#263 Jarek Kutylowski: How DeepL Is Using AI to Break Language Barriers

0:00The difference that D-Bell is making comparing to some of those more consumer oriented solutions on the market is the accuracy part, the integration into workflows, the ability for a company to tweak the language of the translation. towards what they need. I think we're moving into this direction where language is not going to be an issue at all. And I wouldn't want to forget about the kind of human aspect of how much language is embedded into who we are and how we can exchange with others. Having access to technology, which makes it easier for people who are not so well versed in a foreign language, that helps a lot them to be competitive in the work market, but also them being able to express themselves freely and preserve the language.

0:50Hi, I'm Jarek Kutilovsky. I'm a founder and CEO of DeepL. I come from a computer science background. I am super excited about what technology can actually do to make our lives better and how that can have an impact. And out of that passion, I founded DeepL And I'm kind of right now looking into how AI can help businesses become more international, can help in all of those language kind of issues and challenges that we have every day. And yeah, very much looking forward to the conversation. Yeah. I was reading you, DeepL, the tech started at a company called Lingay. Is that right? Lingay. it's uh yeah like deep l is a spin-off out of that out of the company basically and and we've been looking into uh like the translation space or the language space already earlier in that business and then at some point in time 2017 we've noticed hey there is this new amazing technology neural networks the foundation of ai as we know it nowadays and uh you can actually do more about that and there is an opportunity for us to go into into that space yeah uh and so describe uh dpels uh product and tech i'm curious i mean we we spoke briefly earlier about uh that space because ever since transformers hit the scene uh there have been all kinds of companies popping up with translation solutions.

2:39So tell me about DeepL, what kinds of applications you're applied to and what differentiates DeepL from the crowd. Yeah, yeah. I mean, it's been quite a journey, of course. Like the company already exists for eight years and the product is like something totally different nowadays than it was back then in 2017. 2017. We've been around quite a little bit earlier, even the Transformers reached the market. We've been kind of pioneering the way or how you designed neural networks for translation specifically and maybe language in a more general sense. And for that purpose, we've been like doing this really cutting edge AI frontier research in terms of like, what are the model architectures what really works better uh what is a better way of training those models what is the data that you should feed those models uh to always achieve the best quality and accuracy uh in translation but also paired with like you all we really want to make sure that translation has this nuance that it feels natural it feels feels native and and sometimes that is clashing with the accuracy aspect also uh if you're translating marketing materials in a different way than than technical documents or legal documents maybe um and we've been making sure that our ai models actually can capture all of all of that with really the aim on looking at what businesses need um while we are a plg company uh and therefore have grown through this very strong user base uh that we've that we've been able to build uh nevertheless our goal has been always to make sure that we can provide language solutions to businesses in all of those areas where regularly accuracy matters where uh embedding into the workflows of your of your employees matter.

4:43Think both internal use cases where it's about the communication between different offices in different countries, or even external when you're talking to customers that are outside in a different country where you maybe don't have field operations in. Yeah. And this is both text-to-text and speech-to-text. Does it also do speech-to-speech? yeah our our first uh approach to that was really text to text it has evolved then uh towards uh towards documents i think that's something that's also incredibly important for customers like a lot of the the kind of written content is actually out there in the form of documents and and and it's it's it just makes a lot of sense to approach document translation in a slightly different way and since last year we also have our speech to text uh translation out in the market uh we're looking into the speech to speech uh space um we still feel it's a little bit early for that from a technological perspective that actually the um speech to text kind of live captioning during during a meeting um gives you the better user experience and actually is allowing you to better follow on what's happening in a meeting uh but definitely something that we're looking into and kind of waiting on where where is that when is the technology gonna be actually at a point where um it brings a lot of value to the kind of typical business use case yeah uh yeah and the the speech to speech particularly live speech to speech the the issue is depending on the language pair you have to wait until the end of a sentence before you can translate it and that creates a latency issue uh is that right like chinese to german or something where the in german the verbs at the end and chinese it's at the beginning and yeah yeah that that that is a big problem and i think in those languages it is the the biggest one also in other language pairs like some of those problems come up uh come up too so there is like a little bit of an inherent latency embedded in the problem in some ways yeah unless unless we are able to predict what you're gonna uh what you're gonna say and uh that is also a fascinating field of research because quite often the models can do quite a good prediction on what the end of the sentence is gonna be nevertheless there's gonna be mistakes there's going to be kind of curveballs coming up and uh and then they will go uh then we'll have to deal with that but there's there's a few approaches of of which we are thinking about how you can actually uh do that better with lower latency uh because latency is is kind of the the biggest problem in this whole uh this whole space yeah but as long as you're doing uh the output in text and it's not you're not expecting live translation it gives the model time uh to to to look at the entire sentence and and that are are you using uh uh pure transformer models i've been talking to people about mamba models uh that have a a longer uh memory in effect What's the architecture that you guys are using?

8:17Yeah, we're actually not disclosing the architectures of our models. That's like kind of a little bit of our secret sauce. We've transitioned to LLMs, but there's our purpose-built LLMs that are built for translation over the course of last year, kind of going with the flow of the advances in the technology. and that has given us once again a bigger boost there. There's also models involved for speech recognition, of course, when we're thinking about speech-to-text there. And we're trying to make sure that our translation models are also really tailored to how you need to translate spoken language, just because it's a little bit different than what you'd usually write, and the models need to be able to cope with that too.

9:05Yeah. So the product, how do you sell the product? I mean, is it an API? Is it a usage-based model? Or is this a piece of software that people embed in their own systems? it's it's actually like the part where i said like it has evolved over the last years um it has evolved into a complete platform there and and the way that we're thinking about that is really enabling a whole enterprise to solve all of their uh translation and language uh challenges that they that they might have and that starts with like a seat-based subscription for all of those of your employees that really need access to this instant translation like all of those who are exposed in some ways to to an international uh context uh who don't want to go to an agency who don't want to uh work maybe with an internal localization team that's kind of there's always like a turnaround time uh for that if you want to have the translation immediately uh then uh then can be equipped with that and it starts with that um giving all of the employees access to to the platform but then when you think about this further uh there is some workflows that an enterprise is going to be running that can benefit from a deeper embedding of the technology into into those flows a typical one would be like a customer support um use case where there is inquiries for cast from customers coming in from different parts of the world and you always want to route those to the agent responding to those in the in the relevant language and therefore kind of being able to to serve your customers uh better so uh so from that perspective we think about that in a kind of holistical way uh there's different ways to to approach the technology itself from an end user based application over to our DeepL Write product, which helps you express yourself in your native language with DeepL Voice, speech to text.

11:24And then obviously, all of those services can be also accessed via the API to embed into your workflows. Yeah.

11:36Give me what's the most common use case that you're serving? And frankly, this machine translation has revolutionized anyone who's working across languages. I spend a lot of my time in China. I can read basic Chinese, but it's slow and painful. but now I read Chinese websites with no problem because built into my browser is a translation function. I think it's from Google. I don't know. And it does a pretty good job. I mean, That's what I was saying.

12:29Why would companies use DeepL as opposed to just using the Google plugin, for example? I can only confirm that this development of technology has been amazing. Like 10 years ago, we wouldn't have thought of all of that being really possible. And like one of the kind of latest points for me to really realize was with me sitting in a customer conversation in Japan, where until now we actually either had an interpreter in the room or I had members of my sales team kind of trying to help me and interpret while we've been going. but it's always cumbersome that really doesn't work that nicely and it's just a bad experience and right now with technology I had access to like the full breadth of the conversation I could express myself freely knowing that the technology is going to be able to pick up on the nuances in in my English and uh and translate them correctly into into Japanese uh that's that's that's been pretty amazing and and i think um kind of the difference that the bell is making comparing to maybe some of those more consumer oriented solutions on on the market is is the accuracy part is the integrations into into workflows is the ability even for a company to tweak their their language of the translation towards what they need uh whether that's compliance rules or whether that's um certain aspects of how the brand voice looks like uh but then going like really also deeply into terminology and if you think once again about legal text about uh technical texts that matters quite a lot because it reduces the amount of work that you have correcting the text and that plays a big role for our customers in terms of ROI because kind of any additional editing that must be done on top of the AI's output, that's incredibly expensive comparing to the technology costs.

14:44What's a typical use case? If a company has an office in Japan, uh does it can will for example will english language emails automatically be translated into japanese or if you're uh sharing a document is the document uh automatically translated it i mean give me a couple of uh use cases yeah yeah i mean uh i think i think the example is is is great and and i think like one of my uh one of my favorite customers is is an automotive automotive company which has all of the r d really in japan uh whereas large parts of the customer base and the markets are in the us in europe and uh there is really this language barrier between r d and between the fields and obviously like they need to talk to each other um and in those cases it really depends also on the customer there's those who have translation really embedded into their own systems in the company so that as you say like every email that gets sent out and it's a different language it gets automatically translated documents that are posted into uh into a system uh be it like a contract management system for example are being automatically translated there's tons of plugins that work with the bell in those systems and integrations that do that automatically but sometimes it is really the kind of the user who either uses our browser extensions our desktop apps or just goes purely to the website and takes the the content of the current document that they've received translates it and uses it for its own consumption or even publishing it within the company.

16:41Yeah. And in that case where it's documents, as you said, technical documents and that sort of thing, do you build an ontology for each customer between the language pairs? Or how does that work so that you can ensure accuracy? Yeah. There's terminology basis that customers can embed into DBL or that you can even ask DBL to kind of interfere from your existing translations as a company you usually come in with a history of maybe a massive set of of documents haven't been already translated in the past and and you can ask the AI that that we provide to to give you like to to extract the terminology out of those documents like what is what is special basically about your language comparing to maybe the average uh that you would want to check and then embed into further translations that are being uh produced by DeepL but also you can come in with your with your pre uh pre-made pre-categorized glossaries uh terminology basis just plug them in into the AI it's going to know those words it's going to be able to um uh properly use them in in all of those sentences it's going to be able to even uh sometimes figure out when uh there's false friends in those translations when in certain areas you actually don't want to use the terminology that is embedded into the glossary because it doesn't make any sense really uh and then deviate from uh what it was supposed to do in that situation so so you have a lot of control but at the same time the ai remains incredibly smart about the language being used and makes sure that that whatever you're translating maintains both this deep out typical accuracy but also this this nuance and the look and feel of somebody native has been has been writing that yeah you were saying it's a platform uh meaning uh i mean but but you also mentioned uh browser uh extension uh so how does how does a company work with with deepel do do they work through a DeepL website?

19:20Or yeah, just describe that. I think translation is such a broad use case and therefore there's also different ways of applying that and different ways of applying it to different workflows. And we as a provider of this kind of complete solution package of this platform are trying to accommodate to also what our customers really want uh there so which is why I mentioned like there's the browser extension for those who who really wants to work in their browser but you don't have to do that you can go to the website so there's different means and that really depends on the company how their technological stack looks like how um how much uh how much they want to deeply embed translation into some of those workflows or whether it's just a casual kind of long tail uh workflow in in that particular case do you see wide adoption of uh this technology among multinationals that have offices i mean you know as i said i worked for a long time in china i've worked in europe and And generally, multinational companies have senior staff that speak English, and even mid-tier staff speak English.

20:41So English had become the lingua franca of business. with this technology do you see first of all do you see multinationals adopting it and is it expanding their footprint do you think or do you think it's leading to less emphasis on the use of English by non-native English speakers yeah I mean like multinationals are a big part of our customer base and uh and I think sometimes like we expect everyone to accommodate that the company's lift language is English uh but we sometimes uh underestimate how big a problem or a bit of a kind of maybe even efficiency challenge challenge that is for for some of the employees that are uh that are out there uh sitting in a non-english speaking uh country and for them quite often it is just much more efficient even if they can speak the language uh to write it in their native language then translate it or it just kind of gives them this confidence in in being able to express themselves in a way that is much more professional um that is uh that is on a comparable level as their peers who are maybe coming from the us or from uk or or another uh english-speaking country so there is there's a big emotional component to that uh quite often uh but there's also just this pure productivity efficiency aspect that is uh that is still uh really really important and and we're seeing that across multinationals and in general international companies that whenever there's a bit of spotlight on getting that diverse employee base engaged in conversations making sure that they participate then the technological solution to that becomes quite often the choice very, very quickly.

23:04Yeah. You've been in this end of the AI world for a while. What fascinates me about it is the gradual breakdown of language barriers between cultures and the increased understanding between cultures that is likely to come from that. how do you view that i mean do you do you think that that it accelerates or expands business communication and you see a day when language will no longer be an issue at all i i think i think we're moving into this direction where language is not gonna be an an issue at all and and and finally it i think solutions like the bells really help even maintaining those local languages uh because uh yes you have to speak english in your in your business world but maybe you don't have to put so much emphasis on that and and you can live in your local language much much uh better um also i think over the last year's customers expectations when it comes to the language in which they're being handled and and spoken communicated to uh have grown uh so all those multinational companies that we've been just talking about they cannot afford anymore to just offer customer service and customer support in english they will have to accommodate to the local markets this is just how the demand is growing people know that those solutions are out there and it's possible and the competitors are just frankly doing that uh so uh so i think technology is actually contributing very much to uh to the local strengths in uh in in different languages and and i really like that aspect i think that that brings in this this diversity there uh but at the same time it makes our life so much easier uh when it comes to to to international uh business and and global trade and we've seen a very strong correlation in our customer base depending on how strong trade relationships are between countries and also then the demand on language translation for a country as a whole.

25:28How many language pairs do you cover? We cover over 30 languages right now. So the pairs is going to be like the square product of those. uh so it's a it's it's it's a large uh space uh we we are mainly focusing really on quality and specifically for those languages that are the most important in the in the in the business uh area and were um also the amount of data available and the focus that we can put on those languages um allows us to achieve this kind of typical deval quality that our customers and our users uh know and expect from from the service yeah uh the um uh what's the the smallest language in terms of uh the population speaking population that you have or one of the more uh uh exotic i don't mean exotic because i hate that word actually That's actually, yeah, that's a good question.

26:36I don't have an answer for you on that. Like, it's gonna be like one of the smaller European languages for sure. But I haven't like really checked on which one of those is really the smallest one in terms of number of speakers or the population base. Yeah. And are you growing the number of languages you cover or is your focus on, uh you know on verticals on maybe medical uh or or uh you know scientific uh uh areas where the language becomes more specialized yeah actually both i mean that the language portfolio is is expanding all of the time and we're bringing in new languages uh like every year uh this year is to be a little bit more about asian languages uh there's still a huge market uh there uh there's there's more and more collaborations and corporations that we're seeing within our customer base that span between asian countries uh directly and not only between kind of europe and the us and and asia uh so there's more and more demand for uh japanese to vietnamese translation for example um the japanese korean uh translation so that's that's something that's that's kind of coming up as actually the uh economic uh ties between um asian countries are also strengthening um at least from our perspective over uh over the last uh over the last years um at the same time we're looking uh also as a company more and more into into into different verticals in in the particular language space that's uh that's there um we can we can take maybe life sciences as a good example of that with like very specific specific terminology uh there um even fda related requirements on how translation uh needs to needs to look like and and we're working with our customers to broaden our product also in that direction and sometimes that's really like hard ai work uh to make sure that the models are capable of this this particular um language type uh and sometimes it's really also industry specific integrations and and understanding how does translation plug in into into some of those workflows uh specific for those areas yeah actually that's uh fascinating i hadn't i mean i I sort of assume that your user base is primarily native English speaking companies trying to reach or extend their or make their communications easier with non-English, native English speaking companies.

29:31but there's a lot of the world is doesn't touch English at all. It's between non-English language pairs. What's the biggest pair that you see there with the highest demand for? I mean, like English German is still our strongest. And that's also kind of because maybe that's one like our home turf that we're operating in. but that's not only because of that if you look at the distribution of translations being found on the internet in different language pairs then English German is actually one of the biggest ones if not the biggest one yeah it's actually just because of the economic ties between Germany and the rest of the world being pretty strong it's a country that's been exporting for so many years and had this very deep uh trade relationships uh with the rest of the world and and most of what was happening externally kind of really goes into or is being translated into into English um Chinese Japanese are also uh then always into English are also incredibly strong pairs um and then pretty quickly we're gonna see like some non-English pairs uh Japanese Korean is one of the important ones, German-French, just because that's like kind of directly on the border.

31:01And there's no need for using English as an intermediary language in the conversations between like a German company and their French customer. So you're going to see a lot of direct German-French material going through the Eval. Yeah. And that's interesting because until now, Now, English has been the intermediary language. I mean, a German company, speaking to a French company, generally they'll speak English, in my experience.

31:37What verticals do you see the most demand in? I mean, you mentioned life sciences. It's mostly really those that have the highest dependency on translation quality, but also compliance aspects where we're kind of every word matters um where debel is pretty strongly represented within uh legal firms uh within professional services companies everywhere where there is like a big numbers of documents being exchanged where there is a lot of white collar work knowledge workers uh happening and at the same time where where the stakes are high and and I think legal life sciences um professional services but also manufacturing uh because really manufacturing companies very often span different continents they spend different uh different languages at that makes them um relevant for us and the kind of language barrier being important for them.

32:45Yeah.

32:49Do you service the European Commission or any of the EU bodies? I think not directly to a large extent. I think certain EU bodies might be using DBEL either through us or through partners there. But we're working with other governments and have actually government contracts, especially for countries which are multilingual. That's a big market for us, just given their internal needs for even publishing documentation and materials and the legal ramifications on what they're obliged to do. yeah yeah that's why i mentioned uh the eu i mean don't they have to publish every document in i don't five languages or something i can't remember but um yeah and and i would guess uh on latency i mean you're not doing uh real time right now but uh but you did uh say instant so So what kind of latency are we talking about?

34:04It's like when it comes both to text and speech, I think we're thinking quite often 200 milliseconds, something in that range. That's not a lot. And I think that really helps you being able to, when it's about speech to text, to follow very well that conversation. It's not zero though. No, that is still latency and we're working very hard to get that down. And that is both work on the level of making the models faster, smaller, and being able to just be more efficient, but also deploying the models as near to the user in geographically distributed data centers. And that still matters. Yeah, although 200 milliseconds is pretty fast still.

35:03So are people using D-PEL not only for text-to-text, in which case latency isn't as critical, but for subtitling, for example, in a Zoom call? Yeah. yeah with with with deep voice that's exactly the the case it's it's real time a meeting uh subtitling uh and and therefore like this is this is really critical when it comes when it comes to speed this is this is essentially real time um in that case but also with text to text latency matters and for example in in customer support chat uh use cases uh we are we're fighting for every 100 milliseconds there too uh because that all adds up for a great customer experience and uh whatever whatever is the inherent latency of the response of somebody whom you're talking to you don't really want to add up a lot on the translation itself so minimizing that is an important part and and actually something that our customers really care about and and are asking uh for yeah um and the on that 200 milliseconds.

36:26I mean, as we said at the beginning, depending on the language pairs, there's kind of a fixed latency because depending on how fast the speaker is speaking, because you need to get to the end of a sentence before you can accurately translate the full sentence. So how do you bring down that latency? Is it through prediction?

37:02It's a mix. There's a bit of just simply sometimes waiting for the end of the sentence and then translating as quickly as possible. but there's also an aspect of prediction in that and sometimes the kind of advantage that you have with have with subtitles is that if you're wrong with your translation you can still correct yourself like you can basically flip the subtitle a bit and we're trying to minimize that as much as possible because that is a bit confusing if that happens too often but in rare cases when this or really occurs when the prediction is not correct, you can pull that. And that helps to bring down the overall latency of the whole conversation.

37:53Yeah. Does DeepL provide its tech as a component for other software systems that are then being sold onward? I mean, I'm thinking of, you know, I talked to a guy recently who's developing a live translation plug-in for Zoom. And, I mean, is it possible that what he's doing is taking DeepL and adding some bells and whistles and rebranding it? I mean, do you do that, provide that as part of a tech stack? Yeah, totally. I mean, that's kind of basically the API play that we have. And I talked about that being like relevant for enterprises to put in translation into their internal workflows. But it applies in the same way to tech companies who just want to have translation as part of their product.

38:58product and under the hood a lot of the translation that you're going to be seeing in in some tools is going to be actually provided by by the bell in in the bank crowd so uh we've been at the forefront of this translation space uh for for quite a while and and have been able to build out those relationships with ISVs with with companies that that build great products uh but also need a translation component uh in those and this ranges really from software that is intended for very dedicated small audiences like for lawyers uh for example um across to general productivity suites which uh cater to like basically all of all-knowledge workers out there and also need that translation aspect as part of what they're doing.

39:59And is the, I would guess, but I want to hear it from you, is your market expanding quickly or are there so many players in the market that it's slowing down growth? it's it's i mean the the the big players and the competition is something that we've uh that we've been born into uh it is it is basically something that we've that we've lived from from the beginning on um and that competition is out there that's going to be always out there uh it's it's part of our dna to to be able to cope uh with that um i think what is exciting about this market is that the possibilities and the abilities of of the ai are ever increasing and every year we are basically seeing new advances and those advances like every time we build something new they really enable new use cases where machine translation can come in there's there's use cases where five years ago you wouldn't ever think about using machine translation for that because you think it's just too critical and right now our customers are finding out hey that's that's that's like totally enough and we can speed up our development cycles we can reach new customer bases we can be just so much more efficient with with the technology and and therefore i think the market is really expanding and also expanding for us pretty pretty rapidly yeah do you um Do you work with any news organizations?

41:41And I asked because, you know, I was at the New York Times for much of my career. And I actually ran or established the Chinese language platforms. and you know we used human translation we had a team of translators sat every day they translated 30 odd articles in Beijing and I remember being really eager for machine translation to start working and I left before that happened. But it seems that this would be perfect for news organizations to broaden their footprint or their audience. So are you working with news organizations? And if so, can you talk about any that are using the tech? Yeah, it is one of our customer bases.

42:43and many of them i cannot talk about um the one that i can talk about is actually nikkei and in japan uh who have been translating all of their articles or like making them available to to a more international uh reader base uh with with technology and and i always found that fascinating in a way that uh this this kind of ability to read up on what's happening in another country in their own with their own local press that always gives so much better perspective on what's actually happening there rather than going to your local newspaper and and and and getting like a third party view on what's happening there.

43:34And that's both kind of enabled by some of those publishers who are doing that with DeBell. And on the other hand, there's obviously also DeBell users who are using our browser extension and kind of reading up on their own without the kind of publisher's direct support. Yeah. The news organizations that are using it, do they just put a link on their website if you want to read this in English, French, German, or click here? Or are they publishing entire sites in English, if they're, say, Japanese, to give an English language audience access? us yeah the answer is a clear both uh so we've we've seen we've seen both uh there are there are publishers who just give like offer that as an on-demand functionality uh for the users uh but there's also publishers who just want to like run sub sites of whatever they're offering is and do that in a different language and uh sometimes it's it's like fully automatic translation sometimes depending on the visibility of the article and the criticality of that there's a human aspect in in that and there's a review going on.

44:54We've seen basically all of those and like our technology enables all of those use cases. It's up for the publisher, for the customer really to decide on how they want to interact with an international audience. And that really very much also depends on the publication and the country you're you're in. Yeah. And is it priced so that I mean, you said a subscription is, is, are there subscription tiers based on, and you also mentioned seats, you know, depending on how many people in an organization. So can you describe the, the sort of pricing matrix a bit? Yeah. Yeah. Basically two models there.

45:41I mean, when it comes to equipping your employees with just kind of easy access to the tool to do their own translations, then it's seat based, as I mentioned. But as soon as it goes into the direction of the API and integrating that into workflows, integrating into other products, then it's going to be the typical API consumption based usage based model where we price per basically character translated. Yeah. Okay. I'm running out of things to ask. What would you like listeners to know that i haven't talked about yeah i i would like everyone to be really fascinated by how ai can can really really revamp that space and how we communicate with with each other um i think there is a there's a very pragmatic uh aspect to that and like how how well is the communication in my company working how well can i reach new markets um how how good am i by competing in um in this in this global world as a as a company uh but there's also really this emotional aspect of that and and we had employees um from our customers on stage with in our customer events uh who've been who've been really feeling differently about how they can participate in um in this global work environment uh since they have access to this technology and uh with all that efficiency i wouldn't want to forget about the the kind of human aspect of how much language is embedded into who we are and and how we can exchange with with others so and and therefore like i'm super excited about what the technology brings in the future and how we're uh going to be able to to grow together even a little bit more yeah um yeah and you mentioned uh i'm not sure this what you're talking about but how it it protects or preserves uh languages that are otherwise being eroded by english i mean i i've spent time in the arctic uh and there's you know inuit and and some fascinating languages up there but uh but they're shrinking yeah and and something like this would allow them to operate on a global scale without losing their native language.

48:21Definitely. I think it's going to be more challenging for those extremely smaller languages which you're just talking about. But there's even this mid-level of actually big countries with their local big languages. I've been born and raised in Poland, so Polish is my native language. And even there, I see everybody's learning English, and that's good. They still should, I guess. But having access to technology, which makes it easier for people who are not so well-versed in a foreign language, that helps a lot them to be competitive in the work market, but also them being able to express themselves freely and preserve the language.

49:08yeah so looking forward what what does deep l have in store for us and how do you see the world evolving with language barriers becoming increasingly porous yeah i mean we've talked quite a bit about speech translation and i think that's that's still an incredibly interesting avenue i think we are a little bit right now at the at a level where we've been with text translation in maybe 2018 or so it's it's still the early days and and and i'm excited to see how that really plays out in in real life usage um in in all of those multinational businesses um i think when it comes to text translation i'm i'm still really excited about the path of of making sure that every company can speak their own language and how customizable ai can become in an in an easy way we've spoken a little bit of terminology um today that's that's kind of the beginnings uh I think we're gonna be able to uh in a much better way make sure that all of those businesses are unique out there and also unique in the way that they're um that they're communicating and that they I really understands uh what is being communicated so I think a lot of advancements that are coming up in in this area um and of course uh like quality advancements in all of the languages that we have out there uh there's there's still a long way to go i think we've come to an excellent translation quality for the major languages that that we know uh but for smaller language pairs there's still work to do and and we really should not forget those Yeah.

50:50And just on that, on the quality of translation, is it that the translation, when you run into quality issues, is it that the translation is understandable but is clearly not from a native speaker? or is it that there are mistranslations that could change the meaning of sentences? The mistranslation case is getting rarer and rarer as the technology develops. Obviously, we cannot fully rule that out as much as we also cannot rule that out with human translation mistakes just happen there. I think exactly what you're saying. It's like sometimes it just doesn't sound that great yet. Right. And we got to get the technology to the level where it always just feels supernatural and as if a native speaker would have written that.

52:05Yeah. And, I mean, we'll wrap up, but I'm just, as you're talking, i'm thinking that it would be great to have a site that translates all the the world's uh major media in one place so you know you could pick a topic or maybe that exists do you know if that exists i i don't i don't know of that existing i don't know if the publishers would be so happy about like an aggregator like that uh i think that's that's that's a discussion of a different type there um but i think i mean in general it is it is really possible like take the bell take the the bells browser extension put that into your browser and just go to to all of those different uh media places out there in the world and and start browsing their content that's feasible already yeah you don't sell the browser extension as a standalone product do you So there's a free tier there.

53:06So you just can download the browser extension and start using that. And depending on your usage patterns and how much and what you need, there's going to be an affordable plan there to use that. Okay, great. Well, I'm going to download it as soon as we're done with this call. Perfect. And Eddie can assist if there's anything on the team. Sure.