How DeepL Built a Translation Powerhouse with AI with CEO Jarek Kutylowski

Hi, thank you for having me. It's a pleasure. So I'm Jarak, CEO and founder of DeepL. And DeepL is a company that has actually started a little bit before the AI hype. We launched in 2017 and we've been using AI to tackle the language problem in the world. We are specialized in translation, specifically for businesses, like in all of those use cases where you have customers in a different country, where your company is maybe spread across the whole world, and we're trying to provide solutions which just help you cross that language barrier as good as possible. And AI has made amazing strides in making that so much simpler. Yeah, that's basically us.

Yeah, I think it was a totally different space when we started, as you say. It's been a fast-changing moment and we were really lucky to start in 2017. I think that was the moment when everything turned to neural machine translation. And I think we chose that moment really wisely because everybody had to throw away basically what they had been doing until now. Everyone had to switch over to neural and at this point in time, I think for a startup there was this opportunity to go ahead and build models that are excelling what was out there maybe in academia or what the others had been doing. Back then it was a lot of own architectures, kind of creating just the best model type that can suit translation. Like the transformer came out very quickly, but we found there's actually better architectures for translation specifically. On the one hand you need to kind of generate text, of course, but you have to also stick to what you're seeing as the source text. You have to both maintain a certain level of accuracy because the translation needs to be near to the source text, but at the same time you want to write in the target language natively. You don't want to let the model do word-by-word translations. You want to give it a little bit more creativity. So this mix of both models that are good at copying and also at writing itself, monolingual, bilingual, that was something we've been working on for quite a while and that has only continued. Model sizes are much bigger right now, reinforcement learning and all of those techniques are coming in, also allowing those models to do more than just plain translation from sentence to sentence. So it's been quite a journey. I think the advantage that we have really comes from the fact that we're focused on this one area and those models that we built, even though they might be really competing with the large ones on size, they're still really very much focused on this use case that we're building them for.

Yeah, we're looking at those and we're using them as pre-trained models. We're putting still a lot of compute on top of that. There is just this advantage of training on specialized curated data that we've built up over the years and also making sure that we have a proper distribution of all of the different languages. Those models need to be able to tackle maybe not only English and German but also a few more languages, the kind of smaller ones. And for that you really also have to have the data and give the model the training steps to look at this data.

That's really a good question. I think all of our research somehow, we tend to think about that it's academic and it's sometimes really, really model-driven. But it always has to be super applicable and always has to really go into the product, and it also means that a lot of that is engineering. I think it's like maybe 50/50 would be a good way of describing that. Performance is super important, compute is expensive, and for training, we're always a step ahead of the whole market. Back then when we came out in 2017, we had to start to build our own data centers because we essentially couldn't get the GPU compute and we had to build our own frameworks for how do we put the training workloads onto the data center. So a lot of that is kind of pioneering and that just increases your engineering workload because you cannot take the off-the-shelf product that's already out there on the market.

Yeah, like the kind of the first machines I've kind of racked myself personally. That's been pretty cool actually.

I mean luckily there was a good kind of correlation between the growth of the company and our revenue streams and the compute that we needed to build what we had to build. So we've been able to finance that pretty well. But yeah, it's been becoming larger and larger and especially with the advent of the DGX generation at Nvidia and now with Blackwell, this is substantial cost. But we also consider that being essential for us to maintain an edge and be able to train those large-scale models.

Yeah, I think everybody can scrape the web. You can be better at this, you can be worse at this. We've been doing that for quite a while already and being able to both find parallel corpora, bilingual, on the one hand, but also monolingual. That's important too, especially if you're thinking about languages where you cannot find that much bilingual data, then supplementing with monolingual becomes pretty important. It is an effort and you got to know what you're doing. I think in 2017 that was even harder. Right now you have a lot of already pre-crawled corpora on the internet. It's a little bit easier. You can try to start and kick off with those. I think extracting the data maybe out of websites, etc., it's just a lot of engineering work. Sometimes actually pretty fun algorithmic work even to do that efficiently. If you have a huge website and want to really, a full, extremely large domain and want to match which sentence matches the other, it's computationally not that simple sometimes if you want to do that cheaply. But an exciting problem to solve.

I think an important part is taking context into account. So quite often if you really look nowadays at a sentence without any context, just looking at one sentence, then even a great human translator or you and me, we cannot do a better job without really knowing what this is all about. You have to take into account what kind of document that is, maybe sometimes what is this company about that is translating that, in order to get this one sentence perfect. And with that, you give the model so much more power to do that. So I think that's a truly important one. Keeping the models fresh, of course. Data changes, language changes, things develop and you want to have the models be able to do that. And then this fine-tuning of the model, how much you want to give it a focus on accuracy versus on fluent writing. And those things are sometimes really contrary to themselves. Especially if you have languages that are really different from each other, there might be this clash of whether do you want to make it sound nice or do you want it to be really correct in a way.

Yeah, that won't be scalable. So we're not, at least not in large-scale, training models per customer. We're trying to find ways on how we can give the models the right context, how we can inject context-specific or customer-specific information for the particular use case and for this particular customer without having to retrain everything. And with the hundreds of thousands of customers that we have, that's just the only way of doing that. I think there's many companies, not only in the translation space but in general in AI, who are trying to train models per customer. I don't think that this is a particularly great way unless you have really very specialized situations in which there's an ROI on that big investment.

I think the models, we have to pre-tune them in order for them to be able to pick up what kind of language that is. And then for the technical application, this customer, they're going to maybe upload their terminology that they want to have used in their translations to our models so that it's always consistent across the whole technical documentation base. That's not going to be so important, for example, in the marketing case, when you really want fluency, when the creativity of the model, and sometimes choosing something else out of the probability distribution is going to actually make for a great translation. Whereas if you want it really to be consistent, you're going to maybe just control that on your own.

I think the whole language industry and the whole language problem has changed so much over the last eight years when we've been out in the market. That has been both driven by the availability of the technology and the ability to just throw in something into the translator and get an answer so quickly. It's sometimes even not about the cost. It's really sometimes about the speed in which you get those translations. And the demand from the market, I think, has been growing. Customers demand to have customer support in their own language. They want to see materials being localized when they want to buy. It's not such an easy market anymore if you're just speaking English as a company. So I think that has driven a lot of our customers to embrace that. Really, I think one of the biggest changes was that, and we're going to see that in AI in general, is that some of those customers of translation, even within a company like the legal department or the marketing department, they started to really self-serve on those solutions. It's not a centralized function in many companies anymore. They just go out to a provider like us, they start using our product on their own, they integrate it into their tools and do not have to rely on an external agency maybe or somebody who'd be doing translations in a traditional way. And that has changed this whole access and therefore also makes up for much more content, much more volume being translated in general.

There's a lot of hunger for quality. I think depending on which quality level you are at, you're always unlocking new use cases for being tackled by machine translation. Like whatever is enough for this single one-to-one email that you're sending to your colleague that's sitting in another office in let's say Taiwan, you don't care so much. It's honestly going to be fine. If you're then thinking about translating a contract or translating your terms and conditions and putting them onto your website in 20 different languages, that matters a little bit more and a mistake there might have really legal consequences. So if you're able to do that automatically and you're able to simplify this whole workflow a bit more, that makes for a big difference. And then in a lot of workflows, there's still a human in there checking the translation, post-editing it as we would call it. And the easier you can make this job, the less edits are necessary, the less changes are necessary. This really impacts the time needed for that process. And if you think that this person checking is a paralegal, there is really a hefty hourly salary that is associated with this process. So there's really a big return on investment on any incremental quality improvement that you can do.

I think that they will definitely be there. I think the amount or the content of translations that are going to be done by humans only is going to be severely reduced. A lot of the kind of boilerplate and boring work of translation, that's going to be all done by AI. A large part of that is already done right now and in the future even more so. I think humans are going to be still incredibly important in this process to kind of guarantee especially in those high-compliance use cases. If you think about life sciences companies and financial institutions, there's really the need even now to have multiple human translators on a single piece of text and it's going to definitely continue. I think we also have to be realistic that on the simplest cases of translation, in the most common languages, and in cases where quality doesn't matter so much, I think AI on itself is going to be doing an amazing job by itself.

Not for production, not for inference. That would just be not scalable. We're working with thousands of translators and humans to train the models and to kind of give us the feedback and the quality assurance and all of that. But you can't employ that, at least not in those volumes that we are translating, during inference time.

I do think that models are definitely more reliable and more accurate in a sense like they're not going to be doing those mistakes that we as humans from time to time do just because our brain slips. So that is an advantage for the models and that's going to be an advantage even more so in the future. I think the models still do not understand the world as we do and there is a difference there. With all of those great reasoning models and with the LLMs that we're using for language, we see they kind of get the world just because of all of the text that they've seen. But this knowledge, this understanding is not as deep as with us humans. And therefore sometimes in those very tricky situations, they just cannot distinguish on what was meant there, what was the intention of that particular text. And which is why I said context helps because it gives you more of this. But even sometimes that's not enough. And then you're sometimes running into those edge cases, you're running into half-broken sentences which due to some kind of text parsing are slightly weirder, you're looking into very short texts because they've been written for an app, the user interface of this, and the models sometimes get confused by that honestly. They haven't seen that in the training material so much or they've just seen that very rarely and they cannot cope with this added complexity.

Yeah, it's been coming up. It's not like it's not there. The models are encouraged to be a little bit creative and you have to give them the freedom to just kind of write on their own. And sometimes if they don't know what they should be doing, they start making stuff up. So you have to employ control over that. I think within translation it's a little bit easier because you can always cross-check and go back to the original text and then kind of even post-factum maybe sometimes do some evaluation of whether this has gone astray or not. And the creativity space that you're giving those models is then slightly lower than in a general-purpose LLM which is just generating text. But you have to be wary of that and we in general have seen that specialized models, and that's kind of one of the differences, hallucinate less than general-purpose GenAI models when they're being employed for translation.

It's just a newer market. I'm super excited about speech actually because it makes such a big difference. That's something that we've just put out on the market last year. I think the tech just wasn't there yet for it to be productized in such a good way, to be able to make users just happy with the output. And it has just come to that level where it's really practically applicable. I think we've gotten accustomed, as you say, to great text translation over the years by now. So it's not making that much of an impression to us. I think speech translation is just this new amazing thing that has come up. I was on my own in customer conversations in Asia, in Japan, where we'd usually have my sales team help translate a little bit or we would have even an interpreter in the room. And it's always a little bit cumbersome. You don't get really fully what's happening in the room or you get it with a ton of delay. And with now speech translation technology, you're like fully embedded, fully immersed into the conversation. It's not as good as if you really speak that language, of course. But it's pretty damn near, I have to say.

I mean yeah, you've got the speech recognition part which is super important. Then the language we're speaking in is just different than how we're writing and it's much, it's just not as clean. We don't have so much time to think about what we're saying compared to when we write something and therefore it tends to be just a little bit garbled. You don't know where the sentence starts, you don't know where the sentence ends. So speech recognition can kind of solve for parts of that because those models are really trained to package that stream of words into some coherent sentences. But still, I think the model has to cope with more and the quality of the source input is also lower just because speech recognition makes its own mistakes and then the translation model has to somehow figure out what should I do then, does that word even really match here or should I maybe substitute with something that's just more probable at this point.

That's not a big focus. I think right now the main focus is really lying on latency and just making the translation as real-time as possible. We know this is incredibly important for the user experience. The quicker you get the translation, the better you stay in the flow, the more you can match the mimics of the speaker that you're talking to with what you see in terms of translation. So the conversation becomes much better then and that's one of the most important parts. And then just the pure translation quality, being able to catch up on all of those company-specific terminology, making sure that you do not miss on the proper name of the CEO of that company, all of those things are super important to make a good impression on the users.

I'm very much looking forward to that honestly. I think this way we can really get all of this great cultural diversity that is out there in the world and the different working styles and the strengths of different countries and we can really mix and match that through our global supply chains and the way that we are working. Those of us who really speak well English have been incredibly privileged in this international world I would say, and we have the ability to now have many, many more people join this community. And then maybe even in the process of participating really learn that language and become fluent by themselves. But in the first moment, giving them the confidence that they can speak up in that meeting, they can participate when they have an idea, which quite often honestly does not occur if you're not really proficient in that language.

I mean yeah, for me there's a limit to that at some level. I think I would really still want to have friends and speak to people whom I really understand the language on my own. And I think that also brings us much nearer from a cultural perspective because the language is usually even tailored to the cultural history of a country and there's so much embedded in that. So I think there's realistically a limit to what the AI can do here. Especially in all of those private situations, I cannot imagine living with a partner and speaking through a phone with them for my whole life. That just doesn't work. But for all of those business situations, I think it's just going to be purely great.

I mean the question is usable for which purpose? But yeah, there is a gradient in how good translation quality is depending on the different language pairs. It's availability of data on the one side, and then it's also obviously the amount of work that companies like us or even what happens at academia or with our competitors can put into these particular language pairs. It's just a question of once again business return on investment and we're trying to make sure that we cover the languages best that our customers need and that they're requesting from us. So definitely there is this tier one of languages that are the biggest global languages. Then there is a second tier of slightly smaller languages where there's already quite a lot of material where you get really good results. Polish would be a good example, which is where I was born, and it's a decently large language with a good amount of training material. But then if you go into really, really smaller languages, that's going to be much harder and that's going to take more time to really get those on the board at the same quality level. I think also probably we're going to have to become smarter in how we train models and not require so much data for them in order to get those languages to the level that we're expecting to get really fluent.

We've been building a lot of separate models actually, and in the last time we've been starting to consolidate that at least into chunks of models, or models that can handle a group of languages. It's also a little bit different depending on when you're thinking about text or voice. If it's speech translation, then once again latency comes into play. Super small model sizes or smaller model sizes are important and then they might not be able to cope with all of the different languages at the same time. Just the parameter count is not enough.

Oh yeah, totally, you can do that if that makes sense. Yeah, you can do that.

Yeah, they work better. And on the other side it's also much easier on the engineering and deployment side if you don't have hundreds of models to cope with and like version them and train them independently. So it's just easier for us. Groups of languages, especially if they're similar, that helps a lot. And then if you have a group of similar languages that do not have enough data, they fuel each other and make it easier. Over the time we've been looking at many features of those models and how they map, what happens in those models with some of those linguistic nuances and our understanding of language. And sometimes these are really kind of funny things that you can find and how certain things match to each other and how you find clusters of meanings and how that all really sits near to each other. But then at the end it's so many dimensions that if you want to kind of try to sum it up and understand really what happens from end to end, that at some point in time it just gets to just far too complicated.

For us it was pretty much everything just we started so early on that as I said we had to build a lot of the stack. Even things like how do you distribute requests that are coming in from the users to the different GPUs that you have available, like you always have to kind of strike the balance on how big batch sizes do you make. You want to utilize your GPUs well but also you want to maintain low latency for your users. So you have to make sure that you have the tech which kind of groups those, understands what those requests are, sends them off to GPUs. Now in 2025 there's more common technology and it's much simpler to do. Back in the years that was definitely trickier. If you have a wide range of models, depending on language pairs, depending on the load that you're getting on the system for different language pairs, you might want to spin up new models, like spin up more models for Japanese because it's the time zone for Japanese language, and spin down other models. So we have to build the tech for scheduling all of that and reacting to load changes. GPU compute is really different than CPU compute and there's been quite a few funny algorithmic challenges to really solve in that too from an engineering perspective.

I think we don't have a problem of getting the GPUs. I mean we've gone through a few moments in time when just getting the GPUs even if you have infinite money was super hard. I think we're not at this point right now. The supply works and of course more compute would be great. But I think we would probably also need more researchers, more brains basically to also utilize that at some point. It's not only about raw computing power, although that's also important of course.

It's actually all Nvidia. We're obviously looking at the other ones and kind of testing, benchmarking. Migration is hard though, like that's not a big secret. And speed matters in this industry. So really doing big migrations is not that easy and especially as we're running our own individual architectures on the models, it's also not that easy to just go to an off-the-shelf provider of inference and just stick your model in there. There's going to be just so much more migration overhead for that. So yeah, for the time being sticking to Nvidia, but also really looking at the alternatives all of the time. We're seeing the market catching up, I would say. There's a lot of new fascinating stuff coming up there.

Yeah, I think we have to change there. I think we have to understand much better what the translation is being used for. It's less so about kind of translating the sentence from A to B, but understanding what is the full workflow in an enterprise with this translation. Is there going to be somebody reviewing that? Is that actually a second version of a translation which has been done already earlier, and that first one has been done by AI but then there was a revision by a human which introduced some changes, and then it makes sense to feed all of that input into the model in order to enable it to be even more accurate. So it's becoming more about the enterprise workflow and how you can embed the AI into that. How you can do actually more deeper product research, but also fueled by AI, in order to solve the higher-order problem and not just the simplistic translation case. And it's frankly not that trivial because we're coming from a super horizontal product as we are with just translation being embedded in so many different use cases, and you have to be smart about picking and choosing the most important ones and where you can also add much more value to those.

I mean that's just kind of part of our ongoing product discovery, like understanding what our customers are using translation for and then kind of embedding functionalities into the models that drive those and then exposing them in the proper way. So it's just like part of our normal product development cycle I'd say.

Lucas, it's been perfect. Thank you very much.