 Hi, thank you everyone for being here. This is Marcelo Soria, and I'm Alejandro Vidal. And we are going to talk about all the things that are not data science. And typically, people don't hear about them. But we think that are quite critical, quite important to create products. So we are going to start explaining why this talk using a metaphor, similarly with a mobile phone device. So maybe some of you know this video. A phone and an internet communicator, an iPod, a phone. These are not three separate devices. This is one device. We are calling it iPhone. Today, Apple is going to reinvent the phone. And here it is. OK, I recommend everyone to watch this, because it's 10 years ago. And it's amazing to see what people was crazy about that moment. We are not going to present an iPhone. But we are going to talk things that are related to that with data. So back into 2008, compared with now, at that moment, it was a very unique experience. In fact, they used revents the phone as an expression to say, we are creating something totally new. Also, they need new words to talk about it, because previously, they didn't have any word to talk. For that reason, they used a kind of mix of three ideas, three concepts, a phone, an iPod, with a way screen, an internet communicator. The tab, the swipe, other words that at that moment didn't exist. So probably at that moment, you had highly specialized profiles. You can see how many UX designers of mobile apps were in 2008, probably only a few of them. Now it's a very common profile to work in more or less every company. So at that moment, probably, it was really hard to create products with this technology, because you have very highly specialized profiles. It was kind of a rare experience, so you don't have examples to take ideas from them. Now is the old new normal. At this moment, more or less, we are talking about five millions of mobile apps, which is a huge number. Everyone is familiar with it. We create new concepts on top of them. I mean, everyone knows more or less what this mobile phone, how do you interact with it, the tabs, the swipes, the notifications, the vibrations, more or less everyone in the team knows what is a mobile device, although you are not a designer. So you can create products on top of them. The whole team has a first-hand experience with that. So you can reason and think new products easily. That's the difference between 2008 and now. So we think that the same is happening right now with AI. Products with AI are quite rare. I'm not sure how many years ago, maybe three years ago, two years ago. Now is the moment of changing the increasing ratio of new products. But we have new materials to work with. For example, we can talk about categorization. We can talk about recommendation that those concepts didn't exist before. We have highly specialized profiles as data scientists here. How many of you consider yourself a data scientist? Raise your hands. OK, come on. Everyone is a data scientist now. OK, so it's really hard to create products with AI. The same way, in 2008, it was really hard to create a mobile device experience. So in the future, or at least now, AI is a new, new normal. You have to educate your people to get those ideas known by everyone so people can make new concepts, new concepts, ideas on top of them. The whole team should know about the new materials. So a data scientist is a specialist. He's not a gatekeeper of knowledge. He's not the owner of the knowledge. He should be a specialist to help people to achieve their goals, but everyone on the team has to know about data science. So the team can imagine and design new experience. For that reason, we are here to think about how can we solve those problems. We have here a small video. And to unlock the phone, I just take my finger and slide it across. And we'll see that again. We wanted something that you couldn't do by accident in your pocket and just slide it across. People got crazy at that moment only because of the sliding thing. But if you think about that, it's not about the sliding. It's not actually because before iPhone, there was different touchable interfaces. And you can slide things on them. It's not about the sliding. It's about how people pocket their devices. So they have to create an idea to avoid the typical pocket dialing. So they create this experience. But it's not about the idea. So other small fragment. And I scroll. OK. Maybe it sounds kind of normal right now, this experience. But at that moment, it was crazy. So it's not about, again, touch screens. It's not about the technology. Because that technology existed before the iPhone. It's about natural gesturing. You can move things. Like you move things in the real world. Because you are a physical human in a physical world, in a physical environment. So we think that the same change is happening right now for AI-based products to a new normal. So Marcelo is going to explain what's exactly an AI-based product. Exactly. So the question that we're going to hear is what do we mean by this mobile metaphor then? What is an AI-based product in the same way that all of a sudden phones, smart phones, just became mobile phones again because they dropped the name, the surname of smart, or a touchable interface, or whatever, which was a technology. But that was not so important. The important thing was that they did things better. And they kind of adapted to the real needs of users. So what's an AI-based product today? We should really start dropping the AI part. Or at least acknowledge that the AI part is not the centerpiece of it. And it's kind of normal that we all use it now as some kind of marketing tool to talk about it. Because it's the new thing in the market, not so new maybe. But it for sure comes with some advantages. So a product, any product today, is kind of somehow digital. And any digital product today has some kind of data baked into it. And if it's got some kind of data baked into it and you want it to really adapt to its environment, then it's going to be using some sort of AI advanced or not advanced. So basically, an AI-based product is becoming the new normal. Let's just drop the AI thing from that in the same way that the mobile phone became that. So we're just going to take this product as an example of a product that is using AI somehow to become better, to adapt better to whatever the user's needs are. And I guess we can all agree that this is a product that is really using some kind of data power into it. It's how we all listen to music nowadays. So the thing is that Spotify is not really so much about the recommendation engine. It's not really so much about how it's using, powerfully using AI to do things. Because it's probably not so advanced in some ways the algorithms it uses. But basically, and those algorithms actually existed also long before Spotify. I mean, we had recommendations algorithms long before. And we had maybe, I don't know, different versions or I don't know. But the thing is that Spotify, it's all about learning. It's all about learning about how people actually listen to music. It's learning about how people really want to do the things that they want to do. So the basic, let's say, paradigm shift that comes with this AI thing is that the products now learn, digital products or any product, like we said. Because digital is another of those surnames that we should drop. They're learning now. And they are learning from whatever we teach them. And the thing is that when we used to design products earlier, we used things like this. We were doing these screen workflows. We were trying to say, oh, my product has got these options. And from here, you go to here. And then you've got this fixed set of options. And you can do this fixed set of things. And this journey, let's say, across your application does no longer hold. Because your product now is not fixed. I mean, this is like you were doing, I don't know. Yeah, you were just pre-programming whatever thing, fixed structure. You build a model of something. And you cannot change the parts of it. It's like you build a city with Lego with no moving parts. But now, all of a sudden, the city has a life. Now, all of a sudden, your product is not predictable. Your product is learning from your behavior. Your product is offering you something new every day, basically. And therefore, you cannot really decide and thinking that everything's going to be always the same. And in this way, this is an example. This is Alex's Spotify. And you can see if you want to follow him, you've got his username somewhere there. Pretty interesting music he listens to. This is what he sees every day when he opens his Spotify app on his phone. And what will happen if we remove all the things that the application has learned from that interface? We are left with nothing. Because the whole product is learned from his tastes, from the way he actually listens to music, which is, by the way, the way he leaves. Because maybe today he's feeling happy because he's speaking at Big Data Spain. Maybe yesterday he was frightened because of that. We don't know. But then if you look at the music he was listening to, it was kind of instilling some kind of mood. And this thing, if you remove all that, if you remove, let's say, the human part of it, there's nothing there. You're left with just a home button, which takes you to nothing, because the whole home content is personalized. Then you can search something, which, OK, you could maybe find some music. But then it wouldn't be personalized search because you wouldn't have any of that learning part there. And then you could still access your playlists. But maybe those playlists would be dull and boring because you wouldn't have really discovered new music, thanks to all those learning. And actually, it's like we said earlier, it's not the same Spotify every day. It basically changes. And it does so all the time. It's not just every day. It does so evolving continuously and continuously. And this is a very important thing when we think about the algorithms that we do as data scientists, when they influence the products, we have to understand that they're going to be influencing them in this way. And also, at the same time, the experience is probabilistic. Also, for users, it's a different thing. So when we design for these new experiences, we have to think about these things. We have to think that users are going to be expecting probably something that's changing, but probably not too much, because then it's going to be also difficult for them. In the design processes, when we deal with these data-based products, we don't need to use design personas anymore. This is a technique that designers were being using for many years, where they build these archetypes of users, because it was a simpler way to approach the needs of those users. We've got the needy user. We've got the very nervous user. We've got those things. But now we know how each and every user is working. So we need to start designing for this, and we need to take this into account. When we, as algorithm designers, are designing for a UX designer, and we need to start interacting in new ways to bring this to products, because flows in those applications are not going to be deterministic anymore. There are many more examples out there, like Netflix, everything's personalized. If we take the personalized content out of the way, there's basically nothing. And well, one of the ways to approach design for these types of products is to think about UX as a container for dynamic experiences versus the previously used paradigm of a deterministic flow, which we can still kind of use. But maybe there is some kind of deterministic flow with some adaptive thing built into it. And this is just one of the ways we could be doing fully adaptive interfaces, where the app that you see and the app that I see are completely different experiences, different colors, different everything, maybe, because we've got different mindsets. And now Alex is going to talk to us about how products behave. So the huge change here is not about the algorithm. The huge change is your product in some way is kind of a living thing. We know that there is no artificial general intelligence, but Spotify is different. It behaves in a different ways for everyone. You can't predict what is going to happen next day, exactly, as a user, but also as an owner of the product. If I am the owner of Spotify, I can't guess what is going to happen tomorrow, exactly. So it behaves in a more different way because it learns. So let's talk again about Spotify. The first version of the algorithm for recommendation systems was the Discovery Weekly. As you can see here, it was a playlist that they bring to you each Monday with more or less 30 songs. Do you remember when this comes out? More or less? I remember very well. People started to get very crazy about that because some of them were huge fans of the playlist and some of us were not so fun of it. But it was a huge change because before, we didn't see anything similar to that. And if you think about that, as a data scientist or as an engineerian, maybe it's a naive solution. You have an algorithm. You run a batch on Sundays and the morning on Monday. You have a new playlist for everyone. This is the naive implementation of a recommendation algorithm. But Spotify, at that moment, probably observed the behavior of the product in the wild with real people, with actual people listening to music. And they observed that their solution is not the best solution for the problem that they are trying to solve, which is discover new music. So if you have some background on statistics or data science, you can think about that as a sampling. You have experience, which is dynamic. It's different for everyone. You have to sample different options from your actual users and try to guess what is happening in some way that you are starting to learn about the living thing that you are creating. You can think also about mass-scale anthropology here because you are looking at the behavior of the humans that are working with your app or are interacting with your app. So when you learn about the behavior of the humans that are using your app, the product evolves because it adapts to the new ideas, to the new behavior, to the new concepts. So Spotify changed a little bit more a few years later and they create the daily mixes. So everyone has six daily mixes, which is more or less the same idea of Discovery Weekly. But as you can see here, I don't know if you know some of the artists that you can read there. But for example, Daily Mix 3, this is mine, for example, is kind of hipster music from Spanish artists, more or less, if I have to put a name to that place. The first one is electronic music because I listen to music not all the time, not all the Mondays, a lot of different things. Maybe I want to listen the same music today, or maybe I'm nervous, I want to listen to jazz, or maybe for having dinner, I want to other kind of music. So they learn how people listen to music and they change the algorithm, not because they can change it, not for the sake of the algorithm because they learn how people listen to music. It's not about the AI, it's about how people listen to music. That's the important part of Spotify. Let's try to get some ideas how Spotify is working. You can Google this one and you can go to the Spotify API and try to get our recommendations based on SITs and they create these definitions for the API which is create a playlist-style listening experience. They are not talking about the algorithm, they're talking about the experience that you can create with this API which is more or less our recommendation system but they use kind of SITs to guess more or less what kind of music do you want to listen at that moment. So the funny thing for me is that table is one of the tables of the documentation. So if I want to show only one thing, maybe I want to show this parameter which is danceability because people listen to music for dancing and that's not a genre, that's not jazz, that's not pop, that's not a genre, that's an idea. So they create an algorithm to detect which music is danceable because that is how people listen to music. So they have all the parameters that are very curious. I recommend to Google it and find out other ideas from Spotify. Again, it's not about AI, it's about how people listen to music and as you can see with Marcello and as you can see Marcello is expecting kind of is behaving in an expected way. So Marcello is going to talk about what can happen in a living product. Exactly, just as he kind of controls if I'm going to be dancing at stage or not. The thing is once your product acquires a behavior at some point maybe it will start behaving in things in ways that you wouldn't expect, right? And you need to be really looking into that. There's an example, many of you may have seen it. The DeepMind team at Google, they trained a network to make a humanoid-like kind of creature run as long and fast as it could and with 21 joints or something like this and with some constraints and then the network learned how to do it. And it's really interesting to see the way that it's using the arms it's going to be doing like this and you wouldn't expect your algorithm to do that because you're human and you're running in different way. And the thing is, yeah, and then it falls and then it tries to do things and it's acquiring very strange behaviors which you of course didn't code because the thing learned on its own. And the thing is at some point maybe your recommendation algorithm starts doing something like this in your product and you need to think about when that will happen because you will need to maybe be looking into how that happens so that you can kind of control that in your product. So the thing is, you will have to start reverse engineering your product because even though you built it, you don't really know why it's doing the things it's doing. And the problem with that is that at some point maybe you will not even be able to reverse engineer it because it will be so complex. You will have so many layers of I don't know deep learning networks or whatever that it's gonna be kind of difficult for you to do that. I mean, you will have to kind of create like artificial intelligence doctors and magnetic functional resonance imaging or something like that to really peak into the algorithms and say, oh, what's going on here? Because it's got some strange behavior. It's got like a headache. We need to think about the concept of graceful degradation. This is when your product in deterministic products when one of the features fails when one of the functions fails your product has to still provide the user with some useful set of features because it still needs to do what it was meant to do. But when your whole product is an algorithm that is learning and that it's starting to behave oddly maybe it becomes ill for some reason or crazy or we don't know the concept of graceful degradation is different. So your interface needs to understand maybe at some point that this thing can happen and you need to prepare it for that. And to do that, you need to be measuring absolutely everything that is happening with your product. It's not just KPIs. It's not just measuring how the user is using the product but rather you need to measure how the product is behaving which is a new kind of a new concept and we need to start thinking into these things as our products evolve. And now Alex is going to make a small exercise with all of us. Okay, so how many of you has at this moment Google Maps in your phones? Raise your hands, okay, nice. Very nice market, sir. So if you, let's suppose that I'm one of the product managers at Google and I ask you as data scientists or as engineers to create some tool or some ideas to help people to pick restaurants. That's the problem to solve, okay. I have a lot of restaurants in Madrid and I want to pick one of them for tonight, okay. So I'm gonna give you some options although you can make some algorithms. Try to think the main algorithm to solve that problem, okay. You have to think a way to help people to pick a restaurant and you have to think about the best option to solve that problem. First option, our recommendation engine, maybe a collaborative filtering way of solving it. Two other way, other technique for recommendation. C, a natural language processing algorithm. D, image classification, okay. One moment to think about it. Okay, raise your hands if you think that this is the first option. A, okay, don't be shy. Okay, B, other recommendation technique for A is win. C, a natural language processing algorithm. Okay, third one. Image classification, okay, quite popular. There's some of you who didn't vote. Come on, take a side, take a side. So more or less we use this example because it's a kind of red herring problem. Probably the typical way of thinking for a data scientist or if you read a book about different techniques in data science, you can hear about recommendation algorithm that more or less match exactly with the problem that we want to solve, but as you can see. Which is what most of you voted, maybe not very much different, but most of you chose A, I mean, maybe you were not looking at each other. Okay, so let's see what happens in Google Maps to solve that problem. The first solution that they use, what this one. Okay, the recommendation system is the name solution for the problem is maybe a data science way of thinking. But what actually happens, the first thing that they implement was if I have a lot of pictures that people are uploading to Google Maps. For example, from my Chinese restaurant here, I have here a lot of pictures and they create this experience, which is an image classification problem. Like they only pick pictures from food of the place. So you can look at the things that they are cooking and if you like the idea, so it's a super cool place or it's only a squeeze sandwiches or whatever, or paella or whatever you want. And the other classification here is atmosphere because you want to know about the place if you can go with your partner or you can go with your friends or you can go with a big group. So a very easy solution which is maybe not the most straightforward idea at the beginning, it was one of the ways of solving that, okay? So we have a problem that seems like a recommendation system, but you can solve with the image classification algorithm. Okay, so again, it's not about the algorithm, it's not about the AI, it's not about how data science or data scientists think about how to solve the problem, it's about, neither is about using the best algorithm. In fact, it's not at all relevant artificial intelligence because maybe you can solve this problem with other approaches. It's about how people pick restaurants. So if you pick a restaurant with the food, you need something to help people to see the food that they are going to eat. Just one quick comment on that. The recommendation algorithm route is of course very interesting. If I could state what I'm looking for, but the problem is when I want to go out for a restaurant, maybe I don't even know what I'm looking for and you give me the typical filters, oh, do you want Chinese or do you want Italian? I don't know, I want it to be good. Oh, so you want the start rating of people, I don't know if I can trust people. Do you know that, so there's a lot of, and it's not even about that. I mean, if I pick five stars, lots of other things, there's too many options and I cannot decide, but you give me photographs, my brain knows how to do that. So my brain is used to picking restaurants in that way. So help me do things the way I want to do things. And AI is a very good way to do that because AI is trying to learn from the way we do things. It's not pre-programmed. It's kind of, if you gave me filters, you are pre-programming me a way of thinking and that's something that I don't want as a user. I want to move freely with this product in the same way that I take the iPhone and I do this to swipe up the list or to unlock the phone or something, which is natural to me. I want to pick a restaurant in the same way, a way that is natural to me, not thinking about, oh, is it Chinese or has it got, that's too complex. Sorry. Yeah, okay, it's the same idea with the Spotify example that they create something to recommend music that are for dancing. There is not that general, there is not that typical way of thing. There is no Chinese restaurant. Maybe the category here is restaurant that I can go with my partner, okay? So, the other example? Yeah, there's another example which is very telling. Yeah, this is a living example that Marcelo is going to try. This is a live example. Let's see, this is AI applied in real time and it's actually very well applied AI in real time. Let me say that. So, this is Google as lights by the way? Yeah, so I click here and then all of a sudden, oh, sorry. So, like I said, this is just an artificial intelligence application being used in real time because there is a real need here from users. Maybe if we turn this on, we can help our hearing impaired audience to follow what we are saying. And without this, it would be very expensive because we need to have translators here and I don't mean that we need to get rid of them. Please don't take that as a take away from the conference. But the thing is for smaller conferences maybe that don't have a budget, they can use this option. And this is really very well placed AI in a product. And, well, that's all I wanted to say. I'm just going to switch it off because I don't want to, okay, to be leaning forward all the time. But the thing is, again, it's not, oh, we've got this cool thing, we're going to push it anywhere in the product because, you know, trying to talk into products is something that people try to do nowadays very much. And well, maybe people don't want to talk to products because it kind of feels a bit awkward. But there's applications where you really can use things because there's a real user need again in the way people actually do things, not because you force them to talk to a product. So don't try to talk to your dishwasher, try to create an experience based on a need, and you can think about algorithms to achieve that. But don't try to use every algorithm in the world in every application. Maybe we don't need talking dishwasher, okay? So here, the problem, the actual problem to solve is how can we find the right solution because it's not easy as you can see, it's not straightforward as you can see, it's not the typical way of thinking as you can see. So Marcelo is going to tell us some ideas on some solutions to do that. Thank you. So basically, we need to, as product designers, as product teams, we need to really think like people act, not, we don't need to think like people think because people don't think the way they think that they think, they act in a different way. So basically, we just need to observe, which is something that is very interesting and maybe not so much done from technical teams. We just basically need to think like people actually leave because what we want to do with our products is to get integrated into their lives. So if our products don't think how they live, and if our products don't learn how people live, and this is the nice thing about AI, that is that they can learn how people live, how they listen to music, how they do things with Google Slides, then we're not going to get in there. Design was typically the discipline that was doing this for many years. It was observing people, it was trying to understand what they do. And by design, we mean a lot of different things. It's not just putting colors and designing lines and circles and points, but basically it's under that. I mean, it's a very big umbrella. We could consider that there's anthropology, there's sociology, psychology, there's art, there's many disciplines that have to do with how humans actually live and with how humans actually behave. And our products need to learn from that. And our teams, our scientific teams, our technical teams need to interact with these people. We need to come into a product design process that actually understands what AI can do. And where all the different profiles have a, let's say a dialogue about how people actually basically live, which is what we've been talking about. We don't need to just push AI into products just for the sake of it. And that's why we need to have combined teams where you've got AI people, where you've got business people, of course, because you need to make money, where you've got design people who understand people and where you have got even people. I mean, you've got to take your product, take it out there, get feedback from people, from real people. And before you do your product, before you do your concept, I know this is probably not just AI, let's say specific. This is the way you, you should have been designing products all your life. But the thing is now with AI, it kind of seems like it's like with, it's so powerful that we're just gonna throw AI at it and it's gonna magically work very nice. No, you need to really think about how this product is going to fit in society. And to do that, you need to have these multi, these multi-profile things. Don't make this product. Don't do this thing. Oh, we're gonna make a phone with an iPod. Let's just put a phone, dialing phone in iPod. It makes no sense, all right? And there is this very beautiful project by Google, this Google duplex that can reserve a table, it can book a table for you, calling for you. And then you'll know that, well, some people were not so happy that an AI system will call, will phone and they would talk to an AI systems and they would be fooled and well, now they changed the product. And the AI systems will identify themselves as, oh, hello, I'm an AI system and I want to book something for Mr. Alex. And that's when you actually interact with society, you get this type of feedback and you need to do it and we need to do it to understand that data product. And that's not only about the feedback about your users. It's something like, there is a living thing which is your algorithm. I mean, when you put into your wild, maybe you get behaviors that you didn't expect. For example, if you create a risk scoring algorithm, which is a kind of big thing in a market industry, one example was there is algorithm starting to score better to people that had a lot of credit cards. So people start to notice that. So they start to get a lot of cards to get a better risk scoring. So the algorithm learns again after that behavior of the humans that it was because of the algorithm. The algorithm again learns the opposite behavior that if you had a lot of credit cards, maybe you are not the best risk scoring idea. So the algorithm change, so people start to hack again the algorithm. So it's not only about the feedback, it's that you have a living thing and you are putting it into a real environment. So the environment is going to react to that. Exactly. So instead of being that, I'd be this one because they kind of made it kind of right because they understand, oh, we need to get to do things the way people want to do it and blah, blah, blah. So they've got a pretty nice valuation nowadays and it's basically because they find really very well the context of their users and they can go to it. And then they of course have a massive database on how people listen to music, which kind of means how people live in some way and that's very valuable. Alex. Yeah, by the way, when I asked for the Google Maps example a few months ago, at least in Spain, they add an actual recommendation system. And for example, you can see here 100% score for this place for my tests. So although the first solution was image classification, maybe there are more than one right solutions to your problem and you have to think at least the order of them because one of them is easy to implement. So you have to start with that. But again, later you can think about other solutions. So yeah, there is a recommendation algorithm in Google Maps to do that, but that was not the first solution. Okay, so as the takeaways, I'd say we've been hammering this idea all the talk, but it's not just AI. Never think about AI as the ultimate solution. It's the whole product design cycle and you need to have all the different people together. They're multidisciplinary teams that need to understand each other and this is very hard to get right. Believe us, I mean, we're doing this at BUVA and it's pretty hard to get there. It requires a lot of effort and a lot of education and a lot of, yeah, doing all these things. Again, you need to educate your data teams that data is not the only thing that design in a broad sense, let's say is something that they need to consider and you need to educate everyone else on what AI or data can actually do for them. Not just the design teams, but also your product management teams, your, I don't know, business teams, your legal teams even sometimes because they need to understand these things. Again, they all have to go together to build a successful product. These are all things that we are doing at BUVA data where we work, so we're building these cross-disciplinary teams. Alex, do you like intersections? Yes, I do. In fact, for the reason we work here, so thank you. Thank you. Questions? Any questions? Pretty welcome to do it. Maybe your AIs have questions for us. So no questions, then thanks again and enjoy the rest of the conference. Thank you.