 Good morning and I'm glad we have this 11-21 time slot so that hopefully everyone has recovered from whatever it is that you were doing last night and I guess slush is a little bit like Vegas what happens here stays here but it's it's great to be back at slush this is actually the 10th anniversary of my first slush and before we get into it I just want to make sure we recognize all the volunteers that make this happen it's my favorite conference of the year and to think that college students are doing all this on their own is really something special but we are here today to talk and I'm pleased to introduce you to Yark Kurilowski the CEO of DeepL I need to refer to my notes because he's a very accomplished person and the company that he's built is doing some amazing things so as our host mentioned Yark holds a PhD in computer science and if that wasn't difficult enough he picked an emphasis in math to go with that he's clearly too modest a brag so I will do it for him and DeepL is based in Cologne Germany under his leadership the team at DeepL have built the world's most accurate language translation engine and I am going to take a deep breath this is verified through third parties but they now cover 31 languages including Bulgarian, Czech, Danish, Dutch, English, Estonian, Finnish which is quite popular here of course French, German, Greek, Hungarian, Indonesian, Italian, Japanese, Korean, Latvian, Lithuanian, Mandarin, Norwegian, Polish, Portuguese twice both Portuguese Portuguese and Brazilian Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, Turkish and Ukrainian and you're trying to take up all my talk you're I'm sorry about that if your language isn't offered yet though I'm sure it's on the way but maybe let's start with a little about you to get the sort of the backdrop tell us about what makes you so passionate about starting an AI company that deals specifically with language yeah I think it's both and and by the way it's a pleasure to being here it's my first time at Slash not my 10th time so I'm pretty excited Finland is great and I love the snow I think translation and machine translation is just an amazing combination nowadays of technology and languages and I've been born in Poland I've lived half of my life in Germany and obviously I had to learn some English and I have a basic commandment of French let's put it let's put it that way so so I've been dealing quite a lot with with languages actually when I came to Germany at the age of 12 I think I was being thrown into the school and I didn't know a word in German I had to the teacher asked me to spell my name and I could not spell my my own last name which is not very easy but I could not spell it in German properly so so that was kind of the moment when I definitely realized languages are important and but also really tricky so so that probably has set me up for for understanding the problem itself and after that in like the time when we've been starting working on deep-air when we're starting me launching that that there there was just those signs that there is an amazing breakthroughs coming in with neural networks and how those can be applied to this problem and how we can actually through technology solve this huge problem that is that is a really a big one for humanity and that we especially here in Europe really see each and every day even when we come to Finland right right so without getting too technical I do think it's worth diving in and because I think your background is makes you the perfect leader for this company the the computer science background the math background but also the very personal experience about learning a new language and under the gun so to speak what is the linkage what you know for AI to make translation more effective and more powerful because you know the Rosetta Stone was doing translation thousands of years ago what's what's evolved in AI to make this something so powerful today yeah I mean first and foremost I think neural networks and the AI as we know it right now is just an amazing and beautiful combination of maths and computer science it's just like really on the edge of both of those if you and I had the had an honor to to kind of go to a university that that was really very theoretical where a lot of emphasis was put on on kind of the the theoretical foundations of how things work and not only kind of the applied how to code in Java and I think that's that's really important I think I think if you want to do cutting edge research or cutting edge products I would kind of advise it to everybody in the room here to go for that you won't learn how to code in Java and therefore you might need like two years more to get into your job but at the end you're going to understand how how these things work and and and how they how they really how they really built and neural networks like if you think of those of those constructs a little bit like also brains that are not made up from chemical components like ours are but I really built from from mathematical building blocks then you really have to in order to to improve those in order to make sure that those work even more efficiently and in our case produce better translations you have to take a look at how are those are trained how really the math works inside and how you make sure that those train even more efficiently or that they can kind of grasp the the problems or the intricacies of language much better and then this is combined also really with computer science because you're really trying to make AI work on a few years ago at least on GPUs that have been built for gaming and not for not for AI so you're trying to put that computation over there do that in parallel like with thousands of those GPUs which fail like after a few hours of operations and you got to restart them and make it work so so there's a lot of really intricate engineering going around around the around the around the mathematical part got it so let's take some of that theory in into practice and so personally I've been making investments on behalf of my firm in Europe for many years now and while I do speak a little Spanish ablum poco de espanol I actually really appreciate that in Europe in the tech ecosystem people are speaking English it makes me very comfortable and language of course is key to communication personally for business etc but if any of us have learned different languages in school you know that what you learn in a textbook in a school setting doesn't always translate when you're sort of in the wild and I guess I'll give you an example because the nuance and even small words can make a big difference so you know part of my language for a moment but if I say to someone hey you're the shit I'm actually giving them big compliment right but I say hey you're shit that's a very different meaning and that one word can lose a lot or convey a lot and so as you think about translating you know those words and making sure particularly in settings that are maybe sensitive how do you continue to build a model that's accurate and as objectively better particularly when we have some large competitors in the business including one that's got a you know a big logo right on the other side of this stage yeah I think I mean there's tons of factors that go into a great neural network that is that is built for translation and that starts maybe even with the balance of how accurately do you want to translate versus how fluent you want to be in the target language and there's there's obvious trade-offs there like if you're if you're translating a technical document then you are caring far more about the accuracy it's it's it's gotta really convey 100% the same meaning on the other hand if you're translating in marketing text for a website it just needs to be fluent in the other language so kind of getting that trade of right and getting the training data right for those different cases is important I think what is really important in general in neural networks and AI training is you usually when we are training those models we are training them on tons of data and mostly it's internet crawl data because this is humanity's largest database and those networks kind of consume everything that we've written so they tend to kind of put out the average of that and that's really not great so after that you have to really focus those models on making sure that they are actually producing what they what they need and we know reinforcement learning we know a lot of other methods that are kind of showing those models hey you've seen now all of that you know how to do bad translation you know how to do good translation you know how to do great translation and now I'm asking you please do great translation don't do those bad ones so there's there's a lot of work in terms of how we are training what we are training but also really in the architectures of those models how they're able to to read the text and consume and put out something something new so let's translate that to the business for a second and I'm gonna eventually ask a provocative question here but if you're a consumer it's pretty easy to go to depel.com or download the app and you can translate what you might need in your daily life and that's free but there are also other free alternatives out there like a Google translate and others so why should someone pay for depel and and along those lines what does help you build actually a very healthy and impressive enterprise business yeah I think I think we have tons of free users I mean let's first acknowledge that and this is great for for a company like ours over the first years we really didn't spend any money at marketing we just had a great free service that was kind of everybody was telling their friends that depel is around and that just making their lives easier so that was that was a great opportunity for us to get market reach and we've used that and we're still using that and when we're kind of still joking that our free service our biggest competitor it's it's not Google translate it's actually our free service because there's probably quite a few people over here who using the free translator and I don't know how many of you are paying for that that's fine that's good but there when we are when we are looking when I'm looking back we've been always pretty particular about making sure that we know that the products really works on the core on a commercial in a commercial setting that we can sell it and like kind of a few months after the first launch of the service itself we've we've added the option to to to kind of go for for depel pro and that has allowed people to consume more of a translation like if you're dealing with it each and every day if depel is the first step you're opening in the morning then obviously you're going to benefit from higher quality translation from things like security from from the ability to customize your translations more so there's far more beyond the free service itself but it is it is a very tough question I think for every company that is doing a freemium model to distinguish what is available in free and therefore what drives your viral growth versus how hard are you monetizing right right and this is obviously a big problem that people have had for you know millennia as long as people have been speaking and therefore attracts new customers you know a year ago we weren't really talking about chat GPT of course the obvious obligatory chat GPT question and now they've got translation built in as well so that the market's getting more competitive how do you respond to that because I think originally when we know we were talking about putting this chat together it was about Goliath but now there's more than one Goliath yeah I think I think it's not very very much of a different setup for us as a company I mean with one Goliath and two Goliaths I mean what's the difference maybe it's need more slingshots yeah yeah I think that I think that's I think that's fine I think I think competition drives us forward also as a company and it's really a good motivator from that perspective but yes the technology is going forward and it's even not the competition itself is from a commercial perspective but the technological competition how this whole field of AI is advancing that is putting our teams on the edge that's that's definitely a hard one and and with kind of the advent of LLMs we had to revamp a little bit of our of our research agenda and put our teams on to making sure that we can also master all of that and that we can put LLMs into into operations in in our translation fortunately all of those technologies are really near to each other so whether you're translating or that you're building an LLM that is kind of like chadgivity that's pretty similar so I think I think I think we're we're kind of in a very fortunate place either way got it now also we think about AI and large competitors I think there's a in the popular press there's a debate is AI a expansionary force or is it's kind of a contractionary force you know our jobs being created markets being spanned or jobs being lost one of the favorite stories I have about one of the companies in our portfolio using DPAL is that they were you know paying millions of dollars a year to translate their product and services into five languages and now with DPAL I think they're doing 20 more languages for a fraction of the cost do you think you know DPAL is expanding the market contracting the market is that generally true for AI I think I think in general AI is expanding the market quite a lot and and I think of us as really democratizing access to translation because think of a lot of our use cases I mean sometimes we just have a quick translation to do and like I don't know in the very early stages of DPAL I had to translate our terms and conditions into English that's been yourself yeah it's been crafted by a German lawyer I had to either to give it to that lawyer and they're gonna take like a lot of money for that or I'd have to give it to an agency and I'm not that great of a translator I kind of speak fine English but I'm not great at doing the job itself and I just threw it in and DPAL and it kind of worked I mean so so I had that instant access to that and I didn't have to ask anybody for that translation I didn't have to go anywhere it took half an hour during a break at the McDonald's actually and that that makes our lives really easier if we get that access on the other hand all of the professionals everybody who is working in the language industry is really utilizing AI for their own job I mean this has been our early adopters users everybody who was a translator it just makes their lives also so much quicker and through that combination I think we can own only through that we can cope with this huge increasing demand for content in this world like companies are building more and more content marketing is more and more important you have to talk to customers like the whole business is getting larger and therefore also through that the localization industry can kind of cope with this with these increasing demands got it okay and I want to go back to me you said about the different foundational models the proximity to them you are building your own models that is a core part of your IP should everyone be doing that who should be borrowing from the available public models who should be building their own well we didn't have really any other another opportunity another chance that there was no models available in the market like 2017 when we've been coming out to the market there was no transformer there was no and not a lot of all of that we have been built after that so so we had to start with that I think it's at the core of deep else DNA to kind of build our own stuff sometimes we're building a little bit too much but but in general if you want to really advance a certain area and and and be the best one then probably you have to have control of that whole stack of technology you have to own the models you have to be able to understand them and you have to be able to kind of build them in a way that that they're gonna be just better than your competition if your product idea does not depend on you having the best technology then obviously this does not make sense that this is a huge investment you have to have we are running kind of an academic research team within the company so that doesn't make sense if your strengths lie in product and go to market but for us this technology is kind of at the core of the company and and therefore for us that was never a question of whether we want to build it or not yeah got it well let's and let's talk about that point because in order to build and you have this research team you you have to recruit and maybe another way to think about the sort of David versus Goliath theme and we talked about this a little bit is you know the the perception around where leadership in AI is coming from is it coming from San Francisco or the States there are clearly some companies deep bell amongst them in Europe that are receiving a lot of funding and attention what is your view on that sort of the geographic balance between the two well I mean we have to admit the fact that there's more technological companies and also more AI companies in the US that's that's just a statistic I think and and that does not mean that great companies cannot come off out of Europe and that does not mean that AI companies cannot challenge the big players like the Googles of this world I mean open AI has been a small company I mean it's in the US yes but it's it's it's been a small company it's growing very quickly now but they they have taken up and challenged a Goliath too and in a sense so so definitely there's ways of new companies coming up and I think Europe has a chance doing that I think in general just due to the fact that we have a slightly smaller ecosystem over here which is which is on a great trajectory but it's still slightly smaller we have to work a little bit harder to accomplish what what we need we just don't have that much of experience in tech that has grown in Silicon Valley since like decades and we are a little bit more geographically distributed all over Europe which does not help in kind of networking effects so so to a large extent we can but we have to put some work into that and we also have to stay patient I mean it's gonna take a few years more until we're there mm-hmm and same vein you know recruit misses maybe this is a question that can that the other founders not first the room can take from take advice from in your answer so you're recruiting you know the talented PhDs that are helping to build a model are in high demand at other companies too what is it that you're looking for what is it that resonates when you're trying to say hey instead of going to a Google come join us at deep L what is it what is it that clicks what is it that they look for what is it that you look for I think I think we're building something that is just great in terms of the mission I think connecting people and making them understand each other is it's just a really nice purpose honestly and this is also why I'm why I'm at deep L and and and why I want to to be at deep L and and this is important for a lot of researchers and this is important for a lot of the in general our employees they want to make the world a little bit of a better place we can't solve all of the problems but at least we can we can get you a good coffee in Finland so so so I think that's that's really helpful and I think Europe has a lot of talent in terms of all that I mean I've been asked yesterday why is there so many Polish names in AI nowadays and apparently yes there's there's there's a lot of word that there is a lot of Poles there and then also at Open AI and at Google honestly and I think that boils down to this kind of theoretically founded education to a large extent that was my answer to that question back then and I think that also helps us in kind of the recruiting like we have we have a lot of countries that are really setting focus on that in Europe and this this this builds like kind of the talent that we really need yeah and I would I would have to admit if I can poke fun at my own country the fact that your business is about translation is probably better suited to be built in Europe because as we all know the vast majority of Americans don't know that there's any other language out there except for English I didn't want to say that I'm leaving this up to Eric and yeah maybe to wrap up because I think also a lot of times people like to think about AI what's going to happen in the future maybe a little bit that's a disruptive question if you think about entertainment and translation you know right now Disney Plus is in a lot of languages they hire staff you can see it in the credits at the end of every episode of all the different voice actors and actresses and then you have some multilingual actors you know a couple Germans come to mind Diane Kruger Michael fastbender excellent actors can do work across geographies are we getting to a world where and now that the writers strike is over in the states are we getting to a world where maybe the writers can put the script out deep l translates it across languages so you don't need an agency anymore and then you know an 11 labs takes the text and makes it voice and that becomes the soundtrack on some Dolly or Synthesia video and that's just what we're going to consume is as media content I think the world is probably moving quite strongly into that direction I don't want to say that this is going to be there in the next year I think there's parts of translation that are easier and there's parts of translations that are harder and actually movies are really hard because like if you if you're looking at it at a journal in the New York Times I mean it's it's really nicely written it's completed it contains all of the context so if you're translating that like deep l is gonna be practically flawless like it's gonna be a perfect translation of that if you look at the movie like the text and the subtitles they're just so usually very very very very small without context like you have to watch the video itself to understand what's happening with the text so this is a really hard translation but yes I think like looking at the trajectory and how things have been developing and how translation quality has been rising over over the last years I do feel strongly that that we're gonna be able to to really have a lot of the translation being done by computers and by AI so maybe we'll actually have more content to watch because it'll cost less to create and yeah and kind of you're not gonna be watching movies in English with subtitles anymore that's true you're gonna have them dubbed as in as they're doing in Germany for that's true that's true well I think that takes us about up to our time so I just want to say thank you gracias merci danke schön kitos and thank you in any other ways but really appreciate the time Yarek and congrats on the success so far and more to come thank you very much thank you and it's been a great pleasure being here at slush with you guys