'I’m Used to Producing Distilled Knowledge'

Ivan Rubachev works in a HSE University laboratory established jointly with Yandex Research, where he focuses on machine learning with tabular data. In this interview with the HSE Young Scientists project, he discusses why following a vibe can be better than goal-setting, explains the concept of the Neural Turing Machine, and argues why withholding scientific knowledge is counterproductive.

How I Started in Science

In some ways, it was accidental; in others, it wasn’t. I never felt that I should become a scientist or a programmer because it was lucrative or important at the time. I was simply listening to myself.

Back in school, when I was preparing to enrol in the Faculty of Computer Science at HSE University, I watched interviews and listened to podcasts in which people from companies like Yandex talked about ML (machine learning). I didn’t really know what it was, but I was curious. I also liked the overall vibe of the Faculty of Computer Science, so I decided to apply there.

Then, as a student, I did an internship at Yandex.Maps, where I focused on ML and 3D. In my fourth year, I was offered a job there and worked on developing a map data warehouse. There were opportunities for career growth in development, but I didn’t like how that prospect made me feel. At the same time, I truly enjoyed being at university, and that made me realise that maybe I was drawn to science.

I attended a scientific seminar organised by researchers from the Bayesian group, where I learned how to read scientific papers. We were deeply immersed in the subject and followed developments at the very frontier of science. Today, AI and ML are on everyone’s lips, but at those seminars back then, we were discussing papers considered pivotal in 2020.

Towards the end of my studies, many of my fellow students plunged headlong into their jobs and completed their university courses with only the minimum required performance. I, however, attended all seminars and considered staying in academia. I had enjoyed being part of this community from the very beginning, and that feeling only grew stronger over time.

What My Term Papers Were About

In those golden days, ML and deep learning research did not yet demand large amounts of resources. The concept of ‘scaling’ had not yet become popular—the idea that resources should be poured into expanding existing neural networks rather than into basic research and exploration of new directions.

I came across a DeepMind article about simulating computers using neural networks. It sounded strange, since neural networks themselves run on computers.

However, in theoretical computer science—and in computer science more broadly—there are models such as Turing machines, or just computer models with memory. In these models, you can write data to a memory cell, read from it, and thereby execute any program. It is therefore possible to try to simulate the execution of arbitrary programs within a neural network. In the articles, this approach was referred to as the ‘Neural Turing Machine.’

I experimented with these machines and enjoyed it greatly. I read follow-up articles from Google DeepMind on the topic and, in my own toy experiments, tried to teach a neural network to reverse a string. There was no practical purpose behind this work—I was driven purely by scientific curiosity. It was fun and quite challenging, and I succeeded, eventually writing a term paper on this topic in my fourth year.

Today, our field has a neural network architecture known as the Transformer. What I was doing with Neural Turing Machines was very close to Transformers, although I didn’t realise it at the time. In retrospect, it’s amusing to see that I was conceptually close to ideas that later became central to the field—although I couldn’t yet see their significance, I genuinely enjoyed the process of working on them. Perhaps this is when I realised that science would be fun for me. I enjoy reading papers and digging into methods, trying to bring the ideas described in articles to life and make them work.

How I Got into Yandex Research

I applied and was accepted. So I actually moved from Yandex to Yandex Lab at HSE University. My gut told me it was a good decision, and the vibe felt right. Still, I spent about a year searching for my field of study. In my third year, I trained neural networks to play Five-in-a-Row (similar to Tic-Tac-Toe, but more complex), and in my fourth year, I experimented with Neural Turing Machines. After completing my bachelor’s, I joined Yandex Research to work on deep learning and ML in a more scientific sense. But which topic should I focus on, and how should I choose an academic supervisor?

The laboratory worked on a variety of tasks—image generation, predicting uncertainty levels, and sharing computing resources to run the huge neural networks that were just coming at us. Similar projects had already been carried out in other fields: volunteers used their devices to analyse genomes or processed sky images on networks of home computers.

I chose image generation. I approached Artem Babenko, Laboratory Head at the FCS and my future academic supervisor, and we began experimenting together. At that time, Yandex Research was publishing exciting articles on image editing using the generative models that were popular at the time, so naturally, I was also involved. However, I couldn’t claim much success, and then COVID hit.

When I was about to join Yandex Research, I expected it to be incredibly exciting. My expectations were met—and that scared me. I worried about whether I was good enough. This fear led to procrastination and made it difficult to work and function normally, a situation worsened by the fact that all communication was happening over Zoom. But when I finally went to the office and started working on something in person, everything improved. My fear of people faded, and a sense of belonging began to take shape.

The Focus of My Research

Tabular data. It’s much harder to make this topic sound exciting compared to image generation. You can’t exactly call it 'sexy'—it’s all about numbers, in every sense. And it’s about ML. We’re all involved in deep learning these days, myself included.

But machine learning—which includes deep learning, among other approaches—covers a much broader range of tasks and methods. That’s what initially attracted me to the FCS. It’s about creating programs that can improve by learning from data. Using ML, for example, you can classify people applying for a bank loan, predict whether a user will like an advertisement, or forecast movements in the stock market.

When solving problems, we work with tables containing many numbers, along with a target variable that we aim to predict. One of my students remarked that, in the past, the only economically useful ML products came from classical solutions to such problems—but now, about half of the products are based on this approach, while the other half involve LLMs. The main output of our research is scientific articles, but these articles often include code, since we are practice-oriented.

However, my favourite articles are those that contain ideas with universal applicability. Readers can understand the concept, write their own code, and have it work for them. In this sense, we are less like developers—we don’t create machine learning libraries—we are researchers. Our 'product' consists of articles, accompanied by code demonstrating that our experimental findings truly work and can be applied by others anywhere.

My Opinion on the Commercialisation of Science

We publish the results of our research openly. Personally, I don’t see any benefit in withholding scientific knowledge or technological know-how, although this is becoming the norm in our field. Much of the current development is happening thanks to China, which is not hiding anything. Russia also has many open publications. However, the frontier of science related to large language models is much more closed. OpenAI and Google DeepMind are not like Alibaba. The founder of DeepSeek writes articles and posts them on arXiv. This approach resonates with me much more.

Yandex’s open-source CatBoost library is used by a vast number of scientists worldwide and consistently ranks among the most cited in articles. Its popularity extends beyond academia: my students use it to predict stocks and market movements in high-frequency trading, and they are very enthusiastic about it.

In English-speaking circles, people often talk about a 'moat'—literally, like the moat around a castle. The idea is to identify what protects the earning opportunities of scientists and businesses. In my view, the moat isn’t about hiding knowledge. You can do useful work and remain open at the same time. And the fact that I can have an impact through openness at the point where I am now is encouraging.

How the Russian and European Approaches in Our Field Differ

Recently, at a conference in Copenhagen, I attended an AI for Tabular Data workshop and had extensive conversations with European professors. Our 'products' are the same: the knowledge shared in our articles and the accompanying working code. But in my view, our approaches to research and the types of end users we focus on are different.

We ended up having a long debate. They argued that our output is an article and a piece of code. We pointed out that their output was more like: 'install our library, click a button, and machine learning will work for you.'

While reading their articles, I constantly sense this difference. I’m not used to it. I’m used to producing distilled knowledge—something transferable and applicable. In my world, people take ideas further and work through them. In their world, a lot of resources are spent on things that are useful and important, but there is much less emphasis on the underlying science.

What I Take Pride In

The fact that over five years, our small team of about five people has made a visible impact in our subfield. We’ve been noticed—our papers are read, and we are recognised at conferences. I’m proud that, together with our European colleagues, we have formed a part of the scientific community. And tabular data, once an almost non-existent and rather obscure topic that only occasionally featured at conferences, has taken shape as something truly exciting.

Startups have begun to emerge in our field, and I can feel growing interest from people at various institutions. They tell me, 'You’re working on a hype topic.' That’s surprising to hear, because when we submit papers to conferences, there still isn’t a 'Tabular Data' category in the 'Research Area' field. We usually have to select 'Other' or simply 'General Machine Learning.'

I plan to write a post about it but haven’t gotten around to it yet. It’s only a matter of time, and we need to keep drawing attention to our topic. There is already a lot to be proud of—but in my view, this is far from the end; it’s only the beginning.

The Subject of My Dissertation

Some scientists put forward bold, highly unexpected theories. Others, building on theory, produce remarkably sophisticated technical implementations. And still others clear the forest, so to speak, by doing meticulous work in their fields—carefully comparing methods and systematising knowledge.

This kind of work doesn’t inspire everyone—far from it—but it does inspire me. Of course, I’d love to create something state-of-the-art and write a paper that overturns everything we thought we knew. But in practice, my work often comes down to carefully arranging cubes into boxes. It produces new knowledge and real value, but it’s a different kind of impact.

My dissertation is largely based on one of our most recent conference papers. In it, we propose a new benchmark for the field and highlight aspects that researchers like us—working on tabular data in an academic context—often find lacking in many settings.

Working with data can be monotonous and is sometimes dismissed as trivial or unnecessary. Yet, in 2025—precisely because the area seemed unattractive—we were writing papers based on income data from the 1998 US census, as it was the best and most recent dataset available in the public domain.

It was both gratifying and sad to hear at the Copenhagen conference that when people need a large dataset, our benchmark is often the only one they find. We have a tremendous opportunity, thanks to the many big tech companies that can take their domains with useful ML applications and release datasets. Recently, the team at Yandex.Music posted a massive dataset containing billions of anonymous interactions between users and tracks. Scientists around the world can now use it to improve recommendation algorithms.

My Dream

To live a happy life.

Who I Would Like to Meet

At first, I thought it would be amazing to meet the Einstein of our field—someone like David Huffman or John von Neumann. But then I realised I would probably be too intimidated to ask them anything intelligent. Recently, my wife and I read two autobiographies: I was reading Sakharov’s, and she was reading Kapitsa’s. I would truly love to meet both of them—if they were willing to indulge my questions and have conversations not only about science, but about life.

A Typical Day for Me

It depends on the time of year. When a conference deadline is approaching, my days are usually spent running experiments and trying to pull the story together. It feels a bit like the frantic rush before an exam session.

If a conference has just ended, I shift into an exploratory mode: I read papers, think, make calls, and discuss ideas. In the morning, I might teach a class; then I’ll write code for a few hours; ideally, I’ll run an experiment in the evening and go to bed, only to wake up and check the results. Somewhere between these three processes, I exist.

Whether I Have Experienced Burnout

I would call it running out of steam. I feel a bit low, and that’s when the search for meaning begins: is tabular data really here to stay, or will it all come to an end soon? At that point, it helps to slow down, stop trying to control these thoughts, and simply live with them for a while. Eventually, these thoughts helped me better articulate what we are doing and think about the future with more clarity and confidence. One of my hobbies is surfing the internet—not scrolling through TikToks, but following links. I pick a topic and explore it in depth. This helps me cope when I’m feeling run down.

My Interests Besides Science

Music and the internet. I started playing the guitar during COVID, but I also love listening to music—in large quantities. There are things I like right away and things I don’t at first, but if I spend enough time with them, I often grow to appreciate them. For example, I’m not a fan of heavy music like death metal, where vocalists scream in an inhuman voice (known as 'growling'). At first, I couldn’t take it at all, but I had a theory that I simply hadn’t immersed myself in the genre deeply enough.

Eventually, death metal gave in when I approached it from an angle I could relate to—through experimental rock and jazz. In 2025, critics praised the album Lonely People with Power by Deafheaven. That album spoke to me, and I ended up listening to it several times.

As for Russian musicians, I’ve been really enjoying Rushana lately. She appeared on The Voice, released a single, and then disappeared from the scene. Recently, she performed at a kvartirnik (a small, private concert—Ed.) at the FCS, and overall she seems to be experiencing a renaissance as an indie musician. Now her performances are less commercialised, which I appreciate. In general, I’d love to see more interesting developments in our music—from lyrics to composition to artistry.

What I Have Been Reading Lately

I recently re-read a fascinating book that might be risky for me but valuable to many others. It really resonated with me. The title is Why Greatness Cannot Be Planned: The Myth of the Objective, written by Joel Lehman and Kenneth Stanley; they are computer scientists, but the book is about life in a broader sense.

Many things in our world are built around goals. A scientist is expected to publish numerous papers, defend a PhD, and eventually become a professor—success is framed as something measurable to strive toward. But what if your goal is simply to be happy, to meet the love of your life, or to make a discovery in science?

The authors argue that setting goals in life often gets in the way, because we can never know in advance which steps will be right or where we should go. Instead, you need to follow a certain vibe—the same way I chose which faculty to enrol in or decided to join Yandex Research. People often say you should visualise the outcome and then act—or not act—accordingly. But it seems to me that sometimes you pursue things simply because you sense there is something worth exploring there.

Advice for Aspiring Scientists

If you’re feeling anxious—whether a paper isn’t coming together, an idea won’t take shape, or domestic science seems under threat—read the book I mentioned. Relax a little, go with the flow, but keep an oar in hand.

My Favourite Place in Moscow

Dizengof, a restaurant on Pokrovsky Bulvar—a place I frequented two or three years ago: sometimes after classes as a student, and later after classes I taught as a lecturer. It was always crowded, noisy, and full of life. Dizengof became part of our lives, taking the place of the café At Lara’s, when the FCS moved from Kochnovsky Proezd to Pokrovka. Despite its plainness, I loved that building and still feel nostalgic for it.