Founder and CEO, DeepL SE
Why Specialized AI Models Are Challenging the Frontier Labs — With DeepL CEO Jarek Kutylowski
Jarek Kutylowski is the CEO of DeepL. Kutylowski joins Big Technology Podcast to discuss why specialized AI models are beginning to challenge the industry’s biggest general-purpose systems. He explains how purpose-built models can deliver better accuracy, lower latency, and reduced costs, and why companies are increasingly using model routers to choose the right AI for each task. We also explore how real-time translation could help businesses expand across borders, why voice represents AI's next frontier, and whether glasses and other wearables could give models a better understanding of the p...
About Jaroslaw Kutylowski
Jaroslaw Kutylowski, CEO of DeepL, appeared on the Big Technology Podcast on July 8, 2026, to discuss the rise of specialized AI models. He argued that purpose-built models can offer better accuracy, lower latency, and reduced costs compared to large general-purpose systems, and noted that companies are increasingly using model routers to select the appropriate AI for each task. Kutylowski also highlighted real-time translation as a tool that could help businesses expand across borders, and described voice as the next frontier for AI. Kutylowski stated that AI translation tools like DeepL can reduce the upfront investment needed for companies to enter new markets by handling documentation, sales communication, and customer service in multiple languages. He described the ability for every person to talk to another person in the world as a "beautiful application of AI" that is worthwhile from both a business and human perspective.
Transcript (49 segments)
A
Alex0:00
Big AI models have gotten a lot of attention, but specialized models are on the rise. Let's talk about how these purpose-built models are challenging the status quo with DeepL's CEO, Jaroslaw Kutylowski, who joins us today in a conversation brought to you by DeepL. Jaroslaw, welcome.
J
Jaroslaw Kutylowski0:17
It's a pleasure being here with you, Alex. Thank you.
A
Alex0:19
Thanks for being here. Okay, so DeepL is obviously building specialized models. We'll get into it in a moment, but I first want to go through the bigger question about specialized models versus bigger models. The conversation recently has all been about these big foundational models. Bigger is better, the more compute and data you add, the better results you get on a generalized path. But there is a notion now that smaller specialized models are starting to challenge the bigger models, outperform them in some way, and be the models that are routed to when a model router decides what to go with. So talk a little bit about why we're seeing this rise in specialized models and where you see this going.
J
Jaroslaw Kutylowski1:08
The fact that generalized models get a lot of attention is because they can do a plethora of different things at the same time. For many of the tasks we do every day, there is not enough incentive to create something specialized. But there are areas where specialized models can outperform those big models, not only on quality or performance, but also in the triangle of performance, quality, latency, speed, and price. In real applications, all three parameters matter a lot, especially if you're pushing a lot of data through those models. If your company is translating millions of documents or hours of real-time audio, it makes sense to employ models that are better at that task and more cost-effective. So in business applications, you'll see teams relying on specialized models.
A
Alex3:02
Okay, but let me ask you a question. This entire generative AI era kicked off in part because of the transformer model, which was built specifically to translate language. DeepL does a lot of language translation, helping businesses like an American company wanting to operate in Brazil translate their website and customer service. You've built specialized models to do it. But if the entire generative AI moment is built on a model innovation meant to translate language, why would we need something specialized to do that as opposed to the bigger models with that foundational innovation baked in?
J
Jaroslaw Kutylowski3:52
You're right, the transformer model and language translation are very tightly coupled. But early on it was clear that transformer models can do more, and when they do more, they lose a bit of the capability they had when made only for translation. The parameters need to be divided among many things, so if you keep the model strictly focused on one task like language translation, it can perform better and more consistently. Generalized models tend to be better or worse depending on input, while specialized models have better consistency. Quality assurance is built in to ensure whether it's an email, marketing material, or a technical patent application, the language quality is as good as possible. Also, large language models are usually trained mainly on English and a few languages, so to go truly multilingual, you need models and data that can cover all those languages.
A
Alex5:52
So it's interesting. The larger large language models have that transformer capability baked in, but when you query them for language, you're also querying a model that can advise on fitness, science, and medicine. It becomes so sprawling that as it tries to boil the ocean, it might end up with worse performance in specific areas. That's why there's a movement towards more specialized models. Am I getting that right?
J
Jaroslaw Kutylowski6:31
Yes, specifically the super important second step of training models. It's not only about giving them access to a large amount of training data. The reinforcement learning aspect that comes in heavily nowadays shows those models what we actually want them to do in more detail. This enables them to solve real-life tasks like preparing a diet or a recipe. If you run this reinforcement learning step on too many tasks, the model will be very broad. But if you focus on making sure the model understands it needs to provide the best possible translation for a given text, it will be able to do that better.
A
Alex7:42
So to go with the big one, you sacrifice accuracy and latency as opposed to going to a smaller specialized model where you can be sure you'll be more accurate and less latent.
J
Jaroslaw Kutylowski8:04
Exactly. Accuracy comes from specialization. Latency and speed are natural functions of model size and the ability to process it faster on the same hardware. Especially in real-time applications like live speech translation, latency is key to the user experience. It's often a combination of accuracy and speed that delivers a good experience.
A
Alex8:49
Yeah, if you're waiting for a live translation and the model takes a couple of minutes, it ruins the conversation.
J
Jaroslaw Kutylowski9:01
You can't have any engaging discussion. In use cases where people use real-time translation in negotiations and business conversations, the back-and-forth is extremely important. So you're building specialized models in language, but there are also specialized models in areas like law that are starting to come out. Harvey recently announced they're going to build their own models. Do you see this as something where now that we've seen the capabilities of big language models, the next step will be almost every discipline seeing these specialized models built out?
I don't know if it will be every discipline. There needs to be a balance. Is the use case big enough? Is the business model big enough to invest? Training these models is still expensive and complicated. So it will be limited to the most important ones, but those will definitely move in this direction to cover edge cases and deliver applications that run every day in a business.
A
Alex10:36
So is there any trade-off if I opt to use a more specialized model as opposed to going with the bigger model, or is it all upside?
J
Jaroslaw Kutylowski10:51
I think it's actually all upside. Other than potentially having another vendor, the question is how to do model routing within a company – how to pick vendors and suppliers for different areas. The convenience of working with one big model vendor is there, but when you start thinking about doing a particular task really well and it's core to your business, specialized models make a lot of sense.
A
Alex11:36
You brought up something I'm hearing a lot about: model routing. Companies and people using these tools have seen that asking the biggest model for everything is expensive and not necessarily the best way. So now model routing is in vogue, where a routing layer decides whether a question is better for a flash model, a high-powered model, or a specialized model. Can you talk about why model routing has become such a popular term and where it's going?
J
Jaroslaw Kutylowski12:31
The first reason is the sticker shock for many businesses looking at expenses from using big high-powered models. Everybody wants to rationalize how they're using those models. In language translation, this idea has been around for a couple of years, with model routers already in place deciding which models excel at given language pairs or content types. Going forward, it's not just about cost but also performance, ensuring a certain task is handled in the best way.
A
Alex13:47
So where do you see the AI world going? Is it that there will be a router, and big foundational models handle some tasks, but the majority go to smaller or flash models or specialized models? Or is it still mostly big models handling stuff with occasional routing elsewhere?
J
Jaroslaw Kutylowski14:13
Right now, a lot of common business tasks are well tackled by large models, primarily coding and software engineering. That's likely to stay there. But we're seeing more specialized models going into other directions. If new models come up that are very specialized for coding, or when the trade-off between price and capability shifts, we might see more use cases routed away.
A
Alex15:10
When we talk about whether this AI moment can be sustainable, I think it will live or die by its ability to make real change and create business opportunities. For DeepL, you're helping a company that might function in English to set up shop in a country speaking a different language seamlessly. How do you do that?
J
Jaroslaw Kutylowski16:03
A lot goes into that. For internal use, large international organizations have language barriers. Studies show how much gets lost in context and productivity suffers. We help companies forget about those boundaries and hire the best people regardless of where they sit. For external appearance, to serve customers in new markets, companies need to translate documentation, sales, and customer service. AI tools like DeepL allow them to start doing that much easier, removing the need for upfront investment in local staff. It also helps optimize product delivery cycles where localization was the bottleneck.
A
Alex19:38
How much has the growth in capabilities of LLMs enabled you to do this job? Talk about how better LLMs have enabled DeepL to go from point A to where you are today.
J
Jaroslaw Kutylowski19:55
There is a big stack of use cases with varying complexity. At the bottom are spam emails where you don't need the best translation. At the top are regulated documentation like medical leaflets with legal liabilities. As model quality has risen over the years, we've unlocked more and more of those use cases. It's complicated for customers to know if the model is good enough, and it's our responsibility to help them find the right quality level and optimize the setup.
A
Alex21:39
Where were you a couple of years ago compared to today in terms of being able to attack those use cases?
J
Jaroslaw Kutylowski21:44
It's night and day. We launched DeepL in 2017, which was the point where neural networks started coming up for translation. That was a major step. Before that, solutions didn't work. After 2017, things got better and better. Now for well-written documents in major language pairs like English to French or Spanish, the translation is pretty much flawless. Sometimes it even exposes problems in the source document, catching mistakes made during writing.
A
Alex23:07
I think I need to use it for my writing to make sure I'm being clear.
J
Jaroslaw Kutylowski23:11
It really helps because if you look at something you've written, your brain is skewed to accept it. If you translate it to another language and see it's clunky, you notice there's a problem. The models can figure out a lot and help improve the text, but there's a level of ambiguity they might not resolve.
A
Alex23:50
You can look at something you've written 10 times and not catch a typo, but then you shift perspective and realize it made no sense. When you translate into German, there's no mercy.
J
Jaroslaw Kutylowski24:11
There's no mercy. It's a complicated one.
A
Alex24:15
For poor grammar. I'm in the middle of learning German, so congratulations.
J
Jaroslaw Kutylowski24:19
I've learned this the hard way. Thank you. Not yet congrats. So language is one thing. Let's say I'm a US company wanting to operate in Brazil. I can get my website and customer service in Brazilian Portuguese, but there are also laws, regulations, customs. Does language get you half the way or how far can this take you?
I think it gets you pretty far because you can start leveraging local partners. You can engage a Brazilian law firm and communicate with them. You can start much quicker. There are other aspects to set up in a new market, but you can get there. Even at DeepL, we talk to journalists and customers in local markets using our technology for translation.
A
Alex26:00
That brings me to something I've been thinking about. We've talked about customer service, communicating across borders, and latency. How much of AI do you think will move towards voice from text? I was speaking with Greg Brockman, and he was excited about bidirectional voice models where conversation is seamless. To me, the path forward includes a lot of voice, and ground zero is translation. Where do you think the voice side of AI is going, especially for your use case?
J
Jaroslaw Kutylowski27:10
When it comes to translation and languages, this is really the next frontier. We've gone far in text and documents. A lot is possible. But we're still in the early years for real-time speech, making conversations engaging and fast with quality. There's a lot that can be done. I'm looking forward to a world where those conversations happen. For voice as an input interface, I'm split. Conversational models are great for situations like driving or cooking. But there have been many attempts at voice input over the last decade that never really caught on. Maybe we humans don't want to talk to technology that much, preferring writing because we have more time to think. Adoption hasn't been as fast as technology enabled.
A
Alex29:20
So maybe it's a form-of-use issue rather than a technology issue.
J
Jaroslaw Kutylowski29:27
Yes. My wife says these models creep her out. I'm fine with it, but she doesn't like that.
A
Alex29:41
Where do you stand on the AI device? Do you think an AI wearable or whatever OpenAI is brewing will be successful?
J
Jaroslaw Kutylowski29:49
I think it makes a lot of sense. Getting devices as small and near to us as possible, especially for language translation, makes sense. I'm a big advocate that for real-time translation we have all the devices we need, like phones. But having more data and devices embedded with us all the time can gather more context, which is pretty cool.
A
Alex30:35
But can't the phone play that role?
J
Jaroslaw Kutylowski30:38
To an extent. For serious AI, even the best phones can only run extremely small models. We need a bigger device. But for gathering data and context, the phone is too restricted physically. Having glasses or sensors in our body could help determine what is happening in the physical world. Often when models get things wrong, it's because they don't understand the real-world context.
A
Alex31:33
Because they're limited to the prompt they just got.
J
Jaroslaw Kutylowski31:37
Totally.
A
Alex31:39
I want to end with this. We're in a moment where AI capabilities are increasing, and many people are becoming AI pill or have AI psychosis. But there is also growing uneasiness about technology, worries about data centers, jobs, and reality altering. I'm curious about your perspective on where this is heading and if there's a world where AI flips the narrative among skeptics.
J
Jaroslaw Kutylowski32:22
We are at the beginning of a huge transformation. This might be the biggest technological revolution because intelligence is core to everything we are. Having another form of intelligence is a completely new thing. A lot will change. It's not just about jobs but about capabilities. For example, SpaceX can launch rockets, but we can't yet build spaceships for large numbers of people. The engineering hurdles are too big for humanity alone. AI can be the solution. But I see how these changes can be unnerving. Sometimes I feel intimidated by the speed of progress. However, language translation is an amazing use case of AI. Civilizations are built on communication. The ability for every person on earth to talk to another is a beautiful application, both from a business and human perspective.
A
Alex34:29
Agreed. There are certainly people in my life I would love to have seamless translation with. Communications and ideas get lost without it. It's one of those positive visions we can all believe in. Jaroslaw, great speaking with you. I learned a lot. Thank you for coming on the show.
J
Jaroslaw Kutylowski34:50
Alex, thank you for having me. It's been a pleasure.
A
Alex34:53
Definitely. And if folks want to learn more about DeepL, where do they go?
J
Jaroslaw Kutylowski34:57
They go to deepl.com. That's super easy.
A
Alex35:01
There you go. Thank you, Jaroslaw. Thank you, everybody, for watching, and we'll see you next time here on the channel.