That's correct.

I think I'm rather in the middle. I'm not in any extreme positions. I think it's a very important new tool for, well, many different disciplines, many different domains. But I would not go to the extreme and say that it will finish work as we know it or it's not going to have an impact. I think it's really in the middle; it's an important tool, it has a lot of potential, but as it always happens, when you have such a new powerful, complex tool available, there's a lot of fear surrounding it. I mean, it happened with electricity, it happened with television, it happened with the internet, and it's happening now with AI. I think it's sort of how we handle new things.

I'm on the excited side. Again, it's a new tool. I think we don't really yet know up to which point it can become useful. We're probing right; we're trying to apply it to many different domains. We're trying to find the best way to make it work in different domains, and in some cases, it's already providing results that are astounding. In other cases, we don't see much relevance. We're learning how to use it. But it's a very good moment for a new technology's practical deployment at least, and you discover little by little what the actual implications of it are.

I would say it was a bit earlier than that, but not that much. I've been working a lot with AI applied to mobile networks. But then, as a user of AI in other domains as well, I think the real first applications that had my jaw drop were applications to videos, video editing. The capability of taking some facial expressions and applying them in a video with a new face, for example. When I saw that published in a magazine many years ago, like 10 or so, of course, it was very early stages. But when I saw that, I said, Hey, how can this be possible? It's amazing. And this is, I think, the moment when I realized that or at least a moment that really struck me from AI.

It's IEEE has this magazine, monthly magazine that is called Spectrum. And it provides insights on future technologies, new things that are coming up. And this is where I saw the technology. At that time, again, it was 10 years plus ago, so at that time, it was not really working. It was more of a concept. But they were showing that with very simple cases, they could make things work with a lot of effort and so forth. So, it was not yet the era of deep fakes, but those are a consequence of that.

I'm actually one of the scared people as well when it comes to these kinds of tools, the power they have, and the fact that they are so accessible. So, I'm strongly in favor of regulating AI. This comes from an ethical side and also, I would say, from the side of controlling in some way what is produced. If you make a tool that can generate deepfakes openly available to anybody, I mean, we're going to see something similar to what happened with social media. Everybody can tell whatever they want, and it becomes a war of tweets or whatever, where nobody knows where the reality is. And deepfakes go along that direction. And I think that if we don't have a way to control what is fake and what is not, this is going to generate trouble for sure. So, we need some technology, some way to force somehow deepfakes, similar types of products, let's say, being watermarked or somehow being stamped in a way that you know what is true and what is not. Because we already have troubles now with things that are not fakes. If you add that technology on top of it, things could get out of control. So, regulation, definitely, yes.

I agree, it's not easy. I live in Europe, and I've been through all the data regulation that we have at the European level that then has been inherited by the different countries in Europe. I think Europe is leading the way, at least in terms of data privacy regulations, and it's not simple. Many times, it's cumbersome, many times, the results are not those that were intended. It's just adding overhead and not actually preserving the privacy of the user. So already, the data side alone is very complicated to regulate. And of course, as you say, when you talk about AI, it's multiple dimensions. So, I'm not saying it's easy. What I'm saying is necessary. I'm not a law expert, or I don't have all the needed expertise to tell people where this should go. And in terms of time scales, we need to act fast because if we wait for 10 years before we regulate this, I mean, I can easily see situations where people get really confused about what's going on around the world because of it.

Yes. So the company develops AI driven solutions for network operators. So our, let's say, target customer is typically a mobile network operator, talking to, you know, large players like, I don't know, Orange, Telefonica, AT&T, and so on and so forth. Those are the industry that manage the cellular network that allows mobile phones to stay connected to the internet all the time. And this architecture, these mobile networks are extremely complex infrastructures, pervasive, that are deployed over countries, and whose management is becoming more and more complex also because, of course, as you know, mobile applications are becoming increasingly complex. They have requirements in terms of latency capacity that are essentially becoming every single year more stringent and operators need to keep up with those requirements, so they improve their networks constantly, new technology are deployed, but this also makes the networks much more complicated to manage. And this is where we come into play essentially.So our goal is to develop solutions that help operators manage their mobile networks.And the whole concept of the company stems from the fact that mobile networks are very complicated systems, very large ones, but they are also systems where there is athe richness of data is quite, quite amazing.So you know, imagine all the traffic that flows through the network. I'm not even talking about the user topic, I'm talking about controlling traffic, right? So the information that has changed within the network to allow users to use the network.All this information about tens of millions of users flows constantly through the network and, you know, AI saves some data. So when you have that amount of data, the concept is you can use the data to transform the network in sort of at least partially autonomous system that observes what's going on in terms of traffic demands, learns from the patterns in the traffic demands and optimizes its own operation based on that data, relieving the operator from a lot, a lot of manual engineering work that is happening today.

Yeah, so this is at many different levels because, again, the networks are very complex infrastructures that are separated into domains. You have the cell towers, essentially. You have a transport domain that is all the transport layer that allows bits of information to flow between the radio part and the core part that is the third big domain that is where all the, let's say, accounting happens or the internal routing, the interfacing to the internet and so on and so forth. So it's multiple domains and radio transport and core are really the basic ones. As I was saying, networks are becoming more and more complex. So the domain is being disaggregated. The same is happening in the core. These domains are being transformed in many micro domains that interact with each other. So depending on which domain or micro domain you're looking at, when I speak about the resources, that means that can mean completely different things. So to give you an example, and for us, NetAI is one of the key, let's say, examples because it's something we are working on a lot. The radio access you have are these cells that provide coverage to the users. And the cells are very redundant because of capacity reasons. So the way we use our mobile phones is very diverse over space and time, as you can imagine. Very simple example: overnight, very few people are using the cellular networks while during the peak hours at a train station, you have plenty, plenty of users. So you have these huge fluctuations in the spatial temporal mobile traffic and the operator needs to dimension their radio access, their cell deployment to the peak, right? You want to be able to serve the users when you have the maximum demand. The point is, this peak happens twice a day, maybe, and that's it for a few tens of minutes. And the rest of the time, you have a huge radio access that is underutilized. And with the current technologies, until today, I would say, operators didn't do much about that. And the cells were essentially active all the time, which implied a huge which implies a huge waste of energy because cells need consume a lot of power actually the radio access is the part of the network that consumes the most power. In that case, in that specific example, when I speak about the resources, I speak of the coverage provided by the cells. So nowadays operators start the technology, let's say the radio access starts to enable the operator to switch off cells based on some policy or logic. And this is where we come into play because our technology, as an example, can learn how and when to switch off cells in a way that coverage is guaranteed, all users are served, and yet the energy consumption of the network is minimized. And that's just one example, right?I could give you another if we moved from this domain to other domains. So it really depends on the domain. Got it.

Yeah, again, now just to go back a little bit. Of course, nowadays, there are situations in the network where the operator doesn't have the capacity or the interest to control things, control resources, or manage them. So as you were saying, or in the example I made about the coverage at the radio access, they don't manage no, so they just leave things as they are. In other domains, there is control. If you have anomalies as an example, operators gather continuously a lot of thousands of key performance indicators from their network, and there are some typically empirical approaches to raise alarms. So whenever there is something that malfunction because some KPIs, some key performance indicator, you know, go off scale or whatever the empirical rule tells engineers, the network engineers are informed, are warned, and then it's manual work. They go and check what's going on at the switch at the cell station um and and take take countermeasures but these these uh you know human in the loop kind of management that is extremely extremely slow and very expensive as well so there is of course there is some level of management this is just what i wanted to mention but it depends on the domain depends on the task and the AIin my opinion has the potential to bring that management in a pervasive way through the whole the whole network. And just to come to the example you were making, you know, like sports events. So yeah, I think it's similar to what I was saying before. It's a bit of an extreme case because when you have stadiums, usually you have these huge peaks that are however predictable very much because they are scheduled. So in those cases, you know, operators already have often some very, let's say, adaptive solutions, like they know that there will be a match on that specific day. So what they do is that they have dedicated cells that are available there and that are only turned on when the match is scheduled. So, in that case, this is already happening today you know without any need for AI because the case is so straightforward I would say that the operator can allow to send a person to switch on off or even control remotely the cells that cover the stadiums and just turn them on whenever it is needed but in many other cases if you think about a whole city with fluctuations that are partly predictable partly random that happens over time and space, the management becomes much more complicated. And this is where the big gains, let's say, in energy savings could be enabled by AI.

Right. And even even, you know, everyday dynamics that are relatively periodic, not relatively, they always happen the same way. Still, the operator doesn't have today the capacity to, you know, learn them, even if they are not not surprising, like the fact that you have these peaks of traffic at commuting hours in transportation hub, places like train station, bus station, so on and so forth. This is something we observe in mobile traffic, this is very common. But operators don't do anything about it. Because it will be a very consuming effort to understand how these peaks happen at each train station, what are the different periods of the day, where they should switch on, switch off. I mean, doing this by hand is just unreliable and too time consuming. So even if those patterns happen every day, operators right today, they don't do anything about it. And this is where AI can help.

I think there is a lot of potential and, you know, of course, as we founded the company, we believe the same. We believe there is a lot of potential. And we see that, at least relatively to the tests we could run until today, it actually helps a lot.

Well, telcos are very, very, very big companies. And as such, they are also well, they are very big and they are, I would say, very old in the sense that, you know, for today's kind of market and economy where you have companies that are, you know, huge companies like Google that were born 20 years ago. Telcos are old, right? Telcos were there in the 80s, 70s, some go all the way back to the 40s. So those are huge elephants that are very very slow and you know i i think that developing this kind of technology internally, it's, it's, it's really not their thing. They, even when it comes to technologies that are not so disruptive as AI, which somehow comes, you know, from a different domain, even when it comes to either the base, I'm a networking person, my PhD is networking, my master is networking. So even when you talk about network technologies, like softwaretechnologies for mobile networks, they are slow in adopting even those. Those are their core technologies. Those are the technologies that update and improve their architectures. And still, adapting them takes them forever. So you can imagine when it comes to AI, which is not something they're familiar with, I mean, there is a lot, a lot of effort to convince them that this is useful for them. And I mean, we're still fighting. I'm not saying that this is easy. Actually, it's the most difficult challenge we have now. Proving that what we do is worth something for them.

So it all started, I mean, both we are three co-founders in the company, Paul, Mark and myself. And Paul and myself, we come from academia. Paul is a professor at the University of Edinburgh. Well, I'm now a professor in Madrid. And we've been knowing each other for a very long time. And we've been working together recently, well, relatively recently, since say 2018, 17, 18, we started to work on AI applied to network management. And through our academic work, we got in touch with mobile network operators and started, those were at the time academic collaborations with them. We were lucky enough that they granted us access to their data. And we started playing with that. And we saw that, as I was saying, there is an interest in applying AI tools to this kind of data to manage the network. And this is where we started from and say, OK, we had real data. We could test that AI made sense for this data. Of course, not not in production or not even in a prototype system. It was, I would say, mere simulations. But still, the data was real, right? So it was the actual data that was flowing through the network. And this is where we started from. We sort of validated the concept using real measurements. And then we started to think about the company. Mark, who's the third co-founder, he's, let's say, he's been in the business of startups for a long time. He successfully exited a number of those. So he helped us with the first steps and we moved forward, exploiting the collaborations with operators that we had developed during the, let's say, academic times. I would say that it took us around two years to get to the point where we could convince an operator to run an actual proof of concept. And this is why I was saying that it's very difficult to convince operators of the value of AI. And yeah, essentially, this company is also relatively new. It was founded a little bit more than two years ago. So we are now at the point where we started demonstrating our technology live in real production deployments. And yeah, we're doing that with a couple of operators. I don't think I can disclose the names, but we're going live, yes.

So yeah, so one of these proof of concept is about energy savings. The other is more about analyzing the KPIs that come from the network and using them for core network management. So about the energy consumption, I was saying, you know, the radio access is really the most expensive part energy wise for the operator and we're talking of substantial costs to the point that you know when when 5g arrived many operators were skeptical because 5g base stations are more energy hungry than 4g ones and as soon as they knew that they were saying yeah okay this is not going to work because already with 4g we are barely you know having a sustainable network with 5G we will not be able to then of course it has some advantages so it took over as it is natural but the energy costs are huge and the telco ecosystem as a whole, the mobile network ecosystem as a whole, I think it has a weight on the total emission of CO2 connected with the energy consumption that is around 2% globally. So it's not negligible at all. And most of it comes from the radio access. So I think that being able to cut down that cost you know, also has an impact in terms of environmental sustainability of the whole communication system. And so we're running this proof of concept and our preliminary tests indicate that we could achieve savings up to 40% to be confirmed by further tests. But if that were real, it would be a huge success because dropping costs by 50% or 40% at the radio access is really a lot of money for the operator.

Yeah, exactly. So I can tell you, in Europe, that is, I think, denser than the US in terms of city layout, at least. One operator, one single operator in one country, has tens, many tens of 1000s of, of cells deployed. And this is only for G, that you need to, you know, add, okay, 3g cells now are being dismissed, but 2g are still there, because they are useful for many, many, many things. 5g are coming.

Yeah.

Yeah. I would say so. Of course, if you talk with the network engineers, they would still be a bit fearful about the technology, but...so first of all i would say uh you know there is i see it's very hard that AI will be deployed without any expert approval from the network engineer. So, you know, what typically happens when you try to deploy this technology at the mobile network premises is that there is a first period of months, many months where the technology is deployed, but it's not really active, you know, like you monitor the data, you let your AI run, and you let the AI generate decisions. But then decisions, they are not enacted. The decisions are just displayed to the network engineers who look at them and sort of validate them, right? And then the engineers keep control over the network and decide whether, okay, decision of the AI makes sense, doesn't make sense. There is, of course, a back and forth. And usually it's a process that is needed to convince the engineer that the AI is taking the correct decisions, right? And once the engineer is convinced about that, this is the moment where they allow you to deploy the technology, right? At least in a proof of concept and see what happens in practice. So there is a long period where the AI and the engineers coexist. Plus I think no operators, at least for the next. 10 years will trust AI without some human supervision. So I think that the current approach is where operators still, where network engineers still monitor what's going on in the network, they will still be there as a sort of redundancy system. The goal of the AI is not to reduce costs in terms of management, because those are not big costs for the operator. The goal is really energy reduction, or as an example, something that could be very useful is, and I see, you know, proposal about that is that a big cost for the operator is support service, you know, people calling because the network is not working well, they have this problem, they have that problem. One interesting application of AI in that sense is looking at what's going on in the network and if there is something, an anomaly or something like that, inform the user, send some SMS message saying, hey, we're noticing some level of disruption in your network, don't worry, we're fixing it so that they don't call. And this relieves pressure on the service, customer service support which is a big cost because then this is a call center. This is many, many people answering the phone all the time. So in that case, yes, in that case, I could could, let's say, reduce the amount of jobs in call centers. But it's also maybe a job that I don't associate 100% to the mobile network operations. It's a bit different. Yeah, I mean, there's definitely a lot of people looking at customer service applications and user management, user and client customer applications. That's interesting. So you're applying this thing.

There's also some, yeah, some potential there. Although, I think that, you know, the internet, it's, it's a unique infrastructure that is very, very distributed, you know, by its own nature, there is no such thing like one entity controlling the internet. And these is such that the protocols and algorithms that rule the internet, that decide how things go on in the internet, they are also very much distributed. And I think this creates a situation where the impact of AI is not that clear. You want to rely on those distributed systems that even with protocols that have very simple, relatively simple rules as we have today, they create a system in the end that is so so distributed and so complex that we don't understand it anymore. So if you add the AI on top of that, I think it's going to be to be complicated. Well, AI has potential is maybe in in network systems that are in the Internet, but are, let's say, more more circumstantial somehow and more controllable like like data centers. If you want to control how your traffic is engineered, so how the routing occurs in your data center in a way to maximize the use of resources. Yeah, this is where AI can come to your rescue. Or, well, mobile network operators, as I was saying, because also those are sort of silos, you know, like structures, infrastructures that are controlled by a single entity and where you can manage things. If it comes to the internet as a whole, well, I see much less impact.

Well, yeah, so we're running this proof of concept, and this is for us a very exciting and important moment. I honestly think and hope that if this goes well, if we can really show that in the portion of the network we are allowed to control, we save the amount of energy we are seeing in our offline tests. Well, this will be a big landmark for us, but hopefully for AI networking as a whole to be a very, very first demonstration that this works in practice. It's a huge gains. So Yeah, I mean, I think the best wish I can make for ourselves in 2024 is to have this successful completion of proof of concepts because this will really open the door to many, many things for us and for the community as a whole.

yeah i mean uh it's uh it's not four five globally but uh it's maybe four five per country uh so you know it's uh it's not a very, very few player like in the, let's say Google type of ecosystem, but still it's not a lot of players. And so when you are successful in a proof of concept with one of them, I think it creates a sort of cascading effect where they can show some advantages, powered by AI and hopefully powered by NetAI. And then this draws the attention of the other, you know, competitors, maybe not in the same country, because of course this creates strange dynamics, but in other countries. And so, you know, it's difficult to penetrate the ecosystem, but I think once you succeed, the voice about your success spreads pretty quick.

Thanks a lot. It was a pleasure. Yeah.