“We are like the handsome guys in the forefront”: how the AI ​​startup Toloka, which broke away from Yandex, lives

On the channel “This is Ossetian!” An interview was published with the founder and CEO of Toloka Olga Megorskaya. Toloka is one of the IT companies that are part of Arkady Volozh’s Nebius Group, the same international part of Yandex that, after a long division of assets, remained with Volozh and focused on AI infrastructure.

Toloka began as an internal Yandex service that tagged data, and today it is an independent international company that has attracted more than $70 million in investments, including from Jeff Bezos himself.

In the interview, Olga tells her personal story for the first time. At the age of 18, she was seriously injured, learned to walk again, and for a long time could not even leave the house. It was then that she began working as an assessor at Yandex, assessing the relevance of search results. Over time, this area grew into a separate product, and then into an international company, with which Microsoft, Anthropic and other tech giants now work.

Toloka does what is usually hidden “under the hood” of AI: it collects and labels human data for training neural networks. This is the same work that a person does and without which neither ChatGPT nor AI search would exist.

We talked about how Toloka appeared, why artificial intelligence is still impossible without people, how to pitch a startup to Bezos himself, how to manage thousands of employees using mathematics, how the company exited the Russian business, why Megorskaya does not believe in superintelligence and whether Arkady Volozh mentors her.

We are publishing excerpts from the interview, but watch it in its entirety here .

What is Toloka

— Olya, thank you for agreeing. I have a first and very important question: is it still “TOloka”, “Toloka” or “Toloka”? I listened to several videos, and the pronunciation is different everywhere.

- Yes, this is a great question. Almost any conversation about TolOka begins with this.

- TolOka?

— Yes, that’s right, TolOka. Then we have professional jargon: internally we often say “TolokA”, because the word itself is originally Belarusian, it was invented as the name of the service by a team from Minsk. And in Belarusian it sounds closer to “TolokA”. Therefore, the correct name is TolOka. But sometimes we get confused among ourselves at TolokA.

The word is from Slavic languages, although it is less common in Russian. It means a tradition when people come together to do a big job that one cannot do alone: ​​harvest a crop, build a bridge, and so on. This was called “toloka” (or in Belarusian “toloka”). It seemed to us that this was a very accurate metaphor, and we took it as the title.

— What does Toloka do? As far as I understand, this has changed in 10 years.

- I'll try to explain. I'll start with the basics. How does machine learning work? You take a large set of examples, show it to the machine, and it forms its own logic based on them. And then applies this logic to new cases. And here the key question arises: where to get the training data? Sometimes they already exist. For example, if you are building a model of credit trust in a borrower at a bank, you collect a large data set of facts: this person took out a loan and paid it back, this person did not pay it back. This is a training data set; no additional markup is required. Or, for example, when a model was taught to play Go: two models play with each other, and this is how a large synthetic data set is generated: which moves lead to victory and which ones lead to loss.

But often this is not enough. Because AI aims to automate human judgments and actions. And when you need human judgments, there is nowhere to get them - you need to directly take people and ask them to mark them. Some of the first problems solved in this way were related to web search. You have a billion documents and a search query, and you need to not only find relevant documents, but sort them so that the most relevant ones are in the top. The user doesn’t look beyond the first five lines. How do you know which are the “best”? The only way is to train the algorithms on a large number of examples: show people the request and the document and ask whether it is suitable or not. For a long time this could not be automated. The same is with driverless cars: in order for them to understand where the traffic light is, where the pedestrian is, and where the bicycle is, it was necessary to ask a person to circle them in the picture.

And here is the great irony of artificial intelligence, which is rarely talked about: under the hood of all artificial intelligence there are literally hundreds of thousands of people working who share their human expertise in order to train this very AI.

- Still?

- Still.

“I remember there used to be cars with sensors, and there were people in them...

— Cars drive around, collecting images from lidars. Naturally, there is not a person sitting inside the car who is quickly doing something, but an algorithm is already working that recognizes objects in real time. But in order to train him, it is necessary for the people at the computer to feed him a sufficient number of examples in advance so that he can learn from them.

— So, all this happened manually?

— Training data is collected manually. Then the algorithm is trained on them, and it works on its own. This doesn't mean that every task is solved with human help, but that each is trained on a data set that would be impossible to collect without human input. And the question “still or not?” - this is just about the evolution of AI and the role of human markup in it.

- Let's go into history a little. When did it all start? About ten years ago?

— If we talk about search and ranking, around 2007–2008, both Google and Yandex began to actively use machine learning. And since about 2009, data tagging has become a separate, understandable task.

First it was search, then computer vision - receipt recognition, and so on. Then new products appeared. But the principle is the same everywhere. If you want to train a voice assistant, you need thousands of hours of speech recorded by voice and transcribed into text and vice versa. Drones - we have already discussed. Almost any application of machine learning that implies that you are automating some actions that were previously done by a person, you need a person in this place to train and show examples on which she would learn.

— So Toloka was born as an internal service?

- Yes. It originated within Yandex as a tool for tagging data for search. Gradually we began to support other services. And there was always the same task - scaling. There is never enough data. And the more there are and the better quality they are, the predictably better the final algorithm becomes.

- But this was an internal department of Yandex. At what point did it become a separate entity?

— We started inside. And we read quite a lot of articles, then wrote our own, developed new approaches - simply because no one had systematically dealt with this before us, and it was very interesting. At some point, we asked ourselves the question: how to control the quality of people's answers when they are performing quite complex tasks? Is there a difference between just testing them with test questions versus just testing them by asking a few people and getting their consensus opinion. What if people know they are only being tested by consensus? What if they don't know? How many people are needed and so on?

We conducted experiments: we gathered people into groups of two or three people and recorded how they exchanged information. As a result, they formed a framework that began to be used in data markup for search - it provided higher quality for tasks of increasing complexity.

In general, we constantly solved the problem of scaling: there is not enough data, and so are people. And any human operations are always a bottleneck; it is always slower than any automation. At the same time, we have accumulated a lot of internal expertise: mathematics, methods of aggregating estimates, quality control and everything else. We packaged all this into a product - Toloka - and made it an open platform. This was an important point: anyone can register as a performer and anyone as a customer, post tasks, use our quality control methods and collect data sets on their own.

- But for money?

- Yes. But it was an important step in terms of democratization of the industry. Before this, the examination was carried out internally. If someone outside needed a data set, he came to us, and we did it through our assessors.

When we opened the platform, we saw sharp growth. On the one hand, a huge number of people came to us - up to 10 million registered users.

- Are these just assessors?

- Yes, we called them tolokers. And, on the other hand, many customers came. We saw that people themselves come up with very cool approaches to data collection. Our clients were ML engineers, researchers - those who needed training data.

I remember how I used to say: “I’m doing Yandex Toloka,” and no one understood what it was. And then at conferences you ask: “Who knows Toloka?” - and almost all hands raise. It turned out that most ML specialists in the industry started using it at some point. In fact, it was the only available source of human markup.

— Where did these ML specialists work? In companies?

- Yes. We are now talking primarily about the Russian market, but, of course, not only Yandex was involved in machine learning. There were other teams that also needed data, and they started using Toloka. At the same time, we traveled a lot to international conferences. And it is important that these were not commercial B2B events, but scientific conferences where articles were published and technology development was discussed.

— Let me translate it into human terms: web summits are more about marketing.

- Yes.

- And here - aimed at research.

- Exactly. Very narrow audience. But the international community of AI researchers is actually quite compact, and Toloka quickly became known outside of Russia.

— Did you collect the data manually in Russia?

— People were from different countries.

— India, Türkiye and so on?

- Yes, from anywhere.

— So this is crowdsourcing?

- Yes. The idea of ​​crowdsourcing is that you give access to the system to almost any person with a minimum entry threshold. And then it's like controlling wind energy: everyone makes a minimal effort. Someone came in for 15 minutes, completed a simple task—like tracing an object in an image—and left.

Your task is to turn these small impulses into useful results with the help of quality control systems. And customers, coming to Toloka, built data collection processes for their tasks using our tools.

The question was how Toloka set out on its own. We started inside Yandex. Then, as an assessor service, we accumulated expertise on how to label data for machine learning. Then they launched Toloka as an open platform and gave it access to everyone who wanted to use and collect data sets. Then product-led growth worked: people from the bottom up began to use it themselves and brought the tool to their companies. Plus, Yandex developers moved to other, including foreign, companies and continued to work with Toloka there.

This is how we began to expand into the international market. By 2021, revenue from external clients (outside Yandex) was almost equal to internal revenue, with a significant portion coming from international clients, including big tech companies from America.

After this, Toloka was separated into a separate business unit. And when the division of Yandex began, the next logical step was the transition to the Nebius group of companies.

— Have you separated into a separate structure?

- Yes.

— A separate legal entity that first belonged to Yandex shareholders, and then to a new structure after the division?

- Yes, that's right.

Personal story

- It's time to talk about personal things. There aren't many women in this industry, let's just say. In principle, there are few women IT specialists, and there are also few women with technical education. Tell us, what is your background and education?

— I studied at St. Petersburg University, at the Faculty of Economics, majoring in mathematical methods and models in economics.

— We could be colleagues.

- You too?

- You know, I didn’t master mathematical methods because I became interested in journalism. These two disciplines came into temporary conflict. But why the economy?

— On the one hand, I always liked mathematics. But on the other hand, I didn’t want to go completely into some kind of hardcore mathematics, closed in on itself. I was interested in the application of mathematics in various aspects. The very concept of a mathematical model in economics was interesting to me, because it’s about how to try to formalize and package any incomprehensible garbage happening in the world into a mathematical model. This is probably why I chose this specialty.

—Are you a hereditary technician?

- Yes. My parents are probably even more techies than I am. I'd rather be a little...

- A renegade?

- I gave up, yes.

— Are your parents engineers?

- Yes, both are engineers. Dad worked at a research institute, mom was probably one of the first generations of real programmers, back in the 70s.

- How did you get carried away? After all, mathematical modeling is either economics as a science, or finance, investment, or something like that. This is much more logical than suddenly ending up in language models and Yandex.

— Yes, the path was such that after university I went to work in my specialty. At that time there was an investment bank, Keith Finance, where I dealt with risks and built econometric models. She worked there for probably about a year. But at some point I realized that this is still not the world to which I want to belong.

- Why?

- Don't know. At first it was very interesting. At that time, the bank was quite advanced; they introduced new technologies. I remember that they invited a consultant from Barclays who talked about how to build risk management processes. And I was just working with him, because suddenly it turned out that I understood mathematics and knew English. In that environment it was a rather rare combination. It was interesting. But then the financial crisis happened, everything began to collapse...

— The bank, in fact, went bankrupt in 2009.

- Yes, yes. In 2008, when everything started to fall apart, there was no time for consultants or anything else. And overall I thought that I didn’t have a soul match with this industry. Then I wrote to the guys in Yandex. By that time, I had already worked part-time as an assessor. She said: “I want to be a search analyst, take me.” They answered me: “We are not looking for analysts now, but we need a person who will deal with assessors.” Because, as we said, no one particularly likes to do this. That's how I ended up in Yandex.

For me it was some kind of magical place. He had a vibe like the Strugatskys in “Monday Begins on Saturday”: people do some incomprehensible, amazing, almost magical things. It was a world that I wanted to belong to, that was interesting to understand. That's why I went there.

— Why did you start working as an assessor in the first place?

— I have a non-trivial personal story. I can probably tell you. I rarely talk about this, because it’s not always clear in what language to tell it so as not to sound too melodramatic, but also not to devalue this experience.

I was an ordinary person: I studied at the university, went skiing, went hiking, and led the ordinary life of a young man. And when I was 18 years old, an accident occurred: I fell from a great height and broke my spine. Very strong, catastrophic. When I found myself on the ground and came to my senses, I realized that I was completely paralyzed and could not feel my body at all. It's like it's separated from you.

But as unlucky as I was to find myself in this situation, I was just as lucky to receive all the support that could be received. The doctors performed an absolutely amazing, unique operation. In fact, my entire back was rebuilt from titanium, because several vertebrae were simply missing. I had an amazing rehabilitation doctor who worked with me for several years - literally every day for several hours. Essentially, I learned again not only to walk, but to do everything in general. My friends helped me a lot, my family completely restructured my life to support me at that moment. Even professors from the university came to my home to take exams so that I could complete the session that I did not manage to pass.

In general, the whole world helped me as best it could. And that’s probably why everything ended well in the end. This support has been incredibly important. This is, of course, a very interesting experience. It seems to me that I haven’t talked about this at all for twenty years. And now, the older I get, the more it seems to me that this is important. Because we sometimes get locked into our own bubble - AI, technology, investment, whatever. But the world is not limited to this, and life develops differently.

This story brought me into contact with people I would never have otherwise encountered in my life. If she had continued to live in her prosperous bubble, she would have stayed there. And here next to you, for example, is a boy programmer who was hit on the head in the entrance, his phone was taken away and half of his skull was blown off. Or an outstanding architect who suffered a stroke. Or a dwarf woman - very cheerful, noisy, with beautiful pillows, because without them she cannot reach. We even went to a special driving school together and learned to drive an Oka.

- Because she is tame?

- Yes, yes. There are an incredible number of levers that need to be pulled at the same time. To be honest, I never learned to drive such a car. And this woman needed several pillows to even reach the steering wheel. At the same time, she was a very interesting, intellectually developed person, but it was objectively difficult for her to find a job. Or a teenage boy with cerebral palsy - an absolutely bright mind, locked in a body that does not listen to him.

This is a very important experience for me. Because it quickly became clear: you can get a lot of help in terms of treatment and physical rehabilitation; society knows how to help in this sense. But that's only half the equation. The second half begins later, when classes end, the guests leave, and you are left alone with yourself and your situation. And it's really hard to deal with. These are not made-up stories of people being physically rehabilitated and then using their newfound control over their bodies to walk out the window. And at some point I just realized that free time kills, you need to keep yourself busy with something constantly so as not to sit idle.

I started looking for part-time jobs available to a person who is paralyzed and confined to an apartment. I translated texts, wrote articles, tutored mathematics, and did something else. To be honest, it was all pretty boring. But I didn’t have much choice - I still had to do something. And then they told me that Yandex was recruiting search assessors. The word itself seemed completely wild back then: “assessor” - what is it anyway? But I thought: okay, I'll try. And it turned out that it was interesting.

What did the Yandex assessors do then? They tagged search queries - they trained a ranking model to distinguish good search results from bad ones. And when you look at search queries, the whole world opens up before you. For several years you sit locked in a room - and suddenly everything passes before your eyes. I’ve probably never been such a knowledgeable person as I was at that moment when I was tagging Yandex search queries. And you need to constantly make small decisions, but you need to think about each one.

This story helped me personally a lot - to gain self-confidence, start earning money, and stop feeling completely dependent on others. Therefore, when I had already recovered, this part remained very important to me. For me, for example, it is important that I definitely know several people who found themselves in difficult life situations and whom I helped get settled - people who otherwise might not have found a way to realize themselves.

Nor is this a story that has ended. Because, as we have already discussed, the demands on experts are constantly growing, but people’s lives are not getting any easier. Many left for other countries with professorial degrees and lost their jobs there. And through projects in Toloka, they had the opportunity not to carry boxes, but to monetize their expertise.

“It seems to me that there is no melodrama here.” This is a super inspiring story and I think it will really help a lot of people. In general, when we lose our temper and share such things, others feel better. Because there is a feeling: “I am not alone.” Even if nothing like this has happened to you yourself, it still gives you a great surge of strength. And you think that everything can be overcome.

Profitability, investment, Bezos

— Tell me about your role. You came as a manager of a direction, but at the same time you are a founder, this is your brainchild. Usually it’s like this: the founder starts a company, he has all the ownership, and then he shares it with investors. What about here?

- It's more of a combination. I’m probably not ready to go into details right now...

- Share?

- Yes. But in general, the structure reflects both the fact that it is the brainchild of a team and the fact that a significant part of it belongs to shareholders.

— Nebius shareholders, right? So you are both a shareholder and a director?

- Yes.

— You said that the company had revenue from Yandex as an internal customer and from external clients. Can you tell us a little about finances? How much does Toloka earn and is it a profitable business?

— We do not disclose the amount of revenue at the moment. At some point we published orders of magnitude totaling tens of millions of dollars.

- Per year?

- Yes. This is the state we are in now. The market is growing quite strongly because the demand for data is not going away.

“It’s growing,” she sighed heavily.

- Yes.

- Is there any profit? Or not yet?

— We haven’t fully achieved positive results yet.

— So, what are investments needed for?

— For further product development and expansion. We have now made great progress in food history. It was important for us to attract additional strategic partners for the sake of value and market understanding. That’s why we entered the round with additional investments, including from Jeff Bezos.

— I looked at Crunchbase, and I was surprised that among your investors is [CTO of the Canadian marketplace Shopify, former member of the board of directors of Yandex Mikhail] Parakhin. This looks like an angel investment. Can you tell us about it and about the investment itself? Was it before Amazon?

- No, it was the same round. The family office of Jeff Bezos, Nebius and Mikhail Parakhin took part in it. Mikhail was the CTO of Yandex. When we were working on Toloka, we provided all Yandex projects with data markup. More broadly, there are not many people in the world who deeply understand AI. If you single out those who understand the essence, complexity and importance of data in AI, there will still be some subset left. And if we further highlight those who are truly interested in the topic of data tagging by people, they can be counted on one hand. Mikhail is one of them. And it was important for me to attract like-minded people to the company who look at problems in a similar way and want to develop in the same direction. Mikhail is probably No. 1 of possible such people in the whole world. And so I think that we were very lucky to attract him to participate in Toloka.

— Did Nebius also invest?

- Yes, sure.

— And Bezos’ family office?

- Yes.

— Tell us how you found Bezos’ family office. What's the story behind this?

— We had a common understanding with Nebius that investments are not needed for the sake of investments as such, but for the sake of an additional impetus for development. So we were looking for investors who could really help with this. Both Mikhail and Bezos’ family office are just such examples.

- How did you two meet? Did someone contact you?

— Mikhail introduced us. And then the discussions began. It was quite an interesting process because Amazon has a famous theme called the 6-pager. Do you know about this?

- No, I don’t know at all.

“It’s an important part of Amazon’s culture. Before each meeting, a text document is written. In it you describe what you want to discuss, what to get, what you expect from this meeting, and so on.

— So, 6 slides?

— These are not slides, but a text document. They came up with this, probably 20 years ago. The one who organized this meeting brings the document. They often sit up straight and spend the first 5 minutes of the meeting silently reading this document. For everyone else, it seems a little creepy.

- Yes, it's strange.

“But everyone definitely respected the essence, agreed on what they really wanted to discuss at this meeting, and then made some decisions and moved on.” This is a rather specific “Amazon” topic.

— What was in your 6-pager?

— We talked about our business, about technologies for managing people’s efforts at scale, and about where this is developing. We were just at the transition: we started with simple tasks for 10 million people from the crowd, and then began to work with highly professional people - with academic degrees and so on. But, regardless of the level, people remain people - and these are not the most reliable agents.

— People are always a problem.

- Yes. They don't do exactly what you ask. They don’t read the instructions; to be honest, I rarely read them either. And this is always a big problem. I thought that when we moved to highly qualified experts, it would become easier.

- It seems to me, on the contrary, worse.

- So it turned out that no. And we discussed this at the meeting. And Bezos said, “Exactly. That’s why I make everyone read the documents right at the meeting - because no one reads them in advance.”

- Wait, did he say that himself?

- Yes, it was a conversation with him.

— So you had Bezos at the meeting?

— Yes, via Zoom, but yes.

- What impression did he make?

- Very good. In this format, you first of all feel that you are being appreciated. And you can judge by the questions they ask you.

“But you still read something.”

“It was important to me that the questions Bezos asked were about exactly what was important to me. It doesn't always happen this way. We talked about Toloka, about technology, and most of all we discussed the technological part - managing people. This really resonated with me because that’s the point.

— So he personally makes investment decisions for his family office? Not a committee?

— As far as I understand, yes.

- And what is he looking at? For the future of AI, for the market?

“It’s probably better to ask him.” But it seems to me that we have the same vision of two things: the problem of data quality and the potential that is hidden behind it. An industry that does this most unloved and dirty work. Still, they try to distance themselves: “Fu-fu, people, I won’t touch this, I’ll give it to someone else.” By solving this particular problem, we actually train muscles in ourselves that are not trained anywhere else in the industry.

— So technology companies are trying to minimize this part?

— AI is a technology industry, largely about introverts who want to minimize human interaction. But at the same time, everyone understands that there is nowhere without this.

- Evil.

- Yes. And there are a very small number of players who see engineering beauty in this.

To me, managing human effort is a purely mathematical matter. Very interesting. A person is a much more complex system than any algorithm. He has multi-level motivation that you don't know, it's a black box. But you need to build a system so that his contribution produces a useful result.

— How much did you raise in the end?

— We talked publicly about $72 million.

- What about the rating?

- We won’t reveal it.

- Well, does it feel like hundreds of millions?

- Maybe.

“But you’re not a Unicorn yet?”

— We did not have a public assessment.

Departure from Russia and exit from the Russian “Yandex”

— I understand that you lived in Israel for some time. Why?

— We first moved to Israel because we focused on the international development of Toloka, and it became impossible to continue this from Russia. As part of the Yandex section, we moved as a team, and quite a large part of it ended up in Tel Aviv, where there was an office.

- He still exists.

- Yes. It was a natural step. We didn't live there long - maybe about eight months. Then we moved to Amsterdam because we initially planned to go to Europe.

— How did you make the decision to leave? As I understand it, within Yandex, everyone decided for themselves where to stay.

“For me, this was not some kind of separate moment of choice. This was a continuation of the path we had already taken. We built Toloka, made it a separate business unit, brought it to a state where a significant part of the revenue came from Western clients, and focused on international development. Therefore, the move was a consequence of the same path. If there was a point of choice, it would have been much earlier.

— That is, until 2022?

- Yes. It was a choice to build a world-class project. And all other steps are a consequence of this.

— But at the same time, Yandex still needed data marking.

- Certainly.

- And what happened to this? Some kind of curly cutting?

— Toloka was part of the key infrastructure of Yandex, and it was important for us that nothing fell anywhere. Therefore, we spent quite a long time doing this “curly cutting”: it was necessary to preserve the infrastructure inside Yandex and at the same time move the business function outside.

— Was it a conflict process?

- No, not conflicting. It was a work process. Another thing is that it was difficult for us as a business.

- Why?

— The specificity of Toloka is that we are based on a marketplace of people. And when you “cut” it, it suffers disproportionately more than a regular product.

— So the connections are being broken?

- Yes. Connections are being broken, the network effect of the marketplace is being disrupted. The number of available performers sharply decreases, there is less data on the basis of which you match them, rank them, and so on. Overall it was a technologically complex process. But we brought it to the end.

— It seems that most of your revenue was foreign, and most of the performers were from Russia.

- At some point this really was the case.

— And you had to put it all back together?

- Yes. And not only the performers, but also part of the proceeds.

- And where did you get it?

— In a sense, we were lucky: this coincided with the beginning of the GenAI revolution, when the need for low-skilled mass crowdsourcers became noticeably less, and the need for high-quality experts became noticeably greater. But we had almost none of them, so we still started from scratch. And perhaps it was even better to do it from scratch than to try to divide the old system.

— Did you transport anyone involved in marking?

- No.

— Only company employees?

- Yes.

— By the way, how many people do you have now?

- About a hundred.

— And these are mostly Russian speakers?

— Most of it, yes, but not all. Approximately 40% are non-Russian speakers.

— That is, about 60% are Russian-speaking?

- Yes.

— And these are mostly former Yandex employees?

- Yes.

— Have they relocated in 2022?

- Yes.

- How did this happen? Were they given relocation packages?

— This is the heroic work of the HR team, to be honest.

-Where is the team now?

— We have two main locations: Amsterdam and Belgrade. There are a little more people in Belgrade, a little less in Amsterdam. Plus there are distributed employees in different countries.

Disputed clients

— Listen, I can’t help but ask a question about some of your past clients. As far as I understand, the platform is open, and almost any organization, even the Pentagon, can place orders on it.

— Now we have certain rules and KYC processes. That is, it is not true that anyone can just come and immediately start posting tasks, there is a verification process. But overall, yes, it is an open platform.

— There was a story that Roskomnadzor placed orders with you. And also - that some publishing house tried to identify literature “promoting LGBT” through you.

— To be honest, I don’t know about the second one. About the first one, the situation there was a little different. We have never had commercial relations with Roskomnadzor. As far as I remember, we had a so-called sandbox - a “sandbox”, where you could simply test how the platform worked. There was no access to performers, no payments. And, as far as I understand, they tested their approaches there.

— Did you have KYC after?

— It is constantly evolving. We are updating it to comply with legal requirements and restrictions.

- Listen, but these are not legal restrictions. The next story I wanted to ask about was two companies that were involved in facial recognition. Well, what about an IT startup? The question is that his customers are, for example, the Russian government, which recognizes faces. But it may not be the Russian government, but the government of Venezuela or North Korea that uses LLM to fight the opposition. Where is the line when, roughly speaking, the order needs to be dropped? Or is this a commercial, normal commercial story?

— For us, the main border is legal grounds. We conduct fairly in-depth KYC checks: we look at the company itself, its structure, and its subcontractors. If we are talking about sanctioned companies or countries, such orders are not allowed.

— That is, the restrictions are primarily in the legal field: if the customer is under sanctions, you do not work with him?

- Yes, that's right.

- And you check this?

- Yes.

— But they started actively checking after they got into trouble?

— There have always been checks, it’s just that new parameters were added with sanctions.

— Are there ethical boundaries?

— Since the platform is open, markers themselves choose what tasks to do. And sometimes they themselves signal: “Something here looks suspicious or unethical.” We check such cases. If there is a violation of our terms and conditions, we remove the client from the system.

But in general, I think it is important to rely on formalized criteria. Otherwise it is very difficult to draw boundaries. At the same time, the final decision still remains with people: they themselves choose whether they want to complete the task or not.

- But people have different beliefs.

- Certainly.

- Some are ok, some are not. It’s the same as with Oppenheimer’s story: the task was scientific, ambitious, and the result was a weapon of mass destruction.

- This is a big philosophical question in general - and about AI as well. Just like with nuclear energy, you can’t say for sure whether it’s good or bad.

Watch the full interview here .