Agents at work 18: AI Translation at Enterprise Scale w/ Olga Beregovaya (Smartling) — Agents at work

Hello everybody and welcome back to another episode of Agents at Work. I am Jordi, your host, and today we are with Olga from SmartLink. Hi, Olga.

Hello. Hi, Jordi. Thanks for having me.

You're welcome. My pleasure. You are working for one of the big names on translation. When people need to be introduced to SmartLink, how do you usually do that?

We are AI-first, agentic translation delivery platform.

I see. And this is mainly for content?

Oh, well, we need a definition of content, right? The content types, the content, we are witnessing the content explosion and content types are so diverse that I would define absolutely everything out there as content. So in that case, in that case, yes, it is for content. And I think one thing I did not mention that it's enterprise scale and large enterprise scale. So it's not, yeah, it's not just a not proven, not repeatable translation process, it actually is catering to the largest enterprises imaginable.

I remember in 2010 with Google Translator that it was already one of the main uses of AI back then, but it was different technology. How long has Smartling been here and how was the starting point of the company?

If my math doesn't deceive me, the company was founded in 2009.

So it has been here for a while, so I guess that a lot of things have changed. What about your background? Were you this long in the translation industry?

Actually, I only know two things in life. I know how to run translation technology and as a side gig, I also know how to produce music. So my background in translation technology stems back once I graduated from grad school, I stepped into translation technology and translation automation and have stayed there ever since.

I see. And how was those early days? Like, how did the translation business look like?

The translation business was really funny. That was the time when the concept of translation memories, which is basically Levenshtein distance-based, pretty mathematical, pretty simple matching, string matching, string matching approach, I would say, or algorithm. That's when translation memory was first introduced, right? And translation memory, for those of us who are not in the technology industry, is basically, as I said, mathematical, Levenshtein distance-based. We are reusing previous content. And interestingly, because before that translation was done in spreadsheets and whatever that was available, right? The version of Word at the time, there was a major pushback on like, hey, why would we want to automate? Why would we want to lose money? And why would we want to let mathematics tap into our production process? So we start from translation memories. That's also the time when computer-assisted tools, computer-assisted translation tools, CAT tools, tools were introduced. So back in the day, it was, as I said, the work surface was pretty heavy. It was a lot of string matching and a lot of, like, very rule-based lookups, and at the time, there was very little of machine translation in the picture. Then a couple of years later, machine translation got added to the mix, and we're talking rule-based machine translation, and those, again, those of us who've been in translation industry for a long time know that rule-based had a lot of constraints and might not, if anything, might not have been fit fit for purpose for enterprise-scale translation. And again, if we remember like AltaVista BabelFish, that was a little bit, I would say, fit for purpose for public translation, but as we know, it got deprecated pretty fast. But when we tried for post-editing rule-based machine translation, that was absolutely, you know, the quality of the output was definitely not there. So that's the landscape, a little bit of MAS, a little bit of rule-based machine translation, and new, brand new environment for translators to work in.

Yeah, I remember even as a kid, like the main use of translation that I saw was in business in those instructions when you bought a new appliance for your home. And then many times there, it was really funny because it said things that didn't make any sense. And I guess because it was translated automatically from some other language by these kind of heuristics.

I remember actually a translation, automated translation of a user manual for some operational machinery where a human to manage it would do the metrics. Should have been about 130 centimeters tall and would have hands or arms of 2 meters each based on the translation in order to be able to operate that machinery. So yet indeed, the use was— DOD was a huge use, right? Government was a huge use case, and then enterprise translation for user manuals, knowledge bases was the use case back in the day. I agree with you completely.

How was your career path from your starting point to today that you are VP of AI at Smartling?

The career— I think my career path, like, literally followed the trajectory of natural language processing technology, which means I started as a structural linguist, and I would not want it to be confused with translation industry because what I studied was really the rules of language, the patterns of language, different languages, like morphology, syntax, phonetics. Prosody and like that kind of thing. So my first job is natural language processing was building dependency parsers and glossaries and dictionaries for rule-based machine translation and for what was a search engine at the time, once the search engines actually started existing. And then as the natural language processing technologies started evolving into statistical technologies, Statistical machine translation, and then further on, there was more and more AI or deep learning or like general machine learning, I would say, injected into the system. Then naturally my career took me there.

Usually how I think with new technology, it's like you can do it like better, faster, or cheaper. And with language, what I noticed is that at the beginning was like faster and cheaper, I guess that's why people were using it, but still not better than, you know, like a human translator. Do you think that we are at the point where, you know, like automatic translation, it's as good as humans or not yet?

I would say that like here at Smartling, better, faster, cheaper is actually the conversation we have with our customers. But when I argue with the R&D department that I manage, they usually, okay, Olga, what do you want? If you want faster, we need to prune our prompts to like X number of tokens. And we really need to optimize our pipeline, and which exact, uh, agentic step do you want me to cut out? Or if you want better, sorry, give me 3,000 tokens for a prompt and I'll deliver better. The cheaper would always be in the equation because it's definitely 4x or 10x cheaper than human translation. So I would say, yes, we're almost there. Yes, it's fit for purpose, and often, uh, our translation, our agent-based translation is better than just vanilla LLM translation or neural machine translation, but But I would say that the triangle still exists, not to the extent it existed 3 years ago, but that triangle, like, okay, where, like, where's your top priority? That's definitely there. Having said that, our edit distance for certain content types, more structured content types, is actually shrinking, and the human parity can easily be at very high percentage, like 70%, especially for languages that are easier for natural language processing. Now, what I can get—

what—

hi, me. What we, Smartling, can get out of Latin American Spanish, which renders itself beautifully to automated translation, and what we can get out of Southeast Asian languages would obviously be two different things.

And why do you think that it's that?

There are multiple reasons. I mean, first of all, it's— the world is about data, right? And as we see it now, the world is yet again about labeled data and annotated data because people are discovering, and large language model developers, foundational model developers, are discovering that unsupervised learning can only get you this far, right? So there is a lot of money if you see what's happening with Scale AI or what other companies, well, let's go with TELUS and Scale AI. They were, they're making ton of money off natural language or LLM developers because there is shortage of data. So when we go back to shortage of data and good clean data because the garbage in, garbage out concept has not gone anywhere. Obviously we're talking about lower resource languages and higher resource languages. And if you think about the number of Latin American Spanish speakers in the world, and you speak, you compare it to the number of Urdu speakers in the world, the size of corpus is natively smaller, just statistically. So one of the things is maybe lack of training data. The other one is obviously grammatical complexity, right? Romance languages are just They're just grammatically simpler than say Southeast Asian or like say, uh, Finno-Ugric languages. So I guess it's between complexity, knowledge of those languages in development teams, and sparsity of training data.

And for people who have been following the moment of LLMs, like you have, we have these generative machines now that many people say like, well, they can be used for translation, but before we get there. What technology was used for doing translation, let's say, in 2020?

I would say that it was still heavily statistical or hybrid statistical with injection of rule-based, but people definitely started dabbling with transformer models. So, but when you go to like enterprise-scale translation, I would say that 2020 was probably still the time when statistical translation was prevailing. Now, I'm only talking from my enterprise experience perspective, but that was definitely when neural machine translation started taking place over. Statistical and models are being migrated. And again, datasets were being migrated to neural machine translation.

I see. Something that it's a good moment to define right now is like what enterprise scale means, because there's like a lot of gap there. Many people work on startups and it's like, well, yeah, we have like 1,000 users. It's pretty enterprise. And you're like, well, depending like how much do they use you exactly? So here, when we say enterprise. We are talking about Disney, I think, is one of your users, IBM, and I think that many others. So how big is the corpus that you need to translate? Do you have any metrics, metrics that you can share?

The way we approach it is we like to build on top of foundational models or generic neural machine translation engine, and the reason is that the generic broader domain corpus is already there. Now, a healthy starting point where you can actually see impact from the weights of customer-specific purpose is usually 10,000 segments or translation units, which in our universe would be about 10 words on average. So I would say 100,000 words is probably a healthy place where you can start to see the impact of, you can start to see the impact of where you can skew generalized models towards customer-specific. Now, to answer your question about what's an enterprise, if you look to our, if you go to our website and you look at our at our customer base, we are talking not necessarily 100, uh, 10,000 or 1,000 employees, we're talking hundreds of thousands employees. So, since we're able to deliver to IBM, since we're able to deliver to Disney, that should give you an idea of the scale of the operation. Now, when I say enterprise scale, there are several dimensions to it. It's gotta be repeatable, right? Because one-offs, what works in the lab might not necessarily work in an enterprise scale system. It's going to be like, when we say scalable, you gotta be able to mitigate latency. Are you gotta be able to manage that huge corpus of data. And I would say repeatability and predictability are definitely a huge factor alongside compliance and data governance. And we like to think of ourselves, I said the data is king now, but when you think about it, there is so much messy data out there, even on the enterprise scale, right? You have user-generated content. You have, no offense to our engineering audience, but you have content developed by like 30 engineers at the time, and that knowledge base developed by 30 people who are not technical writers. You have amazing content and instructions and knowledge sharing, but it's not necessarily the best training data for a translation. So I would say ability to manage, centralize, and clean data, predictability, repeatability, ability to mitigate potential slowdowns and latency and compliance is what makes you an enterprise-scale translation platform. And actually one more, ability to cater to different content types. Because if you think of a large enterprise, you have your UI, you have your UA, user assistance, right? You have your legal content, you have your marketing content, you have your user-generated content, and you have your multimedia. So a healthy enterprise-scale platform should be able to handle all those use cases as well.

When the users knock at your door, are they already like big enterprises and then they come with a specific problem, or you foster them since, you know, like this is more startup that wants to go global. Since day one or maybe like day 365, like the first year, because startups now are going faster and faster globally thanks to like things like Stripe and things like that. So, I guess that translation is something that, you know, like younger companies need now more than they needed before.

Actually say it's not just younger companies. It is a lot of enterprise companies that actually are struggling with older processes. So, whether it's a young company or whether it's a large enterprise and the majority of our customers actually are large enterprises because this is what we do best, we still come in as trusted advisors. We still come in as somebody that, and quite often the customer does not even know what their problem is. They just come in and like say, you know, we're authoring in spreadsheets or even using Copilot. I was talking to a VP, SVP of global content for Fortune, not even 100, probably Fortune 50 company. And he said, like, what do I do? Everybody's now able to use Copilot, but that means that if we have 300,000 employees, I have 300,000 voices in my company. So then we say, hey, you have a problem. Let us centralize it for you, right? So quite often our discovery session actually helps our customers understand what the problem is. There is something that's called a Localization Maturity Model, which was developed by Common Sense Advisory, which is one of the oldest and largest analytics companies in our industry. So it also depends on what kind of customer comes in. If they're lower on localization maturity model, probably you're looking at a bunch of spreadsheets and Google Docs or Word documents. If a company is higher in localization maturity model, usually we have a very educated customer coming in.

For those cases, do you have like categories of, you know, like this is like the top 5 things that usually people are doing wrong and we can help with?

Automation. Automation is a big one. That's where there are a lot of manual processes at place. And this is obviously something our agentic workflows, workflows help solve. So automation is one. And again, there is automation is one of the KPIs, how automated your processes are and how much money and time can you actually save through automation. Second one, again, if I were to rank them, centralization, again, centralization of linguistic assets and centralization of processes, because quite often what happens is country offices would usually work with local providers. And might as well have their own homegrown technology, and different content types would be parsed out in different ways. So centralization is probably a second one. I still have 3 more to go. Compliance, again, that's another one where people would paste, and that's like here, the age of, the age of AI and the age of large language models and very accessible through browsers or through applications. That's another one where people would just paste their corporate, their enterprise content into ChatGPT windows. What happens? You just exposed it. You don't have an SLA in place, and you basically just, A, introduced the risk, and B, you just gave your data away as training data for the models. So that's where we come in and say, again, hold on. We're gonna perform the governance, risk assessment, and compliance analysis for you. So here's the third one. Two more. Brand voice. I think brand voice comes along with centralization. Because all companies have marketing brand books, all companies have their requirements, in-geo requirements, and the brand needs to be consistent across locales. And again, if you have like 30 processes across different departments, the brand voice can be dissolved. It can be completely different between different departments. So that's another one, ability to parse out and send different content types through AI-enabled translation process. There'll be like, what you do, what you need to parse out a UI string is dramatically different from what you need to do if your legal department gives you a flat PDF, not even a parsable, like a scanned OCR PDF, and you obviously need the facilities to handle that.

When you go to these companies, are you usually substituting something that they built in-house? Like they just had like some small team that took care of these and, you know, like grew to the point that was not sustainable. Or they were using other tools or other companies, and they said like, well, we need a better, bigger partner here because they outgrow those, those solutions.

Actually, buy versus build is one of the biggest debates in the world right now, right? It's not even just us. And actually, one of my colleagues just did research, and buy seems to be the prevailing approach now after build has proven to be not financially viable. And takes a lot of workforce. And if you think like, hey, I'm gonna buy this thing and then you just start— and multilingual thing— and then you just start adding up the headcount, okay? 2 data scientists, 2 computational linguists, couple of engineers to build all the integrations. If you really want to be truly multilingual, you need linguists, in-country linguists to validate the output. So once you start adding it all up, it's like, why would I do this? So this is another conversation we have with our customers. Like, yeah, you had a great idea, you like built something that's integrated with your Git or any other repo and you plugged in GPT, but are you really getting work? Like, is it really your money worth and are you really getting the quality and the speed you're counting on? So usually we advise against build and more towards buy, but the thing is AI and again, like large language models is so accessible now that we also work in the hybrid scenario. Like, hey, you figured out how to fine-tune your model? Great. Now we're going to plug it into our platform and compensate for the missing pieces.

That makes a lot of sense. With LLMs, there is this problem that they look magical at first, so you just try something and it looks like it works and you're like, wow, this is amazing. It's going to be able to do everything, but then you try to plug it in production and you're like, damn it, it's too yappy. It's adding things to the translation that it should not. It's adding emojis. Where are these dashes coming from? Staying on brand, I think that it's almost impossible with out-of-the-box LLMs. So it's good that people try it first, so they go to you already, like, with the lessons learned of at least what it doesn't work.

And again, that's either localization or translation maturity model, and we love clients that come with experience that say like, hey, it worked for, like, English. But once we go into the multilingual space, it actually doesn't work, so help us out. So that's our favorite customer, like a customer that actually tried something out and say, hey, it doesn't work. Like, I'm a native speaker of Russian, and my GPT, as much as I trained it, like yesterday started talking to me like a drunk sailor and also a heavily hallucinating drunk sailor. And you never know, right? You never know. It's very— unless you have really complex prompting, You never know what the model is going to spit out, how factually accurate it's going to be, and how much it's going to hallucinate on you, and whether it's gonna deliver the right tone of voice. And again, this is where companies like ours come in. If you have several computational linguists and data scientists putting a lot of effort into very complex prompting with RAG and like with MCP, we can deliver the level of quality and the level of predictability that you, you cannot just get by plugging in a model.

Yeah. And inside the company, who owns this? Like from your customers, when you say like, well, translation, localization, what department actually owns this? Because you were talking like, well, you need to translate things about compliance. You need to translate things about legal. You need to translate things about PR, engineers. So, you know, like who usually is a touchpoint in your client's company and how they organize when you say like, well, we help centralizing all this. Well, who is taking care of this inside those companies?

Before I joined the supplier and technology developer side, I actually worked for a large enterprise, and if somebody looks at my LinkedIn profile, there will be Autodesk for 7 years, and back in the day, the ownership of localization could be completely random, and the way you would look at translation and localization would be just a line item on your P&L and a cost center, and then whoever takes on that expense, it could be product, it could be engineering, it could be Marketing, could be local sales. So that was an absolute whatever. Whoever adopts it and whoever is willing to spend extra few million. And now I would say changed dramatically. Now it mostly localization and translation initiatives and departments would report into product or engineering for two reasons. First of all, the localization and translation efforts are more and more immersed into the actual development environment. So first one is the processes are more immersed, and second again is AI is so much more accessible and available that actually product is just a healthy place where it would sit because there is now translation as a feature is one of the latest trends. So logically that would sit within, within the dev department or the product department. And one other thing is localization used to be an afterthought, like you've completed your process and then like, hey, throw it over the fence and see what happens. And I remember there was very little influence upstream. Like I was literally scolded in the developer, in the room full of developers, because I was still like, okay, who are you to tell us how to handle our variables? And why are you even like, we're already, I mean, we're coders. So now it's completely different. Now the influence upstream and translation first is a concept.

I mean, it gained its place because people saw, for example, I guess that for stores or e-commerce, actually translating converts more. Companies that want to gain new markets, they have to be like really like grow really, really aggressively. So they see that there's a lot of value here. So I think that even developers now say like, well, yes, it changed how I develop, but I have to because it brings so much value that I cannot oppose to this.

Yeah, and like, I mean, the developers would know that once you need to ship global products, the concept of pseudo-localization is pretty big, right? Pseudo-translation, where you just plug in machine translation or an LLM, and actually you can see what your content is going to look like in various languages, right? It's going to expand into a chain of German compounds of like whatever, 40 syllables, or it's going to shrink completely for Chinese and Korean. And now again, I think developers develop with global content and globalization on mind and the tooling that they have for that, actually, again, Smartling included, has evolved dramatically. There is another dimension too that I did not mention. When you look at the developers' teams, they're rarely diverse enough and multilingual enough to be able to account for complexities and particularities of different languages. So again, that's changing where dev teams actually learn to trust subject matter experts. Like, there is beauty of right to left, right? Or code switch, right? Where you're just switching, like, all of a sudden, half of your string is Japanese and like, hey, why don't we throw some English in? And that's not necessarily the expertise that dev teams would have. And this is where a trusted advisor like ourselves would come in and call this out.

Yeah, they look like small things at first and then you see how big it is. Like you start craving and you say, well, but this, this never ends. So I think that people like really, really quickly see like, yeah, this is expertise that needs to be bought or bought in. Like we bring someone who already has this expertise or it's not something that you can develop in a couple of sprints from like 30 blog posts saying like, okay, now I'm the expert on this. Let's, let's just do it.

I mean, we still are more than willing to plug it into our customer sprints. But we like to come in as trusted advisors, right? Because that just, that just helps. And another thing, actually, it's really funny about diversity. There is a lot of research. If you introduce different age groups, different nationalities, and genders into your dev team, you actually have different perspectives. And it's even more applicable into developing large language models because this is, I mean, they're inherently multilingual. And this is where you do want those different perspectives. And I do know that firsthand, we had, um, unfortunately she left the company, uh, Chinese data scientist on our team, female Chinese data scientist, and the way she was looking at things and the recommendations she was given were dramatically different what would come to my mind, for instance. So, here's another aspect of that.

And when you go global, your users are going to be like much more diverse than however diverse is your team. So, you know, like you need as much as possible because, uh, the world is pretty complex out there. Out there and your customers are all over the place.

Right. And actually, to your point, now there is a super interesting concept that we're also working with, which is called hyper-localization, right? Just like out in the world, there is hyper-personalization. There is hyper-localization where it's not just Latin American Spanish anymore. You need to be able to tweak your output to specific demographics in Chile. Or in Peru, and again, the modern, the state of modern multilingual AI actually allows for that catering to very specific niche audiences.

Are these things then, like with things like hyperlocalization, done like as a preprocessor? So you basically take all your data and then you translate it and, you know, like now it's clean, it can go out there, or there are starting to be things that are done at the edge of, well, it's just in time, maybe it's even done on the device of the user and things like that.

I would say, like, for instance, our process is fully agentic, and what we're discovering the best thing is one model or one function of the model checking and vetting that previous output of another model or the same model. So the way we like to hyperlocalize and think about it is what's called automated post-editing, where you've already put a lot of effort into generating the best translation first. But then when you plug in that automated post-editing step, this is the best place to do your hyperlocalization and adapt it. Check, check for hallucinations, see what you need to mitigate, and also say like, okay, great, you did it for Spanish Spain, right? Castilian Spanish, and now adapt it to ABCD, taking these things into consideration.

Okay, and that's done just in time every time that the user wants to consume it, or it gets stored and they like reuse.

Both. Actually interesting you would say that, cuz one of the features is actually exactly this. How do you store AI output? Cuz traditionally you would only store the, uh, human output.

Mm-hmm.

But now if AI output has gone through the vetting mechanism and actually is vetted to be of human quality, why not store it and reuse it? It saves us money. It saves our customers money. Why? Because you don't need to spend tokens anymore. You don't need to like run it through the whole pipeline again. So that can equally be stored, but if we see that it still requires human input, then what's a product of human in the loop, that's what would be stored in the system.

I see. And as someone like working at the frontier of these, like, what are you seeing right now? Like, is it like prices going down, quality going up? What is the next frontier? Like, do you think that we're, yeah, like we are still like, it's going to reduce a lot of the price. It's going to get much, much better. We're going to be able to do it faster. Like, you know, what are the KPIs that someone who is in the weeds?

As I said, the KPIs are the same: faster, better, cheaper. So it's definitely getting dramatically cheaper, and the conversations about translation being commoditized have been around probably for the last 15 years, right? As the quality of statistical machine translation, other NLP functions went up, obviously it was it's absolutely apparent that the price is gonna go down. And again, if we go to the Smartling example, we're able to cut the costs dramatically and cut the customer price dramatically because the quality of AI output is getting better. So the cheaper KPI is definitely there, and we do, we're not oblivious to the fact that the cost is gonna go down, the price is gonna go down even more, right? The higher the, the higher the quality of automated output, The lower your reliance on human in the loop, obviously the lower the costs. Now, same applies to the speed and average throughput for a human translator is 2,000 words a day. If you go AI first, you can go all the way up to 6,000, 7,000 words a day. So obviously the cheaper is there and the quality is the underlying, it's like the core, right? That's the foundation everything is built on.

I see. And on multi-language LLMs, how do people approach it? Like, is it better to use only one language? Language for all your content and then later on, like, translate it? There are like some tasks that actually make it, no, no, you need to regenerate with this technology from scratch. How are these things developed?

Obviously, and we can see that across different corporations and across our client base and across different applications, there are two schools of thought, right? And one school of thought is you start with English and you can use AI to generate your copy, but there is another problem with that. If you use AI to generate your copy, but your copy has not been checked for hallucination and your copy has not been human-vetted, then— and I actually was talking to one customer and they said like, hey, it's all awesome except we get unvetted AI-generated copy, then we need to correct it, and then it goes into different languages. But if you have clean English copy and you've done source analysis, hallucination mitigation, source pre-editing and optimization, then the translation approach is still very valid because you still have your fully vetted English source or, sorry, I mean source language source. If it's like, say the company is based in Sweden or Germany, and then you've already cleaned it up and then it renders itself nicely to translation. But as you know, there is a second school of thought is where, and we, we call it death of the source, where the source language doesn't even exist, right? Where you start from, like you can write a multilingual prompt that actually caters to different languages, and then you just generate all languages at the same time. Each of the approaches has their pros and cons. Most probably the second approach is going to win eventually, but there is also an issue of if you look at foundational models, they are predominantly English-centric, right? Just because of the presence of the training data. So still at this time and day, I would say that it's better to translate because you can build smaller models or you can fine-tune your models and train them for the needs of specific languages. And then you probably will get more accurate results than from generation, but it's just a matter of time.

When enterprises are at this scale, does their data match the normal distribution, like the one that the LLMs are already trained on? So you, for example, say like, well, it's mainly English, for example, or there's a lot of dominant English language out there. Do your companies like have the same problem or the same distribution?

Actually, that's a fantastic question. The KPI changed, the KPI of localization changed, and in the past it was mostly quality, linguistic quality of the output, and we have things like, for instance, multidimensional quality metric. That's a very standard thing for us, a very standard thing for our industry where you really judge by linguistic quality. The KPIs changed dramatically. Now they're mapped directly to business outcomes, right? And how well is your translation, how much is your translation impacting the sales, the conversion in a specific region. And the reason why I'm bringing it up is because it ties directly to data. Depending on what the company is doing in that region, right, it maps directly to how much data they will have, right? If it's a like region with huge customer base, obviously absolutely everything is going to be translated, right? If the company is just warming up to a specific market or made a decision that they don't need to translate as much, you translate a much smaller scope of different content coming from different applications. It's tied specifically to markets, and that impacts the content distribution. And this is where you go into AI again. Hey, I still need to train my models, and you go into the world of synthetic data or hiring linguists to build up enough of the corpus to actually produce even results across different languages.

I am really new into this world, so I don't know how developed the market is out there, but I could imagine that even like global brands, they don't have the same branding all over the world. So I would assume that maybe like Nike says, just do it everywhere, but there are like peculiarities from, you know, like South America versus North America versus Asia. Does your platform help there?

I mean, again, fantastic question, and this is where automated post-editing and adherence to customers' global brand books and style guides comes into play because you're absolutely right. What's good for English, like, let's go with formality for German and Japanese. What's good for English is by no stretch good for Germany, right? You really want to differentiate between formal and informal. And there's another tweak to that. Again, I'm quoting one of our data scientists, like, informal can sound like a, again, drunk sailor. And formal can be like, uh, you are writing a letter to King of England, and neither can be nuanced enough for a specific language. And this is exactly where, as I said, we come into play and we understand the nuances of each individual language. And this is where the way we use linguistic assets like local terminology, our legacy content, and style guides actually allow us to cater to that specific market. And the days of like one-to-one literal translation, which was the case like maybe 10, 15 years ago, those days are definitely over.

I see. It has like a huge business impact because I remember the growth team of Facebook when they went to the Japanese market, they actually had like some troubles growing there and it's like, we don't understand like why in the rest of the world we're like, you know, like growing this fast and in Japan we are not. And I don't remember exactly what cultural thing was and I am not now sure if it was about like the phrasing. Of invite your friends versus share with your friends or something like that. Because in there it was like, well, it's really important that if I share something or if I invite someone here, I am endorsing it. So I don't feel comfortable with like inviting them to Facebook. But if you say like share, they were like, oh yeah, share it with everyone. So I have no problem. And it actually had like, like they were lagging in Japan for a while. Until they corrected this thing. And, uh, there is no one in the world and your company that can help, you know, saying like, well, we have seen this so many times that it's actually like really, really helpful for, you know, like global, global brands.

To this, my response would be again, back to hyper-localization. And now again, the modern state of AI and the model, uh, the modern state even more so of generative AI actually allows you to do two things: fact-checking and actually making sure that you have enough vocabulary, either inherently from generic models or through fine-tuning or through— I mean, RAG is a huge thing in our platform, right? The more you're able to fetch from validated locale-specific sources, the more are the odds that you actually produce relevant content. So again, I would say that there are so many ways of mitigating it right now that you really are able to help your customers avoid such mistakes. The models are not perfect, right? AI is not perfect, but if you have the right mechanism to mitigate those hallucinations, to fact-check, and to RAG into reliable sources, usually you can help your customers avoid this. There is another one. We all know it's common knowledge that there was a stop list issued for China, China, by the government, where you cannot use superlatives, like you just cannot say our product is the best product, right? And things like this can be plugged into your workflow and they can be plugged into your training data or into your prompt, or like, hey, reference this. You just cannot do it in this geo. And that's something you were not able to do before unless you just build a huge glossary and just brute force it into your content. And that's obviously not the way to go anymore.

Yeah, it's a really complex business area. So if you can partner with someone who already does this for a living, you can save like a lot of headaches at enterprise scale. So when did you step up or you were already hired as VP of AI at your company? At Smartlink?

It's been a little bit over 3 years.

And you already started as VP of AI?

It's really funny. Again, it follows the trajectory of where the world of translation is going. I was hired as VP of machine translation.

Okay.

And that was literally my role for the first year, pushing machine translation, custom models for our clients, and just building up the practice of delivering machine translation, raw machine translation, or post-edited human machine translation, but the more AI came into our automation, and our translation processes actually went to the CEO and it's like, the title doesn't make sense anymore. Let's really talk about VP of AI because it's just not machine translation alone anymore. So I would say that's, yeah, that's how my current title came to be.

How was the GPT moment when you are working at this company for, you know, like a big company and then out of the blue, there's this new technology that for you may not be new, but everyone is discovering. And it's like, wow, okay, like, are we, is this going to be our substitute? Are people going to need translation more than ever? How was this internally?

I mean, first of all, when GPT came out, I was still working for my previous company and GPT-1 was like, eh, you know, nice, interesting, not quite accurate. If you remember the cutoff for the first version was like 2021, I think, uh, for GPT-1, the cutoff on data, it could not fetch from external sources. I was like, ah, you know, Nice, interesting, but neural machine translation is definitely still the way to go, and we didn't quite see that room for it in multilingual space. GPT-2, and it's actually trainable, and the corpus for GPT-2 was much broader than for GPT-1, and that's where I started like, huh, it actually, it actually will have room in the whole translation, like translation pipeline, and when I moved over to Smartling, one of the reasons why I moved over, like it is a native native SaaS platform that was designed for handling natural language in all of its manifestations. And this is the year of GPT-3, 3.5, and this is where it was. Yeah, it definitely has a home and not just for translation, but for all the other transformations and now all the other agent-based processes. So now, very important, you say it's substitution. Not yet. We do— one of our fortes is we benchmark across 8 different automated metrics like BLEU, COMET, Meteor, MetricX, but we also use human judgment and the correlation with human judgment. And we still see that neural machine translation has its good home for certain languages, even more so lower resource languages, and for very high-risk content, just because models are not non-deterministic and neural machine translation will give you that comfort of predictability. Maybe the cost of fluency, right? Because it mostly operates on a sentence or phrase level, but we're not saying substitute, we're saying augment, at least for the, in the foreseeable future.

I see. And in multimodal, do you see this being like the new path? Like most of the people interacting with machines and with content through voice and audio, or you think that it's going to still be like text and does it change the technology that we need to handle this? These two modalities?

It changes it dramatically. I would say that the initially, at the very least in the translation space, we were talking about text-based. It's easier to handle. Text is the natural language, large language models or small language models, any size models. They're called language for a reason, right? Because they handle human language, but now multimodality is definitely coming to play. Over the world and in localization and translation industry the same. We are maybe catching up a little bit. Maybe our industry is a little bit slower than other industries, but multimodality is definitely becoming a thing. And now automated transcription, right? It's already there. ASR, automated speech recognition, is already there. And again, it's essential for AI-first translation platforms like ourselves to be able to capture it and deliver it. Concept of digital twins. Again, that's another one, right? You need to be able to cater to different digital personas and develop those different personas. There is something that's extremely costly in our industry, which is parsing out PDFs. And again, like using the whole OCR, and now you can get much more out of AI-based technology for handling complex formats. There is another very expensive thing called desktop publishing. Where you actually like have a Photoshop output and then you start modifying on-screen text and images and everything else. And now you can actually generate multilingual, multilingual images. What it means is the reliance on human-based desktop publishing is much lower now. So multimodality has definitely made its way in. Again, ASR and then text-to-speech. Text-to-speech is another one. Right.

And for speech, is it hard? Like, does it make it harder that it's like real time?

No, I mean, it would. It depends on how you stack up your models, right? If you handle multimodality within a single model, and then obviously you need much more in your prompting to handle multiple formats. So I would say you expect a little bit more latency and a little bit more delay just because of how much you need to handle, how much you need to handle to generate automated speech. But again, there is real-time and real-time, and there is— it's one of another debates in our company. What is real-time? Is the customer willing to wait a couple of seconds, a couple more seconds and get higher quality across different modalities? Or do you want it instant where you obviously need to bypass certain AI steps? So there's still a trade-off, but again, our prediction and our roadmap caters to that. That trade-off should be shrinking more and more and it should be becoming more and more real-time.

I am really excited to see like how these things evolve because they are going to literally be around everything that we do. So, you know, it's not something that is like down in some, you know, like Substack level that people don't do and just benefit. It's going to be like literally like how we relate with machines. So really cool to see like how these things go.

I think what changed also is in the past, again, you run your ASR, automated speech recognition, then you edit your transcription. And often using humans or using machines. So you already have two steps. Then you perform machine translation, and then you go text-to-speech and generate your speech. And what does it mean? You train language model, you train acoustic model. So there was like a whole hell breaking loose, right? A multi-step process. And now you can actually— the beauty of it, again, if you speak about the modern state of models, you can actually, with some training and some proper prompting, you can actually do it within a single model. Model, right? I mean, when you wake up and like you, I was trying to get a hybrid of, I needed a crossbreed of Giraffe and Dash Hunt. Don't ask me why, but I needed to do it. Actually, ChatGPT was able to deliver it like after certain iterations, it was able to do it for me just through text prompting. So you obviously can do it all within the same, the same model right now. And that cuts out all those multiple steps.

I see. You have like a long career on machine translation and natural language processing. Do you feel that it's accelerating? Or because it's something that has been improving since ever. So I remember since 2005, first like Google Translate until today, you could see like year after year, like, well, yeah, it's already there. And then 5 years pass by and like, dude, it's much better than before. And then 5 years pass again and you're like, wow, it's getting like really, really good. So do you feel any kind of acceleration or is that the pace that we were before?

It's really funny you would ask that. Again, if you think about it, you go from rules right? Not just machine translation, but any kind of natural language processing or any, like most of the computer-based applications. Then you go to statistical, and then you have a dramatic breakthrough in transformers, which are also keep on evolving, right? Transformer models have been around since 2017, and then you start adding parameters, right? And then you go trillion, and this is where you go dramatic breakthrough. Again, going back to our benchmarking, but hey dude, It's tremendously better. For instance, we didn't see much difference between 4.0 and 5.0. When 5.0 came out, again, I'm in the GPT world, as much as we're using all the models out there, like there is Mixtral in our pipeline. There are a lot of Claude models in our pipeline. There is like place where we have a space for Google Vertex Model Garden, and there is space again where we're using IBM Watson and AWS models. I mean, it's all fit for purpose, but we do not see dramatic breakthrough when it comes to the translation space between previous generation of model and current generation models. We benchmarked 4.0 against 5, and there was like couple blue points and couple of comet points, so it was not dramatic. And I think the reason is the current generation of transformer models, and again, that's probably common shared opinion, they kind of hit the wall. Of how far they can go. And also, they have consumed all the available data, right? So you don't get dramatic. And I mean, the number of parameters is already humongous. So you kind of hit the wall, you've already played all your livers. So how much further can you go? Right? And this is where reasoning models come to play. And like a lot of next generation, the general conversation is like, especially for translation tasks, we've hit the wall and we're not going to see dramatic breakthrough.

Okay, so we are going to see more from platforms and then like deploying them and getting used to them, but we are not going to see out-of-the-box free lunch on translation as we saw like from 2020 to 2022.

No, the dependency on application layer and the capabilities of the application layer are probably more important than ever, and it's really not your LLM or not your whatever the model you it's really what you can get out of it with the right handling. And again, everything goes. MCP, RAG, fine-tuning, changing the way you prompt, pulling from previous linguistic assets through translation memory. So again, no free lunch out of the box, but this is where you would see the growth.

I see. So thank you very much, Olga Beregovaya. We're getting close to the end of the interview. I usually like to close with a couple of questions, one positive and one negative, as a counterpoint. One is what would you like to see more in the world and what would you like to see less?

As I said, from technology perspective, I would definitely want to see either pivoting or iterating on the current transformer approach to developing GenAI. That's for sure. I would definitely want to see smaller models for translation, purpose-built models for, for instance, for translation tasks or any vertical. Models. And I think this is the direction the world should be going in. I read an article, if I'm not mistaken, it was Wall Street Journal, about corporate AGI is likely to evolve faster than generic foundational AGI. Because another thing I would want to see is corporations discovering what a gold mine of data they are sitting on. And I think that's another thing, like where people really realize, purpose-built, vertical-specific, domain-specific models are the way to go. They can be smaller, which makes them faster, but they are built for specific verticals and specific purposes. So this is what I would want to see more of. And again, maybe the humanitarian aspect and outreach to long-tail languages and under-resourced languages, that should definitely be a factor. Yeah. So this is, uh, this is This is what I want to see more of. And what I would want to see less of, and I think we already are getting there, is this GenAI as a silver bullet, LLMs as a silver bullet, and that Wild West and like, hey, we can achieve everything, as you said earlier, out of the box and sporadic deployments. We know that there is an article, although I just saw counterargument and counterarticle, that 95% of AI initiatives initiatives tend to fail and usually because they jumped into like, hey, it's gonna solve our problems. And what I want to see less of is this and more of is measured approach with clear KPIs.

I really like both takes and I hope like happen, happen soon. I think that it's a good counterposition because one is like, guys, there is no silver bullet. At least like we have the technology to develop like really nice systems. That come from the composability of really good parts. So please focus on the parts and then build a system instead of building something that does it all. And then on the other side, like, please stop trying to build a thing that does everything because we are not going to get there. Just build it as a system. So in 6 months, I will check back on you to see how these things are going.

Please do. And also there are so many things specifically on Smartling roadmap roadmap for H2 of 2026, I think 7 months, right? We're 7 months, 7 and a half months away from end of H2. That would be a great time to check back with me because I have a lot, a lot of news to share.

Excellent. Really excited about hear them. If people want to learn more about you and your company, how they can reach?

Honestly, I would say that LinkedIn is the best way because all of our podcasts, all of our blogs, all of our white papers, all of our recent announcements, it all lives in LinkedIn. And equally in the company LinkedIn and in my own. So just adding me on LinkedIn and following Smartling would probably be the easiest way to follow our development. And again, we do a lot of webinars, we do a lot of events, and if you just follow us on LinkedIn, you'll, you'll have access to all of them.

Excellent. So thank you very much, Olga. I enjoyed this a lot.

Thanks so much for having me.

See you soon.