AI, Machine Translation, and the Future of Multilingual Content with Konstantin Savinkov

Konstantin: I think what's bad, and that's what everyone highlights, is that it's a new tech. People don't really know how to use it. They can start sharing wrong data with it. They can over-rely on it. But it happens all the time. When machine translation was first introduced, one big oil company found all their proprietary data in Google search, just because they didn't know they could not paste their confidential information into a free online translator. Much like here, it will take some time.

Brian: Today we are talking with Konstantin Savinkov, CEO and co-founder of Intento. After receiving a PhD in 2008, he led research and development efforts for online content services, then worked as chief operating officer at Bookmate. In 2016, he founded Intento, where they develop an enterprise machine translation hub helping large companies translate 20 times more on the same budget and make their customer and employee experiences instantly multilingual in over 450 languages. Konstantin, welcome to The Native Experience.

Konstantin: Hi, Brian. Thank you for having me.

Brian: Tell us about yourself, what you do, and what you do for fun.

Konstantin: By training, I am an AI person. I got a PhD, it has been a while. Now I run a company called Intento, where we basically help large companies use AI. One of the important insights for me was that although making core AI stuff is fun, it is not something that pushes the needle for large businesses. That is why Intento was focused on the more boring but more important part.

Brian: Tell us about where you are in the world right now.

Konstantin: I am in Berkeley, California, across the bay from Silicon Valley. Where I can, I do some adventure trips. We have very diverse nature in California. Mountains, rivers, ocean. I enjoy this nature with my kids and family.

Brian: If you could speak only one language for the rest of your life, which would it be?

Konstantin: I would fail at choosing. I share such a big joy of learning new languages and finding their differences. You get those phrases that stick to you from one language to another, and it is not easy to go without them once you know them. Probably if I had to, it would be English just for the sake of usefulness. But I get a lot from other languages, so I would not give them up.

Brian: If you could live in any country in the world, where would you go?

Konstantin: It feels like about 10 years ago the world was on a peak of being united. Now, sadly, you can start to imagine what if I had to live in one country. I visited many countries. For short visits, I love Chile and Greenland for the unique combination of people and nature, but I would not be able to live there. I moved to the United States about six years ago. I realised the country needs to be active enough to live in. It cannot be too boring or too calm. The U.S. is perfect for me for the time being.

Brian: What is your favourite foreign language film or book?

Konstantin: The book is called Rumdoodle. It is an English book, similar to Three Men in a Boat by Jerome, but about mountaineering. Since I spent probably 20 years of my life in high-altitude mountaineering, I love this book for the humour and the language.

Brian: Tell me about your mountaineering experience.

Konstantin: Nothing like Mount Everest, but we used to travel with friends in different mountain regions of the world, typically at relatively high altitude around six or seven thousand metres. We climbed some peaks and made some nice trips.

Brian: What are you currently nerding out on?

Konstantin: Everyone nerds out about GPT, and that is where I directly work in the translation industry. Our company helps large companies configure their automatic translation solutions using machine translation. For us, the GPT stuff is twofold. On one hand, we are looking at how to improve the solutions we are building with these new capabilities. On the other hand, we are looking at how much more productive we can be as a business if we use it internally for everything from writing to developing software code. It is totally game-changing everywhere. That is my favourite topic at the moment. During business days I have to run the company, but on weekends I can play with the technology.

Brian: What do you think about ChatGPT, positive and negative?

Konstantin: The technology has been around a couple of years. We got access to the first GPT-3 version in January 2021. Until the update in late November last year, it was very subpar, especially in non-English languages. It was good, but not good enough for people to change the way they work. Then they developed the updated GPT-3, DaVinci 3, and almost simultaneously released 3.5, which is ChatGPT. Suddenly it was better. It was good enough for people to go to a new webpage and copy and paste their daily work.

Konstantin: What is unique about GPT is that it is trained to produce output based on input. If you think about it, that is what we humans do probably 95% of our waking life. So it can be repurposed to do almost everything. The downside is that people don't really know how to use it yet. They can share wrong data or over-rely on it. But that happens with every new technology.

Brian: What does it need to improve first?

Konstantin: In terms of model capabilities, it is sort of enough. Where it needs to improve is the boring stuff: it needs to be cheaper, it needs to be faster. Compared to traditional machine translation, it is super slow at the moment. It takes about three to 10 seconds to process one sentence. We used to have sub-second speeds for AI.

Konstantin: The other area is data engineering. Currently, the data is not very well balanced. If you try to use it for translation outside of a general domain, outside of what people typically write on the internet, quality drops. You can get accurate results for general topics in English, but when it comes to other languages or very specialised domains like specific fields of medicine or legal, it starts to lose accuracy. That is where a lot of improvement is needed.

Brian: How do you think GPT is going to influence machine translation?

Konstantin: Machine translation as a technology is actually pretty good already. But for truly good automatic translation, it lacks certain things. It lacks common sense. It lacks the ability to work in context. If you translate something from English to Spanish, you have to care about tone of voice and gender. Machine translation does not give you levers to add this context. That is where we see a lot of use from GPT: extending machine translation capabilities to achieve really good automatic translation. We already use it in production at Intento for certain customers.

Konstantin: The bigger impact will be in three to five years. It may sound radical, but in many cases we translate just because we cannot create in many languages. Ideally, we would have a copywriter for every language who perfectly understands the business context and target audience, and they would write from scratch. But that is impossible. It is too expensive and not scalable. So we write what we call original content, and by original we often mean American English, which is totally not local enough for many countries. Then we translate it and translators have to figure out the transcreation and changes.

Konstantin: I think the biggest impact from GPT will be that we will have a way to move from pivotal content creation, where we pivot through American content to reach the rest of the world, to creating content directly in the language for the audience we want to reach.

Brian: Are you seeing any other trends in neural machine translation?

Konstantin: Today, most machine translation models are built around language pairs. English to Spanish is one model. Spanish to French is another. When you want custom machine translation, you typically have to train it with English as one of the languages. People have been discussing truly multilingual models where there is no difference what languages you use. When you improve it for one language pair, it improves for other languages as well. GPT takes this to the extreme: a single model working across all languages.

Konstantin: The challenge is change management. If you translate content into 20 languages and every translation boosts the model for all other languages, that is the good side. But humans make errors. If you update your model with bad data, you could break not just one language but all of them. Enterprise solutions cannot have that. That is why big solutions are built in a modular fashion so you can control how errors propagate. It will take a few years to put this into real production.

Brian: Tell us about your cultural background.

Konstantin: I grew up in Russia. I moved to the U.S. about six years ago. Berkeley Skydeck invited us into their accelerator, UC Berkeley's startup accelerator. It was a very radical move. We moved in about two weeks.

Brian: Do you miss Russia?

Konstantin: Of course not, because it is a mess now with the invasion of Ukraine. I do enjoy Berkeley. I found the U.S., or at least the California part, to be good enough to live in and feel alive. Some other countries would be too calm and too boring for me. I really like the unique climate here. At some point it felt like the uniqueness of Silicon Valley was fading with the pandemic, but it is coming back. Live activities have become more valuable because there are fewer of them.

Brian: Which languages do you speak?

Konstantin: I speak English, Russian, and a little bit of Spanish. I was also learning Dari when we were planning a mountaineering trip to Afghanistan. I learned a little bit of Inuit when we were travelling to Greenland because back then Inuit was one of the languages where you did not have any machine translation at all. You had to learn some basic phrases. And a little bit of Chinese when we started working with companies from China. But I forgot all of them except Spanish maybe.

Brian: How has learning languages shaped your perspective?

Konstantin: Every language brings a different perspective. Much like when you learn from other people, you understand there may be a different angle than yours and your world becomes richer. Dari was the only language I learned where in the middle of a sentence, the order of words is not very important. What matters is how you start and end the sentence. That helps you understand that the default word order in languages like English or Russian is not the only way.

Konstantin: I did not learn Chinese enough to speak it, but I understood how words and phrases are built. It helped me understand people from that region speaking English because I could see how they construct sentences.

Brian: Learning languages and cultures has helped me become less selfish.

Konstantin: That is totally true because they create a new angle. They enable this outside look at yourself. I remember seeing Americans as very sensitive, almost to the extreme, and not understanding it. Then after I got my Green Card and lived in the U.S. for three years without leaving much, my first trip abroad made me realise I had become that person too. I started having knee-jerk reactions on things I didn't have before. Neither extreme is valid. What you need to do is understand both perspectives and navigate them. To be empathic, you need to understand the cultural background of a person.

Konstantin: After the pandemic, we became truly global. We ditched offices. Now we have people in probably 16 countries. Being empathic when you do not understand someone's cultural background is super important.

Brian: That is a really good point about not being quick to judge.

Konstantin: Especially with GPT and Midjourney and all these tools that can create messages not rooted in the real world. The very existence of a message or an opinion does not make it valid. It may not even be from a real person. It could be a hallucination of an AI. I think humanity is at an interesting point in time. We detached from reality with everything going remote after the pandemic. Now we see that our digital reality can be bent in any way with this new type of AI. I really like the article by Harari about how language and stories are how we humans exist.

Brian: Culture and heritage come from real people. Technology needs to benefit and assist us, not take over or take away the identity we are trying to communicate. Do you have any final advice on how to deliver a truly native experience?

Konstantin: At the end of the day, it is about putting user experience first. In the translation industry, people used to define quality by linguists' opinion. But for end users, it is very different. We have seen many cases where users disagree with the linguist. Going after linguist opinion can take too much time or be too expensive to deliver value to users.

Konstantin: The core of it is going with the user, getting feedback from them, taking user satisfaction as the most important metric of quality. We have had cases where there was some obscure problem and it was not clear what was wrong, but then we got end-user feedback and finally realised the issue. Then you can fix it at scale for all users. It is about being user-centric and prioritising users over industry rules.

Brian: Konstantin, thank you so much. I loved your perspectives and advice.

Konstantin: Thank you, Brian. I loved chatting.

Brian: Konstantin Savinkov, CEO and co-founder of Intento. Remember, always strive for authenticity and embrace the power of native experiences.

‍