Jarek Kutylowski is the founder and CEO of DeepL, a complicated AI-powered translation tool known for its impressive accuracy and natural-sounding translations.
DeepL is on a mission to interrupt down language barriers for businesses in all places. Over 100,000 businesses and governments and hundreds of thousands of people in 228 global markets trust DeepL’s Language AI platform for human-like translation and higher writing. Designed with enterprise security in mind, firms world wide leverage DeepL’s AI solutions which might be specifically tuned for language to rework business communications, expand markets, and improve productivity. Founded in 2017, DeepL today has over 1,000 passionate employees and is supported by world-renowned investors including Benchmark, IVP, and Index Ventures.
Are you able to walk us through the early vision behind DeepL and the way the corporate’s goals have evolved since its founding?
After I began the corporate back in 2017, we were at a turning point with deep learning. It had really began to realize traction, which inspired the name of our company. We’ve at all times had a vision of empowering organizations to thrive and scale globally with no language barrier standing in the best way, which remains to be our mission today. So we had began experimenting with the technology and its application inside language, and quickly saw how powerful it may very well be – it was clear that neural networks and AI were going to be the following big breakthrough in translation, and that what we could achieve with neural networks would far surpass existing traditional solutions.With this in mind, we decided to package our findings right into a product and share it with the world.
What began as a fun project quickly evolved into something more significant. Just a number of months after officially launching our inaugural product, DeepL Translate, in 2017, we had a few of crucial European and German firms knocking on our door, keen on deploying our solutions. This realization in regards to the relevance and significant importance of our product to larger organizations inspired us to dive deeper into their problems, and invest time really understanding how our technology could solve them. We built out a more concrete business strategy and clear monetization plans, focusing first on releasing the product to the general public and determining the business side later. Fast forward to today, and we’ve grown considerably, expanding our team, our products, and our customer network. We now have hundreds of thousands of individual users of our specialized AI translation and writing tools, and over 100,000+ business and government partners worldwide, including 50% of the Fortune 500 and leaders like Nikkei (owner of the Financial Times), Coursera, Deutsche Bahn, Zendesk and more.
You founded DeepL in a field already dominated by major players like Google and Microsoft. What was your technique to carve out a distinct segment in such a competitive landscape?
We have been up against Big Tech since we first started off, which has kept us agile, modern, and motivated to supply our users one of the best solution on the market. But I challenge anyone claiming they’ve solved the challenge of language for business and enterprise-use-cases with general models – that is where DeepL stands out and is basically unique on the planet of Language AI and translation.
We’re different from general-purpose AI systems in that we’ve built our AI translation and writing solutions to be specialized for language use-cases; this specialization enables us to realize higher accuracy and precision while minimizing hallucinations and misinformation. It’s one thing to translate a menu when you’re on vacation – but most of our customers are businesses and knowledge staff, where the stakes are much higher. Imagine a lawyer who must translate highly nuanced, critical, and confidential legal documents; or a significant media publisher who relies on our platform to translate hundreds of stories a day, in real time, which might be then distributed to hundreds of thousands of world readers. These complex use cases demand a level of accuracy, customization, data privacy, and security that general models can’t provide. Data protection and security are also central to our offering, so we adhere to the very best privacy and security standards including all GDPR requirements, offering state-of-the art data encryption and more.
DeepL has recently launched its first in-house LLM. How does this model differ from other large language models available in the market, and in what context is it considered superior?
Our next-generation translation models are powered by proprietary LLM technology designed specifically for translation and editing, which sets it aside from other models in the marketplace and sets a brand new industry standard for translation quality and performance. Unlike general-purpose models that rely solely on public Web data, DeepL’s LLM advantages from over seven years of proprietary data curated specifically for content creation and translation. It also uses human model tutoring, with hundreds of hand-picked language experts who’re trained to refine and enhance the model’s translation quality.
For this reason, we’re proud that DeepL is widely considered to be essentially the most reliable and preferred Language AI solution for businesses and professionals on the market. In blind tests, skilled translators have found that our next-generation LLM requires significantly fewer edits than other platforms, with Google and ChatGPT requiring between two and thrice as many edits to get the identical quality.
Are you able to explain the method behind training DeepL’s LLM? How much human input is required to take care of accuracy and nuance in translation, and the way do you balance that with the computational elements of AI development?
We train our LLM on a mixture of highest-quality linguistic data, in addition to the expertise of hundreds of hand-picked language experts. This two-pronged approach, combining each computational elements with human feedback, allows us to support our customers at scale while also ensuring that the standard and nuance of our translations remain on point. Human input is basically crucial, primarily in data curation, quality assessment, and providing feedback for continuous improvement. DeepL’s translation quality just would not be where it’s without this aspect.
DeepL has recently expanded into 165 latest markets and added support for 3 latest languages. What was the strategic considering behind targeting these markets and languages, and the way has the expansion impacted DeepL’s user base?
Language is a world problem that touches businesses across almost every industry – whether or not they’re facing language barriers internally amongst colleagues, or externally with customers and in diverse markets where they operate. And as a research-led company, the whole lot we do is informed by our mission to interrupt down language barriers, and the feedback we’re hearing from customers and businesses. Our decision to expand the supply of our products – DeepL Pro is now available across 228 global markets – and add latest language capabilities, just like the recent launch of Traditional Chinese, is rooted in our mission to interrupt down these barriers, guided by customer feedback and market research.
We’re at an exciting moment in AI, where adoption shouldn’t be only a trend but a necessity. Corporations are able to embrace AI but are in search of technology that delivers real value and ROI, and we’re proven to drive significant impact – a 2024 Forrester study revealed that the usage of DeepL delivered 345% ROI for global firms, reducing translation time by 90% while driving a 50% in workload reduction. By growing our reach and capabilities, we will deliver these real-world advantages to much more markets, people, and businesses. Take Traditional Chinese, for instance – it’s the first language for over 33 million people worldwide. So this is basically driving the size of our business and helping us to fulfill rising demand globally. Today, now we have a network of over 100,000+ business customers, and it’s growing fast.
With over 100,000 customers, including major organizations like Deutsche Bahn and Zendesk, what are the first challenges and opportunities you’ve encountered while scaling to fulfill enterprise and governmental needs?
After we’re within the room with a customer, they’re clear on the worth and ROI we will provide to them – – even within the US, where you may think English is the one language spoken, most firms operate with multilingual teams and navigate the complexities of a world marketplace. Numerous people forget that! The worth and impact of our specialized AI translation tools and writing services is obvious. So it isn’t a matter of whether or not they should implement language AI tools, but one interesting query we do hear rather a lot from enterprises is across the personalization capabilities of our service.
Enterprises often adhere to specific brand guidelines and require AI solutions that reflect their unique brand voice and industry terminology. They need to have the option to speak consistently, each internally and externally, at scale – whether or not they’re writing something or translating a document, they need the content to sound like something they’ve written themselves, of their specific brand style. At DeepL, we have designed our tools with this in mind and offer quite a lot of more interactive, customization features including DeepL Glossary, which allows businesses and professionals to customize translations for specific words and phrases; in addition to features like tone setting and more. This level of fine-tuning and the power to make the content their very own is something that is basically vital to our customers.
As DeepL expands and its AI models are used more broadly, what are a few of the key ethical considerations you prioritize when developing language AI systems, particularly when handling sensitive or confidential information?
Any organization considering AI tools should at all times ask these questions when evaluating models and corporations. Many AI solutions don’t provide adequate security, often share user data with third parties, and train their AI models on user text, creating mistrust and making many users hesitant to make use of these tools with work-related documents. As I touched on earlier, DeepL takes a distinct approach – security is core to our mission and product offering.
Our customers come from a wide selection of industries, including highly regulated sectors equivalent to financial services and legal, and the confidentiality and security of their data is crucial. It’s core to our product offering that we adhere to the very best enterprise-grade privacy and security standards to guard user data, including state-of-the-art data encryption and GDPR, ISO 27001 and Soc2 Type 2 compliance. We also don’t use any text from our subscribers to boost or train our AI models, understanding that lots of our users translate sensitive information.
What can we expect next from DeepL when it comes to product development? Are there any major features or innovations that you just’re particularly enthusiastic about?
We’re a research-driven AI company focused on innovating with purpose, and our purpose is to resolve the language problem for people and businesses in all places. Seeking to the long run, we’ll proceed to include customer feedback into our development and concentrate on expanding the capabilities of our Language AI platform to make it an excellent more comprehensive solution for all language-related business needs. Numerous what we’re hearing from customers today is that the more interactivity and personalization we will provide, the higher. So expanding our features and functionality to handle this can be a really interesting area for us, make the DeepL experience much more dynamic and interesting for our users.
In the following 5 to 10 years, where do you see DeepL’s technology fitting throughout the broader AI landscape, especially as AI continues to rapidly evolve?
5 years is basically an extended time for AI development so to me what really matters is the following 12 months! We’re working on bringing the identical revolution and disruption that we did with written language to other multimodal capabilities, which I’m enthusiastic about.
