Education: Teach the Hard Stuff, Not the Job Spec

Evgeny Sokolov starts from the very beginning: the Rostov region, the small town of Belaya Kalitva, an ordinary school, an interest in programming that turned into competitive olympiads and a diploma at the national informatics olympiad, which let him enrol anywhere. He picked the computational mathematics faculty of Moscow State University and landed on a chair working on machine learning — nobody, he says, was saying "artificial intelligence" or "data science" then. He began working in his fourth year, passed through two startups, and spent the longest stretch at Yandex leading a recommender systems group. He started teaching straight after graduating and combined both for years: recording online courses, assembling specialisations and discussing Russia's first online master's programme with Coursera. At some point combining stopped being possible. He explains choosing education with two arguments: impact is easier here, because a programme covers the whole country and you can influence other universities, while capable teachers and managers are in short supply; and it feels socially useful.

He teaches classical machine learning — tabular data, linear models, trees, boosting — rather than neural networks. His main audience is the applied mathematics and computer science students, the faculty's flagship programme that olympiad winners enter. The second is the minor: students must sign up at the end of the first year for four semester-long courses outside their own field, so an orientalist may not take Chinese but may take data analysis. That minor draws 300–350 people a year: economists, orientalists, journalists; many later change career track, and quite a few of its graduates work at T-Bank as analysts and developers. He teaches neural networks there too. Alexander remarks that a course can change a person's trajectory; Evgeny is careful: he would like to believe it, but the students should judge. He gives the scale separately: the faculty now has some 5,000 students, whereas Alexander's own MIPT faculty had a little over a hundred in his year, and MIPT's entire first year was under a thousand.

Evgeny explains industry partnership through two philosophies. The first is practised by online schools training mass specialists: come to a company and extract a very concrete list — not "they must know Spark" but which tasks and settings — and build a programme around exactly those skills. Their approach differs: they take very strong people, for whom spelling out specifics is pointless, and teach hard things so a student grows through struggle and builds erudition. The company's role is then to run difficult courses: from the first year almost every programming course is handed to companies — some to Yandex, some to T-Bank — and in exchange they win students' affection. Alexander likens the first approach to a vocational school and the second to a fundamental education, and recalls how at T-Bank students built on top of off-the-shelf boxes what the market could not really solve: they did not know it was hard. Excellent for a start, but ten years on they add up to a web of connections like the hero of A Beautiful Mind drew — it works, but it will have to be untangled.

Evgeny splits students into two types. The first optimise income and position: in 2008, when he enrolled, internships barely existed, whereas now a schoolchild is hired from sixteen and a first-year olympiad winner passes a junior interview. Hence the pragmatism: a student working full time looks at the second calculus course, counts how often a curvilinear integral helped at work, gets zero — and only the teacher's authority makes him invest; yet third-year machine learning, when models are everywhere at work, gets his maximum. The downside is dropping out of student life and networking. The second type want interesting problems and aim at research, and have to be loaded up: this September the faculty launched research programmes with a heavier load, from January these students get lab supervisors, and a stipend keeps them from leaving for a job. The conclusion is blunt: the student who simply came to study has all but vanished. Alexander sees a product approach here, with archetypes and jobs to be done, and recalls how, after his son was born, he switched to work and fell out of student life.

A separate line is holding a student who judges usefulness in the moment. Evgeny believes a course must be able to answer why it sits in the curriculum: reasoning like that is why differential equations became optional at the faculty — they were rarely used in machine learning, distributed systems and theoretical computer science, although generative models are now bringing them back. Every year he talks his mathematician colleagues into embedding Python lab work in their courses, while holding a defensive line: colleagues from the labs ask him to be tougher and add more linear algebra. They do not want the Western model, where the bachelor's is simple and the hardcore starts in the master's — their graduates report that master's courses abroad feel like their own first year. On soft skills Evgeny relays a colleague's story: charismatic teachers with general theories and matrices failed, while an instruction like "you need to fire someone — one, two, three" landed superbly.

Alexander answers that with practice. To a colleague who asked at a new-joiners meeting what to read about operating one's own systems, he recommended not a book but a drill on a test environment: break the system, let the team fix it, walk through the postmortem — and only then hand over reading material. He suggests running design teaching the same way: students design systems while the teacher injects network delays, failures and dying virtual machines, so it shows whose solution did not fly; he refers to A Philosophy of Software Design, written from semester projects in which the author reviewed student solutions personally. Evgeny sorts teachers into types as well: professional mathematicians, who pass on academic culture; applied researchers, for whom courses bring visibility, students and eventually papers; and people from industry, often the faculty's own graduates, who rotate the most yet drive programme changes hardest. A fourth type he adds is managers in education, himself among them: without them the other three drift apart.