- 8 hours ago
Language has long been a barrier technology promised to break but never quite did. Now it's actually happening. From high-precision translation to real-time voice interpretation, language AI is progressing fast enough to enable use cases beyond office productivity: more inclusive access to information, genuine multilingual collaboration, democratic participation across languages. What does it take to build AI that works across spoken language, not just text? How do you preserve nuance, tone, and intent at scale? And when the barrier falls, who benefits the most and why?
DeepL CEO and cofounder Jarek Kutylowski on the state of language AI, where it's heading, and what it means for organizations, institutions, and individuals.
DeepL CEO and cofounder Jarek Kutylowski on the state of language AI, where it's heading, and what it means for organizations, institutions, and individuals.
Category
🤖
TechTranscript
00:00.
00:43Hello, everyone.
00:44Welcome to the next session here on Vivotech.
00:47And I'm delighted to welcome Jark, CEO and founder of DeepL.
00:50Welcome.
00:50Thank you for coming from Berlin.
00:52Thank you for having me, yeah.
00:53I'm Chris O'Brien, founder and editor of the French Tech Journal, based here in Paris.
00:58Let's just jump right in, because I'm going to assume that everyone probably knows what DeepL is.
01:03If not, you can catch them up a little bit.
01:05Oh, yeah.
01:06But, you know, just to start us off, you know, things are moving so fast these days.
01:12So if we talk about DeepL and its elevator pitch, what does that look like four years ago?
01:19And what does that look like today?
01:23I think there was no elevator pitch four years ago, if I may say so.
01:27Actually, like, the product was just standing for itself.
01:29And you didn't have to really explain that so much.
01:33And also, really, there was not that much competition in the AI market.
01:37And there was not enough noise in the AI market so that you would have to really do that elevator
01:42pitch so much.
01:43You just brought out a great product, and people started using that, and it kind of went viral very, very
01:50quickly for us.
01:51I think the situation for everyone in the market is different right now.
01:55There is just so much AI out there.
01:58There is so much being built, and it is great by nature.
02:02I think we are creating a lot of value in the market overall.
02:04But it obviously also means for a company like ours that the pitch needs to exist, at least, and it
02:10needs to be there.
02:10And for us, it's very much less so about just helping you as a knowledge worker kind of translating an
02:21email,
02:21but going far more into the most critical aspects of what communication means for companies and solving that for them.
02:29Right, because if I think about it as my first use case, like copy-pasting some text in or document
02:34or whatever,
02:36four years ago that seemed revolutionary, and now you think, okay, well, like, lots of places I could go do
02:41that.
02:41Why would I go to DeepL?
02:43You've obviously evolved way past that.
02:46And at the same time, to do that, I mean, it's amazing.
02:51That was incredible at the time.
02:53That felt revolutionary at the time, and you were using AL.
02:56So in terms of how you were building the AI that you were using for DeepL, I'm getting my alphabet
03:03mixed up here.
03:04In terms of how you were using AI to build DeepL, how has that changed then as you've had to
03:10adapt to this landscape,
03:13to the just rapid evolution of everything that's happening right now?
03:18Let's start with the naming.
03:19We wouldn't call it AI back then even.
03:21I mean, it was neural networks that we were talking about, and while it's the same and, like, all of
03:26the AI that we have around nowadays,
03:27it's like neural networks, we were just calling it by another name.
03:30I think the scale of that was smaller.
03:33Everybody was, like, whoever was building models, whether it's us or someone else, those models were smaller,
03:41required maybe less data, required less compute.
03:44So the scale of everything has just increased by a couple of orders of magnitudes, I would say.
03:52Like, in order to create an amazing translation model nowadays, you also have to go the kind of LLM route.
04:01Like, you have to be participating in those models that have hundreds of billions of parameters,
04:05and therefore everything becomes more complicated, and the ROI question on how much research you invest
04:14and what the outcome of that is, that becomes a difference.
04:19It was a no-brainer in 2017.
04:21You'd put in, like, a couple of hundred of thousands of euros maybe into the compute on creating a new
04:27model,
04:27and that's more like on the order of millions and tens of millions and maybe hundreds of millions nowadays.
04:32And you have to find the business case that actually also supports that research.
04:37So I think everybody who is in AI right now needs to have a much stronger sense of,
04:44am I building those models?
04:46Am I putting in that research?
04:48Does my business support that?
04:50Whereas it was far, far, far easier as the sums have been smaller a couple of years ago maybe even.
04:55And so can you talk a little bit about how maybe DeepL's approach to that has evolved,
05:00were you initially right away thinking, okay, we have to build our own models?
05:05Or was it initially, okay, well, we'll use one of these other ones as an API and we'll build off
05:12of that?
05:13Or how has that changed in terms of how you think about your relationship to building your own
05:18versus building on top of someone else's?
05:21No, I think for us it was always pretty clear that we need to own the models for a couple
05:27of reasons.
05:27First is we want to be the best player in the market.
05:30That's just like a very, very, very important part of our DNA.
05:36We want to build the best AI for translation, whether that's across text that is copy-pasted
05:42or whether that's, of course, like sophisticated, multi-paged or hundreds of pages long documents
05:48that we're translating for the legal industry maybe,
05:53or whether that's real-time spoken language where you have to kind of deal with other types of nuances.
06:00So being able to create those models, being able to be the best in the market was always important for
06:05us.
06:05But it also gives us the ability to really adapt the product to what the customers need.
06:12If you own the whole stack, this is our thesis, you will have the possibility to just really take
06:19the customer input that comes in yesterday and build it into the models tomorrow.
06:25Whereas if you're building on top of someone else's technology, you just can't have this iteration loop.
06:30And fast iteration loops have been what has been driving kind of the development of technology over the last decade
06:38maybe.
06:39And, you know, one of the things that strikes me again as a non-technical person,
06:43but someone who's covered tech for three decades, is the speed not just at which things are accelerating and changing,
06:50but the degree to which the conventional wisdom and the thinking just changes almost week to week
06:56about what is the right model, what is the moat, what are the economics of this, what makes sense, what
07:03is the playbook now?
07:04You know, every two weeks it seems like it's turned upside down again.
07:08So as you mentioned you have your thesis and you just articulated that, but at the same time, you know,
07:15there's someone who's going to be thinking this week, oh yeah, he's nailed it.
07:18And then two weeks they're thinking, oh no, they've got it completely wrong.
07:22So how do you kind of navigate that because you can't just turn things upside down every two weeks
07:27and tell the team, okay, now we're going this direction, now we're going that direction.
07:30So how do you kind of pick a direction and kind of navigate while at the same time, you know,
07:36reading the landscape and thinking, okay, what do we need to adapt or what matters?
07:42Yeah.
07:43That's a, that's a hundred million or maybe a hundred billion question actually that you're asking there.
07:48I think there is a lot of debate about that, like how, especially as a company that hasn't been founded
07:56in like 2025 maybe, how do you, how do you survive that pace of change?
08:01There's, there's businesses which have been built maybe like started in the last years
08:06who have experienced that pace, that, that, that kind of different mode of operations,
08:17like from their start on, there was, there was something that, that they've lived always.
08:21And I think for everyone else who was in the market, which is 99.9% of the companies out
08:27there,
08:28this is a much, much trickier thing to navigate.
08:33And it's, and it's sometimes not even really so much about the strategic,
08:38strategic changes that you need to make or the tactical changes that you need to make
08:42because, because the environment or your ecosystem changes, but about the, the ability of the humans
08:49that are in an organization and the founders and the CEOs and the leadership teams
08:54and like everyone in an organization to, to absorb that change.
08:58It is, it is, it is just an emotional burden.
09:01Like people are not, people are not built for the space of change.
09:06And yet we have to, yet we have to survive in this space of change.
09:09And I think that then kind of divides between those that are able to adapt that,
09:16that have figured out a system on how they can, how they can deal with that,
09:22whether that's ignorance to a certain extent or whether that's just like re-adapting so quickly.
09:32Everybody needs to figure out their, their, their, their, their, their mode.
09:35I don't think we can, as humanity survive in that mode for like longer than a couple of years.
09:41I think we're just going to burn out on, on that.
09:44But the reality on the ground is that right now, if you want to be competitive
09:47and most kind of AI companies want to be that, and most companies in general want to do that,
09:54you have to adopt that, you have to adopt that mindset.
09:57And I wouldn't say that I have like a, like a, like a, like a great solution to, to that
10:03problem.
10:03It's just like, you gotta, you gotta work hard on yourself, really.
10:06Yeah, I agree. It's, it's, it's, it's a mental model thing and cultural as much as it is strategic
10:12or technical moment to, to navigate through.
10:15But then speaking of which, I mean, you guys just announced some news yesterday,
10:20talking, speaking of adapting and, and changing the navigating.
10:24So tell me a little bit about what you announced and how that fits into how you are adapting to
10:30this moment.
10:31Yeah, yeah. For the first time in our history and like there's different models of companies,
10:35some companies like are very heavy on acquiring other companies, uh, others do not do that.
10:41Uh, we haven't been doing any of that, uh, in, in, in, since the inception of, of DeepL,
10:46uh, but we just announced that we've brought on a team, uh, from a company called MixHalo into DeepL,
10:51um, over the, uh, like in, in, in, in the last days, basically, uh, based in San Francisco.
10:57So like extending our footprint in the U S kind of strengthening the, uh, the idea that we are in
11:04Europe built company,
11:06but also really very much a, uh, a global company.
11:09And, and the, the fact that the U S is, is an incredibly important, uh, market for us.
11:14And then change means in those situations, quite often also bringing in change agents, bringing in new spirits into,
11:21into the business. And in this case, a company which has been working very much in the real time speech
11:27translation area.
11:29And, and therefore brings like essential know-how to us, uh, but also obviously brings in, uh, like a very,
11:36very, very,
11:37very close look on the market in, in the U S and how they are thinking about voice AI and,
11:43and AI and language in general.
11:45Yeah. And then in, in that sense, I mean, it's partly geographic, partly cultural again,
11:50but then that if, if I'm understanding correctly that the voice AI piece, does that put you in a new
11:55market?
11:56I mean, I know, again, I could speak into something to some extent before,
12:01but this sounds like it's a whole new opportunity for you in that sense.
12:05Yeah. I think, I think that is, uh, that there is, there's kind of two parts to the DeepL business
12:10right now.
12:11Um, from a, if you look at it from a, like maybe human perspective, uh, there is the part where
12:18you,
12:18or where we would concentrate on making sure that even the, the most complicated PDF with like graphics embedded
12:26and incredibly complicated, like terminology that is specific to a business gets perfectly translated,
12:35100% accurate so that you can have full trust in the technology to, to get this done.
12:40And you can scale it out across your organization for whatever your business, for whatever your business does.
12:46Um, that is an amazing use case and something that businesses love.
12:53I think if you look into voice AI and if you look into real time speech translation
12:57and the ability of enabling two human beings to just sit next to each other and talk to each other
13:06and express emotions and express their thoughts,
13:08their, their, their, their, their deepest insight in their native language and therefore being able to transport everything that they
13:16really want to transport in a conversation.
13:18Um, that is being enabled now by, by kind of the, the voice AI aspect and therefore something like I,
13:24I'm personally like really,
13:25truly excited about to, to, to, for the technology to be able to, to bridge a barrier there that it,
13:34that feels like very, very deeply human,
13:37human, but yet needs the AI and the computer to like really empower that.
13:44And so, uh, I, I mean, if I can ask then just a little bit, I know you can't talk
13:49too many in terms of the deals,
13:50but that this idea that's so, uh, so much in the air right now about how much Europe can compete
13:57with, uh, on many different fronts.
14:00And that's a whole other panel in many ways to unpack that.
14:02But, you know, how Europe can compete in AI versus the US, um, as a European company, but one with
14:11a big brand and an established track record and name in this space.
14:15How, was it a challenge to go to Silicon Valley, to go to a San Francisco company and convince the
14:24founders, whatever the dynamic was that a European company was the right fit in that sense?
14:31Or was that not an issue?
14:32Was it more just the classic, you know, uh, due diligence, the, the, the partnership, et cetera, et cetera.
14:40So, uh, not that, not that big of a deal.
14:43I think, uh, at the end, I think, uh, so the, the, the team from X Halo, they've been partnering
14:50with us for, for quite a while already.
14:51And I think this is kind of where quite a lot of kind of exits happen either way.
14:56It's, or like acquisitions and, and, and they, they, they are being built out of partnerships because you know whom
15:04you're talking to and, and you actually know the other party and, and you know what works and what maybe
15:08doesn't work.
15:09So you can make a really, really sound and safe decision on that.
15:12So I think from that perspective, that, that wasn't hard.
15:15I think as a company being like a very global company, we speak Bay area language, I would say.
15:22Uh, and I think if you can do that and we developed that in the point in time when we
15:27started to, to work with, with us investors, like most of our investors are actually American.
15:33And, and, and, and we've been around, like I've been in, in, in, in Silicon Valley all of the time.
15:38So at some point in time, you just kind of develop that fluency of speaking that language and, and being
15:44accepted into the ecosystem and knowing what matters.
15:47And then it's not that big of a, of a problem.
15:51I would still say that for a European company hiring in the U S is always a tad bit more
15:57challenging because everybody's going to ask themselves, like, do I want to work across time zones?
16:03Do I want to put up with the burden of that?
16:06That there's going to be a kind of a cultural clash there.
16:09Uh, but many people also really love that.
16:11Many, many, many people in the U S, um, actually resonate very much with European values and want to also
16:19kind of do something else and, and maybe have this opportunity to, to get to know each, uh, to get
16:25to know that the world a little bit better.
16:27So, so it works, but you sometimes have to like, as a founder, as somebody who is hiring, you have
16:32to go like just a little bit further.
16:35So maybe just to wrap up, we've just got a couple of minutes here, but to come back to our,
16:38our headline question, what happens when language barriers disappear and maybe get a little future facing here for a second,
16:46uh, beyond just sort of the practical use cases in front of us, you know, we imagine a world where,
16:51you know, we have kind of the, whatever, the universal translator from Star Trek, where we don't even think or,
16:58you know, about language barriers.
17:00That's just a thing of the past. And, you know, it seems like the implications are massive across education, government,
17:08hopefully in a, in a positive sense, that's all good.
17:12Uh, but do you think about those things when you're kind of, you know, logging off for the end of
17:17the day and thinking about what that world looks like and, and what, what's going to change kind of on
17:23a larger scale for humanity?
17:25Uh, honestly, all of the time. And I mean, like I live in a, I live in a bilingual household.
17:31Like I'm Polish, I'm speaking Polish to my wife. My kids kind of more speak German because they've grown up
17:37in Germany, but they also go to English speaking school. So they're going to kind of switch into English here
17:42and there sometimes. So it is, it is always present.
17:46And I think this kind of multilinguality is just something that, that defines Europe and, and, and maybe like really
17:53also defines the whole world. And I think that's going to be there.
17:56I really love the aspect of how certain things are really expressed in different ways, in different languages. And that
18:04comes from a historical background of the cultures and how, how, how different the, the, the, the past of certain
18:13countries has even shaped the way that they express themselves. So, so there is a certain beauty in, in that.
18:19Um, but also at the same time, it is a large barrier.
18:23Like if you want to, like if you want to go to a place, like if you want to go
18:25to a place, like if I go to France, like I run into problems. I mean, this is, this is
18:30just what it is. Uh, luckily I have Laurel the most time with me who's, who's helping out, um, when,
18:37uh, when, uh, when something goes wrong.
18:38But I think, I think making sure that as a business or as a person, you just completely can forget
18:47the question of language and you go anywhere and you can communicate to, to anybody else.
18:52Because this is just plainly going to make the world a better place. And I, I have no doubt in
18:57that.
18:58Yeah. Again, as someone who's lived here for 12 years and my French has improved steadily, but you never get
19:03to a point where you're a native speaker because the context, the cultural history, specific situations when you have to
19:10get your car fixed.
19:11I mean, it's a constant, there's levels within levels within levels. So yeah, specializing in just that domain is an
19:18insane challenge.
19:19And I think the technology is actually already there. I think we're just not, sometimes not bold enough to use
19:27the technology in those cases.
19:29So like, I would encourage you and everybody else, like the next time when you go get your car fixed
19:33in France, like take deeper voice with you and try it out.
19:37Oh yeah. Okay. I absolutely will. All right. But I think we're out of time. So, uh, thank you all
19:42very much for being a great audience and please give Jareke a round of applause.
19:46Thank you very much.
Comments