DeepL - AI: Cologne’s Incredible Word-Treasure

By Corinna Baier - Focus Magazine

Translated with DeepL and by Christian Rook

DeepL is one of the few world-class German AI companies. Can founder and chief executive Jaroslaw Kutylowski hold his own against the US tech giants with his translation program?

Office space reveals a lot about a company. There's not much going on this Thursday on the two floors occupied by the Cologne-based startup DeepL. There are colorful, soundproofed boxes in the hallway for making phone calls in peace. On the way to the kitchen, stocked with cereal, oat milk and fruit, you pass an exercise bike bolted to a desk.

Yes, there was a time when everyone wanted the office to be as homey as possible. But in the post-Corona era, even here, only one employee sits in most of the glassed-in rooms, often in front of a TV showing the heads of colleagues. Only one room is well filled. The screen in front of the 15 or so people says, "Welcome." By the end of the year, DeepL still wants to increase its workforce from 700 to at least 800 people. Founder and chief executive Jaroslaw Kutylowski has big plans.

Still better than Google

DeepL is one of the few world-class German start-ups. The company works with artificial intelligence and offers a tool that translates texts within seconds, in 30 languages. In the free version limited to 1500 characters. A paid service is also available.

The software delivers very good results, in many languages still better than Google. Companies, in particular, buy licenses and have their DeepL-version tailored to their own needs. That's what gives DeepL an edge over the big-tech competition, and that's where the value lies. And its value has risen sharply recently.

After a $100 million investor round, DeepL is now worth one billion dollars. Since then, however, the world has turned a little faster again.

OpenAI launched ChatGPT, which triggered a veritable AI hysteria and excited not only children who now have their homework done by the chatbot. The technology is everywhere. Will Kutylowski be able to hold its own against the powerful companies from overseas?

Kutylowski was born in Poland in 1983. His father held a professorship in computer science. And he, too, was a numbers person. He began programming at the age of ten. But his awareness of language was just as pronounced. If only because he understood barriers all too well. When his parents moved to Germany, he had to communicate. Later, he studied theoretical computer science.

He has little in common with the typical tech founders from Silicon Valley or from Berlin. He doesn't want to be cool or superficially impressive. The 40-year-old wears a blue polo shirt, dark jeans, no watch around his wrist, no brands. There's no Tesla in the parking lot either. this way, he actually fits in perfectly with the AI revolution, which is being shaped by rather quiet geniuses and not by big egos.

It's the hour of the real nerds. His doctorate had little to do with computer programs, nothing to do with software; it was about mathematical proofs on paper, says Kutylowski.

During his doctorate in Paderborn, he met the founder of Linguee - the platform that would later become DeepL. Linguee was created in 2009 and is an online dictionary. It distinguished itself early on by not just displaying the literal translation, but also the context in which a term is used. This worked so well, because the data-set on which the platform was based, was excellent. It was fed by documents from the EU Parliament, for example. High-quality texts that were available in different languages. In 2012, Jaroslaw Kutylowski joined as the company’s Head of Technology.

Forward to 2017, when a decisive breakthrough happened in AI development. Now, computing power was finally large enough to perform operations that had been pure theory for the longest time. Neural networks - and thus self-learning systems - were suddenly available. And new opportunities opened up for the Linguee team. After all, they were hoarding just the right data to feed such a system. They had much more specialized data than Google, for example. Kutylowski founded DeepL with this treasure and set about building his new company.

Microsoft's Bing is catching up

"There were only a few people in the market at the time, who knew neural networks well," he recalls. "So, we hired mostly those, who had a great mathematical background or a PhDs in physics." When the translator hit the market, the clear advantage quickly became apparent.

While Google Translate produced partially incomprehensible garbage, DeepL became the new gold standard. Word got around to customers. DeepL likes to emphasize, that they were profitable almost from the start.

Quality remained the standard. Stickers on the office walls still remind people today, that DeepL was the market leader. Even if some experts now see Microsoft's search engine Bing in front. Google has also caught up and now offers 100 languages.

Kutylowski, on the other hand, does not want to cover all languages. DeepL primarily includes those, that are economically relevant, because the company makes the majority of its sales from corporate licenses. Asia in particular has now become a growth driver. The second most important market is Japan. Korean was added most recently.

The DeepL brain that lies dormant in a supercomputer on Iceland - probably also because renewable energy is cheap there - and performs 5.1 quadrillion operations per second, was fed with top texts. They weren't easy to find on the Korean-language Internet, Kutylowski notes. If you ask how exactly he chooses such data, he gets tight-lipped and talks about trade secrets.

DeepL is generally pretty tight-lipped. Kutylowski has hardly given any interviews in recent years. The company also keeps its exact sales secret. Its research department does not publish any scientific papers and does not even appear at relevant professional conferences.

The guardians of the Cologne word-treasure machine remain silent. They also work in secret on "moonshots," ideas that either wont become anything or could one day change the world.

At the same time, it's all about day-to-day business: How can the AI be adapted, so that it no longer needs so much training data? How can we translate the spoken word? And they work with Large Language Models, LLM for short, on which ChatGPT is also based. The chatbot, which creates its own texts and can translate them itself, appeared last November.

"There's more pressure on the market. I don't want to deny that at all," the founder admits. "On the other hand, the current development makes my heart beat faster. From a product perspective, it's great: What else can we do with technology?"

Earlier this year, Kutylowski introduced DeepL Write, which beautifies text. If you wrote an English essay and ran it through DeepL Write, he said, you would improve it by one grade. That's how the 40-year-old explains it. But, will innovations like this be enough in the long run?

Microsoft, which has invested tens of billions of dollars with OpenAI, is building its AI into its own products. Translations are increasingly becoming a sideline business that the big players offer with high quality. Already, OpenAI and DeepL have many customers in common.

"Sure, there are other tools that do instant translations. But, that's usually not what our users want," Kutylowski points out. "Specialized tools will still be widely used. In the professional environment, you need the best quality and product support. I'm not afraid of that."

In addition to competition, however, regulation is also making life difficult for many AI startups. In Brussels, the AI pact has just been adopted. But the DeepL founder also sees this as positive: "The only important thing is that research is not hindered. Because then companies will migrate in the long term. To regulate the use cases, as the AI pact does, I think that makes sense."

As a German-based startup, DeepL has a good reputation for data protection anyway. That can be an advantage. Other young companies tend to suffer. Data protection and a lack of venture capital are cited as reasons why Germany is not in the top league. There are only two German startups among the hundred most important AI startups. Yet local AI research has been among the world's best for years. The TU Munich regularly delivers breakthroughs. The algorithms behind one of the most important image generators, Stable Diffusion, also come from the Bavarian capital.

After all, there are 500 startups working in the field, almost twice as many as a year ago. And they are getting more support. Software company SAP invested millions in Heidelberg-based AI startup Aleph Alpha, which is being billed as the German ChatGPT. But the path out of the lecture hall and into one's own office space is nowhere near as natural as in Silicon Valley, which grew up around Stanford University.

Germany lacks role models, says Kutylowski. Reasons would not have occurred to him in the past either. Moreover, the networks are not yet as developed. When Kutylowski was building his company, he learned the most from other experts who worked at startups like Slack or Dropbox. Investors are also good for that, he explains. They bring people together who then change the world together.