#22 The Real AI Race: Energy, data & sovereignty - David Gurle, Co-founder & CEO of Antimatter

Welcome to the AI Social Club Podcast. I am Marco Bletri, and in this show we dive into the real world impact of AI on businesses, individuals and society. In each episode, I sit down with leaders from across industries and public life to explore how AI is reshaping the way we work, we decide and we live. We go beyond the buzzwords to bring you honest conversations and to make sense of the changes and folding around us. If you believe these conversations matter, please share the podcast, subscribe, or leave a review. It's really the best way to help us reach more people. You may not know David Duel's name, but you've almost certainly been impacted by the technologies his helped build. Over the last three decades, it has been at the forefront of some of the biggest technology shifts of our time. From the early days of Microsoft alongside Bill Gates, to the development of Skype for Business, the creation of the Fentac Unicorn Symphony, and more recently HiveNet and PolyCloud, David has consistently spotted major technological shifts before they became obvious. His latest venture, on T-Matter, may be his most ambitious yet. Instead of building another AI application, he is tackling what he believes is one of the biggest constraints facing artificial intelligence. Energy. In this episode, we explore the hidden infrastructure behind AI. Why data sovereignty is becoming a strategic issue for every organization, how open-source models are changing the industry, and why the real challenge isn't building smarter AI, but making it accessible, sustainable, and independent. It has been one of the richest discussions on this podcast so far. If you want to understand where AI is really heading, this conversation is for you. Hi David, welcome on the show. So you've been building internet infrastructures since people didn't even hear about internet. You build the ancestor of Microsoft Tim, you were an advisor for Bill Gates, you led Skype Enterprise, you as well built a Fentac Unicorn Fentac called Symphony. You built more recently, HiveNet, PolyCloud, and now a few weeks ago you announced a very ambitious project called Anti-Matter. So in this podcast, I want to understand what you're actually seeing that everyone is probably missing. So looking back at your journey, what is the thesis that connects all these experiences? First of all, thanks for having me. It's a pleasure to spend this hour with you and with your audience. I think it's about understanding the momentum of things. I believe that there are a number of patterns which are always consistent and recurring in the way technology comes to our lives. There is always something that exists and there is always something that is in the works to replace something that exists. And I think I have the ability to understand when something that exists, which is at the end of it, or it is close to its end, and when something is about to mature to potentially replace it. And I have that understanding, which I think happens because I'm a big student of technology history and human history. And I try to always find the equation in which they come together. And I believe that our history is massively influenced by technology. And after all, humanity is all about building and using tools to further our survival and our success. And I think understanding how tools shape that characteristic of ours is something that I don't know. I've naturally had a pension for and I think enough experience to see those things before others can. So what's your next bet? For me, the next bet is around two axes after antimatter. One is going to be around urban mobility. I think the urban mobility as we consume it today cannot be continued. And we are arriving to the maximum of number of lanes we cannot. I think we are coming to an era where I know if you watch the movie from Luc Beston, the fifth element. I think at one point we are going to arrive to that, not in the same form factor as he imagined, but not too far. And then the other element is very much going to be around education. I think the education, the way we consume today, all the way from kids to adults and typical how we get trained today, is going to be put upside down with AI based education. The classroom education as we see today is going to disappear and be replaced by something else, hopefully for better. It's always for better. And for me, that is the second axis that I see, you know, the end of an era and the beginning of a new era. Just looking into education because it's a topic that I personally am really interested in too. I think in the right now the generation of students is really in the middle, you know, like they are still educated the old way, but they are preparing for a world that is getting ready for the new way. So how do you see things changing on the side? First of all, I see things changing faster than ever before. And I think that it is certainly not the public institutions who are going to lead the way. They tend to be very stuck in the past for multiple reasons. I think it's going to be led by the parents first and foremost. I think parents now can find a substitute that they couldn't find to teachers. And I think that they're going to be able to take control of the education rather than delegate that to the institutions. And, you know, private institutions are business, you know, are a business. So they are going to catch up to that. And they are going to be the ones I think who are going to cater to this new form of education. And I believe that's going to change, you know, the way kids grow up. Something in that the way I did. Yeah, it's funny because I think the same. And I think my generation of parents, we are, parents are having a way more central role in education than probably your generation. Because you're a bit older than me. And I think we tend to believe before that we could delegate everything. And now we don't have any choice anymore. And when I train people around AI, I always tell them, like, now you have no choice anymore because AI and technology is going to completely change the way your family is organized. So you have to look into this and you have to take your role. If you delegate it completely, your kids are going to get lost. I'm pretty sure. Yes. But there is one topic I wanted to start with actually. So we're going to talk about AI. That's the main topic of this podcast. I want to talk with you about energy. When we talk about AI, we usually talk about technology. We talk about the systems and we think systems are the bottlenecks. And what I heard from you is for you, energy is going to quickly become the bottleneck. That's probably why you built antimatter. You can tell us a bit more about it. Can you explain to us why energy is going to be one of the next challenges? And the reason for that is the. So an artificial intelligence program. Let's call this a frontier model. That's kind of what most people call it as today. Whether it is a world model or a land based, is a mathematical computational heavy construct. And so in the. We are not even celebrating the third year of the Chajjibit. It happened in November 2023. So let's say in just three years, I think we've achieved so much progress that for lots of cognitive type of work, we can substitute AI for human beings. And actually we'll get faster and better results. And so I think the premise that we're going to have a form of intelligence that is good enough for most of our needs is already proven. So therefore we are no longer bound by the limit that the frontier models were at just three years ago. And now people can imagine for the next three years, it's going to be even better, even though the progress that we made in the next three years is going to be incremental less than the one we had made in the last three years. So therefore that's not the limit of the use case, nor the usage that we're going to get out of it, or the benefit we're going to get out of it. So if that's the case, then is the sky limit? In other words, can AI programs from tier models serve 8 billion people today? And giving them a usage of several hours a day, the answer is no. No, definitely not. And the main reason for that is going to be. back to what I said is that these programs are computation heavy and statistical models, probabilistic models, lots of, I would say, recursive checks of the models that they have built and, etc. creates a significant burden on the expectation of result in real time. So if we had an hour to wait after each of our prompt, I will have a very different set of constraints to put on the table. But when we are looking for real time results and more and more real time results, then it becomes depending on the machines. And machines to cope with that growth, which is literally vertical. And with the real time expectations of the responses, have to consume more and more power. And so much so that some of the data centers that were built three, four years ago can no longer be tailored or adopted to the new type of server farms that are needed for the future generation of AI. And so that becomes truly not a problem of even chips, but the power those chips so that we can deliver them the necessary energy so that they can respond to these expectations and this demand. And hence, the bottleneck is today energy. So going back to antimatter, the concept of energy is very interesting. You could either think of energy being limited. Say, okay, there is so much energy this country can produce. And therefore, if you want to build more data centers, therefore add more energy, well, it's going to be more investment, it's going to take time. Or we can't do it for a number of reasons. Could be sustainability reasons, could be with great problems and et cetera. And that's pretty much been the traditional approach, the current approach of the companies who are building those data centers and the governments who are supporting them or helping them to do so. And my view is that it's a waste and not because it is the wrong thing to do, but actually because there is only a certain amount of the produce energy that is used. And so there is a huge amount of produce energy that is not used for a number of reasons. And the approach of antimatter is to observe that need and expectation we talked about. Look at what are the limiting factors, power, energy. And then look at availability of resources and say, okay, is there a better way? We can solve this problem. And our approach is very simple. We take data centers. We'll talk more about what they are. To where energy already exists rather than bring energy where we want to build the data centers. And with that, you are completely reversing the problem. So you are saying, okay, what is the available energy and where? And you realize that there are millions of spots, we call them micro power farms, which are available pretty much all around the world. And they are ranging from one megawatt to five megawatt sometimes even higher. But they are small that it is not interesting to build a large data center or you can build a large data center given the power requirements of new servers. But if you are thinking about a decentralized and a distributed approach to the problem, then you suddenly have a perfect match. And that's what we are doing with antimatter. We go to where electrons exist using a decentralized distributed computing capacity. Can you give us practical examples of where we can find energy and use? Oh, very simple. So let's take Etisalat here. We are in Dubai. They have as you know, all these towers. And those towers are connected to internet. And they have obviously power available there. And they have a parking lot or they have some fence around it. And so, you know, the idea for me was, what if we can put like a 20-foot container, a 40-foot container within that fence perimeter of Etisalat tower. And we powered up between 70 kilowatt to 450 kilowatt an hour consumption and deliver enough capacity to run a local inference model. And I'm sure there are hundreds of such towers across the UE for Etisalat. But I can make the same arguments for cities in France. Cities in France have parking lots left and right. I'm just talking about municipalities and their garage in which they park, you know, all the vehicles that they have to do. And that's what we've done in Can. We put one of the public clouds in a parking lot. And there is, like, they have about 250 kilowatt capacity. So we put a small container there, 104 GPUs. And it's serving the whole community. And so, and those pockets are everywhere. Yeah. Yeah, that's what we actually don't realize. How does the, so we talk about, we'll talk a bit about public clouds. Just for people to understand, how does, like, what does it look like? Is it huge? It's complicated to set up. You say it fits in parking lots. So just to give us a bit of a concrete idea of what it looks like. Well, very concretely, I'm sure that if you have been driving, you know, near Dubai and or elsewhere, you've seen those trucks with those containers. Yeah. This is maritime containers. So they have three standard sizes, a 20-foot and a 40-foot. And I think the 46-foot, it's called the big model. And so we picked the first two, so the 20-foot size container and a 40-foot size container. Within which we have built a data center. Okay. And, you know, if I were to, like, blindfold you and, you know, open the door and close the door and put you inside that container, you will think you are in a large data center. Okay. You know, that's how similar it is in construction. But one major difference beyond the size is that it is not water-cooled. So it's entirely air-cooled. And so therefore you have a much less impact, you know, as it comes to its sustainability and ecological footprint. Now the size is very simple. It is about three feet large to six, seven foot length, sorry, 20-foot length and for 40-foot container is 40 by three. So it fits between, you know, 25 square meter to 100 square meter depending on how you configure it. And so a physical form factor is very easy to see. And because it's standardized, it's relatively easy to procure, you know, at least a metal frame. Yeah. And inside those, we put between 104 GPU for the small version to 360 GPU for the larger version. Okay. And so that is, you know, what we call, hell of a power in order to provide the inference capacity to local markets. Okay. And so you sell the solutions to whom, to tech providers, to governments, to governments, energy providers, tech companies such as Etisalat, and they are not our customer, but I'm just giving an example. You know, companies who are really looking to have their sovereign solution. And we also deploy a few hours. So the idea here is to take advantage of available energy and power up those servers so that we can actually, honestly, democratize access to AI. Yeah. Because what's going on these days is that the frontier models are consuming so much energy and that is expensive for many reasons. So that they are increasing, you know, the prices of monthly fees today. If you want the best of the best, you're going to pay 200, you know, dollars or euros per month. And that means that those who will not pay that much, you know, they are going to be left out and they are already left out. And so we are going, we are seeing already a divergence, right, between those who can afford and those who can not afford. And there are billions of people who are not going to have access to that if this continues this way. And so by, we are building and hoping that by our approach, we are going to be able to lower the price, therefore, decrease the buyer to entry. And with that, it will result in more fair share of the use, especially going after open source models. And our objective is to help humanity to enter this era in a responsible, in a demo critic way so that everybody benefits from this wonderful and marvelous technology. So I understand there is a lot about energy, responsibility, there is a lot about cost, obviously. And then about sovereignty as well. So can you tell us a bit about, I mean that's a topic that is I think really interesting because I believe a lot of maybe not government but a lot of companies still don't realize the importance of it. A lot of people will go by default, which is fine, to the big US cloud provider. I press getters, yeah. But I'm not sure everyone understands what they are actually giving away. I'm not saying it's wrong solutions, but what are you giving away when you go with these hyper scalars? Why should we ask ourselves the question about sovereignty? Can you tell us a bit more about this? Absolutely. First of all, sovereignty has been a topic of fashion only in the last two years. It's recent. It's recent. And I believe that Trump has definitely helped the first of all the European governments. I'll start with them, but in UE and in Saudi, the perception of sovereignty was much stronger than the European ones. And so the topic here has always been a number one versus in Europe. And fundamentally, it's very simple. If we are moving into an era where things are more and more digital. And AI is only accelerating that trend. And there is not going to be a digital world without AI tomorrow. It's not going to happen. Everything is going to be around AI. We talked about education a few minutes ago. And think about it. Every AI needs to process a form of data to be trained. And then every AI is going to process you as a person. And so every prompt you enter, every interaction you have, it's a small trace of you. And this small traces are up and they eventually become you. And so do you want you to be in the hands of someone else, which has other commercial interests, other government obligations than yourself? And I think that's really what's at stake. Governments for strategic reasons, for geopolitical reasons, definitely get it. You know, US is the one of the most sovereign nations. And that's not even talk about China or Russia. And so every nation has obviously a very good fundamental reasons to be sovereign. And what sovereign to means is that the data that is at rest, the data that's at motion, and the data that's processed is yours and yours only. And you have full control over it and nobody else has access to it without your permission. And nor by attempting to have access to it by trying to steal it. So that for me is the true definition of sovereignty. And there's a consequence to that. And the consequence to that is what I call independence. And so in other words, that you may choose not to be sovereign, but it's a choice. And if you are not thinking about that, then you become dependent on this. And that is what most companies are unfortunately doing. What they are doing is that they are going what I call native with hyperscalers. And so hyperscalers have done a fantastic job in trying to compete with each other to develop what I call the native versions of open source software. And they really bet on the laziness of, I would say, you know, human nature to say, you know, why do we have to set up a server? Why do we have to manage a server in our environment? Why do we have to install the software and manage all the patches and all the updates? We'll take care of that. Here's a service for you. And that service is the abstraction layer that they use to get you addicted to their infrastructure. And yeah, it is true that you saved a ton of time and money, not as much as they pretend it is, but they charge, first of all, a much higher price for that particular point. And second of all, is that they are not really adding that much of a value because the open source software keeps evolving. Maybe it's six months, maybe a year behind, but it is not even maybe not that far behind for many of the use cases. And so what happens is that the most companies who do not have the trained IT organizations nor they make IT as a priority. That's a very good point. They are forcing their IT, you know, the media resources they have to rely on hyper-scalers without understanding the long-term implications of sovereignty and cost and dependency. And I think it is time for them to wake up as the governments in Europe had to wake up. They've been triggered by, you know, Trump policies, but fundamentally, they knew about that for a very, very long time. But now they are finally acting on it. And I think it is now time for any company to be at least aware of it and make that choice and not be, I would say, dependent of a choice of others on your data, on your economics. But do you think these countries that are not US and not China, for example, UA in France, we talk about these ones, or smaller countries? Do you think they are capable of this? Like we do a lot, we sign a lot of acts, you know, France 2013, we have a lot of AI acts, etc., etc. Do you think they are capable of relying on these big, on these hyper-scalers? Do you think they are there? They are very capable in the best case scenario. I'll be very transparent in this, and honest, you know, as much as I defends sovereignty and all of the stuff that we discussed, you know, it will be the same thing today to look for a memorabila in a flea market that you are really, you know, keen to have, you know, on the top of your chimney, okay? So if you, I mean, let's think about it. If you want to really get rid of Microsoft Office Suite and not use Google, good luck. You can. Yeah, of course you can, but I mean it is hard. It's hard. We are not making it easy. The market is not making it easy. The governments are not making it easy. And the hyper-scalers are doing everything they can to make it seamless, easy, painless. And so, I mean, every startup who raises, you know, their hand, they get, you know, $100,000 or $200,000 credit from Google or Amazon or Microsoft. Why? Well, it is, it's a drug, you know. Once you are used to those APIs, switching costs becomes much higher. Of course. And they just then hold you hostage with their applications. You know, whether it is Google Sheets or Microsoft Excel or whatever it is that, you know, you use it every day for your email. You are there hostage. And obviously it's your willing hostage and most of the people know the consequences. But yet go and find the best email program that is running on open source, that is entirely sovereign, that you are okay with. Yeah. Which is sometimes a complete, even switching like the, I mean, see if you look at companies that are not tech companies, a beauty company, a luxury company. As you said, IT may not have that much power, that much budget or priority. But it's a complete switch of priority to be like, okay, we will completely like rework on if you look at global companies like L'Oreal, LLVM, you know, these companies that are a bit that are actually relying a lot on technology, but still they obviously use these technologies. It's a completely switch, it's a complete switch of strategy. Yeah, it is. To invest. It's going to take a lot of effort and time. Yeah, and it has to be put as a priority, which I think is not the case. Because when it's, when you're a global company, AI or not, by the way, maybe we talk about AI, and you have had like regional headquarters that are quite solid in Asia, in Europe, in the US, what choice can you make? Do you become, you know, how do you manage sovereignty when you're a global company that is actually implemented in a lot of different regions? Yeah, I mean, the best way to do it is to use your own local resources and encrypt everything. But that requires the right IT organization. And guess what Microsoft does, right? I mean, Microsoft invites your IT team to executive briefing center in Redmond. I know I used to go there all the time as part of my job at Microsoft. You know, dying people and make them meet with top level execs from other companies and build gates during my time, etc. And, you know, it's a very seducing act. And so you create that what I call emotional debt to a large degree. Which is a very good strategy. And then, you know, then people draw from Microsoft draws from them. And then when they send them, you know, the next three year enterprise agreement, guess what? You have that path. that to pay off. There are very few head of IT departments who can say, well, I'm still gonna go with my own choice versus, but I've been induced to do, and I don't care about that particular debt that I have towards Microsoft for taking me and dining me at Enciadol. It's part of reality. We don't talk about it. - It's a defense, not a first. - But it's part of reality, okay? And they have a huge volume of supporting materials. I remember I sold during one of my earlier companies, Symphony Software Stack to the banks, and it's used for real-time communications. And I'm entirely compliant, I'm entirely secure, I'm entirely sovereign. This was absolutely 100% sovereign. They were very happy with it. And then one day, one of the big banks, CIO, says, hey, you know, they went to Microsoft, CIO, and we had this conversation about using Teams, my former product. And I think we're gonna switch to that. I told him, I said, you're kidding, right? And he said, yeah, I know why. Well, because it's absolutely not sovereign. What about the US cloud act? I mean, you are a bank, that means that all the data you have is a French bank. It's gonna be in the hands of the US. Oh, no, but they convinced me that actually it's okay. Yeah, but it takes a lot of education. Yeah, and the answer was a full lie. Yeah. Because just last year, when there was a Senate hearing in France, the top Microsoft executive came and said, in front of that Senate hearing, that no, any data that is run on Microsoft is subject to US cloud act. Of course. And so it was like, they managed to, until loop is the same French, you know? The group CIO of a top French bank, in believing that they are delivering the right security framework. And so when you have so much capacity, it's very hard for, I would say, an average IT person to go and say, okay, you know, I'm going to fight against that trend. No, it's almost impossible. But I'm not surprised because if you even look at an individual level, you can't imagine the number of people who don't realize that when they chat with chat GPT or cloud, do they talk somewhere? You see what I mean? Like it just goes somewhere, it's not on their laptop. And like 70% of the time when I talk about this, people are surprised that there is a big lack of education on the topic. So I'm not even talking about big topics like, I mean, about-- Yeah, I mean, I was thinking about that the other day. And I think it is about disclosure. Yeah. I think each of these companies should disclose where their data is, what's they are doing with it? Yeah, it's just transparency. And transparency and say, okay, you know, this is it. If you use us, this is what we do. And they're going to say, of course, it is there. It is in our ULA, you know, a user license agreement. Who reads an user license agreement? Yeah, for any type of project, of course not. Yeah, and there is the big deal. Then we make as well a difference between, okay, my data is stored there, but it looks safe, but it's not going to train the model and the one of the standards that it means. And it leads me to my next question, actually, I was about a round training the data. If I understand well, correct me if I'm wrong, you are making a, I mean, with antimatter, you are working with inference. Yes. Can you explain us what it means? And what and how it impacts architecture infrastructure? Yeah. So if we were to distill down AI programs, they fall in two categories or two phases, maybe. We'll be the more correct version of saying it. One is you to train the model. The other one is that you use the model. Yeah. So when you use the model, we say it's inference because we are inferring the next thing. So it's like our brain kind of inferring. It's like a logical, you know, deduction of what's coming next, okay? So the world inference, therefore, describes in a generic term the use of any AI program. Yeah. Okay. So antimatter is tuned for AI inference. Not for learning. As the learning requires very high number of computing capacity, all collocated and each server interconnected with very high bandwidth and speed connectivity in order to make that learning happen as fast as possible. And, you know, setting up a micro data center in a small container of 20 or 40 foot size is not, you know, tuned for that type of AI use case, which is learning. However, it really works well for inference because when you are in the inference use case and especially the models we target, which are open source models, they are more frugal in their memory consumption and in their compute consumption than the frontier models. And we'll come to that because that's a very interesting fact. And therefore, they can fit in smaller size than the frontier models which need much larger compute capacity. And with that, we are capable of addressing the compute requirements for such models. Yeah. So I said, I'm going to come back to characteristic inference models. Inference, I would say, programs. So the world of AI, as we see today from an end user point of view, is being split in two universes. The universe of proprietary models. So you mentioned OpenAI, let's call chat GPT because OpenAI has different versions. You know, you mentioned Cloud, but there is Gemini and there is GROC. And then you have the open source models, where deep seek is, where Kuwen is from Alibaba, where Lama is from Meta, etc. And then you have the versions of proprietary models, which are also open source. But in general, six months after. OK. And so what's going on is this, right? The average person, consumer, or average company, is going with proprietary models. And they are paying whatever they are paying either at the token level or at the end user license level. OK. And if you are sensitive cost, you are sensitive performance and cost ratio. And you are sensitive to what we call fine tuning. So in other words, training the model with your data and sovereignty and flexibility, then you are going with open source models. Yeah, 100%. And so open source models are now far more powerful than what they were. And but something magical has happened. Due to the export restrictions of NVIDIA chips into China, the Chinese companies have learned to do more with less. And so they became frugal in nature of using the complete capacity because they didn't have as many as others. And so I'm referring here to chat GBT and Android Pt, etc. And so with that, their models, their open source models, are capable of running in much smaller servers in terms of capacity than their American proprietary counterparts. And they are as good sometimes if not better than these models. But certainly much better in terms of price performance ratio. And that's where the market is heading. And that's what we have built on T-Matter 4. So what in a company, I guess, your client will be-- I mean, the client of this type of solutions in general could be like the CIO or a city or-- what's the question he should ask himself? First, I mean, because most of the companies-- let's maybe I'm wrong, but most of the companies especially here-- I don't think energy is their number one challenge. But so very interesting is. So what is the question? Because to be honest, most of the companies here that I see, they go either, as you said, for proprietary models. Or what they do is they aggregate a number of proprietary models and they say, OK, this is ours, it's a safe, et cetera. But I don't see a lot of companies that are not tech companies that actually go with open source models. Maybe because they don't have education about this, maybe because they don't ask themselves the right questions. So how do you guide them towards this? Yeah. I would say there are three statements they need to make. First, I'm going to use the AI hell out of it. Yeah. Yeah. So that needs to be in the roadmap, in the objective across-- Yeah, that's a very good question. --across all disciplines of the company or departments. And the reason I started with that is because of what's come next. So if that happens, if that's my objective, there is just no way on hell. I should take my data, which is pretty much my IP. Yeah, of course. Give it to anyone else. Because why don't you do it now? Yeah, of course. In this case, just to give your IP to anyone in the world and then let them use it. And so using-- That's just a bleached-- --using a proprietary model. And use your data to train on the proprietary model is a key to doing that. So therefore, the second act is, for whatever AI I'm going to use, I'm only going to use AI in which I can enhance it with my own data. because that's where the real thing is. value is not only using the generic train data. And then you can never stop that. In other words, there is no end to it because once you start, your data keeps accumulating and so on. So that means the third thing, which is then I have to build competence in my organization whether through AI or not or together with human and AI in order to run that thing for me by me and not be dependent on a third party. No, it could be a service organization, IT service organization that's fine. But as it comes to building your AI DNA and making that AI DNA part of every DNA of other products or services you build and you consume, you need that asset and that's yours and yours only. Once you have made that decision, then you have to look at what is the price performance ratio where it's going to fit my need. So for me, these are the steps that they need to follow. So what's the, if you go, if you meet tomorrow, like a business leader, not a tech guy, what's the one question you never ask himself to do one question you should ask himself? Well, I would just say how many people you're going to replace in your organogram with agents? You'd stop by this. Interesting. That's the first question I will ask. Okay. Do you think you're going to have the answer? Hopefully yes, but most likely no. But hopefully it will trigger the right thinking process. Because once you think this way, everything else comes from, comes from that, you know, is a logical suite of thought processes. Yeah, I'm surprised and glad you actually turned it this way. Because I read about you. I don't know when, I think it was recently that you were a bit skeptical about autonomy, like AI autonomy or AGI. Is it still the case? And how, how do you see AI evolving in the company? It's like you're talking about replacements. And it's so thick that every time we talk about replacement, people are like, no, no, so you have two type of people. Some that are like, don't be naive in two years. No one is going to have a job anymore. And the others that are like, we're lying to you. You have a lot of value. You will never be replaced. So how do you position yourself on this? In between. Yeah. In between I believe. Yeah, I'm still very, I'm still very skipped about AGI, AGI, defined as us. Yeah. Right. You know, but we can't, we cannot be reduced to a thought process. Yeah. Yeah. You know, it's certainly a big part of who we are. Yeah, we're not just a set of tasks. But it's not only us. Right. We have, we had a number of other dimensions. And, and I think that when people talk about AGI, maybe, maybe definition of AGI needs to be fine tuned in order to, to build consensus around that. Yeah. But I believe that for any information processing task, any AGI program now onwards is going to be better than us. No doubt about that. Yeah. It's not going to be as human as us. But it's going to be better than us in terms of speed. Yeah. And in terms of quality. Yeah. Now, that doesn't mean we are replaced. Yeah. But that means that we are augmented. Yeah. So in my view, in my view, the agents that I mentioned in our previous minute, one minute ago was it isn't about having a head of HR. I'm picking head of HR because this is the least likely you think that they would use AI. But it is the head of HR running five AI agents for themselves in order to help their compliance, their recruitment, their performance management, etc. Right. And today, if you are not using those, then you are putting your company, I would say it a massive handicap versus the others who are using it. And, and are these agents, if they, if done, are they going to replace people? Yes, they will replace people. There is no doubt about that. Right. I know someone who has built an entire recruitment system using a number of agents. Yeah. And this person from from a HR organization has been doing it now for a year. And, and, you know, they no longer need to use some recruitment agencies. They were using. So, so they have reduced their dependency, they reduced their cost, and actually the the quality of the candidates they came through is better. It's much better. So, so the productivity gain is there. And, and as a result of that, you know, you have not hired someone to do this job because you found someone better in between brackets to do this job. So, that's reality. And, and I think that it's very important to understand that and use it to your advantage. And we have to see this as augmentation first. And once you augment someone's capacity. So, in other words, instead of having two hands, they have four hands. Yeah. Right. So, even one brain, they have five brains. Yeah. Well, you know, your productivity goes up. And the consequences of that is that your organization becomes, if I were to say that if we could measure the density of intelligence versus the human behalf, if it was one to one, right, for every human being, there is only one brain, right? Actually, we are going to go for every human being, there are maybe four, five, six brains. So, in other words, your density of intelligence is going to increase massively for a number of headcount. And I think that's how we should think of it. Yeah. Yeah. And then the next challenge, which is coming, which is coming already for the most mature companies about how do you organize this? How do you maintain this? How do you scale this? Can you tell us for you personally? Because, okay, you're launching a new company, which is a bit of a merger of your previous companies as well. So, it's not really starting from scratch, obviously. But I believe AI is in your processes. You have AI in your systems, etc., etc. So, I'm not talking about the product you say. You say sorry, but about the way you are organized. How much do you integrate AI already or how much do you project? You're going to integrate these agents or any type of AI solutions in the way you work and how much do you think it's actually going to help you scale and accelerate a lot. Yeah. I think it's worth sharing our journey. Yeah. So, if I were to take us back 18 months ago, 18 months ago, so almost 18 months into the roadmap of AI is as three year journey that AI is in. When I did a survey into the company, you are about 50 people back then. About five or six people said we like AI, we use AI, and the other said we don't like AI, we don't use AI, or we won't use AI. Yeah. Yeah. Yeah. It happened a lot a year ago. Yeah. Yeah. And there we ask why. Of course, you want to collect the right intelligence. Most of it was, it was my time. It was like, it is a time-waster for me, rather than prompting and iterating, et cetera, I rather do it myself. And surprisingly, the engineers were about 40 engineers back then. They were like the most resistance, you know, organism, organization against AI. And so, I run the survey just three months ago, even less than three months ago. Now that number is the opposite. You know, we have same people, we are bigger now, 60, but it's the opposite. So, now 18 months later, everybody is using AI. Now, most people would say they use AI second. I want them to use AI first. So, what does it mean? Is that, oh, I write a code and I ask AI to check it for me. You know, I want you to not write a code. I want- It's out first with AI. So, we are not there yet. In some teams, we are in some teams, we are not. There is still some resistance. But at least on software production engine, we have made that shift. But we had to really force it. You know, I mean, force it in other words, not say you have to do it. But we had to create incentives, programs, communication, you know, bunch of things. Yeah, the right environment for them to go. And lots of surveys and discussions and 101s. You know, it was a conscious effort to turn us around, you know, from no usage to what I call second usage today. And our next step is first usage. Which is, okay, from now onwards, we use AI first. So, we want to be AI first company. And why I say that is because, again, things have gotten better. We have a sovereign way of running it. And the team is not excited about it. They created this AI club. Like every once a week, they come together, we are all remote and everybody shares you know, around the coffee, virtually, you know, they're AI Studio of the week. Oh, you know, I've done this. I tried this hack. Have you tried this extra? So this has created a new culture. And that new culture is this, is dragging everyone else along, along with them. So we turned this thing around, but it was a conscious effort to do so. Yeah, I'm glad you mentioned this club because that's something I always tell people when they think it takes so much effort and cost as well to train the people. And sometimes I'm like, as I said, like every week or every month try to put people together and have people sharing about best practices. Already if you do this, the way you had the others. But I'm surprised you talk about resistance a lot. So for you, that was the number one bottleneck inside the company. Not the technology or the tools or the bottleneck was around resistance. Was it because people are scared to be replaced? So was it because they didn't, you said they lose a lot of time. But I mean, after a while you understand it can save you time. So what was the. I think it was trust. Trust was the tool. You know, trust based of time. It wasn't about, oh, it's going to replace me. I mean, maybe it was, but it wasn't explicit in at least the survey responses we got. It was more like, I'm not sure if I'm going to trust this new transport system to take me from A to B. Like imagine a Weimo in the US, it's kind of popular and some people will be still Uber versus a Weimo. I'm not really sure, you know, Weimo is the right way for me to go with the driverless car. So I think we were on that stage of our maturity. We turned it around. Now everybody knows how to trust. Quite quickly for me not enough. And people are now growing into it and they are growing with it. Interesting. So I understand the ambition of Enzymeter. I know it's probably one of your most ambitious projects. Certainly the most today. Yeah, probably. So if you succeed, what will be. I like this question, but I want to know what will be better in the world thanks to Enzymeter. Because you talk a lot about impact. So what is it going to change? Well, everybody can afford to use the latest and greatest of AI programs they like. And so they are not left behind or left out. And the impact of their use isn't killing our planet. Thank you so much, David, for joining us for this discussion. It was a pleasure to have you. Thank you very much, thanks for having me. Thank you for listening to this discussion. If you liked it, don't forget to follow, share and leave us a review. See you next time for more AI insights.matter. [Music]