How DeepL Beat Google With Speed, Not Scale

In 2016, while tech giants were still relying on statistical machine translation, a small team in Germany quietly launched something that would change everything. Their neural network-based translation tool was so superior that it "just blew away everybody else."

That company was DeepL, and today it's valued at $2 billion, serving hundreds of millions of users monthly and powering translation for over 100,000 companies. What started as a PhD student's fascination with breaking down language barriers has become one of Europe's most successful AI startups.

In this interview, founder and CEO Jarek Kutylowski reveals the untold story of how DeepL beat Google at their own game, the critical importance of speed in AI markets, and why specialized AI models might be the future over general-purpose solutions.

Jarek: I grew up in Poland, which was at this point in time kind of a country in a switch between the communist system towards a post-communist one. I got access to technology a little bit later than I think people in the rest of the world, and I was just amazed by how much you can achieve with tech, especially with software.

I moved to Germany with my family, have been thrown into school very, very quickly. I didn't speak a word in German, which made me struggle quite a bit. I still remember the first day in class when I walked in and I couldn't even really spell my name. I had to learn how to survive in an environment maybe a little bit more complicated.

I was lucky to learn German very, very quickly, so that made me feel welcome and belonging in that community quickly. I learned language and communication is super important in this world. If you want to belong to a community, you need to be able to understand each other.

Jarek: Basically, from the very early days when I was fascinated by technology, this was really the path that I have chosen in school and later. In studies, I majored in computer science. I went on to do a PhD in really theoretical computer science. I really enjoyed a very solid theoretical foundation.

I think doing a PhD really also builds up a lot of in-depth process because usually you're really thrown into a field of research that is not yet discovered. You have to really uncover something. You have to go through that process pretty much on your own, at least. For myself, but also for all of the peers that I've seen build up a lot of resilience for the future of their life.

After my PhD, after my really academic years, I spent a little bit of time working for a larger corporation that didn't really fully suit me. And then at some point in time I really realized I want to build something. I want to be part of something bigger and really starting a company, founding a company, kind of embarking on that deeper journey. That was something that really fascinated me.

Jarek: I think in 2016 to 2017 there was this great moment when it has become pretty clear, at least in the academic environment, there is a lot that can be done with neural networks and AI. So that was an excellent point in time in which we started at the beginning really to play around with the technology and see what we can do with that, how that can be applied for some problems that we might be thinking of.

And I had this background in language. I've been living in two countries I knew what it means to speak different languages. I knew how big of a problem that is from a European perspective. Each and every time you want to travel to another country, but most importantly, do business with another country, there is going to be a language barrier.

If you look at Germany, the companies in the country are selling to French customers. They are selling to Italian customers. They're going to be selling to Polish customers, and you can try doing that just by speaking English all of the time. But at the end, every person wants to be addressed in some ways in their local language. They will understand much better what you're offering them.

Jarek: I think for companies it's really hard to establish those new markets. What you have to do is you have to hire people really in this specific region or you have to find people who are qualified to speak in a particular language in your country, and that can mean even like doubling your sales headcount or that may mean a lot of customer service jobs, for example, within your company.

Translation language industry is being told to be like 60 billion. If that is more efficient, if that is more productive, they are going to build better products and be more successful in that market.

Jarek: I think the very early days were pretty specific for DeepL because we knew that this problem of translation, that this is a big one. I think what we didn't know is whether the technology that we're going to be able to build is going to be enough to solve those problems and be better than our competition. I think that was the unknown, but it was pretty clear that there is this big problem that can be solved.

I think what we didn't know really particularly well also was how to embed maybe that technological solution into real life applications, how we can go fully to the market, and for that we've just kind of tried to go the path of least resistance.

Jarek: We built the technology. We put out a service that was super basic that just gave a very bare bones access to the technology itself but at the same time it was also simple to start using as few barriers as possible like no login, nothing like that just to go there and start using the product and for us that was a great way to validate whether this technology and this early product idea actually makes sense and if there is a market opportunity for that.

And out of that we've seen very clear signals that this is actually what people want and this is actually what users need through just purely looking at the usage numbers. It was super simple, I think the next step and the challenge there was to look at whether this is something that people are going to also be willing to pay for, whether there is a monetization pattern for that, and this is something that we then started doing in 2018, introducing new functionality into the product, potentially putting those behind pay walls and seeing whether we can convert customers, we can convert free users into being customers.

Jarek: A lot of that at the very beginning was really based on a gut feeling, and I think at the very beginning you have to have those hypotheses which come out of the founding team. But I think specifically when you're working in such a slightly more consumerish market at the beginning you have to rely a lot on quantitative data rather than on qualitative finds. The more customer focused that came slightly later when we started shifting the product towards a B2B, an enterprise persona as a buyer.

Jarek: I think the biggest challenge as a first time CEO is really making sure that you're making your decisions and that you're kind of pushing the company at the speed that you could, because you don't know what the next step potentially is. You have to rely on a lot of advice on how the company is going to look in the future. You have to find out things on you kind of understand the point which you are in and extrapolate to the next point.

Doing that is incredibly slow. Maybe sometimes if I were now to found another company and do it the second time, I think I could be just much, much faster in that. Other than that, I do not think that we as a company made too many big mistakes. I think speeding it all up would just make such a big difference, I guess.

Jarek: If you're in a startup and especially if you're in a competitive field like ours, like with all of the big tech also having their solutions, you always have to grow and you have to think about growing fast. That is essential for a company, for even the company's motivational health in a way. It always needs to very fast grow all of the time and that made us obviously also scale the business in terms of employees, in terms of the number of customers that we have, in terms of the amount of products that we're offering and all of that was really tailored to that.

In 2018, the business was operating profitably already. That is just a function of how cost effective growth motion is, where you don't have to hire a lot of salespeople. Your customers pretty much come to yourself because they're convinced of the product, but at the same time being financially responsible, we've been lucky in that way as a company, really.

Jarek: I think the biggest lesson is once again it's all about the speed. We've been the first company to the market with an AI-based neural network translation solution which just blew away everybody else and made sure that we gathered our first user base, that we gathered the early adopters, that they started spreading the word of mouth in the world, and it was all about speed.

If we have come up like half a year later, a year later, I don't know if that would have worked so well. So the biggest lesson is really it's all about speed. You have to move fast. You have to figure out what is the next challenge? Approach it, solve it, and then move on.

Jarek: If you decided to grow your company very fast, you need to be aware of the change that is happening there. You have to make sure that you understand what is happening and that you help all of the people in the company be on the change journey.

I think the biggest problem with moving fast in a high growth company is really the fact that everybody in that company has to go through a lot of change because nothing is going to be the same this year as it was the last. Change is hard for us. This is something our brains, they do not really like going through, but also trying to make sure that everybody knows why we have to go through these change processes, why is it important for the company, that everybody is on board with that.

That is incredibly important if you're building an organization, but also if you want to go through the change on your own, and I think context and understanding of the why helps in addition to that really a lot.

Jarek: So there's pretty much two types of AI models that are on the markets right now the very general generative AI models that can do pretty much anything, and the specialized models which really focus on creating one particular solution at the best quality possible.

I think the very big generalized models or those models that you can actually use for pretty much anything, they have surprised us with their ability to do a wide variety of tasks. I think what needs to be understood is that some of those models actually do not perform that well on particular instances on particular use cases of problems that we might have, especially in business, in the case of translation for stable quality and accuracy that is not only on one email, on two emails, but really across the board on a wide range of inputs, and I think this is where specialized models can really shine.

They are usually qualified for a very particular reason, for a very particular use case and can deliver that quality very, very consistently while at the same time being potentially more cost effective and quicker to run, which matters very much in some of the use cases.

Jarek: So I think for businesses it's always good to take a look at what business problem are we trying to solve here, what are the general solutions for that versus what are the specialized solutions and what kinds of advantages they bring and especially I think the world hasn't changed too much in a way that technology itself doesn't yet fully solve the problem. You also have to have the product and the integration and the user experience and the UI solve for having that problem really meaningfully impact your workforce's productivity efficiency. In that case, those specialized solutions will usually come with a full suite that helps solve that holistically.

I think if you're thinking about bringing AI into your business, you really have to start with the basics like what kind of problem do you want to solve? What is maybe at the core of the performance of your company, what is important for your company, and what is potentially working slower where you have a problem. And starting out with that business problem you can try to find out what are the potential AI solutions for that.

I'm not an advocate of trying to find applications for AI just as a technology. I'm an advocate of starting with the problem and then looking for the AI solution that can potentially solve that problem in a very good way, because through this you will be optimizing what is really worth optimizing in your business rather than trying to apply AI to pretty much everything.