DeepL CEO Jarek Kutylowski：DeepL如何借助AI打造翻译巨头

You're listening to Gradient Dissent, a show about making machine learning work in the real world. I'm your host, Lukas Biewald.

Today, I'm talking with Jarek Kutylowski. He is the CEO of DeepL, which is a very successful generative AI company. You might not have heard of it because what they do is translation, and a primary focus of their business is enterprise, but they are making really significant revenue from a very specific generative AI use case.

I think translation is a really interesting category to talk about when we talk about generative AI because it's one of the first categories that's being completely disrupted by AI systems. Many human translation companies have gotten into trouble, and they're starting to shrink as generative AI takes off. I think it's a real bellwether for where a lot of industries are going.

This is an interesting conversation about both the business implications of running a company in this space and the technical implications of how you stay ahead of companies like OpenAI when you have a specific use case. Jarek was very forthcoming with answers to my questions, and I found it super interesting. I hope you enjoy it.

I was really excited to talk to you as the CEO of one of the most interesting generative AI companies that maybe a lot of people haven't heard of, but I have. I think you should introduce your company to our audience.

Hi, thank you for having me. It's a pleasure. I'm Jarek, CEO and founder of DeepL. DeepL is a company that 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.

1. Competing with Google Translate & LLMs

We're specialized in translation, specifically for businesses—for all of those use cases where you have customers in a different country, or where your company may be spread across the whole world. We're trying to provide solutions that help you cross that language barrier as well as possible, and AI has made amazing strides in making that so much simpler. That's basically us.

What's kind of amazing about you is that translation is such a fast-changing space. My background was actually in building translation models back in the day, and I don't think any of it is relevant at all anymore.

You're also going up against Google Translate, and all these language models can do some translation if you ask them. I feel like you're going head-to-head against these juggernauts, but beating them on quality and technology. The underlying technology is so different now in terms of what's state-of-the-art than it was when you started.

Could you talk about how you think your technology advantage works?

I think it was a totally different space when we started, as you say. It's been a fast-changing environment, and we were really lucky to start in 2017. I think that was the moment when everything turned to neural machine translation, and we chose that moment really wisely because everybody had to throw away basically what they had been doing until then. Everyone had to switch over to neural, and at that point in time, I think there was an opportunity for a startup to go ahead and build models that excelled beyond what was out there in academia or what the others had been doing.

2. Pretraining vs. proprietary model strategy

Back then, there were a lot of custom architectures focused on creating the best model type that could suit translation. The Transformer came out very quickly, but we found that there were actually better architectures for translation specifically. On the one hand, you need to generate text, of course, but you also have to stick to what you're seeing in the source text. You have to maintain a certain level of accuracy because the translation needs to be near the source text, but at the same time, you want to write in the target language natively. You don't want to have the model do word-for-word translations. You want to give it a little bit more creativity.

This mix of models that are good at copying and also at writing—monolingual and bilingual—was something that we'd been working on for quite a while, and that has only continued. Model sizes are much bigger right now, and reinforcement learning and all of those techniques are coming in, allowing those models to do more than just plain translation from sentence to sentence.

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. Even though the models that we built might be competing with the large ones on size, they're still very much focused on the use case that we're building them for.

It does seem like pretraining a model on language would inform how translation works. Do you train your models completely from scratch, or would you use a model if Meta wanted to publish a Llama model and spend millions and millions of dollars making it?

We're looking at those models and using them as pretrained models. We're still putting a lot of compute on top of that. There's an advantage to training on specialized, curated data that we've built up over the years, and also to making sure that we have a proper distribution of all the different languages.

These models need to be able to tackle not only English and German but also a few more languages—the smaller ones. For that, you really have to have the data and give the model the training steps to look at this data.

How much of what you do is basic research, and how much is engineering the models?

3. Building GPU data centers in 2017

That's a really good question. We tend to think of all of our research as academic, and it's sometimes really, really model-driven, but it always has to be super applicable and always has to go into the product. That also means that a lot of it is engineering.

Maybe 50/50 would be a good way of describing it. Performance is super important. Compute is expensive, both for training and in general, and we're always a step ahead of the whole market, I would say.

Back when we came out in 2017, we had to start building our own data centers because we essentially couldn't get the GPU compute. We had to build our own frameworks for putting the training workloads onto the data center. A lot of that is pioneering, and the same goes for the rest of the work. That increases your engineering workload because you cannot take the off-the-shelf products that are already out there on the market.

You were building out GPU data centers in 2017?

Yeah. The first machines I really racked myself personally. That was pretty cool, actually.

4. The value of curated bilingual and monolingual data

Did you expect your compute costs to be so large when you started the company? I would think it must be much more compute than what you were buying.

Luckily, there was a good correlation between the growth of the company and our revenue streams and the compute that we needed to build what we had to build. We've been able to finance that pretty well.

But yeah, it's been becoming larger and larger, especially with the advent of the DGX generation at NVIDIA and now with Blackwell. This is a substantial cost, but we also consider it essential for us to maintain an edge and be able to train those large-scale models.

You talk about proprietary data. What is that for you? I think people always thought that Google had an advantage because it was scraping the whole web and must have found a lot of parallel corpora there. What kind of proprietary data do you have?

5. How DeepL measures translation quality

Everybody can scrape the web. You can be better at this, or you can be worse at this. We've been doing that for quite a while already, and being able to find both parallel, bilingual corpora and monolingual data. Monolingual data is important too, especially if you're thinking about languages where you cannot find that much bilingual data. Supplementing with monolingual data becomes pretty important.

It is an effort, and you have 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, so it's a little bit easier. You can try to start and kick off with those.

Extracting the data from websites is just a lot of engineering work. Sometimes it's actually pretty fun algorithmic work to do that efficiently. If you have a huge website and want to crawl an extremely large domain and match which sentence matches the other, it's computationally not that simple sometimes if you want to do that cheaply.

But it’s an exciting problem to solve, really. How do you think about the quality of a translation? I think in the past maybe it was easier, in that the translations were so bad that sometimes they’d be incoherent or just wrong. But it seems like translations have gotten, in my experience, pretty high quality. What do you look at to separate your translation from a competitor’s translation, or what are the metrics that your models aren’t optimizing at this point?

I think an important part is taking context into account. Quite often, if you look nowadays at a sentence without any context—just looking at one sentence—even a great human translator, or you and me, cannot do a better job without really knowing what this is all about. You have to take into account what kind of document it is, and sometimes even what the company doing the translation is about, in order to get that one sentence perfect. Once you do that, you give the model so much more power to do it. I think that’s truly important.

6. Personalization and enterprise-specific tuning

Keeping the models fresh is important, of course. Data changes, language changes, and things develop, and you want the models to be able to keep up with that. Then there’s the fine-tuning of the model: how much you want to focus on accuracy versus fluent writing. Those things are sometimes really contrary to each other. Especially if you have languages that are really different from each other, there might be a clash between whether you want it to sound nice or whether you want it to be really correct.

Interesting. Does every customer get their own fine-tuned model, or how does that work?

That wouldn’t be scalable. We’re not, at least not at large scale, training models per customer. We’re trying to find ways to give the models the right context—to inject context-specific or customer-specific information for the particular use case and customer without having to retrain everything. With the hundreds of thousands of customers that we have, that’s the only way of doing it.

I think there are many companies, not only in the translation space but in AI in general, that are trying to train models per customer. I don’t think that’s a particularly great way to do it unless you have really specialized situations in which there’s an ROI on that big investment.

So how does it work if one customer has an application where they want a more technical translation, where accuracy is really important, and another customer just wants the language to be fluid? Are there three models to choose from, or how does somebody tune aspects like that?

7. Why translation demand is growing

I think we have to pre-tune the models so they’re able to pick up what kind of language it is. For a technical application, a customer might upload the terminology that they want to be used in their translations so that everything is consistent across the whole technical documentation base.

That’s not going to be as important in a marketing case, when you really want fluency, and when the craziness or creativity of the model—and sometimes choosing something else from the probability distribution—is actually going to make for a great translation. Whereas if you want it to be really consistent, you’re going to control that on your own.

Okay. What happens when good translation gets really cheap and easy? Are you seeing businesses operate in different ways once they start to have access to your technology?

I think the whole language industry and the whole language problem have changed so much over the last 8 years that we’ve been out in the market. That has been driven both by the availability of the technology and by the ability to just throw something into the translator and get an answer so quickly. Sometimes it’s not even about the cost; it’s really about the speed at which you get those translations. The demand from the market has also been growing.

Customers demand customer support in their own language. They want to see materials localized when they want to buy. It’s not such an easy market anymore if you’re just speaking English as a company. I think that has driven a lot of our customers to embrace this.

8. ROI of incremental quality gains

One of the biggest changes—and we’re going to see that in AI in general, I think—was that some of the customers of translation, even within a company, such as a legal department or a marketing department, 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 don’t have to rely on an external agency or somebody doing translations in a traditional way. That has changed this whole access model and therefore also makes for much more content and much more volume being translated in general.

Do you think that the quality of translation has gotten over a threshold where it’s less of a differentiator for customers, or do you think most customers are still hungry for even higher-quality translations?

There’s a lot of hunger for quality. Depending on which quality level you’re at, you’re always unlocking new use cases to be tackled by machine translation. Whatever is enough for a single one-to-one email that you’re sending to your colleague in another office, let’s say in Taiwan, you don’t care so much. Honestly, it’s going to be fine. If you’re 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 legal consequences.

9. The role of human translators in the future

If you’re able to do that automatically and simplify the whole workflow, that makes a big difference. In a lot of workflows, there’s still a human checking the translation and post-editing it, as we would call it. The easier you can make this job, the fewer edits and changes are necessary, and that really impacts the time needed for the process.

If you think that the person checking the translation is a paralegal, there’s a really hefty hourly salary associated with this process. So there’s a really big return on investment on any incremental quality improvement that you can make.

One of the things that came up when I was researching your company was a lot of human translators talking about your company and worrying, “Is this going to make me obsolete?” It does seem like we’re on that trajectory, doesn’t it? Do you think there will be human translators 10 years from now?

I think they’ll 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 boilerplate and boring work of translation is going to be done by AI. A large part of that is already done right now, and in the future even more so.

I think humans are still going to be incredibly important in this process to guarantee quality, especially in high-compliance use cases. If you think about life sciences companies and financial institutions, there’s really a need even now to have multiple human translators on a single piece of text, and that’s definitely going to continue.

I think we also have to be realistic that in the simplest cases of translation, in the most common languages, and in cases where quality doesn’t matter as much, AI is going to do an amazing job by itself.

Do you keep humans in the loop for some of your translation applications?

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

Where do humans still outperform the models? You’re talking about a legal use case, and I would imagine a human might also make a typo that a model might make. From what I see of translation model performance, it seems so spectacular that I wonder whether a human would do a better job than a translation model. Of course, I’m not doing this all the time for my job, so maybe I’m way off on that, but my impression is that it’s pretty close. Maybe the models might even be more reliable in some cases.

What are the cases where the model still really needs a human to get it to that level of quality—not just for compliance, but to actually make the translation work for the use case?

I do think that models are definitely more reliable and more accurate in a sense. They’re not going to make the mistakes that we as humans make from time to time just because our brain slips. That’s an advantage for the models, and that’s going to be an advantage even more so in the future, I think.

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 also with the LLMs that we’re using for language, we see that they kind of get the world just because of all the text that they’ve seen. But this knowledge, this understanding, is not as deep as it is with us humans.

10. Hallucinations in translation models

Therefore, sometimes in those very tricky situations, they just cannot distinguish what was meant there—what the intention of that particular text was—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 running into those edge cases.

You’re running into half-broken sentences that, due to some kind of text parsing, are slightly weirder. You’re looking at very short texts written for an app and its user interface, and the models sometimes get confused by that. Honestly, they haven’t seen that so much in the training material, or they’ve just seen it very rarely, and they cannot cope with this added complexity.

That makes sense. What about—and I’m just thinking, I have the experience of talking to a lot of enterprises about LLMs in general, and there are always these sorts of fear stories around hallucination. Is there a parallel hallucination issue in translation?

11. DeepL’s work on speech translation

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 write on their own. Sometimes, if they don’t know what they should be doing, they start making things up, so you have to control that.

I think within translation it’s a little bit easier, because you can always cross-check and go back to the original text. Even post facto, you can sometimes evaluate whether this has gone astray or not. The creative space that you’re giving those models is slightly smaller than in a general-purpose LLM that is just generating text.

But you have to be wary of that. In general, we’ve seen that specialized models—and that’s one of the differences—hallucinate less than general-purpose GenAI models when they’re being employed for translation.

Then I guess you keep talking about text translation, but you also offer speech translation, right? Is speech just a smaller market, or why is your emphasis on text?

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 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 have users just be happy with the output—and it has only 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 on 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 would usually have my sales team help translate a little bit, or we would even have an interpreter in the room. It’s always a little bit cumbersome. You don’t fully get what’s happening in the room, or you get it with a ton of delay.

Now, with speech translation technology, you’re fully embedded and fully immersed in the conversation. It’s not as good as if you really spoke that language, of course, but it’s pretty damn near, I have to say.

Have there been new challenges that have come up with the speech part?

You’ve got the speech-recognition part, which is super important. The language we speak is just different from how we write, and it’s much less clean. We don’t have as much time to think about what we’re saying compared to when we write something, and therefore it tends to be a little bit garbled. You don’t know where the sentence starts, and you don’t know where the sentence ends.

Speech recognition can solve part of that, because those models are really trained to package that stream of words into coherent sentences. But still, I think the model has to cope with more, and the quality of the source input is also lower, because speech recognition makes its own mistakes.

Then the translation model somehow has to figure out: What should I do? Does that word even really match here, or should I maybe substitute it with something that’s just more probable at this point?

Do you try to preserve the rhythm and tone of the speech? Does that somehow carry through, or is that not yet a big focus?

12. The broader impact of global communication

It’s not a big focus. I think right now the main focus is really 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, and the more you can match the speaker’s mimics to what you see in terms of the translation.

The conversation just becomes much better then, and that’s one of the most important parts. Then there’s the pure translation quality: being able to catch all of that company-specific terminology, making sure that you don’t miss the proper name of the CEO of that company. All of those things are super important to make a good impression on users.

Do you think about the impact you’ll have on the world when speech translation is very easy to turn on? It seems like it’ll really change the way businesses work, doesn’t it?

I’m very much looking forward to that, honestly. I think this way we can really get all of the great cultural diversity that is out there in the world—the different working styles and the strengths of different countries—and mix and match that through our global supply chains and the way that we’re working.

Those of us who speak English well have been incredibly privileged in this international world, I would say. We now have the ability to let many, many more people join this community as well, and then maybe, in the process of participating, really learn that language and become fluent by themselves.

13. Handling smaller languages and language pairs

But at the first moment, it gives them the confidence that they can speak up in that meeting and participate when they have an idea, which quite often honestly does not occur if you’re not really proficient in that language.

Totally. Even a world without language barriers, where you could call your friends in a different language, seems pretty amazing, doesn’t it?

Yeah. For me, there’s a limit to that at some level. I would still really want to have friends and speak to people whose language I really understand on my own. I think that also brings us much closer from a cultural perspective, because language is usually tailored to the cultural history of a country, and there’s so much embedded in that.

I think there is realistically a limit to what 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 that’s going to be purely great.

What’s the state of the art in handling less prevalent or less common languages? How much data do you need to collect to make a usable translation model, either for speech or text?

14. Multi-language model consolidation

The question is: usable for which purpose? There is a gradient in how good translation quality is depending on the different language pairs. It’s the availability of data on the one side, and then it’s also obviously the amount of work that companies like us, academia, or our competitors can put into these particular language pairs.

It’s, once again, a question of business return on investment. We’re trying to make sure that we cover the languages best that our customers need and that they’re requesting from us.

Definitely, there is a tier 1 of languages that are the biggest global languages. Then there is a second tier of slightly smaller languages where there is already quite a lot of material, so you get really good results. Polish would be a good example; that's where I was born, and it's a decently large language with a good amount of training material. But if you go into really, really small languages, that's going to be much harder and take more time to get them to the same quality level.

I think we're also probably 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 we expect—to get them really fluent.

Do you build specialized models for every language pair, or is it all combined into 1 gigantic model?

We've been building a lot of separate models, actually, and lately we've been starting to consolidate them, at least into groups of models that can handle a group of languages. It's also a little bit different depending on whether you're thinking about text or voice.

If it's speech translation, latency comes into play. Smaller model sizes are important, and then they might not be able to cope with all of the different languages at the same time. The parameter count is just not enough.

I see. I guess you could do different tokenization strategies for different languages, probably.

Oh yeah, totally. You can do that if it makes sense.

I was amazed by a paper Anthropic put out. I'm not sure if you saw it, but they were showing how it seemed like a similar set of neurons fires in their network for words that mean the same thing in different languages. I always wondered if it worked like that, but it's kind of amazing to see it.

It made me think that maybe these more combined language models would start to work better as these networks get more powerful.

Yeah. They work better. On the other side, it's much easier on the engineering and deployment side if you don't have hundreds of models to cope with, version, and train independently. It's just easier for us.

15. Engineering infrastructure for large-scale inference

Groups of languages, especially if they're similar, help a lot. If you have a group of similar languages that don't have enough data, they fuel each other and make it easier.

Over time, we've been looking at many features of those models and how they map what happens in those models to some of those linguistic nuances and our understanding of language. Sometimes these are really funny things that you can find: how certain things match each other, how you find clusters of meanings, and how that all really sits near to each other.

But at the end, there are so many dimensions that if you want to try to sum it up and understand what happens end to end, at some point it just gets far too complicated.

It occurs to me that you're one of the few companies that has really deployed deep-learning, giant generative models—I guess that's what they call them now—at scale. Can you talk about some of the engineering or operational challenges of making this work? What surprised you as you scaled up the size and volume of inference in these models?

For us, it was pretty much everything. We started so early that, as I said, we had to build a lot of the stack for it. Even things like distributing incoming requests from users to the different GPUs available require striking a balance: How big do you make the batch sizes? You want to utilize your GPUs well, but you also want to maintain low latency for your users.

You have to make sure you have the technology that groups those requests, understands what they are, and sends them off to GPUs. Now, in 2025, there's more common technology for this, and it's much simpler to do; back then, it was definitely trickier.

If you have a wide range of models depending on language pairs and the load you're getting on the system for different language pairs, you might want to spin up new models—spin up more models for Japanese because of the Japanese time zone—and spin down other models. We had to build the technology for scheduling all of that and reacting to load changes.

I think GPU compute is really different from CPU compute, and there have been quite a few funny algorithmic challenges to solve there, too, from an engineering perspective.

Are you one of those companies that's totally compute-constrained? If you had more GPUs, could you generate more revenue immediately?

I don't think we have a problem getting GPUs. We've gone through a few moments when just getting GPUs—even if you had infinite money—was super hard. I don't think we're at that point right now; the supply works.

Of course, more compute would be great, but we'd probably also need more researchers—more brains, basically—to utilize that at some point. It's not only about raw computing power, although that's also important, of course.

Do you run on all NVIDIA GPUs, or have you experimented with some of the more exotic GPUs?

16. Adapting to evolving LLM landscape & enterprise needs

It's actually all NVIDIA. We're obviously looking at the other ones, testing, and benchmarking. Migration is part of it, though; everybody knows that, and it's not a big secret. Speed matters in this industry.

Doing big migrations is not that easy, 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 stick your model in there. There's going to be so much more migration overhead for that.

For the time being, we're sticking to NVIDIA, but we're also looking at the alternatives all the time. We're seeing the market catch up; there's a lot of fascinating new stuff coming up.

Do you think it's going to change what you need to do to stay ahead of the market over time? If LLMs get more and more powerful, could general-purpose LLMs start to eat into your translation market, especially simpler translations that aren't as mission-critical? How do you think about that?

I think we have to change there. We have to understand much better what the translation is being used for. It's less about translating the sentence from A to B and more about understanding the full workflow in an enterprise with this translation.

Is somebody going to review it? Is it actually a second version of a translation that was done earlier? The first one may have been done by AI, but then there was a revision by a human that introduced some changes. 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. You can do deeper product research fueled by AI in order to solve the higher-order problem, not just the simplistic translation case.

Frankly, it's not that trivial because we're coming from a super-horizontal product, with translation embedded in so many different use cases. You have to be smart about picking and choosing the most important ones and where you can also add much more value to those.

Interesting. Have you started to offer services like that?

That's just part of our ongoing product discovery: understanding what our customers are using translation for, then embedding functionalities into the models that drive those workflows and exposing them properly. It's just part of our normal product development cycle, I'd say.

All right. Thank you very much. That's all the questions I have. Jarek Kutylowski

I appreciate your time, Lukas. It's been perfect.

Thank you very much.

Thanks so much for listening to this episode of Gradient Dissent. Please stay tuned for future episodes.