AI in Digital Content (with Alexandar Vassilev & Paulo Nunes)

Thank you very much to be here. Thanks for having me. Very excited to be on the podcast. And yes, I've been at WeTransfer for actually now two years. Two years. But I recently took the helm as CEO. Prior to that, I was the Chief Product and Technology Officer for about a year and eight months or so before taking the mantle. But yeah, very excited to be on the podcast, very excited to talk about AI and obviously to get to do that with you.

I guess guided luck, it's what I used to say all the time. Well, I was born and raised in Bulgaria. At about the age of 18, I packed bags and in the pursuit of technology moved to the US for studying and obviously ended up staying there for quite a while. But ultimately, I was always interested in technology ever since I was a very young child. That was one of the few things that was keeping my attention for longer periods of time, which was very natural for me to go and explore that as an interest and eventually as a career. I moved to the US, spent some time studying there, graduated, had the chance to move to New York. I started working at Google, which was my official first job. Prior to that, I was an entrepreneur. I had a couple of spectacularly failed ventures, but I dipped my feet into entrepreneurship at a very early age. But ended up at Google, where I spent close to 10 years give or take, in various roles. Started on the engineering side, moved into product, did some technical program management in between. And ultimately, my interests were always gearing towards product building, especially in the creative space if you will. I love the drive of allowing folks to bring their creative talents to the masses. So I ended up spending some time building advertising products, but then quickly afterwards I moved to YouTube as I said, and I wrapped up my career at Google with some work on Google Search before moving to Germany from the US to take the helm of Joyn, building it pretty much from the ground up. We can talk more about what Joyn is, but for your listeners who don't know. Ultimately, my journey was always around technology, around product building, a strong passion of pushing the limits of technology and trying to see how we can make people's lives better.

Yeah, in a nutshell for dummies. No, I mean, it's very typical. Google back in what now people refer to as the early days, obviously not the very early days. I joined the company when I think it was about 7,000 employees, and I left when Google was over 100,000 employees. You see the company grew quite a bit in part of my tenure there. But Google was always about solving problems for folks. Monetization pretty much came always second. You know, we famously had the moral of 'do good by the user and the rest will follow,' and the rest did follow. But to your point about data, I think one of the critical elements that allowed us to do great products that are used by billions of people back at the time was that we were very focused on solving problems. But to understand some of those problems, you did have to look at data. Now, some brilliant products like the search engine or Gmail came up from great ideas, but ultimately they represented problems. And as you start building products at Google, it was very easy to solve some of those horizontal problems for many, many people. But understanding where to take this product onwards was based on a lot of looking at data and understanding trends and understanding where the next problem lies. Ultimately, what that allowed was for Google to build products that had mass adoption across the planet. That created an opportunity for the company and for advertisers to reach these folks in a very contextually relevant way. If you think about the early days of search, you search for something and the ability to serve an ad for that one thing you're searching so you can execute your desire, whether that's to purchase or to buy or to read or whatever, it was quite fundamental. Now I'm sure that all your listeners expect that to be part of what we all do, but back then it was quite revolutionary. The internet was still a very difficult place to parse and understand and find things. But luckily, once you gained traction with your products, you can think about how to make the lives of your users better, and that included monetization opportunities to connect them with the right product or services if they need them.

So the very basic way of working, and I'm sure I'm not telling you anything that you don't know but perhaps it's interesting for some of the listeners, is that it starts with you have to map the internet first. To make a search engine quite useful, you have to understand what's out there and you have to understand what do people mean when they type. If you have such a large audience, then it starts there. For example, when you type 'wood cutting' but you misspell it, how does Google know what you really mean? It uses an incredible amount of information that has built over the years with data to understand how people who usually misspell 'wood cutting' do misspell it. It's not necessarily personal information but trends of usage. For example, Google might look at all the results that people search for when they type 'wood cutting', what do they click for? So you start building what we call user journeys, as to where do people want to go next. I'll speak a bit more to an observation that I built by spending a lot of time at Google. But ultimately, you start presenting valuable information and usually folks are on a journey. Now you can type 'what's the weather in San Francisco' or 'what's the weather in Lisbon' and you'll get an answer, or 'what time is it' and you get an answer. But almost always it's some type of journey. You can see how folks are learning. You might type 'wood cutting' first, then 'wood cutting for beginners', then 'best type of wood'. You can see how people build their journeys of knowledge on something like Google to see at what point an ad is relevant. You see ads even on the first query, 'wood cutting', because we use a lot of signals to say people who look for wood cutting may be looking for tools. But as you refine what you really want to learn next, it becomes much easier to surface contextual advertising that is really relevant to what you need. So you might be trying to build your shed in the back, you want to get your hands dirty and build something. Ultimately, as you go down the line, we can contextually give you better advertising that is more relevant to what perhaps you need. So ultimately, it goes back to the results themselves, they're not ads. The search results also learn from what people do. Is that a good page to serve for somebody searching for that particular topic? What's really interesting is that you can take it further. If you Google today, if you're looking for transferring a big file, a business like WeTransfer is in the business of transferring big files. Then as an advertiser, you can target this audience that looks for these particular words because those words represent something that you as a company do as a utility. So there is Google presenting ads, but there's also the opportunity for advertisers to target a particular audience, particular keywords, as you search. Ultimately, the goal is optimizing your online presence to signify to search engines what your property online represents. So it's quite an interesting back and forth, feeding to your point: intent versus utility, and learning from that as you go along.

I think it's involved. I mean, listen, I was there six, seven years ago. If you talked about AI at that time, it was very different. I don't know where the search engine is today, but the original Google search engine was built without necessarily an AI component; it was just a lot of good rankings. If you look at the page and determine if that page is well structured, if the information is useful, that's why even to this day people talk about SEO optimization. It's marketing techniques to make sure their website can tell the engine what exactly that website does. Back then, at Google, you wouldn't necessarily see different results if you're a different person in different parts of the planet. You'd see the same results because if you type 'wood cutting', Google should be able to give you the best wood cutting information regardless of whether you sit in New York or Lisbon. But when it becomes a bit more personal, it gets really interesting. Nowadays, I presume people use Google for information seeking which is very specific. When you type a restaurant name and ask about opening hours, there might be a thousand cities with a thousand restaurants with the same name across the globe. How does Google know to give you the one near you? That's when personal signals like location come in. That leads to the discussion of how you make things relevant by revealing something about you to these companies in order not to spend three hours searching for the restaurant in that particular city. For example, a few weeks ago in San Francisco we were looking for a restaurant called North Beach. North Beach is actually a neighborhood in San Francisco, so typing 'restaurant North Beach' gives you all the restaurants in that region rather than the particular restaurant which was actually not in North Beach but somewhere else. Even Google gets confused sometimes. It's difficult. You start thinking about the complexity of taking a two to three word input, parsing all the information in the world that's online, and deciding what is relevant. It's quite hard, quite complicated. Of course, I spent some time as a product manager at Google Search. When I was there, we had this interesting thing we wanted to think about: how do people learn on Google and how do we make the engine more useful for folks who learn? If you type 'climate change', there's climate change for kids, climate change news, climate change scientific papers, climate change Earth trends, companies. So it becomes interesting how you take these broad topics and help people navigate them to get to the specific things they want. If you think about these journeys as we call them, they start something super broad, then people will read, learn, come back, refine their search to something more specific, then read, learn, come back, refine to much more specific. In every iteration you see the learning of the topic. The question became from my team: can we help with that?

No, it wasn't that. We tended to think about the question: if somebody is typing 'global warming' and then the next thing they type is 'global warming news', what is the most probable next search that they might be doing to refine? The idea became, can we take a complicated topic and help people navigate to the thing they're really interested in by segmenting? That could mean surfacing knowledge cards that are very specific on a specific topic or presenting additional information for you to quickly go down the rabbit hole of learning. That was quite an interesting problem because in the end of the day, we all wanted to think we are very unique, and in many ways we are, but we tend to follow similar paths. If you have enough people follow a similar path, you can look at what the first couple of searches of somebody going down the learning curve are, and you can anticipate what information might be useful down the line for them. Then you can preemptively surface that even before they search. I'll give you another example. If you search 'net worth of Joe Biden' and then 'net worth of Donald Trump', maybe you're going down the path of figuring out the net worth of each president. Very naturally, we should show you the easy way to click on something that will show you the net worth of Barack Obama or George Bush or Clinton, etc. That's where the data becomes quite interesting because you're able to build paths through the vast information that exists, allowing you to service it a bit more quickly to users.

No, when we were thinking about it eight years ago, it wasn't AI. I don't know if that has evolved specifically. But more generically, if you just talk about Google, the vast amount of information from so many people doing it, you could use machine learning models to anticipate that, especially with short-term memory models and Transformers. You could do that nowadays. I'm pretty sure Google is doing it now. But machine learning is now everywhere around us whether we see it or not. I think the utility of parsing large amounts of data and finding patterns, which is kind of how I think about machine learning's superpower, is quite useful in many different ways. But the idea is that we're not that unique in our behavior and patterns. As much as we think we are, sadly not. Even though we are unique in many ways, in terms of some basic knowledge paths that we follow, we follow the same logical path.

It's very similar, it's just a different type of information you're consuming. I don't know what YouTube does per se today, but I always thought it's an interesting experiment. If you have an account with YouTube and you watch a lot of videos, at some point your recommendations will probably get into the realm of what you find interesting. It might have some new things, it might have things you watch many times because the algorithm knows you find that interesting. A good test for your listeners is: don't create an account, go on a fresh browser and start watching videos as you normally would. I guarantee you at some time the algorithm will know and will serve pretty much the same things as on your logged-in account. That's not because they know who you are or you have an account associated with this data, but because the data is such that people follow paths. So eventually, if you take enough steps in a certain path, YouTube or any other platform will start to understand what's the next thing that you probably want to watch.

It's a good question. So when I was at Joyn...

It's a video streaming service in Germany. It was a joint venture between ProSiebenSat.1 and Discovery International, one local and one international broadcasters. The idea was to create a new European-based content service, and our vision was to create it based on what we call local content because Europe is one of those markets where local content is incredibly strong, and there was really no aggregator at that point for content that speaks to you locally. What I mean by that is if you live in Munich, something filmed in Munich or Berlin feels a lot more authentic. The American services are quite amazing at global content, but people in Europe still want to watch things in their language, things that are closer to them, they can identify with. For example, 'The Office' was one type of show in Britain and was very adapted to being different in the US; in Germany it was called 'Stromberg'. They didn't just take the British one and show it everywhere because people are different, cultures are different. So we ended up building quite a successful product, a great brand, with a strong portfolio of aggregating content from partners but also filling gaps by building our own content. Ultimately, it's a video streaming service similar to Netflix that also has a strong focus on live TV, which at that time when we started everybody said was dead, and now everybody is going back to live TV with ads, which is an interesting dynamic. Given that we built a sustainable user base and content pipeline, we had to tackle how to recommend content. The challenge then was that watching long-form content, unlike short-form on YouTube, is different. If you sit in front of a smart TV or even on your laptop or phone to watch long-form content, recommendations are quite challenging because context matters. Not to generalize, but maybe on a rainy day you want to watch a very different type of content than on a sunny day, depending on mood and emotions. That's incredibly difficult to predict. So it's not just the type of content but also the context that people carry with them.

They didn't really add value, so I don't know what they're doing today. They're definitely recommending content, but I don't know based on what. They kind of dropped the whole original idea they had. Maybe because of what you said, because it's just so difficult to predict what people like and what people are in the mood to watch. Yeah, I think it's easy to predict what people like, but it's very difficult to predict the emotional state and the mood in that particular moment. And I mean, again, my knowledge is somewhat stale given I've been out of the streaming business for several years now. But kind of the proven path to recommendation was still the good old actors, genre, settings, plot lines, direction. It had to do a lot less with machine learning or algorithms; it's much more about very simple curation. You just watched, you know, a famous action movie with someone or a comedy or romantic comedy, and then the next one will probably be something along the same genre or you have the options to go down on a similar kind of content with the same actor. And I don't think it's gotten a lot more sophisticated, at least from my experience watching things. But what we did back at Joyn, we also added the human element, and I think that's where some of the power of this, especially on a local level, some of the power of marrying data and recommendation engines plus human curation, has really shown the strength. Because, you know, if you work in Germany or the south of Germany, and you have a creation team there and the mood is something or there's news, you can overlay the engine recommendation with some kind of contextual recommendation that comes from a human, and that becomes a bit more powerful. But people have different emotions at different times, so it's still a problem to predict.

I mean, I'm speculating. I haven't seen it in action, but I'm speculating that it is, in many ways, so it is to understand whether you've built something or created something that actually carries interest. But it's less to say, 'Oh wow, people watch this particular scene, we should have more of this scene or this type of humor.' That becomes very subjective and it's very difficult to measure. Plus, signals are not what I call clean. You might have stopped it for some other reason than just not wanting the content. However, there are elements that these platforms have done very well in terms of minimizing some of the risk of content not connecting with the audience. For example, YouTube even to this day will tell you, 'Don't post an hour-long video on YouTube; that's not the watching behavior. If you want people to finish your video, you should probably make it shorter.' I don't know what the time is now; back in the day it was 15 minutes versus the average attention span. Now it could be an hour, I don't know. But it's similar for Netflix; maybe evening might have different completion rates for certain types of shows. Also, the amount of time you search, we can have a verification of whether you want to watch super long content like a movie or an episode of a series. But back at the time when I was involved in the business, we hadn't cracked it yet. There were a lot of startups back in the day that were talking about tracking with AI elements, inserting advertising in specific ways because we know the users, and highlighting posters. I think that's still there. If you go to a service like Netflix, you might see some actors on the posters, but somebody else might see different actors, and they'll rotate it to make it feel fresh. But I certainly as a user don't feel that the services necessarily know me at every particular moment. I'm sure there's a lot of intelligence behind to make sure that I do find something to watch.

I mean, when I was working at Joyn, what we used to say is the Golden Age of content creation. Simply because music has become such a part of our daily lives, all of us consume much more video than we consumed two years before, or a thousand times more than five years before perhaps. And I think it was just an opportunity for a lot of services to come to market. The challenge is if you're in these businesses, as a user you almost expect to have fresh new things to engage with. So if you come four months in a row to your favorite service and all you see is the same movies, you can do that through new content, but you can also do that through smart recommendations from your library. Ultimately, people want to engage with new content, to feel that things are fresh on the platform, to justify why you're paying a monthly subscription. Also, new content is how you market. You wouldn't see a billboard with a movie from years ago for a service. If you want to stand out, you have to have some flagship content that drives the audience to you. For example, at Joyn we had a very successful series called 'Jerks' which was very popular in Germany. That was an attraction for people to come and check out the service and see what else we have. But you need that content pool. If you have so many services at the same time looking for content, you can understand why so many series get produced and why the need to stay fresh is so high. The fact that some of these series are still hit or miss shows you that it's not an easy problem to solve, even if you are implementing data and looking at data. There's perhaps not enough signals to understand what really makes a good series, outside of people watch it or they didn't. But you also have series that are watched but people don't like them because they just finish them. That's why sometimes second seasons are more difficult to do. You see other opportunities to learn about the audience. People now go on IMDb and check what people are saying about their content, or they go on Twitter, so you have a lot more signals now, but not necessarily within your service. So the sense is that it's a difficult problem to solve. If you marry that with the fact that we have different states and emotions, it's still an unsolved problem. There's a lot of work to do there.

Correct.

No, no. It's just tutorial content. WePresent is an editorial platform, and it's incredibly highly curated. We have a number of very talented individuals who work with artists to collaborate, discover talent, and highlight talent. So we don't necessarily look at it from a perspective of what will drive more clicks. Our focus is slightly different. We want to highlight great talent and great content and great creative moments. So we've taken a bit of a different approach. We're not optimizing for clicks or engagement. That's why WeTransfer won an Oscar this year, and that was something I'm super proud of the team. It was an idea that came internally that ended up partnering with ResetMedia, who is phenomenal. Together we basically made a short action film. He kept an idea for a long time. We worked closely together to make it a reality, and it turned out to be an incredibly great piece of content and ended up winning an Oscar. So WeTransfer is in the small group of technology companies that have won an Oscar, which is not very big, so it's an incredibly proud moment for all of us.

Yeah, absolutely. At WeTransfer, we have this ambition to connect the world with creativity. What we mean by that is to help creative folks build sustainable businesses through their talent. We want to help them connect better with their audience, share and collaborate on their content, on their creative projects with each other and with their clients. So ultimately, we always try to be this agent in between that helps facilitate some of these workflows that creators have. A photographer working with clients, or a content creator who wants more direct engagement and monetization opportunities for her talent with her audience. We started as the urban legend which we finally talked about with the founder. Back in the day, our founder was part of a creative studio in the Netherlands, and the challenge was it was very difficult to teach clients, whether that's Adidas or any other big company, to use FTP back in the day. The only way to exchange content you produced for them was to download it on a USB stick and give it to a motorcycle courier and send them on a 30-minute journey, and do that several times a day. That's how the idea of WeTransfer was born: 'We probably can make that better.' The first WeTransfer product was built on Flash, I think, and ultimately it was an incredible success. The product went viral. We never lost that connection to the creative community; we actually grew it.

Yes, so today what we've built is basically a platform and an ecosystem of capabilities that stem from this strong use case of transfer, which is still how do I share my content with many stakeholders. How do I share with clients in a way that it's very easy for them to understand, very easy to access, and we basically don't stand in between you and your audience. Over the years, we've expanded our capabilities to add on top of transfer other utilities. For example, this year we launched something called Portals and Reviews, which is an opportunity for photographers or other creatives to bring their clients into a branded environment where they can receive feedback, approvals, and other information back from their clients. Because if you think about it, even today, if you send somebody something via email or another medium, you have very little control of who sees it, how they consume it. You have very little focus on how you receive that feedback; you might receive it over text message or email. It's very difficult to get this back-and-forth, and that creates scope creep and time expansion in creative projects. So this expansion, which is now part of our highest tier subscription layer, helps those folks create these environments and bring their clients in the same frictionless way, then interact with them in a way that's easy to digest and easy to go back and forth. This is an example of how we've thought: what makes this very focused is the core transfer, but how do we build on top to help others?

We don't necessarily use data for monetization per se, but I think, as anybody will tell you, to build a great product you do need to understand if you have created any roadblocks, if your user experience is up to par with what you think it is. For that, you need to understand how people use it because you look at it day in and day out because you build it, so you need to instrumentally understand it. However, folks who come in have to get it, and you understand that today. So we are very data-driven, but we are also quite responsible. We are a certified B Corp; we deeply care about how data is used and about privacy. We are very respectful in terms of what data we collect. Ultimately, we look at whether the product works, where are the roadblocks, what can we optimize, which flows are not doing great. If we see opportunities for us to expand, we marry qualitative and quantitative data. For example, as we thought about Portals, we spoke to a lot of users and said, 'Hey, you send files to your clients, what is missing? You say copy feedback, approvals, etc.' So you learn, then you build it and see the quantitative data that shows what people actually are using. That's collected in a more classic way: user tests. Then we compare that data with the instrumentation data from the tools. The good news is that we have an amazing user base. As we ask questions, people take the time to help us understand. We have made it part of what we do every year: go out and speak to creatives across the globe. We produce something called the Ideas Report, where we speak to 6,500 creatives from 180 countries. So we understand trends, see how they think about their business, sustainability, workflows, and we try to understand them very deeply. So we collect it in an old-school way. It works; it's amazing to understand from the people that actually use your product. But again, you have to be lucky. We are very lucky that we have folks who are willing to engage and give us their insights.

Yeah, it's a great question. I think all of us are asking ourselves as we read every day: what is the next iteration of creativity? If you scroll back 15 years, even at YouTube, YouTube was created as a dating site, a video dating site. Nobody had the concept of creatives or creators for that matter. Now, as you can see, this is one of the fastest-growing segments of the global economy. The sheer change in video consumption in the last 10 years, people projecting their talents from YouTube to TikTok to Snap, it's quite interesting. Imagine what's going to happen in 10 years. We're already seeing some trends emerging. Obviously now it's a bit of a crypto winter, but ultimately the idea of NFTs and decentralization, the ability to directly engage with an audience, is interesting to see how it evolves. You also have AIs that are winning art competitions by creating content, and AI is writing content, composing music, generating images, art. So as you think about this, you start to think: what is really creativity, and how is it going to evolve? I spend a lot of time talking to folks, understanding, keeping an eye on these technologies to see where the flow of creativity is going to change in the future. It's very difficult to predict. It's a bit threatening for creatives out there thinking that machines could actually do a podcast. Indeed, some creators worry. On the other hand, maybe we get to a future where the value of a human creation is much higher. Maybe there are things that creators do that are boring, like retouching a picture, that could be done by a machine instead. So I think it's going to be some type of intertwinement between the human element and the machine element. On the other hand, creativity is also about interaction. As a creator, this new way of creativity has been born through social media or video or TikTok because people want this back-and-forth interaction between the creator and the audience. Maybe that's something machines can never really substitute. That's the joy of being around and looking at these things firsthand; you get a front-row seat to the future almost every day. I don't think anybody can predict it.