Education and Reflection: Alexey Tolstykov on Nurturing Researchers in the Age of AI
Yandex SHAD head Alexey Tolstykov on why students must be ready to change, how teachers grow alongside them, and how AI-era courses are graded.
Alexey Tolstykov leads the Yandex School of Data Analysis (SHAD). He discusses how students prepare for change, how research supervision shapes teachers as well as students, and how SHAD adapts its courses to artificial intelligence.
Preparing for Change
Being prepared for change is the most important part of education. This readiness is less common among young students. It is more typical of upper-year undergraduates, specialists, master's students, and PhD students. Tolstykov's work focuses on this group and on prospective students and early-year undergraduates with high potential. Universities usually identify such students through competitive environments, including Olympiads, hackathons, and research projects.
Human thinking limits how people imagine themselves. A person can easily imagine a different appearance, a different location, or different weather. Imagining a different way of thinking is practically impossible. This is a specific property of human consciousness.
Tolstykov argues that people can still prepare for change by observing those around them. A person can learn to reconstruct the way others think and then choose a direction to follow. Real education requires a teacher whom a student wants to follow. In academia, this figure is most often a research supervisor. The student aims to take a step forward from that supervisor's position and to approach how the supervisor solves problems, thinks, and sets goals.
The Teacher Changes Too
In the traditional teacher-student model, the teacher is a constant. The teacher gains nothing from students and does not change. Tolstykov rejects this view for research supervisors. Anyone who has worked in research or science, he argues, sees that students affect their supervisors.
Joint work on unknown problems changes the teacher. When a student moves away from the teacher's path, and not only forward, the teacher recognizes a new method. The teacher feels proud and improves as a result. Tolstykov states that a teacher must not teach only to pull students up to the teacher's level. A strong student who pushes the teacher forward is a common practice in science.
Learning the Rules of the Research Community
Students must understand what a research community considers a valid result. Terminology matters. The same object can carry different names in different fields, and the chosen name determines where an article is sent. Writing for physicists, mathematical physicists, mathematicians, or geometers requires different lexicons.
Understanding what counts as proof is central. Students must also understand which ideas are currently the most active and testable. Research output is growing rapidly. Machine learning conferences now receive tens of thousands of submitted papers.
Reproducing Experiments
A program at Tolstykov's school focuses on reproducing experiments from recent high-quality papers. Each participant group works with a mentor. Mentors are usually specialists, such as engineers from Yandex, or master's or PhD students with strong expertise in the relevant paper.
This year, 45 students took part. They came from the applied specialist program at Sirius University and from local master's programs. Another 30 students attended in person. Participants were divided into groups of five or six. Over two weeks, they examined the experimental design of each paper. They asked how the experiments followed from the authors' ideas and why that part of the work needed verification. They also asked which dataset, computational conditions, and broader task the verification applied to.
The program shows that many machine learning ideas do not clearly transfer to other applications. Participants also learn that not every experiment in modern papers can be reproduced. Tolstykov notes that the reproducibility problem predates modern papers.
The program aims to develop critical and scientific intuition. Tolstykov believes this skill takes hold in a number of participants. In industry, R&D specialists must read many papers, select ideas for implementation, and integrate them into working services. A stronger internal critical filter helps them reject weak ideas and test promising ones. Several participants have returned to report that the program was valuable.
A Historical Example: Luzin's School
Tolstykov cites the mathematician Nikolai Luzin, who founded a school of mathematics at Moscow University in the 1930s. Many prominent mathematicians, including Kolmogorov and Aleksandrov, were his students.
Luzin began work with students by analyzing published papers. He assigned papers and asked students to find errors. Mikhail Suslin, a young mathematician who died of typhus, found an error in a work by Émile Borel and corrected it. The correction led to the concept of Suslin sets. Borel's statement had been incorrect, and the class of sets Borel constructed was insufficient.
Tolstykov also describes a crisis in Luzin's school. Students rose against their teacher. He compares the situation to a model in which an apprentice must defeat a master before becoming a lord. The cycle repeats as the new lord takes on an apprentice.
When asked how a teacher should prepare for defeat by a student, Tolstykov says that the answer is difficult in a universal philosophical sense. In machine learning and artificial intelligence, which he has worked in for almost 20 years, the field remains in a safe position because it is still growing. New tasks and applications keep appearing, and companies are entering new niches. Tolstykov notes that robotic intelligence remains an untouched field, and he cannot cite a single example in his field of a teacher losing a position to a student. He states that teachers are proud when their students surpass them and move forward.
Postgraduate Studies and Research Outcomes
Tolstykov describes postgraduate studies as a system in crisis. Their main output is a dissertation, with a diploma and candidate status as by-products. The question of whether the system actually produces researchers is rarely asked. Previously, the system assumed that outcome by default.
Students have more options now. The Faculty of Mechanics and Mathematics allows students to leave at any year of study. Tolstykov advises against leaving in the first year. Students must at least pass the calculus colloquium before finishing.
Tolstykov argues that a researcher must be the result of education. The space for research is practically unlimited, and technological development creates many opportunities. The transformation into an independent researcher remains necessary.
Three qualities support this transformation:
- Willingness to face the unknown. Students must accept that a new task or an attempt to improve someone else's result may not succeed.
- Acceptance of mistakes. Students must learn when to stop and recognize that an approach is not working. In digital work, the system itself often provides this feedback.
- Self-reflection. Many students skip the step of analyzing their previous actions, such as why they failed or succeeded in an exam. Tolstykov believes universities rarely teach conscious reflection, and he is concerned that education has neglected it.
Tolstykov states that university education has largely become a factory for personnel. Universities focus on whether employers will praise their graduates and whether graduates will find jobs quickly. They rarely assess whether a self-sufficient person capable of moving forward is emerging.
The university's task is to test these research traits early. Students need a chance to take small research steps while working directly with a potential mentor.
Mentors and Project Work
Tolstykov contrasts a classic university of 15, 20, or 50 years ago with the present. The classic model assigned each student one research supervisor. Today, students try several projects with different people, including small groups led by people only a few years older. Within SHAD, these leaders are called project mentors. They often do not hold academic titles such as PhD, Doctor of Sciences, or academician. A person needs only to have done the work and to know how to do it.
Tolstykov observes that the gap between mentor and student in practical activity is much smaller than the gap in academic supervision.
SHAD's project schools create two-week intensive environments that pull students out of their regular studies. Coordination is difficult. Negotiating with partner universities to free two weeks for each student is hard. Groups of 30 to 50 students and professors are scattered across disciplines and classes.
Tolstykov's methodology focuses on persuading students to trust the educational team, set aside their concerns, take a step forward, and then return to their regular routine. Students who experience five, eight, or ten such environments, each with a new mentor, task, and field, can view their education differently. Their path stays the same, but they gain alternatives. Tolstykov describes this as adding portholes to the educational pipeline.
Students can replace elective courses in their curricula with SHAD courses while keeping a theoretical foundation. Most SHAD courses are taught by professors from academia, who make up about half of the instructors. The rest are managers of large groups or units at Yandex, some of whom have worked at the company for 5, 10, or 20 years. They explain their work and show that their practices remain current.
Tolstykov believes this approach builds awareness. Students learn to try different things in a supportive environment that does not punish mistakes. The environment should allow students to understand what went wrong and why. Some actions are negative experiences rather than mistakes, and some are simply done the wrong way.
Flexibility and Readiness
Tolstykov acknowledges that opportunity does not guarantee use. Offering students the chance to adjust their programs does not mean they will do so. Tolstykov notes that SHAD's courses remain demanding. Students cannot choose a program that avoids core subjects and still graduate from SHAD.
He contrasts this with Western education. Some systems require fixed courses. Others let students choose their workload. Both have advantages and disadvantages. SHAD's model sits closer to the fixed-course approach.
Readiness also includes psychological preparation, which is largely absent from most PhD programs. A supervisor rarely helps students with worry, fear about the future, or doubts about their chosen path. Shorter distance between mentor and student makes such conversations possible.
SHAD's Origins and Admissions
SHAD began in 2007 with 50 students, almost all of them senior students from the Faculty of Mechanics and Mathematics. At that time, data analysis was still developing. The founders included Yandex founders, academic advisors, and education specialists connected to universities and Yandex. They selected the strongest students who already showed the required mindset and gave them additional workload. The founders believed these students would succeed in life even if they did not enroll at SHAD.
SHAD is a two-year program. Of the original 50 students, 30 completed it, which forms the first graduating class. Tolstykov does not know whether the 20 who left developed a researcher's mindset and has no evidence either way. Most of the 30 graduates now work outside Yandex and are spread across the world. Tolstykov believes they share confidence that they can act in uncertain environments and remain successful.
SHAD's slogan, "It will be hard, you will like it," reflects its admissions approach. Entrance is difficult, the tasks are demanding, and the workload is heavy. Tolstykov says the team conveys this from the first contact with applicants. Graduates often say the entrance exam was harder than the two years of study. Tolstykov does not claim that students transform fundamentally during the program. He says that he cannot prove lasting change, but the environment and the mindset students bring have value.
Some applicants are not accepted because their motivation does not match the environment. Others apply in a year when they are not ready. Many applicants fail the first, online stage and return the following year. Some reach the second stage. SHAD has three admission stages. Tolstykov mentions students who failed the first stage in one year, the second stage the next, and the interview after that, but who continued and were in their fourth year of study at the time of the discussion. Some students were admitted on their fourth attempt.
Tolstykov argues that the choice made after eleventh grade should not determine a person's future. He criticizes the classic university system, inherited from the Soviet Union, which limits a student's path if they do not achieve required exam scores in that year. Some students give up on their chosen faculty for this reason and do not retake the exams. SHAD tries to keep that door open. Students enter at age 40 or older. The oldest graduate last year was 65.
Admissions peak immediately after the second year of university. Students at that point choose between accelerating their careers through SHAD, possibly while working, and starting their careers directly.
Master's Programs and Specializations
For students on a research track, SHAD runs two joint master's programs. One is the Modern Computer Science program at the Higher School of Economics. The other is the Data Analysis master's program at SHAD's base department at MIPT. In these programs, SHAD courses form the core professional curriculum. Internally, the program is called the advanced track.
Students with strong bachelor's preparation face no rigid frameworks. SHAD offers about 70 to 75 courses per year. Students specialize in big data infrastructure, which Tolstykov describes as the hardware and systems of modern technology, or in intelligent systems, which SHAD categorizes as machine learning or artificial intelligence. SHAD and Yandex monitor these two profiles most closely. This year, SHAD also launched a specialization in Physical AI. Two SHAD alumni, one working in autonomous transport at Yandex and the other in a robotics lab, developed courses for it.
Tolstykov argues that a university or school is strong when its graduates return to advance it. The strongest graduates should invest their energy, time, and motivation in the program. When mentors and teams stay close, both can develop.
Human-Centered Education
Tolstykov agrees that education depends more on who teaches and how than on the specific content. The internal culture of a school determines whether students choose to pursue research. He describes this approach as human-centered.
The Traffic Light Framework for Artificial Intelligence
SHAD applies a traffic light system to its courses. Tolstykov notes that the system is not his invention, but SHAD is testing it.
- Red courses prohibit artificial intelligence. Any suspicion of its use leads to an in-person re-examination. The student and the professor sit together, and the professor asks questions. If the student cannot confirm the submitted work, the case counts as cheating, and SHAD has zero tolerance for it. SHAD does not force anyone to study. The program is free, and students who do not wish to put in the effort may leave.
- Yellow courses are a transitional category. Instructors adapt their assignments for artificial intelligence as much as possible. Tolstykov acknowledges that SHAD cannot yet redesign all course materials for the new paradigm. Students may study materials independently with artificial intelligence, but they should not submit AI-generated answers for grading, because instructors would have to recheck the same answers. Instructors check progress at specific touchpoints, such as the middle review at mid-semester and the final examination. A few assignments require an in-person defense. Students must also submit a reflection block summarizing their work and what they understood. The final grade is validated through an exam. A passing grade is a satisfactory result. Good and excellent grades require more work and a higher level of engagement and ability. Two courses, the algorithms course and the fundamental machine learning course, are colored yellow this semester.
- Green courses are AI-native. They assume artificial intelligence is integrated into the work. Their assignments or research tasks are so extensive that completing them properly within one semester alongside other work is likely impossible. These courses are either very large or focus on developing skills in AI scenarios and AI agents.
Currently, red courses are the majority. Yellow courses outnumber green ones. Tolstykov expects the proportions to shift each year. He is not aware of a better approach among Western, Eastern, or domestic practices, and he invites anyone who knows one to share it.
The Purpose of an Educated Person
Tolstykov does not believe the definition of an educated person has changed. In the new paradigm, an educated person is the same as twenty or even two hundred years ago. Such a person knows not only how something can be done but why it should be done, and why another action should not be done. This person holds a fundamental worldview and values, lives by them, and is ready to share them.
Educated people argue differently. They try to understand the other person's position and may adjust their own view. They do not try only to prove the other person wrong. Exchanging viewpoints remains essential. Understanding a subject requires understanding how others see it.
Tolstykov believes SHAD and Yandex can approach education more responsibly and deeply. The teachers, organizers, students, and graduates understand the technology's internal workings, its limits, and its strengths. Tolstykov notes that even for the most successful modern architecture, the Transformer, no one can explain the output it produces, although the learning process is understood. Understanding the process helps people analyze their own learning.
From the first organizational meeting for first- and second-year students, SHAD does not promise to make students top artificial intelligence engineers. Students who want to build engineering skills can do so. Many assignments and techniques presented at conferences six months or a year earlier are already in SHAD's courses. SHAD also places strong emphasis on the humanistic dimension of education. Tolstykov expresses the hope that a philosophy course, and possibly a rhetoric course, will one day be added to SHAD.
x
Education should prepare students to change; teachers benefit when students surpass them, and mentorship works best with people close to the student.
- Tolstykov argues that a good student can surpass the teacher, and the teacher should feel proud rather than threatened.
- Humans struggle to imagine themselves thinking differently, so preparing for change requires observing and learning from people around you.
- He says true education requires a person one wants to follow, typically a research supervisor.
- Historical example: Nikolai Luzin's school at Moscow University produced Kolmogorov and Aleksandrov; Suslin found an error in Borel's work leading to Suslin sets.
