Ep 46: CEO of DeepL Jarek Kutylowski on Specialized vs. General Models, Beating Google and a Future with Synchronous Translation

I sat down with DeepL cofounder Jarek Kutylowski. DeepL is a comprehensive Language AI platform that enables organizations to communicate effectively across languages, cultures, and markets. Jarek shared a treasure trove of insights on the past, present, and future of AI translation.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Jacob Effron?

  Yarek Kutylowski is the CEO and founder of Deepbell, a company that most recently raised at a $2 billion valuation and supports over 100,000 businesses in AI translation worldwide. We had a really interesting conversation on supervised learning, including his take on why specialized models will beat generalized models, his view on what the world looks like when we have synchronous speech translation, so anybody can understand anyone else, and whether people will learn more or less languages in the future. We talked about how they beat Google and their approach to building their own data centers versus using hyperscalers. And we also hit on data labeling and why they've chosen to do it in-house. I think folks are really going to enjoy this episode. Thanks so much for coming on the podcast. Really appreciate it.

• 0:36Speaker 01

  Yeah, thank you for having me, Jacob. It's great being here.

• 0:38Jacob Effron?

  I mean, I guess to kick things off, I think one thing that's so interesting about Deepbell is you've obviously been doing a lot of cutting-edge AI research long before it was cool and in vogue and the thing of the moment. You know, I'm wondering like what it was like inside the company as this kind of, you know, excitement started to crescendo. And maybe we'll go specifically to the moment of ChatGPT being released and everyone starting to play around with this. Like, what was that day like in Deepbell?

• 1:05Speaker 01

  We definitely felt the excitement at this point in time. That really was a day. I think we really knew what the technology itself can accomplish. I mean, in fact, we've been building language models before too. They've been part of our translation solutions for quite a while. So from an academic perspective, that hasn't been that much of a surprise. But I think it has still been a big surprise in how magic that has felt. And how it also really worked for the awareness and educating a general public about what we are doing and about what AI is doing. And then in terms of the competition in the markets, like we've been up against Google Translate since the beginnings of Deepbell.

• 1:49Speaker 01

  And what that has really taught us is competition is good. It really creates the best outcome for the user. And we've integrated that into this like company DNA. And I think it really has been helping to even run the company. And at the end, It is about the best outcomes for our customers and keeping it in the back of our minds that there are other pretty huge players in the market. That has been a driving force for us and has been always creating this sense of urgency. I don't think that AI translation would be where it is right now without Deep Bell coming into that space. But also the same kind of moment was there with LLM technology coming up. It is amazing.

• 2:33Speaker 01

  It allows us to bring the next level of translation to our users.

• 2:37Jacob Effron?

  What are some examples of that?

• 2:39Speaker 01

  Yeah, I mean, think of the translation not being the simple use case of where you throw in a document and get a document out of that. Think of you want to have a colleague, somebody sitting next to you, working on the translation and perfecting it. And while people tend to want at the beginning to just use AI translation to like quickly translate something and get an output of that. They also at some point in time come to this idea of, hey, I want to personalize. I want to make it better. I want to work with the AI, figure out something that is going to be mine. It's not going to be like an average translation of the internet. It's going to be really my translation. And then then kind of getting that paired and getting a little bit more of this interactivity

• 3:24Speaker 01

  between the human and the AI. I think that is something that is much easier nowadays with with the LLMs having come into that space.

• 3:33Jacob Effron?

  Do you have a go to prompt that you use to like figure out? Because obviously, translation is so nuanced and I'm sure there's some sentences that, you know, any large language model at this point can translate well and others that like, you know, there's nuance in how each language works. Like when you're testing a new, you know, when GPT four comes out or GPT five or whatever, like, is there a kind of go to that you use to figure out like, hey, does do these models yet understand the nuances of like, you know, how to say something in German versus English versus French?

• 3:58Speaker 01

  We have been testing sets of those, honestly, and trying to figure out what what are the best ones and how that really changes also also the outcome. I think I think when you think about those prompts or if you would think about that, then it's really about what you want to achieve there. Is it like more about getting the most accuracy out of that or getting most most out of the fluency out there? That is, all of that is kind of baked in in our models by default. So they try to figure out what kind of text are you writing? Is it is it a more technical text? And therefore it is appropriate to go for full accuracy. Or is it more marketing? And therefore, it would be better to just optimize for native language.

• 4:39Jacob Effron?

  Maybe a good segue just to talk a little bit about like the DeepL product for our listeners. You know, maybe just and also, I think, like, the different use cases you see for translation. And like, I think this idea of you were kind of alluding to it in like, there's the technical side, there's the marketing side, but maybe just how you think about the different buckets of like, what people want in their translations.

• 4:59Speaker 01

  Yeah, translation is super broad. And while we are on this like very general mission to solve this kind of language and translation problems, specifically really for businesses, and obviously with AI all of the time. But I mean, at the end, it's only 18% of the world which speaks English.

• 5:17Speaker 01

  Even in the US, which is like probably the biggest English speaking region, 20% of people speak languages other than English at home. And and kind of helping companies to seamlessly communicate across these languages, that's gonna make their operations both internally much better and much more efficient, but also helps them to go into into new markets. And this is pretty much the two big kind of spaces that we see DeepL being employed in. And the use cases and the and industries may vary quite a lot. It's it's everything from law firms, manufacturing, retail, media, government, pretty much name it. Translation is so broad.

• 6:02Speaker 01

  But what you'd usually see the most is high value use cases where accuracy, quality matters. And just just giving you maybe like two examples, like for for kind of the external side of things. Nikkei, major media company, they're they own the Financial Times, but they're based out of Japan. They're using DeepL to translate thousands of their articles across Japanese, English and Chinese to reach a whole new set of readers, pretty much. But both internally in Japan, but also outside in the world where people want to get informed about what's happening in Japan. And then Like if you look in the internal, um, in the internal use case,

• 6:47Speaker 01

  it's very often multinational, big companies. They're located in different countries, uh, which are quite often in highly regulated industries, where like you have a document coming in in German, uh, but actually the specialist for, for tackling that kind of legal contract is sitting in the U S. And, and how do you make that, how do you make that work? Or like a Japanese car manufacturer that we're working with, they have all of their R and D in Japan, uh, but their customers actually sit in Europe and in the U S. And, and you wanna make sure that the information flow on, on what your customers want, what they're asking about, influence the, the roadmaps, the development of the products. And, and for that, you really have to have this very seamless communication patterns

• 7:33Speaker 01

  between, uh, between those different countries and offices.

• 7:36Jacob Effron?

  One thing I think is just super interesting and maybe to ask it directly is like, how, how did Google not win this category? Like you, you look back at the original transformer paper, and I think the example they had was like a translation of like French to English, right? And so obviously this has been something that, that they've been thinking about a lot. I think it's very relevant to today because people will look at lots of problems and they'll say, God, that incumbent is just so well positioned in that space. You've clearly been a huge success story, uh, like what, what do you attribute? Like what happened as this market unfolded? You know, I'm sure when you were pitching folks in 2017, you know, uh, there were lots of people that were like, yeah, Google's just gonna figure this out. Like, how did this kind of unfold to, to get to the position where you are today?

• 8:13Speaker 01

  A very quick answer, being fast to the market, I would say, and then keeping the pace on that and like really keeping, keeping innovating. And making that answer a little bit longer, um, I think it really comes down to building this strong academic level of research within the company. Um, but also being able to specialize on a, on a particular use case. And then for us, that's really like this translation use case, but high value use cases for businesses. But then I think also like an important part of that really, the fact that we have been established in Europe. I mean, right now we've grown to a pretty global company. Uh, but like being at the center of so many different languages, like when,

• 8:57Speaker 01

  I, I live 50 kilometers or half an, half an hour drive, uh, from Belgium, France and the Netherlands. So like there's so many different languages around. I think that, that, that makes the team understand the problem. And makes the team also be like really motivated for solving that problem. And especially with like engineering research. Like we should not underestimate the power of motivation to solve something that is meaningful and like really important to people and users.

• 9:27Jacob Effron?

  And I guess like, you know, I mean, to that point, maybe like, how do you think about the like the for your research team, the stuff that you guys focus on and then the stuff that you're like, you know what, like the underlying model providers are pouring billions of dollars into training these general models like we don't, we don't need to reinvent the wheel here. Like, well, you know, when it comes to research today, like, how do you guys think about like, these are the areas we should be spending time in versus the areas that, you know, it makes sense to just essentially outsource to folks who are, who are spending money on your behalf?

• 9:57Speaker 01

  Yeah, I mean, we are a very much build it yourself company, I think, I think from that perspective. And that it once again comes back, I think, to the early days of where things just haven't been there when we have started, when you started building. But not only models, but like even the tooling on or data centers or any of that, that just, that just wasn't there. But also like, I do think, and I do believe that having like the ownership of this whole vertical stack from the product, like even starting at the go to market, the product, the, the engineering of that, and then the research, it really

• 10:38Speaker 01

  allows you to kind of trickle down the problems very well. And, and being able to figure out things that you maybe cannot do if you just do prompt engineering. But if you have like all of the models parameters, the training, the architecture, all of that in your control, and also understand very well how those models work, you can figure out somehow better on how to put these pieces together and how at the end to figure out a customer's problem and how to solve it.

• 11:07Jacob Effron?

  Well, I think that's so compelling. And we've heard from a lot of folks that basically, you know, this combination of, of kind of doing your own research, you know, making your own models. And then after deploying an application, that feedback loop just gets really tight and allows you to kind of build a better, more tailored product. Are there any examples of that that come to mind where you're like, if we were just prompt engineering, we never would have been able to fix X or, you know, do Y on the product. Like, I love an example of that feedback loop, maybe between the application out in the wild and then tweaking something on the model side.

• 11:36Speaker 01

  I think what we've built in terms of like the ability for our customers to embed terminology in their models is a good example of that. Like, we've been running that for years already and like none of the other translation providers have been able to integrate it in such a good way. And it's like a crucial thing for a lot of businesses because they want to control the kind of language that their employees or the products, that their tech manuals are using. And you can therefore like influence the translation there. But it's not really an easy task because you have to take all of the time,

• 12:17Speaker 01

  make sure that you follow all of the grammatical rules, even if you're kind of replacing some words. You have to make sure that maybe sometimes even if the customer has said, "Translate A with B always," maybe sometimes it really doesn't make sense because that word A has two different minis and in that sense you don't want to do it here.

• 12:36Jacob Effron?

  And is that like all live, is that logical and coded kind of in a big model? Or do you have like a series of smaller models that you're like chaining together? Like, how does it kind of all work under the hood?

• 12:46Speaker 01

  We run different sets of models. I mean, depending also really on languages and language pairs. And also, given the different data sizes that are available for different language pairs, like there's going to be much more translated material available for German English, for example, comparing to Polish English.

• 13:08Speaker 01

  The model sizes might differ too. And then you might sometimes like group models into

• 13:14Speaker 01

  into certain areas because the languages are similar to each other and the models can learn from each other there. Or you might even like operationally bundle models together because it's going to be easier to just purely run on the infrastructure that you have.

• 13:28Jacob Effron?

  Now, it must be such an interesting challenge for your team to your point of just having, for the different language translations you're doing, there's just such different size data sets, right? And so you probably can run a much bigger parameter model for, you know, German English than Polish English. And, you know, obviously, I think there's, it sounds like, you know, I could imagine a world in the future where, you know, these models learn so much about the inherent basis of language that you'd run it all through one big model because you, you know, the German English model benefits from the Polish English model. But it sounds like maybe we're not quite there yet and it's still helpful to have kind of models that are specific to, you know, a language-to-language pair.

• 14:04Speaker 01

  Or you might just have smaller models than at this point in time, which like from an inference compute point of view, and that that is definitely just given the scale that we are operating at, something that we're looking at always, it might make more sense to have slightly smaller models which are handling individual language pairs.

• 14:23Jacob Effron?

  That makes total sense. You don't need a model that can write Shakespeare sonnets to do your scientific publication translation.

• 14:28Speaker 01

  Exactly, yeah.

• 14:29Jacob Effron?

  Super interesting. I guess like, you know, I'd love to understand, it seems like you still use, you know, a lot of human translators kind of in and around the DeepL offering. Maybe just talk about that a little bit, like where it fits in today, you know, how that's evolved over time, and then, you know, how you see that evolving, I guess, over the next decade.

• 14:44Speaker 01

  I think the influence of human data has been, has been rising all of the time, and I think it's gonna only be getting more and more important. I mean, we've seen with the building and the development of LLMs reinforcement learning, like really kind of coming, coming up. And we've been, we've been even running big data annotation kind of projects and tasks, like something that ScaleA, like you know, ScaleA, of course, having great, having built a great business out of that. We've been running that internally for years already with like human translators, for them to help us both train the models

• 15:25Speaker 01

  on how do we want the translation really to look like, but also looking at the quality assurance of that. And I think that is like extremely important for these kinds of specialized models that we are building, comparing maybe to like the very general ones. You really always like, your customers have a high expectation on quality and a high expectation on the quality being like steady all of the time. So your quality assurance needs also to be, to be top notch. So a lot of human translators we're working with, kind of the interesting and exciting part of that is being, you really want to have people who speak the language very, very, very well. And so you're hiring across the whole globe. You want to have

• 16:09Speaker 01

  the Brazilian Portuguese native speaker doing the Brazilian Portuguese translation, really. So that make, that makes it that really exciting, sometimes also a bit challenging and complicated in terms of, in terms of hiring. But I think that's That is necessary right now and it will also remain very, very much a central area of what we're doing.

• 16:29Jacob Effron?

  I mean, it seems like data labeling is one of those things that obviously makes models so much better. And, you know, I imagine a lot of companies right now are thinking, you know, do we, is that something that we should have in-house? Do we partner with like, you know, Scale or Surge or some of these other folks? Like, obviously, you guys have made the decision to do it in-house for now. Like, how do you think about that generally? Do you think that's what will mostly unfold for folks building kind of specialized models like yourself? Or, you know, like, I guess, does Scale and Surge keep supporting just kind of OpenAI level efforts? Or what happens for, I'm sure we'll see many more companies like you over the next few years. You think they do it in-house or they use some of these vendors?

• 17:05Speaker 01

  I mean, disclaimer upfront, we are starting to also thinking about outsourcing parts of that. So it's probably not that clear of an answer even for us. The thing is, it depends really how much control do you want to have around that whole process? And how much the quality is important for the task that you're doing. And I think that's going to be really different depending on which data labeling task you have. For some of them, maybe it's going to be better to just have a huge pile of data, but like the quality being like at some kind of lower level. Whereas for some of those, you really want to have a small set with like the best possible

• 17:50Speaker 01

  quality. And the ability to even kind of particularly choose single individuals which are doing these jobs and which are working on your tasks, that is going to influence the quality all of the time. And like, we've seen people who've been performing very, very well at those tasks, but then fell sick for a week and the quality dropped for a week. And then they got back to that. So if you're keeping in close touch with that, you can much better sort out those instances and make sure that you always have the data that is required for your own task.

• 18:29Jacob Effron?

  I feel like there's, obviously, this is Initial wave of data labeling was kind of just like a bunch of data. You know, that that was kind of maybe broadly applicable. And now it's such a rush to get specialized data. And so I think it'll be interesting to see whether this high quality specialized data, which is clearly what people want now. If that, you know, if folks like yourself are best suited to both identify who those people are. And then also, to your point, continually analyze like, hey, are we still at the quality bar? Or does that end up happening at the at the data labeling company level?

• 18:56Speaker 01

  And maybe, and maybe even keeping the funnel of those like finding those people. I mean, selecting them and then making sure that that you understand what what quality their work has. And that you get to a common understanding of what the work should really be. But also even like, as I said, starting at the beginning, you have to find people for that. And that is, that is a large hiring endeavor, quite a lot.

• 19:22Jacob Effron?

  Totally. I can imagine that. I can imagine that part one might want to outsource. But certainly you guys have such a good perspective on what a high quality translation looks like, that it's hard to hard to outsource that part. Another thing that's really interesting to me about, you know, since you've been building in this space for so long, is I think a big question a lot of builders have now is just, it feels like I could build something to get around the, you know, the holes that models have today. But in two years, as the models get better, like that stuff that I built might be irrelevant. And maybe I should have just sat on a beach and waited for GPT-5 to come out and then built that. I'm sure over the years of building DeepL, there's like plenty of things you built to make the product work that now, you know, are not needed because these models have gotten better. How do you think about, I mean, any reflections on the stuff that you have built?

• 20:06Jacob Effron?

  And then today, how do you think about that? Like plugging the holes, I guess, these models have versus knowing that they're probably going to get better. And, you know, how do you allocate resources around that?

• 20:15Speaker 01

  I mean, it depends a little bit on what kind of company you want to build. And I think for a company that wants to be cutting edge and that wants to be innovative in a way, this train of thought is really dangerous that you just described. If you want to be at the forefront, you have to innovate. And that really means throwing away a lot of results. Some of those being deprecated by kind of competition and somebody Coming in with a model that maybe solves it all, as you said. But sometimes it's like really even your own research, like you figure out, okay, this idea doesn't work. The other idea works much better. I think research is actually based on these iterations.

• 20:59Speaker 01

  And it's not only the probability of finding a great solution with everything that you do. But even with every failed attempt on creating a new model, architecture or whatever. You gain a better understanding of the problem, which allows you to get to the next level. Like very rarely you can go from here to here without failing a few times in between. You have to just go those tedious steps. And I see that there might be like this inclination to be like financially responsible and cut such investments. Or as you said, like go to Hawaii and drink a coffee. But I don't believe that you can build an innovative company this way.

• 21:44Speaker 01

  Like maybe like call it founder mode on the research side or something like that. You have to have those convictions, I think.

• 21:52Jacob Effron?

  Well, I'm glad we got our founder mode shout out in there. It makes us officially a venture sponsored podcast. Is there like an example of a failed experiment you ran that you feel like taught you something that like actually ended up being super valuable in the product?

• 22:04Speaker 01

  I mean, like obviously on the research from like it's hundreds and hundreds of experiments, of experiments there. I think one bigger project that we've experimented around with and that has been part of the translation industry kind of for years. Was the idea of kind of building custom models for individual customers. That has been, that has been something that historically has been done because. It pretty much was the only possibility to to reach certain quality levels before AI translation came up and maybe like in the the very early years of that. And and we've been trying to figure out ways on how to do that in a scalable way, in a

• 22:49Speaker 01

  repeatable way, where you don't have to have like an AI engineer working on on a particular customer's problem, or a customer not having the the the need to to hire people for for that. I think what we've learned through that is that really the the overall quality of the models and the ability of translation models that are just out of the box has surpassed any gains that you could get from that. And and all of this additional hassle just really doesn't make so much sense. And you can't get to a scalable level there.

• 23:24Jacob Effron?

  Makes total sense. What do you think of like OpenAI's new O1 model? Like, when that comes out, I'm I assume you start thinking of all the things that that that paradigm of model could do in your world. Yeah, what was your reaction to it?

• 23:37Speaker 01

  I think the reasoning part probably makes like it it kind of underlies all of the translation, because the ability for models to understand what's going on out there in the world obviously makes them better to to to translate. You you'd have to know what you're talking about in order to translate really very well. But at the end, like, the reasoning part is probably not at the core of of what is what is so important for for translation itself. So probably a slightly less exciting update, actually, from from our perspective than than what we've seen over the the the other years.

• 24:22Jacob Effron?

  Are there model capabilities that would be meaningful to you? It's like, you know, I assume you're eagerly awaiting, you know, a model that can do, you know, X or Y. Like, what what are the things that you'd be like, wow, if if, you know, GPT-5 or 6 was able to do this, that would be such a game changer for the DeepL product?

• 24:39Speaker 01

  Yeah, I think I think what I'm really what I'm really most excited about is if we can come to a state where we can do more with less compute. I mean, we are we are like right now in this this very like with this with this game of like investing more and more compute into all of the models that we're doing and that's that's obviously bringing a lot of results. But it's like to me, it feels it's a bit too much brute force. Maybe maybe we can just do our models a little bit smarter, kind of work on the architectures. Like we've seen over like with Transformer coming up, we've seen how much architectures,

• 25:21Speaker 01

  not only compute, can make a difference to to what the capabilities of of models are. And I think like maybe even the big Gen AI players right now are not really necessarily incentivized to invest into smaller models as they have this kind of monopoly on on huge compute and it's maybe down to players like us to somehow figure out ways of of how to do things slightly more with slightly less less compute.

• 25:50Jacob Effron?

  You're obviously using such large scale, you know, and so I think you have had an early look into some of the infrastructure required to do this. You know, what have you kind of learned, I guess, over like, I'm sure you guys have gotten way better at this. So so any advice for folks that maybe will will hopefully be getting to your scale at some point in the in the next years and then, you know, as we think about the landscape of tools that are and aren't available, you know, where do you think like there there are good solutions to support this and where are you like, oh, God, I really wish there was someone building something to help me with X because it's annoying when you get to the scale we're at.

• 26:22Speaker 01

  I wish there were more folks building GPUs or you like solutions. You and me both. Like everybody out there. So that's that's that's one point. I think I think when looking at like some kind of some kind of advice, like be

• 26:44Speaker 01

  Be conscious about where you want to run your loads. I think especially at the beginnings and in the current state of the market, hyperscalers are great. And using them as a kickstart is definitely a good idea. But then if you come to scale, maybe try to think about your own data centers. I mean, we've been operating them since the beginning. And but that was that was a clear and easy choice for us. Because there was nothing available as for us. The options are different right now. But even when it comes to cost and your cost efficiency, there is a lot of advantages of running your own.

• 27:31Speaker 01

  And in general, I think that the kind of the platform that we have right now for GPU compute, like the overall tooling that's out there, it's still at pretty, pretty early stages and that compute is really rare. So if you think about general purpose CPU compute, while we've seen a lot of progress in cloud computing and we've been building sophisticated abstraction layers, one on top of the other, that didn't cost us so much, right? I mean, like CPUs are pretty cheap. You can you can run those virtual machines on virtual machines with a higher level language on top of that, some kind of general purpose. And that's that's all fine. That's also sustainable.

• 28:14Speaker 01

  If you think from a GPU perspective, I think we are kind of far more early with that. Computer's scarce. It is it is so powerful. And therefore, if you want to have this big scale, you want to run things sustainably, like honestly, both from an environmental but also commercial point of view. And then you have to optimize there. So so kind of think of. What you can do and what you can't do there.

• 28:37Jacob Effron?

  You know, it sounds like along the journey, you, you know, you obviously early on had to build your own data centers. Now there's better offerings from the hyperscalers. If that had kind of existed in the early days of Deepbell, like, when do you, when do you think is actually the right moment in the company's history? You know, if the hyperscale options have been available to transition to your own? And how do you think about that for, for other companies today as they're thinking about when to, you know, when to partner with hyperscalers first build their own?

• 29:02Speaker 01

  Yeah, I think for the big loads that we have right now, like kind of for our research and like, or like the massive inference loads, like even with our free service. We want to switch even now, just because of the cost aspect. And also something that I didn't mention earlier, it's also sometimes about hardware availability. Like if you want to run and train on the, on the newest hardware available, then you're just gonna be so much faster to the market. If you, if you run your, your own data centers. So, so there's still a lot of, a lot of

• 29:38Speaker 01

  virtue to that. I think what nobody should forget is kind of which costs that really also occurs in terms of your development speed and all of that. Like, running your own is just so much more complicated. And we're moving large parts of our stack towards, towards hyperscalers, towards the hybrid cloud. And only kind of cutting out what needs to be run for either efficiency reasons or maybe even like for security and data protection requirements from our customers. What needs to, what needs to remain on premise with our own data centers.

• 30:17Jacob Effron?

  Super interesting. I feel like, you know, some people that aren't deep in translation like you are, might look at like AI translation and be like, this is like a solved problem. It feels like it's like solved. I'm sure, I know it's far from that, but like, how would you articulate, I guess, like where we are today in terms of, you know, AI translation generally? And like, you know, does it, you know, where doesn't it work? And then, I guess, what's your kind of estimated timeline for when this becomes a solved problem?

• 30:45Speaker 01

  I think quality really differs depending on which language where you're looking at. And, and there's, that's both good and bad. I mean, you have those language pairs that are usually the most used in the world, which actually work very well because you have a lot of training data available for those. And also everybody focuses a little bit on. On on creating uh creating those so so as with overall model capabilities depending on language the same for translation um the more data you have that the better uh things uh things here really are and um I think then in terms of the use cases like there like you can do a lot right now already with translation I I mentioned like newspaper

• 31:27Speaker 01

  articles being translated automatically without any any any human supervision um after after the fact like definitely any kind of one to one communication like sending out an email or something like that I would totally say that this now is uh that this is this is for for the best resourced languages this is this is really solved very very well by now if you think about kind of publishing your marketing website and you're a billion dollar company um yeah maybe having like somebody in the loop there uh to check on that uh is gonna is gonna make a lot of sense if you're building an operating manual for a nuclear power plant um I think it also makes sense to not even have someone who's

• 32:12Speaker 01

  gonna be watching out for the quality but somebody who's gonna take like the responsibility and the accountability for uh for the for the for the for the final result I I mean this is this is kind of this generic problem that we're getting with AI I mean it helps us create a lot of uh unload a lot of our work uh but at the same time we gotta be really cautious about who's at the end responsible uh responsible for that and and as a as a last point I I think um I think we still have like much more potential for like kind of how the user experience and how this interaction with the AI can look like in in translation.

• 32:52Jacob Effron?

  How do you evaluate models like as you're building your own models and they're getting better like I guess how do you know you know it it seems like in some sense there there's probably just like objectively like you know did he say the right word but a lot of it seems like a lot of translation is somewhat subjective too like is you know uh think about different translations of of works of literature there's like people like they're all correct but people like one you know find it much more beautiful than another like how do you solve or what have you learned I guess too about about doing evaluations over the years?

• 33:19Speaker 01

  Yeah, I think that's the part where like in business it's maybe like a little bit easier even.

• 33:23Jacob Effron?

  Yeah, I guess that's fair you're not you're not translating the Odyssey you're you're doing the the nuclear power plant manuals.

• 33:29Speaker 01

  But I mean, literature, that's tricky because I think the translation itself is a is a work of art on its own. I mean, it's really not about the accuracy and it's sometimes even not about the fluency of the language and how native that feels for somebody who's been living in that country for for years. It's it's quite often about what does it mean from a from a linguistic point of view. And I think nobody in translation or like in AI translation is looking at that really as a problem or optimizing any kind of any kind of model over that. But in general, like evaluation, that's that's been quite

• 34:08Speaker 01

  a journey. I would say like, as we started in 2017, like synthetic evaluations like Blue Score, that was that was what was state of the art for for translation. And the problem is like, we very, very, very quickly have reached a level of translation quality, which those metrics cannot really capture.

• 34:31Jacob Effron?

  There's a problem in every in every part of AI right now.

• 34:34Speaker 01

  Yeah, yeah. But I think like here, we've been like very early at this stage, stage already. And like, there's still a place for these kinds of metrics, like, like when you're training models to get like a rough orientation where you are. And like, there's been obvious like iterations on on some of those scores, they've been betting better, they've been getting better. But at the end, like the real test is always human evaluation and what what we're doing. And and once again, this is this thousands of translators that we're working in that come into into play and they're judging those translations on kind of how accurate they are, how nuanced, how native they feel, but also always in

• 35:14Speaker 01

  kind of a comparative way, like always relating like the next gen model or competition or some kind of other model. Getting an absolute level of this translation is ninety seven and this is and this one is eighty five, you can't get it really, it's too unspecific for for that.

• 35:35Jacob Effron?

  Makes sense. I mean, I think, you know, everyone's feeling one of everyone's favorite topics in AI today is like, you know, talking about moats and defensibility and, you know, I feel like we wouldn't be a venture podcast if we didn't talk about that. And so, I mean, I'm curious because, well, one, obviously specific to your industry, but also just generally, like, there's so many AI startups coming out now and, you know, I think they're facing their own, you know, behemoths in, you know, that seem well positioned for this stuff. You know, like, how do you think about that? Like, I guess there's a spectrum of folks being like, none of this matters. It's just like shipping fast and doing right by customers and like, you know, who cares? And then the other folks that are like, look, there are some spaces that you might be better positioned

• 36:15Jacob Effron?

  than others to actually, you know, as a startup. You've obviously lived this, like, and I'm sure you've heard all this discourse about moats and defensibility and your investor pitches and whatnot. Like, how do you think about this whole talk track?

• 36:26Speaker 01

  Yeah, best question ever. I think, like, really, from from our perspective, it's it's a lot about those specialized models versus very, very general ones. And and I think if you think about the ecosystem more broadly, there is definitely very many areas where it's just like financially doesn't make so much sense to build specialized solution for those. It's maybe like the long tail of use cases and you cannot afford to kind of buy a black world cluster for training those within within your own company. You will have to kind of base off standard technology.

• 37:07Speaker 01

  When it's when it comes to those like super big use cases that you have in the world, and I think translation is indisputably one of those, specialized models really make a lot of sense. And what you can build through that is is really this this ready-made solution and a solution that like vertically integrates. We've heard this so often from our customers, like, we don't want to run POCs internally. We want to have some. That just we plug it in, it works. Our employees know it may be the best already because they've been using that for the for the consumer use cases. We just we just go for it. We roll. Not a not a not a big deal. And that's been always a question in tech, I think, and with software companies.

• 37:50Speaker 01

  The the companies with the better products have won. I think it's even a bigger question right now with the advance of like general AI, which which can do everything. But it's actually pretty complicated to get it to this task reliably in a way.

• 38:05Jacob Effron?

  Totally. I mean, I feel like this is the big question, right? Like, you know, what general purpose L L M's as they get better and better, what like, you know, what spaces do they subsume is just because they are learning things across different domains. And obviously, it seems like what you're finding on the ground is actually, it turns out, you know, a model that's just trained on the translation side with really good labeled reinforcement learning on that really outperforms a model that can do a ton of different things. And I'm sure it'll be really interesting to see how that develops over the over the years.

• 38:32Speaker 01

  Yeah, I mean, that's that is basically the essence of that. Yeah.

• 38:36Jacob Effron?

  What's what's next for Deep Bell? Like, you know, I'm obviously, I'm sure you guys have big ambitions. You know, what do you think about like building down the line? Obviously, there's some ancillary areas to what you do, like, you know, there's there's writing, there's, you know, speaking. You know, I guess, yeah, how do you think about where where Deep Bell goes in the future?

• 38:57Speaker 01

  Yeah, I mean, I mean, there's obviously like this part of like general translation quality, like making it even for like even maybe the smaller language parts, like, let's not forget about those countries out there. But I think like you touched upon one of the one of the big points for me, and this is this is really spoken language and and voice. I think this is gonna be like the next frontier and kind of translation and where we are right now. The The way that the language service industry and like the way that we consume even content that comes from abroad has totally changed with the advent of good AI-based text translation. We haven't gotten there yet when it comes to having conversations.

• 39:42Speaker 01

  Like, we still both speak English on this podcast, uh, rather than me speaking in Polish, uh, and, and, and you doing, uh, you doing English. Um, and, and I think at some point in time, this is, uh, this is gonna, uh, this is gonna be changing. We've seen some demos from like the big players there. Uh, we're obviously running quite a bit of our own research on that, but bringing that to a product and like a usable way of, of an integrated everywhere, uh, that is, that is something that I'm, that I'm really excited about. And I think from, from the perspective also from the both history, that's gonna be an important one for us.

• 40:16Jacob Effron?

  I mean, if we, if we do have like real time, you, you, you know, uh, like ubiquitous translation models for speaking, uh, what, what changes happen in the world? Like, I'm sure you thought about this a lot. Like, like, what does that world look like?

• 40:28Speaker 01

  I, I think we'll really start thinking about the world in a different way, uh, especially like on the business space. And I'm particularly interested here about the business space. Uh, I think that technology will be pretty quickly up to a level to, to be able to, uh, for us not think really where, where we're located business-wise. Like, I'm gonna do a tech startup here in Cologne and like get my co-founder in Tokyo and the other one in the US and like, we're gonna just be maybe potentially speaking in, in our own languages and all of that is gonna be happening in, in, in real life. But then also like really on scale for businesses, like for all of those businesses that are,

• 41:10Speaker 01

  um, that are operating in different languages, giving employees access to education, learning resources, knowledge, doing that so, so much more easier, uh, that's gonna, uh, that's gonna, that's gonna help a lot. And, uh, and while like it's gonna, like the AI, I think, is going to be able to bridge all of those gaps, I think there is still going to be, caveat, a lot of interactions where kind of the personal connection and, and the cultural aspect, uh, will kind of give.

• 41:43Speaker 01

  A lot of value to learning those languages.

• 41:45Jacob Effron?

  Yeah, I was gonna ask, like, will kids learn languages in the future? Like, it seems inevitable we'll get to this kind of place at some point. I don't know when, but like, do you think in like thirty years, kids are taking, you know, that extra language in school?

• 41:58Speaker 01

  Yeah, I'm asking mine to learn languages still. I think that goes back to, once again, like, what you want to accomplish with languages and communication. I think like the world, through the fact that we're getting better and better translation capabilities, might actually go more and more into a mode where local languages are gonna become more and more important, because you can still access the most common ones, but maybe it's not gonna be as important for individual languages, for everyone to try speaking, to try speaking English. But then, if you go back to like your private, your personal use cases, I think languages are gonna be still incredibly important.

• 42:43Speaker 01

  And learning about cultures, like, people love that for a reason. That makes a lot of sense. It's also like really good for our brain development when we're in school to learn languages. It's, I mean, we're still learning addition and subtraction, although like computers are far better at that than any of us can ever be. And I think that will, that will, that will make a lot of, that will make a lot of sense. And you really don't want to talk to your partner through a phone for like the twenty years of your relationship. That just doesn't work.

• 43:17Jacob Effron?

  Do you think so, I, I, the, like, do you think people will learn more or less languages in the future? Because obviously, it'll probably be easier, you know, there'll be all sorts of tools to learn them, and at the same time, it's not gonna be a business requirement, but to your point, more of like a personal and learning thing. So if you were to put a stake in the ground, like, in thirty years, does the average person speak more or less languages?

• 43:36Speaker 01

  I guess probably less, I have to say, but there's gonna be more fun with that. I think those people that were gonna do that, they're gonna do that out of this kind of personal interest. And that goes back to like, why do people learn playing chess? I mean.

• 43:51Speaker 01

  It's it's fun, really, and and I think languages are so core to human beings. I think there's gonna be a lot of excitement about that even in in the future.

• 44:01Jacob Effron?

  Yeah, I love the I love the chess analogy because I feel like it's so interesting. Obviously, we have, you know, at this point, AIs are so much better at chess. And, you know, it was like interesting at one point to, you know, when when they were neck and neck. And now, yeah, these AI chess bots can play each other. But people still really care about the the human side of it and the learning and and also just. There's something, you know, there's something fun in the imperfections of it as well. Like, the the exact perfect version is is not actually what people wanted.

• 44:29Speaker 01

  Yeah, and it's an intellectual challenge that we have, like. So much of what we're doing every day is actually for for kind of capturing our interests and and making sure that we have something that we consider valuable to do in in our life. Like, thanks to technology, a lot of the less, like, at least in in in our countries, like, a lot of the the kind of hassle has been already outsourced to technology. And I'm not only talking about AI, but like washing machines and stuff. So so we can put more effort into things that that that just make us feel better.

• 45:06Jacob Effron?

  Yeah. How close are we to these like synchronous speaking models? Like, I guess, when what year, when doing this podcast, are you gonna be speaking in Polish? And I'll be I'll be speaking in English and it will be seamless?

• 45:18Speaker 01

  It's it's always really hard to speculate here, and thank you for for putting me on the spot.

• 45:24Jacob Effron?

  Like, a good host is supposed to do.

• 45:26Speaker 01

  I guess it's gonna take us a few years to to to get there. I expect early, like, early products to to to be out on the markets kind of pretty quickly. But kind of perfect thing that it took us a while for text and spoken language is just so much more complicated. It's it's just like this, it's it's so stream based rather than really being chunked into sentences. And and we speak so carelessly that the AIs are gonna have much more trouble figuring out what we wanna say.

• 46:00Jacob Effron?

  On the technical side, you know, do the main challenges feel, obviously, there's tons of challenges around latency, you know, and getting these these speech models fast enough. You know, there obviously, to your point, it also feels like there's more ambiguity in speech than there is in text. Like, any other technical challenges you think about that, hey, we'll have to overcome this before we get there?

• 46:18Speaker 01

  Yeah, I mean, latent season, important part you mentioned that ambiguity and and just this different different type of language even like if you if you look at a transcription of of a conversation even even if that was like a a properly conducted conversation, I would say, uh, rather than a wide brainstorm in front of a whiteboard, uh, it's just it's just so unstructured. The language is completely different. We are gonna have to teach the models to translate different languages, basically.

• 46:50Jacob Effron?

  There's obviously some ancillary spaces to your world, too. You know, I think of like the voice dubbing space and, you know, language learning apps like what do you think of of kind of what's happening in those spaces and and maybe just generally like what are the other spaces you pay attention to that are maybe like adjacent to to what you guys are doing?

• 47:05Speaker 01

  Yeah, I mean, language learning is pretty exciting, I think. That's nothing that we would focus on as a as a as a company. Even though we hear from a lot of people that are actually using Keepal to learn new languages, just like as a teacher who can speak that language much better than than them. I I'm pretty excited about that technology because until now it was really so expensive to learn to speak to speak a language, like you had to have your in-person teachers. And and learning a foreign language well, and especially like in the spoken variant of that was was a really privileged thing

• 47:44Speaker 01

  to do. I think that can change with the technology. You can really have fluent conversations with models. These conversations actually don't have to be fully accurate. I mean, it's it's it's it's meant for learning. So like a model which hallucinates a bit, that truly doesn't matter at all.

• 48:04Speaker 01

  What is what is important is that you you just get the ability to speak, express yourself, get a little bit of dialogue into that. I'm wondering a little bit about how enjoyable of an experience that's gonna really be.

• 48:20Speaker 01

  Because it's just so much more fun to go for something with somebody to a to a dinner and and kind of try to try to learn the language. Overdid rather than speaking to your phone.

• 48:32Jacob Effron?

  I love it. We always like to end our interviews with a quickfire round where we get your take on a few questions. And so maybe to kick that off, what's one thing that's like overhyped and one thing that's underhyped in the AI world today?

• 48:44Speaker 01

  That's super easy. Overhyped, general models, underhyped, specialized models. I mean, we've been talking about that for a while, but like there's so much more attention to general models. But really the specialized ones is where the value gets created right now. So that's, I think, an important one.

• 49:02Jacob Effron?

  Totally. What's been the biggest surprise like in building AI features in DeepL? Maybe like one thing that you thought would work that didn't, or one thing that you were like, I don't know what if this is going to work and it actually worked quite well.

• 49:13Speaker 01

  Oh my God, that's, that's not so easy. I think beating the tech giants, that was a big surprise. I think having the pace and being able to deliver and back then in 2017 before anyone else. I mean, we obviously betted on that, but was I sure that it's going to work? Probably not.

• 49:37Speaker 01

  And then when it comes to what didn't work, yeah, I think at the beginning, I thought really that the technology alone would be enough. That you just can, can build the tech, you put it out to the world. Maybe me being an engineer, there's an obvious thought to, to, to, to think. But then I've really noticed and we've noticed that if you want to deploy this technology to people, you have to build more. You have to have the product, you have to have the commercials, et cetera, et cetera. Kind of considering, considering the bigger picture is important.

• 50:08Jacob Effron?

  What's something you've changed your mind on since, since building DeepL?

• 50:12Speaker 01

  Okay, I'm going to come with a different angle here, more on the cultural aspects of, of running a company. I've become a big radical Kendra fan, kind of making sure that like you run a company through very direct, very open, very honest communication and I haven't fully understood the power of that at the, at the early stages there. But that's become very, very important to me.

• 50:39Jacob Effron?

  Yeah. Was there like a turning point for you that, that got you there?

• 50:42Speaker 01

  I think that was just a journey and you notice like so, so often that you haven't been fully, fully, fully transparent and how that has hurted both parties actually. So that, that was just like constant learning curve there.

• 50:57Jacob Effron?

  Yeah. What AI startup like outside of your space are you most excited about right now?

• 51:03Speaker 01

  I think you mentioned that at the beginning, perplexity.

• 51:05Jacob Effron?

  Yeah.

• 51:05Speaker 01

  Uh, I think obviously I'm feeling a little bit attached to, to them just because of the fact that they're, they're a small player in a, in a market that is dominated obviously by, by, by Google search, like providing this LLM based search, like what they're doing is great. I'm hearing so many people who, who are actually using that product. And to an extent like this story is a little bit similar to, to, to what we're doing. So feeling a lot of empathy for, for them and, and doing that.

• 51:36Jacob Effron?

  Yeah. The band, the band of folks taking on the giants and constantly being asked about moocs and defensibility and everything. I feel like you guys, I'd assume there's like some, some get together where you guys can all commiserate over the, getting to ask the same damn question every time.

• 51:49Speaker 01

  Yeah, we have, we have, we have some common investors. So, so we see each other quite often on those off sites. So, so we can, we can discuss it.

• 51:56Jacob Effron?

  Yeah. Well, if you hadn't started DeepL and I guess like you were just in, you know, uh, obviously a very entrepreneurial person, like, you know, in this AI era right now, what other like completely different application would you go start if you were, if you were not at DeepL? Anything else pique your interest of what you could do with LLMs?

• 52:13Speaker 01

  I, I would go for medicine. And once again, I'm, I'm not a doctor, so I don't have a clue about that space really, but watching it. And my, my motivation for, for building tech is always to make the world a little bit better. And I hope like with DeepL, we can, we can, we can do a small step towards that. But with medicine, like the possibilities are really so huge and that will have so much impact on, on, on our lives. And it's not only like drug discovery, but like democratizing access to like medicine to in general to doctors. That's, that's, that's, that's going to change the world like really in a significant way, I think.

• 52:54Jacob Effron?

  Well, it feels like, I mean, I don't know if you guys have healthcare customers, but like translation, obviously, and there's a whole medical translation industry in the US. And I think certainly the ability to translate, you know, cutting edge research and clinical protocols and even just communication between folks that speak different languages. It feels like you could, you could really play a part in that.

• 53:11Speaker 01

  Yeah, yeah, we do. We do have, like even at the frontline, really. When, when you have like healthcare workers and you have people coming from, from, from different places, especially in Europe. That is, that is a pretty common use case.

• 53:24Jacob Effron?

  Yeah, makes sense. Um, you know, what company you think would be most interesting to just run AI at right now? If it wasn't, if you weren't at DeepL?

• 53:32Speaker 01

  Once again, easy and yeah. And that's not because, not because of the stock, the equity, it's.

• 53:40Jacob Effron?

  I assume it's to say that you can allocate all the GPUs to your, to, to DeepL, right?

• 53:44Speaker 01

  Yeah, maybe, maybe, you know, no, I mean, I'm an engineer, like building software is pretty cool, but hardware is even cooler. Like, that's, that's, that's made that clear.

• 53:53Jacob Effron?

  Yeah. Well, this has been a fascinating conversation. I'm sure there's, there's all sorts of things folks will want to dig into. Where's the best place for them to learn more about you, DeepL? I'll kind of leave the, the last word to you.

• 54:03Speaker 01

  Yeah. I would say like, ask anyone who has experience translation on a daily basis, like find an expert.

• 54:12Speaker 01

  Find maybe a translator who is in your, who is in your community. They will know DeepL by heart. I'm, I'm sure about that. So that's, that's the best place. And I mean, this is how DeepL has, has really grown, like all this, this viral word of mouth.

• 54:27Speaker 01

  Marketing loops kind of, other than that, obviously visit our website. You can test DeepL Translate and DeepL Write for free. At deepel.com. You don't even have to sign up. It's just like, go to the website and, and, and start off translating.

• 54:42Jacob Effron?

  Yeah. How have you, it feels like you have so much translator love. Like, how have you navigated that? Cause I imagine there's some narrative, one could say, oh, AI is going to replace Elvis. And, and I feel like you've done a great job kind of building a great brand in that community. Any like learnings around that? I feel like lots of founders are dealing with that where people love these hypey narratives of AI, you know, replacing X.

• 55:01Speaker 01

  I think people have just noticed in the industry that their work is going to be just so much more productive. If they let the AI maybe do some of the, of the more mundane part of that and, and I'd really adding value where, where value needs, needs to be added. So, so a lot of kind of the. has come has come from that and and then also like the the amount of content that has been created over the last years and and that people want to localize into all of the different languages that has become so large I think like humanity would just not be able to cope with that in a manual way so it was it was a natural way of where this where this where this only could go

• 55:43Jacob Effron?

  Totally. Well, thanks so much. This has been a fascinating conversation. Really appreciate you you coming on the pod.

• 55:48Speaker 01

  Jacob, thank you very much for having me. It's been a pleasure.