How to succeed in science and survive? An interview with Ivan Oseledets.

— From your biography, I learned that you come from a family of mathematicians on both sides. It seems that in such a situation, it would be difficult to become anything other than a mathematician. Nevertheless, if it weren't for mathematics, where would you want and be able to realize yourself?

— I have always aimed towards mathematics. However, I now feel that I am quite good at managing things. So it's a kind of managerial work, but definitely not related to the humanities — that’s for sure. Organizational work suits me well.

— I also read that getting into MIPT was not easy for you. Could you share a bit about that?

— The story goes like this: in 2000, there was a record number of applicants for various reasons, particularly because there was a baby boom in 1983, and I was born right in that year. In the Soviet Union, women were allowed to take maternity leave while retaining their jobs, and everyone started having children. So I ended up being one of the children from this boom. Overall, the competition was extremely high. I was, of course, applying for the mechanics and mathematics faculty at Moscow State University. The admissions were conducted in three waves. I didn’t get in during the first two waves, started thinking about other options, and submitted my documents to Phystech (MIPT) just in case. As a result, I was accepted both to the mechanics and mathematics faculty and to Phystech, and ultimately chose the latter. That’s the whole story.

— If you had indeed enrolled in the mechanics and mathematics faculty, would you have continued working on the same topics you are engaged in now?

— As is known, there is no such thing as a subjunctive mood in history. However, I was very fortunate to end up at Phystech because — if you look at where the guys who went to the mechanics and mathematics faculty ended up — Phystech turned out to be significantly more dynamic, interesting, and offered more challenging problems than MSU. To answer your question: probably not.

— It is also said about you that you are the youngest Doctor of Sciences in your field, having obtained your degree at the age of 29. Is that correct?

— Yes. I defended my dissertation in computational mathematics[1].

— As I understand, you faced some difficulties and resistance during your defense. What advice would you give to modern young researchers who are defending their work and find themselves in a similar situation? How did you cope with it? What qualities did you need?

— It was a typical academic story: if there is resistance, it usually means that what you are doing is either very bad or very good and causes envy. Therefore, if you encounter such resistance, conduct an analysis to see if everything is indeed that bad and if you are doing nonsense. If not, then you can just be happy for yourself and remain calm. Nothing extraordinary. It’s unpleasant, but not fatal.

— Your wife, Ekaterina, also works in computational mathematics, correct? Do you even conduct joint research?

— Not anymore. She is now involved in artificial intelligence, is a professor, and heads a laboratory.

— What is it like when your wife is a co-author of your articles? How much does work infiltrate your family life? Are you able to separate: family here, work there?

— Of course, work does seep into family life. There are both pros and cons to that. Academic families are always a separate story, and this is not uncommon. Well, we have what we have, and overall we live well with it. We don’t complain much. I would like to disconnect from work, but that doesn’t always happen. On the other hand, you can solve some work-related issues at home.

— In which areas are you currently conducting research, and why did you choose these specific topics?

— Right now, I lead a large research group, where I help talented individuals develop, and we work across several areas. There’s a direction related to tensors, which has given the group significant scientific recognition. This involves applying tensor methods to accelerate machine learning techniques, for compression, and for developing new models.

We do a lot of work with efficient training and inference methods (inference refers to the process of using an already trained neural network — editor's note), looking at how to speed them up, compress them, and make them more accurate. We are also working extensively with diffusion models and large language models.

There is another direction related to accelerating classical modeling methods. A lot has been proposed in this area recently, such as hybrid methods. Now, artificial intelligence is penetrating all fields where modeling is necessary, and we strive to develop such algorithms.

In general, the work can be absolutely varied. It’s very hard to say now: I’m working in one direction and will do so for five years. That approach is impossible. Something is always changing. Something new always arises. We have to adapt.

— How well do you manage to follow global trends in these areas?

— We try our best. It’s not for me to judge, but in some aspects, we are showing ourselves quite well. Although there is always a desire and a feeling that we need to do more, a certain dissatisfaction. But we have interesting results, articles are being published, and applied projects are being developed. In this sense, things are quite good.

— Do you think you will continue to work in these areas in the future? Will they develop? Or will they soon reach some saturation limits, necessitating a switch?

— I believe, of course, that everything will continue to develop. So far, there’s no visible limit. Development is very rapid and somewhat unusual. But again, there’s nothing terrible about it. Saturation will not occur.

— What other problems would you like to work on in the foreseeable future?

— Right now, we have started working on generative design. That is, applying generative modeling for automating processes in construction, in design. For example, how to design electrical wiring with the push of a button? This is a highly demanded and very necessary thing. The only problem is that no one really knows how to do it. It’s always interesting to engage in something like this.

— How do you think the domestic school of computational artificial intelligence is currently strong?

— Overall, we are very smart and resourceful. We just don’t always have enough GPUs (GPU, or graphics processing unit, supports parallel computations and is therefore in high demand in machine learning. — editor's note). But all the people who grow up here, if they don’t stay here, find jobs in very decent places in the West. This means that our workforce has a very high intellectual potential. In some areas, we lack systematic approaches to problem-solving and a deep interpretation, but we are not inferior to anyone in the world.

— Which Russian or Soviet scientists could you name as the founding fathers of the field, particularly artificial intelligence?

— That’s a tricky question. If you allow me, I’ll think about it. Domestic scientists, the founding fathers of artificial intelligence?

— Let’s rephrase that. Who would you name as your teachers?

— Among the well-known figures, of course, there’s Vladimir Vapnik and Alexey Chervonenkis. I would also mention Leonid Vassershtein, also one of the Soviet scientists. Leonid Kantorovich made significant contributions to the theory of optimal transport, which is now used in machine learning. If we go a bit further, there’s Andrey Kolmogorov.

We had a whole school of artificial intelligence, but it was specific and closed. And there were very good results. For example, the same “Buran” — completely automatic landing in 1988, we are only now getting close to something similar. Unfortunately, I don’t know the details about who worked on it and what was going on inside. Only now are we managing to access some information, even historical questions, because we had no such specialty as “computer science,” let alone artificial intelligence, for a long time. Even basic algorithms and so forth were always somewhere on the fringes.

But if we talk about fundamental science, I think I’ve mentioned enough names.

— Is there anything that has happened in Russian artificial intelligence research in recent years that the audience should definitely know about?