DeepL is an example of a major startup success story: present in the EU, US, Canada and Japan, garnering great reviews, including this one from TechCrunch: “Tech giants like Google, Microsoft and Facebook use machine learning for translation, but a small company called DeepL has outdone them all and raised the bar high in this area. Its translation tool is just as fast as its competitors, but far more accurate and detailed than the ones we tried.”
How (claim according to the reviews on your site) did you achieve better results than Google Translate? How do you do it without the resources at the disposal of a powerful corporation?
At DeepL we have a really great team and culture that is focused on achieving the best possible translation results. This focus is what usually allows startups to beat established players in the market, and it’s also the key to our company’s success – we have a very clear vision and we try to make sure that all our efforts go towards making it a reality. So focus, determination and having the best possible people helps to compete even with very wealthy players.
The second important point is timing: when we launched DeepL, the technological landscape of machine translation was at a turning point. As we entered the market, old solutions (such as statistical machine translation) were just starting to give way to solutions based on neural networks – thus leveling the playing field.
At this point you are well established in the market, but in the beginning, like any startup, you had to fight for funding. How did you finance your activities in the initial period?
DeepL is currently profitable, we are already beyond the point where we need to seek external funding. Globally, however, this is quite a rare situation: I don’t know too many companies that have gone through such rapid growth and reached profitability so early. This can only be achieved with very strict budget discipline and … the luck of having a product that grows virally without spending too much on marketing and sales. In addition, we tried to validate our offering on the market very quickly: just a few months after launching the free service, we launched the paid DeepL Pro so that we could start generating revenue as soon as possible.
Translating our experience into advice for startups, we would say: first, minimize your startup costs as much as you can, and try to get as far as you can without external funding, because having a product as mature as possible will give you better access to the funding market. Secondly, maintain budget discipline and get your product to market as early as possible and start researching product-market fit. This will allow you to generate revenue as quickly as possible, but also to see what works and what doesn’t – you reduce the risk of not finding a market for your solution.
What is the main difficulty in creating translation software, is it the ambiguity of words?
It’s not just the different meanings of individual words, but the whole translation process. When translating yourself, and you don’t want the translation to be literal and word-for-word, you first need to grasp the overall meaning of the sentence and think about how to translate that meaning into another language (which may require completely different grammar, splitting the sentence, or using many other linguistic methods). Only then should you focus on individual words and try to complete the translation, occasionally going back to the original sentence to check individual concepts. And this is where methods based on artificial intelligence (using neural networks) reveal their advantage: the models we currently use are able to learn and abstract to the extent that this allows them to follow a natural model, rather than just translating sequentially. The result is a greatly improved ability to formulate translations that are not only correct, but also sound good and natural in the target language.
We can even go further in this thinking (which DeepL does) and not only consider one sentence, but entire paragraphs at a time. This improves translation quality by resolving ambiguities that cannot be resolved at the sentence level – but it also presents a challenge in terms of computational complexity.
How many different possibilities must a translation program “consider” before it decides to display the final translation, what order of magnitude is that?
It’s hard to say unequivocally, because neural networks don’t work that way. What can be said, however, is that translation is a very computationally demanding task: even though our user base is large and we do a huge number of translations every second, our energy requirements to power the servers are on the order of megawatts – numbers that are typical of industrial plants rather than IT startups. This shows how much computation (and therefore power) is required to use AI-based solutions. We use 100% renewable energy to power our translation engines.
What types of documents cannot yet be translated by programs at this point, but must be translated by a physical translator?
This depends very much on the user’s requirements. Depending on the further purpose of the document, a machine-translated document may be immediately usable, or it may require proofreading by a specialist in the field, or even appropriate certification. This applies to both legal and medical documents. In practice, however, many of our users who live abroad translate legal and medical texts into their native language so that they can better understand these documents.
You’re based in Germany, so I’ll take the opportunity to ask you about your advice for startups looking to expand into that market. The two main things they should keep in mind are …?
It’s hard to write a “recipe” for success because every company, every market and every product is completely different. But if I had to list two things to pay attention to, based on our experience, I would say:
- Focus on the product! Think sensibly about what the market will need and don’t get distracted by individual customers. If you want to be truly successful, it’s important to build a scalable product that will appeal to a large group of customers on its own.
Build your team and invest in your core values and the people who work with you. You need to find the best people and invest in their development and a great work environment.
- What makes it much easier to start is to build a product that appeals to individual users first: this is how companies with a bottom-up strategy (like DeepL, Dropbox, Figma) started. Getting your first revenue from individual users (rather than big companies) is much easier, and those users become your advocates in companies in the long run. While this is obviously not the only possible strategy, it makes it much easier to get started without too much funding and without dealing with corporate clients.
One last note: in general, well-established economies in Europe (like Germany) are slightly more complicated markets for B2B solutions, as companies are more “risk-averse” and are likely to lean towards established players in the market (which you won’t be at first). Therefore, taking your first steps may be easier in countries that are more open to new players.
In your opinion, what are the prospects of implementing solutions such as DeepL when dealing with refugees in Poland?
The issue of language barriers has always been one of the biggest challenges for humanity – not only in extreme situations such as the war in Ukraine. Overcoming these barriers is what motivates us – that’s why we provide a free service. So I believe that DeepL can help you communicate with refugees: either through the free service, or through DeepL Pro when the situation requires it. However, I would like to point out that DeepL does not currently support Ukrainian language. We are working on it, but providing the quality we want to offer to our users is not easy and may take some more time.