that this data can be actually used to generate insights? And so, we um we did the following. We thought, okay. So, imagine that an uh an initiative like the uh the Trusty Data Observatory works. So, we can actually find the existence of the data. How do we access the data when the law prevents us from doing it? And so, we um we developed tools, so mathematical tools, computational tools, that allow us to run algorithms on data that we do not see. So, we can talk about the mathematics. We after coffee break or a bit later, but the idea is that when we are interested in insights, we don't really care about the data itself. So, if I want to to get information about you all, say some health information about you, I would care about an average. I would care about a comparison between men and women, between young and older, and so on and so forth. And we protect actually the insights that we generate from the data that we cannot see thanks to some randomness. So, we add some random term to the results. The random term is very small. The insights are still valid, but your personal information and as an individual or as a company is protected. Mathematically, it's called differential privacy, and this is a technology that has been heavily used by the US Census, for example. And we are actually we developed a tool. It's available. It's a public good. So, it's really something we want to share with other offices, with companies if they're interested, with NGOs, with the UN, uh broadly speaking. And we are working and we are at a very advanced proof-of-concept, minimal viable product stage for health data within the program Digi Santé in Switzerland, which is the digital transformation of the health system. So, it's really about reusing data that has already been collected. So, really, if you're interested, I'm happy to talk to you about that. I'm not selling anything. It's really public good. And I'd like to to to conclude with a a a question that would be almost philosophical. We are in a situation where a lot of data is hidden, but we want to use it, and how do we trust data that we cannot see? And so, because we know that data feeds models. If data is crap, your model is going Even if your model is great, your the output is going to be bad, right? We know that. Models are going to inform decisions, and decisions do shape public policy and our lives as citizens. And we cannot make sensitive data accessible. That's not We are We are not going in this direction. So, we want to ensure trust by keeping eyes on the metadata, but eyes off the data. So, we know what there is, but we cannot actually see the data. And we also need I eyes on the methods, the software, and the infrastructure. And so, this is the the the message I wanted to share with you. You can trust what you cannot see. So, this is sort of a paradoxical Saint Thomas, and thanks for your attention. >> So, that that was another dimension in technology. We said we we go up the stack and from infrastructure to data. Just in what I think presented very well, data is an additional cost to what Pauline has presented. I I used to say in the enterprise world, AI is to data what ransomware has been to cybersecurity. Nobody cared about cybersecurity until you have the first ransomware, and you say, "Well, I have to do something." I don't talk about advanced persistent threat the way we'll do it. Here is the same. We have worldwide in the enterprise, in the society a technical debt. And companies are investing on top of what you just saw on the infrastructure, billions in trying to put the house in order uh, around the data. So, it's part of the equation and that's why there is a an urge and it's important for our discussion later on uh, to extract value because you can't continue to spend on uh, cleaning your data, fixing your technical debt if you don't get some benefit. So, there is uh, an urgency to deploy AI services so that you start to get the benefit. So, just to put this in perspective. So, thank you very much. Uh, very interesting. Uh, another dimension to be cybersecurity. I'm going to do it. >> Um, good afternoon uh, ladies and gentlemen. I hope uh, I I'm audible. Uh, my name is Toby Simon. Uh, Patrick was kind to introduce but I was also the co-chair for the G7 last year on advanced technologies. Uh, we had a task force for G7. I was a co-chair on AI and quantum. So, that probably gives me a little reason to be among this distinguished crowd. Uh, uh, excellencies, ladies and gentlemen. Uh, we have entered a new epoch uh, of conflict fought not only in deserts and oceans but in server farms, submarine cables, and lines of code. We are witnessing the emergence of what is called algorithmic deterrence. A new strategic paradigm where nations threaten, coerce, and restrain each other through nuclear not through nuclear warheads but through AI superiority. The Cold War gave us the idea of mutually assured destruction, MAD. The age we are now entering into may give us something far more unpredictable. What we call mutually assured algorithmic algorithmic malfunction. The question before us today is not whether AI will reshape national security. It already has. The question is whether humanity retains the wisdom and political will to govern it before it governs us. Let me give you a context. The race is real and it is accelerating. In January 2025, the Secretary of US Defense announced a a plan and a strategy to make America and in inverted comma AI first war-fighting force, declaring the acceleration of military AI dominance as a national imperative. Concomitantly, the Chinese People's Liberation Army has already demonstrated AI simulation systems that can generate 10,000 warfare scenarios in just 48 seconds, a task that normally would take human commanders 48 hours to plan. I hope we understand the the difference in scale. The PLA's official newspaper declared the traditional principle of winning through tactics will be replaced by winning through algorithms. In China In China's Zhuhai Air Show in 2024, defense manufacturer Norinco unveiled the first AI-enabled synthetic brigade, a combat unit combining armored vehicles, swarming drones, loitering munitions, and electronic warfare system, all coordinated by artificial intelligence. I'm sure most of us know about what is happening in Project Artemis, which the United States is working along with Ukraine. So, let us go to the next part of it where it's about the structural collapse of traditional deterrence. There are four parameters we have put that. One is attribution. I'm sure most of you have followed this very well. It's one is highly attributable, the other is not. The detection time, it's in There is a huge difference. There is a verification issue, and there is a restraint mechanism, all combined. And you see that if you cannot prove who attacked you, you cannot respond proportion proportionately. Without proper proportion response, deterrence collapse entirely. Now, what is this black black box problem? Optimization is not a factor of wisdom. The most dangerous feature of AI's arms race is not power. It is opacity. Researchers have already warned that AI is advancing so quickly that current defense systems cannot adopt in time. They compared AI disruptive potential to the dawn of the nuclear age when atomic weapons forced the creation of an entirely new security system from scratch. But there is a critical difference in AI. We could see a nuclear bomb. We could count warheads. We could build satellites to monitor them. But, the problem in algorithm is how do you see algorithms? You can't see it. A 2021 study of military deployed uh large language models found something very disturbing. AI systems were prone to recommending pro-escalation tactics. Meaning, including actions that provoke arms races and in some simulations called for nuclear weapons deployment without clear logic or motivation. These were not rogue systems. They were doing what they were optimized to do. But, optimization and wit and wisdom are not the same things. This we call as a black box problem of military. The US cybersecurity identified this in 2025 as what technology people call just as soft software understanding gap. Now, it looks so simple, but the consequence of a software understanding gap is so humongous that it is quite frightening. Let me come to this topic of cybersecurity. Uh AI has transformed cyber attacks from a nuisance into a potential act of war, and the escalation dynamics are dangerously under governed. The irony is that AI is able to identify software vulnerabilities and write and write exploit codes at machine speed. AI doesn't sleep, it doesn't need lunch, and it can attempt 10,000s of algorithmic variation attacks that once took a team of researchers a year for just few hundred dollars. They can do this. The asymmetry is alarming. Attackers are unconstrained, defenders are not. What when a defensive AI tool fails in a production environment, people lose their job. When an offensive AI tool fails, the attacker simply pivots the attack and finds another approach. And here is another caveat, a red flag. Cyber security attack analysts predict by 2026-2027 that they will see the emergence of a self-learning agent malware code. It will learn on its own that morphs its behavior in real time based on on the defense it encounters and adapts like a living organism. Now, what is the problem? The problem in all this is attribution is nearly impossible. In cyber domain, the likelihood of identifying and successfully prosecuting a state-based attacker is approximately 0.5%. This creates fundamentally the crisis of deterrence. Traditional deterrence theory rests on two pillars, the ability to attribute an attack to the source and the credibility of your response proportionately. If you cannot I didn't taste that. Uh if you cannot prove who attacked you, you cannot respond proportionately. And if you cannot respond respond proportionately, your deterrence collapses. AI-enabled cyber op operations have fundamentally undermined both these pillars. I would just like to come on cognitive warfare and the attack of minds which uh Christine had somewhere alluded to and this is something you know we faced recently. We trained the Indian military in future warfare. And one of the areas that we trained them was on cognitive warfare before the attack on Pakistan started. And and soon after cognitive war started. So we were asked to come and and do some deep diving on it. But we didn't realize that people were following us around the world. And when we did our conference in March of 2026 jointly with the military we are all the military generals were supposed to come and speak the chiefs. Chief of the defense staff, chief of the army, navy, air force, all were supposed to speak. One day before they all cancelled. And we were shocked. I mean why did you cancel? You were we are doing it together. Why did you cancel this? No answer. And then we come to know that a day before they had put out a video uh based on an earlier video that somebody had put out that India was sending uh troops to the Persian Gulf. And this was on a syn- on a synergia background. Okay? Entirely simulated. And we didn't realize that we got a call from the Ministry of Foreign Affairs saying where did you get this? How did you put this out? Now, it didn't stop there. Because it's an embarrassment for us also. It didn't stop there. Because soon after the Iranian foreign minister calls our foreign minister and says that in the synergia forum there has been a discussion that India is sending its military to uh to Iran. This made the entire military jittery. They said if you come and speak anything to you, they would twist it. Okay, so then we realized how vulnerable, you know, what we were doing became. And in our conference, one of the generals came and his his speech was also twisted. And put out that India is a friend of one country and is going to fight against another country. So, we had to handle so many telephone calls. So, this cognitive warfare is real. And it is going to be, as you see, the battle is a battle of narrations. Uh you have a lot of populist governments around the world now. Each one will say we won the war. So, they will make spins and that will look so real and we cannot stop it. Now, uh in the in the interest of time, I will now uh shorten it. Autonomous weapons and the lethality and the time compression. AI, as you know, compresses decision time. A human commander facing a potential attack has minutes to decide, time enough to call a superior, verify intelligence and seek a diplomatic channel. An autonomous weapon system has milliseconds in a way where AI is integrated into the command and control structure, a crisis that once took days to manage can now be escalated to armed conflicts in seconds. So, the risk we see is that the publication perpetuation of a number of smaller conflicts that are going to come all over the world. And these are going to be quick, short, but they are going to be numerous number of com- And and why is this a challenge? Because highly capable autonomous weapon systems lower the political cost of inflicting con- conflict because there are no body bags, very few body bags. So, as you see in most countries, you know, it's the body bags that create the political problem black hole. In this case, there is none. I would just put a slide on the nuclear nexus and I'm sure most of us know about Stanislav Petrov who was a Russian general who refused to take a command on a on a pop- ported attack from from the United States. Now, in in the in the in the AI warfare scheme, there will be no Petrovs. Uh it will be a world without Petrovs because it's going to be the machine and that's where the huge problem is going to be. We have some key issues on the governance vacuum which we have listed there and I am sure you can quickly read it. Uh let me try to conclude by putting some blueprint for algorithmic stability. Uh one is mandatory human oversight in all lethal decision loops, espe- especially in the nuclear area. Second is crisis communication protocols for the AI age, similar to what the US and Soviet Union built in the Mos- Moscow-Washington hotline after the Cuban missile crisis. Third is the intelligence sharing on AI behavior. This is going to be extremely important. I have put a little footnote there. Anybody needs to know more, we are happy to speak. On the fourth one is binding international norms on military AI. Fifth is democratic oversight of AI procurement because we really need to know where people are procuring this from and it has to be made clear. And I say as I said in the beginning, the biggest threat of AI is its opacity. I I will just conclude that the that to say that every deterrence framework we have built from nuclear map to the conventional military balance rested on a fundamental assumption that the adversary values its own survival. An algorithm does not fear death. Nor does it fear consequences. It executes its objective function to its perfection. If that objective function is misaligned, if it is trained on flawed data, optimized for wrong outcomes, or simply encounter a situation it was never designed for it, it will not pause to reconsider. It will not pick up the phone. It will not look for a dip- diplomatic ramp off. The choice before every nation therefore, and every government, is the same. Do we let algorithms write the future, or do we take charge? The answer must be us. The time must be now. Thank you. >> UH THANK YOU, TOBY, FOR WE'VE BEEN DISCUSSING cybersecurity for a while, and now you tell us uh it's reached a totally complete new dimension. Uh and it's an illustration of what AI It shows also why the discussion between Autonomy and the Department of Defense in the US was so critical. And I think it's good to have enterprise standing for it. Is it sufficient? Toby showed uh no. Uh here, what we've seen, and it's another dimension of AI, it's the first time we introduce a non-deterministic element in structured environment. And that's why you have things such as autonomy. We've seen it. And uh there are God rails, are they respected? Hallucination, it's uh what kind of decision am I making? And you heard about the poor episode uh called deep fake. And uh it's happening here right now uh in Europe as well uh as we speak uh part of the hybrid warfare. Uh so uh a good illustration and uh Francois, floor is yours. >> Good afternoon everybody. So, my friend colleague and boss of the day assigned me a very boring uh topics. Can we compete in the global AI economy and uh or only regulate? I've been very inspired by the speech of the president of Serbia this morning. Very direct. So, question one, can we compete in the global economy? No. Can we only regulate? Yes. Thank you for your attention. >> So, um I'm going to ask you three questions in French and three question of English in English, okay? The French speaking can also respond in English. This first one is if I tell you du pain, du vin du bon sang. Good. If I tell you just do Good. If I tell you métro, boulot Better. Du bon du Yeah, of course. If I tell you to be or not Good. If I tell you liberté, égalité Good. And this one is more difficult. May the force be Good. Excellent. So, like Monsieur uh Jourdain, you're doing LLM's without knowing it. So, I've used those three six examples to explain very complex technology called LLM's, large language models. And because we're doing this since we are our kids without knowing it. Uh Patrick, you've mentioned that ChatGPT is new, GPU is new. I'm sorry, but ChatGPT started in 2019. AI started in 1953. Uh and GPUs I used and sold the first GPUs in 1984. Uh um built by Tektronix. Same as when you look at the evolution of ChatGPT, the first one was like um talking to a kid. The version two two years later was talking to a student uh in a high school. Then the three in college. I would say that four is a student in a PhD. I I was lucky to look at five and six and the equivalent. I would say that uh ChatGPT five will be uh smart PhD and six a chief of staff. So, the big difference between four and five uh is a corpus of information because more than 60% of the information used by ChatGPT is wrong. Uh ChatGPT five or equivalent 100% and the interface you will use you you will use knows you better and better. Six will be absolutely incredible because 100% of the copies of information will be checked and rechecked. It will cost a lot of money, you know. So, ChatGPT is like a drug, you know. Uh they give you for free. At some point you'll have to pay and the bill will be will be huge. But, most interesting, the interface of ChatGPT will be like a chief of staff. That means, it will be a kind of agent knowing you exactly like a chief of staff. I was lucky to have six in my life, and they knew me better than my wife or myself because those person were anticipating, you know, all kind of all kind of movement. I was very interested Pauline by your presentation on data center. Many years ago I I used to be in this business and build a lot. In fact, technology has three pillars. I say this all the time, but it's good to to understand that. It's hardware, telecommunication, and software. In hardware, it start by the chips. Uh then, you have the GPUs, of course, by aggregation of chips. You have storage, you have data centers, you have the new do-dos, you know, where the teens sleep with it all the time. Then, you have telecommunication. Telecommunication is quite easy. It's fiber. Uh it's satellite, and also it's spectrum. And last, uh software. I hate this word ecosystem because it's used very badly. As soon as you have a group of 10 people, it's an ecosystem. No. The etymology of ecosystem is when you have an aggregation of nodes, of people, or or groups, if you take one out, uh it dies. That's exactly how it works in in these three pillars. Uh and what is very difficult to anticipate is the speed of development of those technology. Uh when we were younger, you all know the Moore's law, invented by Intel, doubles every 18 months. Now, the development of technology is absolutely scary. Let's assume Pauline, you build three data centers. And let's assume now a new technology allows you to move from 1 gigabit to 100 gigabit. It will completely change the topology of the network. Let's assume now we have a chips that can analyze the weak signal of the the organs of the body. So, as an example, if you learn that they do a wet wet if all of a sudden it makes do do, that means you have a micro cancer, I would say. Um So, every time you you do a a leapfrog into a technology, it can be the speed, it can be algorithm, it can be a quantum chips coming. It change everything, the topology, the use of proceed um of the technology, how it's distributed. So, it's very difficult to predict uh the the the technology. I do AI since 1982. So, during 40 years my diploma was in my uh my office and all of the sudden I became very popular to discuss about all those um uh technology, but it's not new. What has been new is three things. The incredible development of technology, the fact that ChatGPT has been a tool for popularize it's a popularization of technology, and many people have used uh AI as a discussion when they go to TV because very few people understand what's what is AI. So, to your question um can we compete in the global economy? I'm afraid not. Uh because when you look I'm talking about Europe. Because when I look at those three pillars what if you look at chips, if you look at software uh you you you showed Pauline at the incredible investment in capex and opex of all those huge hyper scalar, we nowhere near in Europe. So, and I'm not trying to emulate what President of Serbia said this morning or to be sad. I'm just being realistic. Where we can be very smart is how to use AI, how to use this technology, how to use to make the citizen life better. I do a small ad for my conference in Laval 27-29. You're all invited, by the way, on the beach. And the theme is tech for citizen. So, we're going to look during 2 days and a half on the beach every part of our life. Life is at work, with the family, learning, enjoying. And is technology going too far? Not enough? And how does it work? So, regarding regulation, you know, you know the famous joke that in the US they have a GAFAM, Google, Amazon, and so on. Uh in Asia they have BATX. Um and in Europe we have LGPD. So, we're very good at regulate. We're not great at innovate. And what we're not good at scaling up. So, here I'm quite suspicious about the ability of Europe to have a one game plan in order to uh move forward. In the US it's not an example, but I'm half American, not in my accent, which I try to keep French, but in my mindset. In the US you have four uh different area uh for innovation. You have media in Los Angeles, you have uh tech in San Francisco, you have fundamental research in Boston, and finance in New York. Uh and there is a incredible pact between the governments. I'm not saying it's good or bad. It's not my point. Between schools, between research, between companies. When you look at Europe, there is no one center where even in France, you know, I was part of Valerie Pecresse commission for universities and everybody wanted to reunify the research around themselves like a politician, you know, we need to be unified possible around me. That's exactly what happened in Europe. So it is very difficult to create one game plan for Europe when you have 27 or 28 countries. So I strongly believe on the positive side that if we do well we can apply technology at least if we have a good cloud and sovereign stacks to make a better life for people because the European the social democrat values of Europe are not only turned to profit or the triple me me me me which is more the Anglo-Saxon way. So a good balance between innovation and deployment can be very interesting. I like to to finish by the the the what's next in fact. So what can you put the slide please? Thank you. Thank you. Thank you very much. So I would like to to finish by three things. First of all you know our brain is divided in two the left which is computerization and right is more emotion. Clearly today the left brain is totally outsmart by the machine. I say this on the stage last year. The speed of transmission of information in the brain is 100 m per seconds. When I touch this stuff, it goes at 60 m by second in my brain. Uh now, when you look into the machine, it's 300 k 300,000 km by seconds, and the latents is 10 nanoseconds. The latents of human being is 1/10 of the second. It's, you know, the fast spot uh start at 100 m. Uh the the the the the runner receive the noise, goes into the brain, and it's 1/10 of a second. So, it's very unfr- unibalance. So, we need to get used to it because in few nanoseconds, you have access to thousands of information, which are stored as long as as you said, uh Christine, it's uh secure and as it's not fake. Uh the second thing is the LLMs or the A- general AI, and you were very right, by the way, Patrick, to say it's not not only text. It's music, it's arts. Uh Uh I'm a jazz pianist, by the way, so I've asked ChatGPT to analyze my my style. So, I was surprised because it says a bit of Bill Evans, it's a bit of other pianists uh I like. Um and then I ask her, "What do I need to improve?" And the response say, "A lot." My wife is a painter. She analyze also her style. Says a bit of Rothko, Nicolas de Staël, and but the system was good enough to help her to finish one page she couldn't understand. So, the LLMs, in fact, the evolution of LLMs you will see in ChatGPT 5 and 6 will be a kind of avatar of yourself, will be a kind of personal agents that will get take, you know, all the information like a super um chief of staff. Last, the biggest revolution is yet to is not yet come. It's what we call and you've mentioned this Patrick already, LQM large quantitative model. So, the entry point into the system are LLMs. You will receive datas. And those datas are going to be compute um and address a lots of market. So, on the slide, uh every time people say, "Where do you invest, Francois? Where should I invest?" I say, you know, in your family, in your house, and you have some money left. This is how the market in the future uh will um And I think the value creation, the money spent is going to be absolutely incredible on financial service, uh on drug discovery, uh what what one thing I'm very interested in right now is uh the proactive maintenance of the body. When you buy when you have a car or you take a train, when you take a plane, whatever, you don't take your car if you do not you have flat tire or you have no electricity or if you take a plane that has 3,000 captors or a train 500 captors, they send signals which are correlated and they analyze say, you go, you don't go. For us, the weak signal we have, we tired, we have fever, we have and it's too late. So, technology, the biggest leapfrog in technology will be to develop a sensor that will be into the body, capture the weak signals of the organs, correlate them, AI will cancel the noise, and then compare to 100 million examples and say, "Hey, be careful because we hear the signal on this organs, so you might have a micro nodules." So, if you take finance, if you take biopharma, if you take uh leisure, travel, whatever, the LQMs, in fact, is aggregation of all the the the physics, thermodynamics, chemistry, whatever, that will use LLMs to have dialogue and then compute and back and forth. So, the question, what are we going to do with you with our life? I think it's our kids to respond because it's not yet to come. Thank you very much. >> Thank you, Francois. This will be a nice transition. Yes, please. >> Just one question. The um LQM thing, what is the abbreviation for >> large quantitative models. >> Yeah. No. >> Yeah. >> Typically, I'll just give an example. >> Okay. >> Uh we created a company with few friends, you know, in 2022 called Sandbox AQ. It's a spin-off of Google. You have 72% of PhD. I joined company 4 months before the the spin-off. Uh we we focus on quantum technology and AI. By the way, we raised 950 million dollars on quantum. But, quantum became less popular because every 6 months the issue of the technology was postponed by 12 months. So, the second round we did uh we we did and the third one we are going to do we say we are AI company that we will powered in the future by quantum. So, typical example, drug discovery take 8 to 10 years. So, the algorithm in um the quantum software algorithm allows you to do test in 10 days between 10 days and 1 year. So, it's it's a huge um We've done also very interesting research for two very high-end French brand luxury. I can quote, fortunately, but you can believe. For those, I call this the posh of vegan who wants to have a very nice uh bags, but not built uh not killing animals. So, we were able to create uh artificial or new materials in 3 days instead of 3 years. So, there is a plenty of uh example where the technology will help for the health and big believer for the health and also to to to be uh uh much more uh uh eco-friendly in term of preserving nature and CO2. >> So, thank you uh Francois. So, I think the very good reminder uh going back to Ada Lovelace, but we went from complex input uh simple input to simple output in AI where when we started uh after we had the mathematics in place. Uh AI is three things: mathematics, compute power, and uh data. So, that's what you need plus the energy today. So, uh and then we had uh complex input, simple output. No nobody cared because you didn't understand the label. That's was the example of Francois explaining how labeling worked and that's very good. And then GenAI is complex input, complex output, and all of them sudden we can visualize and see something and interact. Uh the other thing I think it's very important what you said at the end. Uh quantum will put AI on steroid on top, and there is another race which is not the topic of this workshop, but we you can elaborate on this afterwards. And it's coming uh uh China has announced uh quantum communication in space and verifiable. Uh you talked about health and I know you work on this with Sandbox's quantum sensing for health. And, uh, China has announced recently, unverified yet, uh, 2,000 cubit, uh, which I've not seen. Uh, today, if you look at the IBM and Google, who are the leaders in the space and the spin-off, they are in the 100 plus cubit for fraction. Uh, but it allows you to start to do so. It seems we start it will take some time. But, when it will come, you don't need to send a constellation of data center in space. You send three, and you're done. Uh, so, it's it's another thing that makes the this AI question so critical, because it will, uh, it will it will not slow down. It will it will accelerate. It's coming. And this is a very powerful technology. >> Yeah, you said one very important things. I I came to Silicon Valley in, uh, August, '98. Uh, so, and I could see the, uh, the economy going up, uh, and down. And the the I'm on board of many companies and invest many companies. And the biggest mistakes that the entrepreneur are doing right now, they have a vertical view of a technology. And as you remind, and I said, when you have three pillar, when you touch something here, it touch here, here. So, you need to be totally paranoid of what, um, what the others doing. I give you a last example is on, um, on the the gray and white area. Uh, the white area and gray area in the world are area where you don't have access to network. Uh, if you add $1 billion of capex, which is allowed to people in remote or, uh, place to get access to the network. The market cap of the company will go up from 10 to 40 billion. So, there is a huge pressure to have access to remote part of the globe. That was consideration. If tomorrow as an example, you have a new an antenna that will consume 1,000 less energy and very small, it will totally change the two orders. So, it's very difficult to predict, um, you know, the future because everything, uh, and everything is, uh, correlated. And so, you need to be very smart or kind of guru to predict what's going on. >> Thank you. >> Thank you and good transition, uh, for the uh, remind that presentation, uh, where we will highlight more, uh, the social, uh, societal usage of AI and which, uh, positive perspective on it starting with, uh, Daniel. >> Um, thank you. Thank you, Patrick. Thank you for the slides. So, uh, next slide, please. Ah, sorry. I do it. Okay, so, um, I'm an obedient kind of person, docile, and I was asked to talk about this, the human factor, preserving agency and intelligence in the era of AGI. Fair enough. But, I realized soon enough that I could not actually talk about that. And so, instead, what I'm going to talk about is this, the human factor, preserving agency and intelligence in the area of the notion of AGI which is not the same. So, uh just because I'm going to say certain things that may sound sort of unusual, um first of all, you can buy my book or you can wait for the pocket book edition which is going to come out in the fall and which is revised and augmented because this one is to some outdated in certain ways but not the basic philosophy. But also, you can look at this uh very important paper, what I regard as a very important paper by Arvin Narayanan and Sayash Kapoor, AI as normal technology. It's a new way of looking at what's happening in AI and although I didn't know about these guys, I mean, the book came out before I read this uh article, I was struck by the fact that we're sort of moving in the same direction. And so, if you think that my ideas are really uh bizarre, uh you can you can refer you can at least try and read this excellent and very long paper followed by other papers. Okay, so what's the difference between the first title, the one that I cannot fulfill, and the second one? Well, the first title was saying, you know, AGI the technology is a threat to human agency and intelligence. But the second degree interpretation, which is what I'm going to talk about, is AGI the notion is a threat to human agency and intelligence. And compare this with ghosts in Scottish manners. Uh the creatures are a threat to tourists in Scotland. But ghosts in Scottish manners, the stories are a threat to the tourist industry in Scotland. And I think we're I'm interested in the second thing. Okay, so uh I realized with age, I thought that the this movie The Usual Suspects is something that everyone has seen recently, and then I realized that the movie came out 31 years ago. And so it may not be have remained in your memories. Anyway, in The Usual Suspects, there's a character called Keyser Söze, and he is uh ill-defined. He's a presence that's felt but never seen. He's absolutely terrifying. He's unbelievably powerful, and he can do anything. Now, I think AGI is on the same status level uh ontological level as Keyser Söze. He's ill-defined, but its presence is felt in some frontiers LLMs, and in the dreams of the pioneers of AI, but never seen. AGI's absolutely terrifying. AGI's unbelievably powerful, and AGI can do anything, at least anything cognitive. So, I claim that there's no AGI. And uh I'm going to not going to develop this because I was told I had very few um just a few min- short time to uh explain, but I'll try to give you some idea. First of all, AGI is ill-defined. Uh if you look at the literature, you have tens of ideas of tens of definitions of AGI. For example, you have some sort of this general thing, AGI can do anything a human can do. Oh, really? I mean, it's not very likely that, you know, you could have an algorithm that can do anything a human can do. What does do mean? Uh and sometimes it's more like AGI can do anything that a human can do and has economic value or AGI can you can can replace a human in all cognitive activities. But what is a cognitive activity? All of these are interesting interesting interesting questions and their answers are given. But none of them I find satisfy me personally. So I also claim and of course that's going to ruffle some feathers. I also claim that AGI in any of the senses in which it is seen as real, authentic, genuine, full human intelligence, that's nowhere to be seen. It's not on the horizon. It's at best a fond hope. In fact some of the creators of deep learning and LLMs have said so much. They said, "Well, it's nice deal deep learning and LLM, but if we really want to reach human intelligence, which I take as a substitute to AGI, although you know, human intelligence in an algorithm is a bizarre idea, but still the the the the rough idea is that what what what these people were saying a few years back. I haven't heard that as recently, but still they I can find quotes. You can find quotes in the literature saying that if we really want to reach full intelligence, then we're going to need some sort of new idea. And some people say, "Well, that's you know, we'll see if this new idea comes about." And others are so convinced that it must happen that of course this new idea will come about. Now, I have I develop in my book uh to believe that it's a senseless pursuit. And in fact, just as Narayan uh Na- Narayan say, uh actually, most of the profession agrees with this. Most of the profession has more or less tacitly toned down their claim that we're going to reach full human intelligence. They're interested in something else that I'm interested also in and that I believe in very much, but it is not the realization of human intelligence. Still, the Keyser Söze effect rises from the ashes again and again. Uh it's hard if you start interacting with LLMs, even the cheap cheap version of LLMs, and you know, if you can afford the $20 uh a month or the $200 a month, you get even better results. It's extremely hard to resist this idea that there's something really strange going on, unexpected going on. And it's it's true, it's unexpected. I mean, the the the the the LLMs, uh what the LLMs can actually do was not something that was planned. It's something that researchers hit upon, and they were surprised. And they were surprised. They didn't expect it to be able to do this. So, what accounts for the illusion? And I think it's a mix of what I call the LLM magic. Now, what is the LLM magic? Now, contrary to what you will hear many people, knowledgeable AI specialists, you know, will say, there's no magic in LLM. It's, you know, it's mechanical. It's just lots of data that get um mashed up together. And really, there's nothing to nothing to be so struck by. I disagree strongly. I think there's a magic in LLM that is we still don't have a scientific account of why LLMs work when they work. And therefore when they don't work, we must be able to predict if we want to be able to rely on them, we want to be able to predict when they don't work. So, I think we're there's still lacking is I'm not saying we have no idea of how LLMs work because obviously it's our engineers who actually build them and supposedly when you build something you know why it works. But in this case, it's one of these exceptions. It's not new. I mean, when you know, metals were were dealt with were used thousands of years before we had any sort of notion of metallurgy or the basic structure of matter or anything like that. So, we're not quite there, but we're in somewhere like that somewhere in that area where we can do certain amazing things. We can create amazing effects very powerful effects. I'm not denying at all that they're very powerful. In fact, I'm myself a victim every time I interact with an LLM. I say, "My God." And it it's not so much when I ask some very technical question on an area that I know quite well because actually the facts can be found rather easily. I mean, deep theoretical questions. But what stuns me is and I think I think Patrick said so with his idea of chief of staff. What stuns me is that within a few moments of changes, the system seems to know where I'm at and is answering with more and more sense of relevance. And that's what I call the LLM magic. So, I think there's a mix of LLM magic, which is still unaccounted for, and then there's the limitless ingenuity of tens of thousands of engineers experimenting, tinkering, and fixing gaps. Fixing gaps, you know, when things don't For example, uh everyone knows that the early models of GPT were terrible at arithmetic. And now, of course, the modern more recent models are wonderful at arithmetic. But how how did they do it? Maybe they just filled the gap. Maybe they retrained it on a number of arithmetic truths, and suddenly it worked. But that's not theoretically satisfying. That's tinkering, essentially. So, why does it matter that there's no AGI? Oh, I thought I was doing well. I'm sorry. Okay, I'll be quick. So, first of all, the level of anguish about AGI goes down. There's no dragon to slay, there's no takeover by maleficent AI systems. AGI, the notion, is the golden calf of the AI community. It's a threat to AI. It shoulders it off the track. And once the calf is ground to powder, as it is in the Bible, then the AI community can focus on real opportunities, which is augment and not replace. And again, I think that the vast majority of the profession is really into the augmentation paradigm rather than the replacement paradigm. And then you can worry about real risks and threats, which are many, as we've as we've noticed. And human intelligence is back. Why? Because we were so mesmerized by the possibilities of AI that we forgot that, you know, human intelligence is at the base of at the origin of these systems, and it's right on time to help us deal with with the problems we have. So, quickly, the good news is that humanity is getting on top of AI by shedding the superstitious fear of AGI and by letting intelligence do its job. I won't detail it. We're talking about that. It's rather familiar. One one one important thing is that human intelligence is able to divide and conquer. Take problems one by one. For example, the disappearance of jobs is one problem. Misuse is another or dual use is another. Runaway self-improving system is another problem. The problems that were mentioned in the previous talks two talks are to be dealt with and are beginning to be dealt with one by one. Still there concerns, there contrary winds. First of all, there still the grip of the AGI narrative that AGI will happen and the first country or firm to reach it will own the world on and this grip is really still operative on a large segment of public opinion, political leaders, industry, finance, and some genuine experts. There's the competitive dynamics, invested interests, of course, when you have hundreds of billions of dollars invested, you better you better fulfill the dream. Hype and confusing and something that wasn't mentioned at all today or yesterday, growing public resistance to AI such leading to polarization and politicization. That's a real concern. Um so, facing existential risks. So, very quickly uh misalignment, cognitive atrophy, volatile atrophy, and corrosion. You can cross out the first. I don't believe there's going to be any problem of super AI sort of getting on top of us. Then there's cognitive atrophy, that's decay of intelligence, loss of agency, and corrosion of sociality by dissolution of trust. Let me quickly give you an idea of how in general AGI the notion of AGI threatens human intelligence. Humans count on AI systems to help them with tasks that require intelligence. No doubt about it. And with uneven results, sometimes good, sometimes bad, sometimes in between. And there's realities most people sort of run a reality check and they use AI systems AI systems with caution and only on certain kinds of tasks. But others who believe in AGI predict that the results will keep improving. And therefore they prepare themselves to offload more and more of their intelligence to AI systems leading to self-fulfilling prophecy that human intelligence loses out to AI systems. What AI does can do is left undone and human intelligence atrophies. In other words, instead of AI rising up to human intelligence, human intelligence stoops to artificial intelligence and shrinks. So it's the same more or less for agency and for lack of time I won't run you through this, but it's again a very simple reasoning that you get. And I think it's even more important for human agency actually. And the the the the response to this is to you know, to think of human-centered AI. Where where where human-centered AI is not just that AI has to be built from the ground up with aim of serving humanity, being beneficent, bringing real benefits, fulfilling real needs, and being non-maleficent. that's part of it. But more ambitiously and more deeply, AIs has to be built in such a way as to respect the foundations of human existence, agency, cognitive integrity, full and free exercise of intelligence, trust, sociality, normativity. Or another in other word in just one one word, AI must be built in such a way as to follow the grain of humanness. Humanness or humanity. So, the key is to uphold and never lose sight of these values. And let it be in all applications and deployments, let it be the compass. Thank you. >> Thank you. Thank you, Daniel. I realize it's been a long session of in technology, there is something we call TMI, which is too much information. So, I don't want to have a information overload. We have a two presentation, but before we move we I would like to ask if based on what you've seen and by the way, on the last point you said not to augment but replace. But from a geopolitic perspective, you've seen this morning that our demographies in the Western world are declining and there is a strong appetite to replace some of this demography with AI and robots so that you can maintain the current dominance even with a declining population. So, here it's another question. So, are there questions before we move to the last presentation, an example on what AI can do that you would like to address based on these or comments, please? >> I have a question for Daniel. Uh you Everybody knows what is artificial AGI. Because it's an acronym. It's It means artificial general intelligence. How do you compare this with superintelligence? That's a topic we're working on in the US and Stanford right now. Um And which in your view is a more personal related to us between the intrinsic intelligence and a personal intelligence? >> Thank you for these questions. Now, a superintelligence is obviously not AGI is generally thought of as human-like or human-level. These are the two main sort of human-like intelligence or human-level intelligence. But a superintelligence is obviously not an AGI, right? Because a superintelligence is not a human intelligence, it's superintelligence. So, a superintelligence could easily be reached by not going through the stage of the mere human artificial intelligence, human level, or human task, but directly over it. And in some some aspects, of course, I mean, the suggestion is that already human super I mean artificial superintelligence has already been reached in certain areas or about to be reached. And as about your second question, you're you're getting close to my theory, which I haven't had time to expand on, which is that I think that human intelligence has something to do with the ability to deal in a human way in human situations in every concrete case. And that is really very very far from anything that AI can do. Although AI, of course, can solve problems, and many times we deal with our situation with our situations by solving problems. But there are many many cases in which we deal intelligently with the situation we're confronting without solving any problem. >> Thank you. Other questions or comments >> I have a question to Pauline. Um you mentioned that the GAFAM are using their huge cash reserves to invest in infrastructure for data centers. Uh okay, however, they seem to continue to invest quite a lot. At one stage they will have to go very far and cash will be running out. Uh isn't there a risk that there's a discrepancy between the investment and the return on investment and which could be sanctioned by market correction over time. >> Uh thank you very much. Um I think yes, there's definitely a risk and actually the the last reports from from all the the the the GAFAMs, we have seen that the market tends to sanction those who have invested a lot but have not been able to show uh already a very very good return on these investments. So, for example, Meta uh has because they're able to already generate return on these investments through all the ads that they do through social networks, uh they have been pretty rewarded by the market while Micro- Microsoft, even though it's uh currently uh through uh through Copilot, Microsoft 365, Microsoft has a very good position within enterprises. Uh it the and and was showing very good results. It was still sanctioned by a 10% decrease in share share price when during the the the the last reports. So, so yes, there is going concerns of the market. Um but if you look at the the ratings of these companies, they are all still triple A's. Then Google and and and Micro- Microsoft are generally acknowledged to be uh so sort of the best corporates in the world. Everyone wants them as their counterpart. All the banks want to finance them. The only real one The only one that is really punished by the markets is Oracle. Because Oracle has had a very very high because to to that. So so so So yes, there is a risk, but for now uh these companies are still the the the strongest in the world and and they need to show that they can sustain the race because otherwise the the risk is is is higher, I think. >> Thank you. Uh other questions um Yes. Please. >> Just a very quick comment. You used the word that I've not paid attention to, which is humanness. And I think it is interesting to discuss the question that has been in the media a lot about the threat to humanity And I'm wondering whether in certain cases we are not confronted with a much more daily problem, which is the diminution of humanness in our activities. The fact that things being trending towards efficiency efficiency efficiency efficiency is basically making us cogs in a in a machine where you are in symbiotic relationship with the tools that we've developed. So that instead of augmenting ourselves, we are becoming a component that serves the tools. And so paying attention to the human nature in the way we develop AI is probably one element to avoid being reduced to the less human part of our identity, I guess. Just a grabbing the word. >> Yeah, I thank you. It will be covered most likely in Jay's presentation. So I will refer to his presentation if you don't mind. You had another question and then we move on. >> Yes. >> Yeah. >> Can you hear me? >> Yes. >> Yes, okay. >> Maybe it's better It's a question for for Daniel because I understand that the AGI is less than a level of intelligence but a way of thinking very similar to the human. And my question is quite simple because you say that we can't speak about intelligence for the AGI since they cannot have a new idea. And my question is what is a new idea? Because I think we it's very difficult even for a human to have a new idea because it's always something we reuse from the past or from experience or for a source of data. So, what is a new idea? >> It's It's a nice question because it's the the basic question of creativity whether there can be machine creativity or whether um and whether there's in fact human creativity. You have You have I see the view of creativity that you have sort of re-combination and re-mining and so on. I'm not convinced by that. I think there's some really new ideas. But in this particular case, it's just that there were a number of statements by the prime movers of the connectionism of deep learning that they said, "Well, we're not there yet and we need a new idea." A new idea in the scientific sense. For example, you know, suddenly Newton had a new idea about how to deal with forces at a distance. And then or Darwin had a new idea about how my species are different and so on. And are And are adapted. So, it's it's this this idea that we need something and I think the notion of a new idea is is naive because it's it seems to suggest that there is a a single key to intelligence and I don't think there's a single key to intelligence. >> Practically and those who who are active I don't know if you Can you hear me? Yeah, Arthur, uh talking of creativity and drug discovery, which is more systematic creativity, if I may say. You will kill me for this, but don't worry. >> Thank you, Patrick. Uh good afternoon. Can you hear me? Yeah. Good afternoon, everyone. Uh it is so hard to speak after you, Daniel. Uh I loved your Kaiser Söze analogy. You're a master storyteller and thinking. So, thank you, Patrick, for the order and thank you for tasking me with the utopian vision of using AI to cure disease in 8 minutes, dangerously close to the gala dinner. So, I'll do my best. Uh my biotech leverages the power of gene editing to develop cell and gene therapies for cancer and other indications. So, I'm seeing firsthand the impact of AI on medical research and drug development, but I would like to warn you about the hype. While there is excellent progress in some areas, we're very far from a transformative revolution. So, just a bit of context. Drug development is the slowest, most expensive, uncertain, but also the most personal human endeavor. Would you invest a billion dollars over a decade for a candidate cure that has 90% chance of failure? Would your answer change if a family member could benefit in the future? In the last decades, while some diseases have been cured or substantially managed, many others, particularly neurology and psychiatry, subtypes of oncology, rare diseases, and complex chronic conditions, still lack effective treatments. AI is beginning to improve part of the system, but it's not yet transforming outcomes at scale. So, first, AI is making drug discovery more efficient. François, you mentioned it. It's very clear that AI has had a big impact in early stage research, target identification, drug design. Its ability to screen and exploit large data set and parameter spaces has led to identifying candidates that would not have been able to emerge through traditional research. Patrick, you mentioned Demis Hassabis. AlphaFold, of course, has made protein structure prediction broadly accessible. You also mentioned the role of AI on dramatically improving gene editing efficiency, that's us and others. We use AI every day to develop our gene editors in silico before bringing the most promising candidates to wet lab testing. This accelerates hypothesis generation and de-risks development candidates before they enter the clinic, but they do not by themselves create new drugs. Companies like Schrödinger or Insilico Medicine have shown that AI can help design candidates therapies more quickly, and some have entered clinical trial. This month, Anthropic entered the fray by buying New York-based Coefficient Bio. These are meaningful steps, but candidates still face the same clinical risks and costs before coming to the markets. Similarly, our gene editors need to be tested in cellular models and then in humans before becoming therapeutic realities. So, the primary difficulty here is that AI has not yet revolutionized clinical development, and clinical trials remain the major bottleneck. Phase three trials represent the largest cost, sometimes up to 50% of the billion I mentioned at the beginning, depending on modality and therapeutic areas. And the improvements coming from AI here are at the operational, not the fundamental level today. So, you have companies like Tempus, Flatiron Health, that are aggregating clinical and molecular data to better define patient population, to refine inclusion and exclusion criteria, or to identify subgroups, but this is not yet a full scale. What is starting to emerge more interestingly is the development of synthetic control arms. This is something we're experiencing life with FDA today, which leverages real-world evidence to remove the need for placebo or standard of care control. This can lead to faster, cheaper, more efficient clinical trials. Uh this is promising, but the better uh the biggest bottleneck does not come from the technology, it comes from the regulators. And and regulation only will will able to unlock these bottlenecks. Third, AI is delivering tangible value in diagnostics. So, the most obvious use case is imaging, where AI-assisted screening and mammography and dermatology is now showing the ability to reduce false positive, and in some cases detect cancer before a physician can detect them. Algorithms leveraging retinal imagery are being deployed at scale to the to detect a range of cardiology or metabolic disorders in a non-invasive way. On the genomic side, you have the Plan France Médecine Génomique in France, UK Genomics England, and similar equivalent in other countries that are integrating genetic sequencing into care. That allows to reduce diagnostic delay and develop new cure. So, I was particularly excited this year by this N-of-1 genetic medicine, so baby KJ uh in the US that suffered from CPS1 deficiency uh was treated at the Children's Hospital in Philadelphia with a bespoke genetic cure that was designed exclusively for him, and that was under the FDA blessing. So, this kind of N-of-1 genetic treatment is the way forward, but will require adequate uh regulatory pathways. So, what remains to be solved for AI to deliver cure at scale? Pauline showed us that computer is there or will be there, but Christine shows us that data is critical. And interestingly, I will argue that the three major bottlenecks in healthcare are the same and have been the same since the dawn of ages. The first one is data fragmentation. Healthcare data is the textbook case for incredibly siloed, poorly structured, and difficult to integrate data. Who in this room has access to a digital copy of their entire medical records since they were born? There's someone like who would love to meet them. Uh this is the dark data that you alluded to. And this is why your work is so important, and this is the most urgent step to address as we said it many times models are as good as the data they've been trained on. Second is regulatory adaptation. So, FDA and European authorities have begun drafting pathway for AI-based tools, integrating AI in their guidance and drug approval pathways. This is still very slow, and the swift and successful evolution of regulatory pathway is the single most important step to dramatically reduce drug development costs and accelerate the adoption of innovative therapies. This will require trust, and I will argue that the man statutory human oversight that Toby you mentioned in the defense sector also strongly applies in the life science sector. And then last, economic alignment through novel business models. So, allowing these AI tools to capture a portion of healthcare costs through value-based models will be critical to ensure that the adequate level of incentives is in place for widespread adoption. This will remain very challenging given the rigidity of public and private payers on both side of the Atlantic and the complexity of designing payment for long-term health outcomes and not immediate benefits. So, obviously AI is not yet the ultimate engine for curing diseases. I'm still very excited about its potential. The technology and the compute are there. We need data structures, innovative regulatory frameworks, and novel economic models, and hopefully we can transform our approach to disease. >> Thank you, Arthur. Thank you, Arthur. You put all the presentation in perspective. You've done Thank you. So, it's very good illustration in one very specific area how the different topic that we've addressed uh at the end come together. And last presentation Jay our US representative here. So, you have a double mandate. Illustrate in your area what you do and then also give us some of your other perspective as well. We're quite of a European if not French biased in this conversation so far. Thank you. >> Thank you, Patrick. And and good afternoon to you all. I'm Jay Truesdale. Uh my role is CEO of a risk intelligence firm which was founded 27 years ago by members of the US government that sought to take the methodologies of information gathering and analysis and apply them to a corporate setting to provide a symmetric uh an asymmetric understanding of the competitive landscape as well as potential opportunities as as as well as risk mitigation approaches. And this entire business model it was it was fundamentally based on on human source inputs. It was based on getting access to information that could not be accessed uh anywhere else other than by talking to people and speaking in confidence in a manner that not only elicited that information but ensured that all parties who were involved in those conversations were somehow benefiting. My my business falls within the the broader category of professional services firms. Professional services firms arguably are being the most impacted by certain applications of of AI. If you look at professional services firms and as a percentage of GDP in the OECD countries, they constitute between 10 and 15% of overall contributions to the GDP. And indirectly, if you think about how these professional services firms have impact on their clients, from a client perspective, they account for roughly 70 to 75% of of GDP. So, this is a an industry that's massively vulnerable, but also potentially in a position to benefit greatly from from the advent and advances of generative AI. I'd like to talk a bit about how we think about this. Daniel, maybe to kind of put on my philosopher hat because I very much appreciate how you were able to help us think about just the concepts here. We think about um information in terms of both knowledge and understanding. In the world of AI, we believe that probably 60% give or take of of knowledge exists within the bounds of what can be known within AI or AI-like instruments. There's probably an an additional 20% of information that's not yet accessible by AI for a variety of reasons. Either this is proprietary information, it's behind firewalls or paywalls, or it's information derived from uh countries, jurisdictions, locales that are not part of the data that's inputted into these LLMs and and other models. In fact, um historically, 90% of the data that is being used to train AI models, 90%, is in English language. That's remarkable. And the remainder uh is is really split among five other languages, the bulk of that remainder being Mandarin and Spanish. So, when you think about what is knowable, what we're really talking about is what is knowable in English language by those who are both native speakers and non-native speakers who have used English as the medium by which data is inputted into to LLMs. So, for that 20% that is not yet knowable, this is not just proprietary information. This is information that's not yet at all been considered by those that are using uh um data training uh models. So, I want to just put that out there cuz it's a point that we haven't yet talked about. Maybe coming from a native speaker, coming from an American, coming from the country that's uh obviously preeminent in this field, it's something that we recognize very much as a gap in uh in in uh in in the knowable information category. The last 20% is this human-sourced element, which of course is what has historically differentiated my company, our ability to get access to that information and be able to translate that to commercial considerations. The other side of information, of course, again from a philosoph- philosophical point of view, is not just what is knowable, but what is understandable. Now, um this is where I think uh professional services firms will be uh will be used, and where uh to Francois's point about having your chief of staff, where professional services firms will increasingly curate the knowledge that they're able to gather, both what is knowable and what they are able to build from a proprietary perspective. And this is really where I think uh from a a layperson's point of view, and I don't know if you all maybe put yourself in that category similarly, but this is where I think AI could be incredibly valuable, incredibly useful. And so, I just wanted to share a couple of things we've done as a firm to manage both what is knowable and what is potentially better understood or or curatable. We think of this as as both defensive and offensive. On the defensive side, we're highly cognizant of governance, highly cognizant of ensuring that the data we collect, but the data that we're also provided by our clients, is managed in a way such that it cannot be shared outside of the bounded framework in which we operate. Obviously, that has positives and negatives. You have to create means by which to input additional data beyond that framework, which takes effort, it's highly highly inefficient to some degree, and it costs quite a lot of money. Um but, we believe you have to do this because otherwise, without these proper governance mechanisms in place, you run into concerns around data confidentiality, around leakage, around accuracy, trustworthiness, and and bias. And so, what we've sought to do is maintain the highest standards of governance and compliance. And effectively, this is what the entire professional services world is moving towards from a defensive point of view, focusing on these core elements of ensuring that we protect our our clients and protect our information. From an offensive point of view, or you know, trying to think about what the opportunities are for firms like ours. Um we have really leaned heavily into agentic AI, and that hasn't really been discussed yet in uh in this forum, but agentic AI is effectively like your chief of staff. It can take a complex problem, break it down its into its component parts, assess which of those component parts need to be done in which way, how you prioritize them, how you then resource against them, and how you integrate a whole process flow around them. So, that you can achieve your discernible objectives within the given time frame and with within the resources that you have. This is, in a way, what a good chief of staff should do. Um but this is something that AI tools absolutely can deliver. And what we've done in in this respect is we've created, from an internal point of view, processes by which we have built out increasingly efficient ways of taking complex data uh around the core client problem and delivering that in a much more transparent and consistent fashion across all of our teams. It's created clarity, it's created process certainty, it's created um an understanding of what the what the outcomes need to be on the part of of the client. Um it's also kind of moved our mindset from human labor to human oversight. Um the the our analysts who are responsible for these processes recognize that they if if the if the structures are not right, if the architecture is not right, the outputs are not going to be correct. So, for every problem, for every project, they have to the most important part of the work is at the outset to ensure that the architecture of how AI is used is done in a manner that's not only going to protect client confidences, but is going to lead to the outcomes that need to be um need to be achieved. So, that is that is one. Using agentic AI from a kind of chief of staff point of view to ensure that the architecture is correct, but also the outcomes can be done in a manner that's that's highly efficient. The other way in which we've used AI from an opportunity point of view in our professional services context is to uh both access different types of knowledge, knowledge that isn't necessarily available currently with the data that's provided to LLMs, but also to to curate the knowledge that we are able to to bring in. We for example, a common question that our industry had had asked of us to to think about is where is this policy going within let's say the the context of the current administration or how might this regulatory framework affect my industry and my business in this specific jurisdiction. And so, what we've done is we've actually created an AI-based tool to solve those problems. I guess the question is is how. It wasn't very difficult. In fact, conceptually it's not difficult at all. We we built something which we call reg- regulation navigator, which is industry agnostic. It was built first in the United States and Western Europe again because English language data was the most prevalent. Um but it allowed us to take a snapshot of every regulation, every regulatory structure across every industry in those jurisdictions. So, that's interesting. But what's even more interesting is our clients can upload their policies, and they can compare their policies to the regulation in those jurisdictions, and they can pretty quickly, almost instantaneously, see if they're compliant. And if they're not compliant, they can see where their gaps. That's interesting. But, that doesn't really help you beyond what a law firm can do or what your own staff can do if they're closely monitoring the the the regulatory context. What's really differentiated, I think, and this is really driven by AI, is scraping the available information plus the additional information we're able to put in to our model through our proprietary sources and begin to project how could the regulatory context change within that jurisdiction, within that industry? And, of course, there are all kinds of signals. There's political debate. There's NGO activism. There may be a crisis within that society or a major emergency. But, if you amalgamate this information, you train your model correctly, you can be begin to project from a professional services point of view, how that problem might evolve over time rather than how you look at it from a static point of view. So, again, we're using AI from professional services context to make sure that we're amazingly focused on the governance side, making sure that the data we have already from a client perspective is fully secure, making sure that our inputs are not just bound by the knowledge that exists based on the data that's used to train LLMs, and then we're using our our kind of capabilities brought to us by our internal experts, but also brought to us by some of the best practices we've talked about today, to think about humans have to how humans remain in the loop through Agentech, but also through the curation and architecting of tools that are designed to forecast and look into the future. >> Thank you, Jim. >> One question. >> And one comment. Uh I I I I'm sorry because I took the commitment for you so you don't have to take it. Answer the question on humanness that was raised before. I think it's it's you you touch upon it but not completely. How do you see the impact in your work or through other activities? You see the impact of technology on on basically on what you define before as the humanness factor and its evolution. >> Yes, I mean I I think the we've all had the experience when we've used AI where what you get back feels almost human but it's not quite right. There's some discrepancy. And the way in which we manage that discrepancy, first of all is we try to understand why it's not quite right. So is the data that's being input inputted somehow incorrect or is it not fully complete? Uh if it's in non-comprehensive, why is it non-comprehensive? Are we missing a certain perspective? That requires a human at at the at at the at the front end to be able to identify and address. So that's one aspect. The other aspect is once we get back that information and we're assessing its validity, the human trained analyst with judgment and with experience, ideally with wisdom, is able to understand actually know that's not the full story, not just because of what was inputted but because this doesn't connect to another important feature that actually is the real issue that we need to address. So where the human is for us is at the front end ensuring that what's designed, what's what's what's modeled has as much input as possible and architected the correct way. But once it comes out on the other side, we we validate using wisdom, using judgment to ensure that what we actually suggest is not just that 70 plus percent answer that comes out of the model but it's much closer to the 100% that that we demand. >> I would I would let if you don't mind I'd one dimension probably behind your question. I think we we lost one battle with the social media. We lost a battle. Society didn't see what was happening. We were warned by maybe not Daniel but by colleagues from him. We didn't see it and 20 years later boom, we have an issue. We have a step back in the humanness. How do you learn if you don't have the experience? How do you and I think it's part of your question and I think this is an aggravating factor in managing AI. Maybe you want to react and then have another comment on the US domain. >> Yeah, actually I'm really glad that you mentioned this because I I would have wanted to raise it myself. I don't know if people are familiar with the word of the work of Tristan Harris uh for the Institute for Humane Technology. And actually he made exactly this argument. He said the first encounter that we had as a society uh and the title is here on society of China's is uh was with um algorithmic recommendation on social media. And I've worked for 10 years on one of the topics which was the regulation of platforms and the and the regulation of the content on platforms and we completely missed the angle because we focused on removing pieces of content instead of regulating the algorithm recommendation that were producing the polarization. And there's a word that we haven't used but I think it's interesting to bring in which is the term emergence. And one of the challenges is that we do not know what the deployment of agent uh at scale is going to produce. True agents, agentic AI is going to be chief of staff but it will be also thousands of different profiles specialized in one competence or another, more or less specialized. And we need to think in terms of billions, and probably hundreds of billions of agents that will be that will be functioning and interacting between themselves. I don't know if you followed the the experiment that has just taken place recently called Malt Book, which is basically a social media platform that was open and dedicated just for agents. There are a lot of flaws with Malt Book, but that's another issue. But the notion that there will be emergent behaviors from billions of agents that we will not completely control individually is something that needs to be studied because if you look at the finished the work of Thomas Schelling, the the Nobel Prize winner in 2011, I think, looking at micro actions and macro behaviors, we don't know what the deployment at scale will be, and sandboxing will probably be an an important element to explore this. And to finish, there's one thing on the humanness, which is the anthropomorphization of the of the LLMs. And I really recommend everybody to read the constitution of Claude, which is a document that they issued recently, and which is an unbelievable document in terms of the behavior and the relationship that they have with their own creation. It's a fascinating document. Thank you. >> Thank thank you. Any reaction on humanness? Otherwise, I have another topic I would like to before we do the round of closing. It has been a long afternoon. Any comment on it? Because I think it's a fundamental topic that we didn't put it, but this one and it's also part of our weakness in the answer. The the other thing I as Jay said, give us the US perspective. He didn't do it really except for one thing. He he presented himself saying on the defense side and on the offense side. And I don't hear it in Europe. We never We don't articulate like this. We are systematically on the defense side. We don't think we we don't articulate. This morning I heard a panel who comment that was the same. We if we don't my view, but I open for discussion, if we don't shift our mindset thinking yeah, okay, defensive part of regular and the offense which is not aggressivity and so the same word. On the we miss the point here. And and when I am I'm in the US, you all all a lot of us are regularly in the US, you will hear it. You say, "Where is the opportunity? Where do I grow? What is my my take on this?" We're all on the defense side uh our culture. And that's my what I see. Open for discussion comment. >> I'll just I'll just offer just quick response to what you said. You you referenced earlier the White House's new AI the national strategy that just came out on on AI, which was just revealed uh about a month ago. This follows on an interim strategy that was put out in July of last year. And the fact that there is a interim and a and a then a final version is not unusual, but what I think will be the case is that there'll probably be another AI strategy that will come out even during this administration because the focus areas are so rapidly shifting and the government realizes it needs to stay on top and ensure that all of the agencies that are empowered to address this have the right level of guidance necessary to implement it. So, I'll say I'll just on the one hand it's an iterative process in the US. The government obviously is taking it super seriously and say what you want about the Trump administration, this is not only rhetorically a top priority, but they are putting real resources into this from both a defensive and offensive point of view. Second point I'll say, if you look at the national strategy and I I was going to pull it up and I I could kind of dig into it here cuz I I have the long version, but I remember the short version, there are five main bullets, four of them are defensive. It's very interesting. It may be just a political framing, but it also shows the degree to which the government is concerned in a manner not inconsistent with Europeans about the risks related to AI or the importance of governance around AI. But, I will say the emphasis on innovation is is extremely important and even though there are only five, you know that innovation, even if it's embedded as number four, is the top priority and the one which the administration is really pushing the most. So, I I just wanted to make those two quick points to in response. >> Thank you. Any other comments on um if not, uh I would say yeah, please. >> just just uh maybe to to compare to to Europe, uh I I I I I think there are two issues in Europe. First, the venture capital market in Europe is in incredibly smaller than in the US. 80% of the capital the venture capital funding is in the US. And so, even when you do have uh young startups that innovate and that and that scale, like like Mistral, for example, Mistral immediately had majority capital in the US until ASML invested last year. So, that that's an issue for for innovation in in in in Europe. And the second one, I think it was that it's true that the European Union was much less offensive than than the US for for a long time. And if you look now at the the EU AI Gigafactory program that was launched last year, so the EU launched a call for interest to subsidize five AI Gigafactory in the in the European Union and with 20 billion subsidies. Um so, that was a very a huge program. Seven More than 76 consortium throughout the European Union answered. But in order to be awarded the subsidy, you need to show that also your clients Uh so, the clients of the Gigafactory will be sovereign. But the issue is that 70% of the European corporates are already on on US clouds. So, when you try to to to ask big industrials, big corporates in Europe to come into into sovereign AI Gigafactories, they're not interested. They're not going to move all their compute, all their cloud, all their data. They've just do the They've just completed the move to cloud on on on US services. They're not going back to to to Europe. It's a It's a It's a huge burden. So, that's I think even if now we are thinking about the importance of having more infrastructure and and keeping the data in the territory, uh the the the providers are all all US partners. >> I I I If I can give also one glimmer of hope in the health sector for Europe, because even if I live in New York, I'm a European at heart. Um so, if you take health data, this is where the fact that the public health system in Europe is an absolutely critical competitive mode versus a very fragmented private US system. Healthcare data in the US is also fragmented within insurers, payers, everyone is having different like Cigna, Aetna, all the different hospital systems. In France, for example, you have the Cnam, the the social security health data. Unfortunately, there's some limitations in access and I can go on a tangent on the Commission Informatique et Libertés if I was allowed, but Patrick is going to kill me. Uh but we have this wealth of structured, organized, centralized data in Europe that is a huge competitive mode versus the US and I wish we would use it more. >> Uh I won't conclude. I I agree on the my point on the mindset is that it's a mindset. It's not that we don't have assets. We have assets. And and the change in the competition policy uh goes in that. You remember the when they forbid the merger of Alstom and Siemens, that was stupid totally because the problem was not the rolling stock, but the data set. But now, a little bit late, but better late than never. >> Uh I'd like to come back on Christian, you've mentioned something which was very fashionable few years ago, data. You know, we had quantum, we had AI, we had cloud, and we talk about big data, data. It was, you know, the golden mine and whatever. Uh and when I look when listen to you Arthur, is drug discovery is a fascinating business because it 8 to 10 years is billions with a high potential of failure. So, how do you in 8 to 10 years uh life cycle on a on a molecule, keep the data safe because all the Sanofi of the world, UCB, Bayer, uh Novartis, I've talked to, they're very paranoid. about the data. They want to have the data in-house because I don't believe the cloud even the private cloud or secure cloud or sovereign cloud is safe enough because it's the IP of the company which represent dozens of billion of dollars is is really at risk. So, how do you see this Christine and and Arthur? >> Okay, so so thanks. The first first thing is that okay, maybe it was fashionable data a while ago. I'm I'm sad it's not a fashionable but fashionable anymore. It will come back. It always It's like fashion. It It comes back. And so I I would give you an answer that is about maybe trying to treat data as a first class citizen. Meaning that it it does require an investment to collect it. Once it's collected, it does require an investment to keep it to be able to reuse it because even big companies can create their own silos. And so it's about I would say having a long-term plan for what we're going to do with this data. Can we Can we factor it in the way we think about our financial future? And can we find a way to somehow an amortization of the way we're going to to keep the data. The question about the cloud uh Maybe if there were an answer that were like a public cloud compatible with the needs of of companies, then it would be cost efficient. If there's not, then I guess a private cloud might be a solution or or secure data long-term data storage. But it's um Yes, so if if data were to come to make a comeback in in fashion, then it would be a question to be addressed. And we would be happy, well, I'm speaking in the name of of an administration, but with with connections with the research world and the in the in the research and development world to make it in such a way that it can be beneficial also to the economy broadly speaking. I wish I could work with a company like yours to actually help you create insights. >> I I I love your question because it's the mirror of my point on the CNAM, the same way the CNAM is never going to give access to Sanofi to the SNIRAM. Sanofi is never going to give access to Novartis or or to others to to their own data. Where I'm super optimistic is looking at technological solutions where you can train the model without necessarily having access to the data and then extracting back your model and the user of the model will not see the granular data. So, it has a ton of issues on de-anonymization and and and and unwanted outputs, but I'm a firm believer that if research goes is better like evolves in this space, the data can remain locked in the vaults of CNAM or Sanofi or whatever, but you can still have the models like hopping on the all this data and then people using these models without having access to the raw data. We're not there yet, but I'm hoping that >> May I just add something to your to your point? So, we focus a lot on on privacy immensely and you say like train the models without actually well, focusing on the data part, but we need to protect also the trained models and we know how to do it actually. So, we we can work on protecting the data in ways that that you mentioned, but we can also think back as mathematicians, as statisticians how we protect the models and we can share the models and we know how to do it. We could literally trade on neutralized data. We know how to do that. >> Uh, we'll need to close the the session for today. It's uh No, uh at 7:00 I've been told, uh please join us on the terrace for the cocktail and gala dinner. Now, that's a message from the organization. You can tell them I've done my uh duty. So, to to close uh the the session, I would like to ask the participant in 30 seconds rapidly, uh you take after this conversation on the title power struggle and social chain challenges. >> Uh, it's long. >> Pauline. >> Um, I I would say now the concentration of of not only capital but knowledge uh uh and and fragmentation of the world is particularly a risk for me in AI. We are talking about entropy constitution and the the difference between entropy and and open AI and how they dealt with the Department of State, but the truth is none of us can really assess which one is is doing the best it can because none of us really understand how these models work. So, >> Okay. You had the first shot. So, concentration and fragmentation. So, noted. >> Also, >> Uh, being in this business for since 1977, uh I think we are in a critical moment. I think things are going too fast in my view, while I'm an entrepreneur, uh because when you give the tools to people where the only objective is to make profit, it creates a big spread between the normal people and the the happy few. So, and it's very difficult to create um some governance, you know? When you look at uh I was in the Silicon Valley when internet was launched and internet was an incredible tool, but there was no governance around that. So, it was an exchange uh platform for students. Uh and then it became an exchange between people and then it created um trillions of dollars. Uh so, the the net neutrality has been a nice attempt, you know, to distribute the value. So, I think it's and there is no way to to fight against the hyper scalar, the the billions of of dollars. So, it's like for me, it's I'm in this business, so I like it, but I'm always uh focused also is it good for people? And so, I do another ad for my conference in Laval, this is tech for citizens. Are we going too far, too fast? And uh how many people will be uh let down? It reminds me, remember when the IT came >> Thierry will shoot at me if you continue. >> I don't care. >> Yeah, of course it's me. >> When we had the Go Force, when the IBM put IBM, so there was 500 people left on the ground. So, it's it's good. Sorry if we don't >> So, speed and synchronicity, if I can summarize, Toby. >> I think uh what worries me most is the cognitive element of what AI can do on that place and praise on the mind of people. This goes beyond any military threats. It comes to the heart of civil society. Uh it can spin narratives very quickly. And the rate at which uh disinformation and misinformation goes out, there's no way to control it. So, all societies are going to be vulnerable. Uh especially in a world where there is going to be a bigger difference between the rich the rich is going to become richer and the poor is going to become poorer. And there are going to be a lot of people who are dissatisfied, upset, and uh governments are unable to control this. This was actually the job of the government to balance this. But, now they are not because they are run by populist around the world. They are not looking at that. So, I think there is going to be a huge issue from a societal point of view. And everything that we talk about AI and all is a subset. The larger thing is how do we all live? Thank you. >> Okay. So, deep fake and divide. Daniel. >> Um I think humanity is painting itself in a corner. I mean, you said Patrick that it is unsustainable, right? But, still we're we're we're everyone's working at, you know, constructing more and more data centers and mobilizing more and more resources in a planet where resources are scarce. I mean, you know, in hospitals you cannot take care of all the patients. You let some of them die. I mean, that's how it goes. We're not that rich. So, I think there's um either we sort of close our eyes and just go until we crash in the wall and there will be lots of uh lots of deaths of entre- entrepreneurs and uh and people. Or we somehow hold back and take this really seriously rather than give it just lip service. And I would can make two suggestions. One both of them one of them's unrealistic and the other problematic. The unrealistic one is that I think that in order to over to sort of counterbalance the incredible megaphone of the industry, the AI industry, which is has wagered hundreds of billions of dollars if not more and therefore has a vested interest in defending the idea that we should just go full head ahead full head along. In order to counteract it, we would need something like a a powerful voice that would be carried by something like a world council on AI. I know it's Yeah, I know I know there are objections to that. But nonetheless, I think it's it's important. Right now, there are lots of institutes that are about human-centered and you know, preserving values and so on. But they're sort of isolated little voices in the in in in in in the landscape and we need a stronger voice. So that's one suggestion and I know it's you're going to be skeptical, but I think I want to uphold it. And the other thing is that we must deepen our theoretical understanding of AI. Emerging effects, I mean, it's just a label on a problem, but it's a real it's a real phenomenon that we have to understand better. And so I think we must get to the point where we're comfortable, for example, with both uh both assertions following assertions. A, LLMs don't understand a thing. And LLM understand a lot of things. And we don't yet have the right conceptual framework to understand to hold on to those tools. >> Okay. >> Thank you. So, understanding and humaneness. >> Understanding humaneness and and and and and and >> Oh. >> You have two Okay. >> I'll choose one of you to answer first. >> I would say from hoarding to using. So, uh there was when data was popular, there was dataism. Now, there may be infrastructurism. So, we can do more and more and more and more. Uh but maybe instead of hoarding, we can use. And we can see whether the use has an impact that is positive and that we can actually measure. Um with it it's it's starting to be seen in clinical AI, for example, that investment is not exactly uh returning what we expect. And uh yes, so I would say that hoarding versus using. >> Perfect. Thank you. So, I'm not a Silicon Valley guru, but I'm a big uh believer in science and medicine and and what it has ultimately brought us. I'm very convinced that AI will bring us new cures and new medicine. It will also be responsible for dramatic mental health issues in our children and we'll have to juggle through both. And I think the two ways to do that properly is one, make data fashionable again. I think we talked about it many times. Uh and then the second thing is make sure that we have the appropriate regulatory frameworks and make them evolve at a speed that is much faster than what we have today because technology will uh outpace it otherwise. >> Thanks. >> For me, the throughput in this conversation is around two concepts. One is governance and the other is innovation. We talked about these two concepts at the macro level with the initial speakers providing us, me personally, tremendous insight. Both in terms of what the challenge is on the governance side and perhaps the overemphasis to some degree on the governance side in certain jurisdictions, but also the challenge on the innovation side in that finance, access to capital, data are increasingly concentrated in those that are winning and crowding out those that would otherwise seek to get into the game. And that's concentrated certain jurisdiction certain jurisdictions, which puts even more impetus on governance stemming from those jurisdictions in in my view. We also talked about it the micro level where I believe the information and and access to these tools is much more diffuse and where uh the ability to take advantage of these tools is effectively um out there for those who are willing to invest their resources and time into learning how to do this. So, I'm much more optimistic on the micro level than I am on the macro level. And therefore, I I was able to talk I think at some some length on the micro side with that optimism in mind, but I don't want you to leave with the impression that I am ignoring some of the major challenges we talked about that at that macro level. Thank you. >> Thank you. Thank you. Thank you. Thank you for my colleagues on the panel for your contribution. I think it was very rich. I hope for the participants, you get a a little bit of the picture. We we did in a structured manner more technology at the start from the bottom of the stack to the application layer where you find the business logic and trying to make sense of this very complex topic. Not to paraphrase Thierry this morning, well the beginning it's an easy one. So, and then still I think there is an urgency because of the means deployed and I very much concur with Jay. Probably today, and no offense to the world coordination, it will be at micro level where we will find where we find people like I mentioned Demis Hassabis or Modi Modi on Anthropic where we and if we can support these people who move the lines and and with the right mindset, the right ethic, and and and go and try to navigate, put your money also behind these people that only behind the others who don't have the same particular uh views. And and probably short term, that's what we can do and then try to build in parallel some more consensus because we've seen it didn't work so much. Thank you very much for participating.