In 2012 Durov wrote an essay about the school of the future: «7 элементов системы образования XXI века». Seven elements. I read it fourteen years late, and almost everything in it has been built since. So this is an audit. What worked, what didn’t, and what it means for how I study now.
The diagnosis
Durov’s diagnosis: school is a factory. Bells, shifts, kids sorted by year of manufacture. It produces standard workers nobody needs anymore. A kid who can’t sit through it gets an ADHD label and Ritalin. The factory blames the raw material.
His enemy was the media industry. One blockbuster gets more talent than a whole school subject, and a chalkboard loses that fight every time. So the school has to change or die. His fix, seven elements: lectures by star teachers, game simulators, cinematic tests, weekly public rankings, educational blockbusters, historically accurate game worlds, and no state curriculum at all. Parents choose, schools compete.
The diagnosis was right. Then we got fourteen years of data on the fix.
Fourteen years later
Star lectures → MOOCs. Failed. Thrun, the Stanford professor from the essay, founded Udacity. Coursera and edX are the same model. Completion rate: single digits. Thrun himself called it a lousy product after San Jose State, where online students did worse than the classroom. Best content in the world, free, and people don’t finish it. Content was never the problem.
Gamification → Duolingo. Half. Streaks work. A 500-day streak proves loyalty to the app, not the language. Engagement and learning are different metrics, and the money is in the first one. Durov’s cinematic tests dissolved into the same apps, with the same problem.
Rankings → Codeforces. Worked. A Codeforces rating says more about a programmer than a diploma. Durov called it. But it works on people who chose to show up. China put rankings on every schoolchild, got top PISA scores and enough pressure that in 2021 the state banned exams for the youngest along with publishing results.
Digital school → COVID. Failed. Two years of school without a building. Same result everywhere: a third of a year lost on average, strong families fine, weak families hit hardest and never catching up. Yes, that was a zoomified factory school, not Durov’s design. It still failed exactly where his would: the structure and the adult disappeared, and the middle of the class switched off.
Educational cinema → YouTube. Different. No Hollywood. Instead 3Blue1Brown, Veritasium, Kurzgesagt: cinema-grade science made by independent creators on ad money. Right demand, wrong producer.
Game worlds → Discovery Tour. Worked. Ubisoft shipped a combat-free educational mode for Assassin’s Creed, written with historians, used by schools. The exact franchise from the essay.
A teacher for every question → the LLM. Not a star with video answers. A model, instant and personal. It fixes what killed MOOCs, feedback, and adds a problem Durov didn’t see: the answer without the effort.
Three kids
Build the school by the essay and run three kids through it.
Kid A is gifted, with parents who pick the program. They flourish. The school is built for them. It’s telling that the essay ends with Durov’s own experimental gymnasium, four languages. He generalized himself.
Kid B is average. No fire. In a normal school the structure carries them: be there at 8:30, hand in the homework, a teacher notices when you slip. In Durov’s school a star lecture competes with YouTube, and YouTube has the bigger budget. A month later they’re in the bottom third of a public table and they stop. That’s the 95% who never finish a MOOC.
Kid C has parents with other problems. Nobody picks a program. The state curriculum Durov called a totalitarian relic was their floor. Take it away and birth decides everything.
Durov’s school raises the variance. The top wins big, the middle sinks, the bottom falls through. The middle is two thirds of everyone.
Two mistakes
School loses to Hollywood, so make school Hollywood. No. Learning is mostly holding attention on what isn’t interesting yet. Make every step pleasant and the kid learns to demand pleasant. Hard skills are never pleasant at the entrance.
And school isn’t a content channel. It’s where you learn to be among people and under an adult who isn’t your parent. COVID showed what happens without it.
What would work
Split the school into functions and give each its own tool.
Knowledge: the machine. Star lectures, simulators, game worlds. Durov wins this one outright. On top, an LLM tutor in Socratic mode: it guides, it doesn’t hand you the answer. Progress by mastery, not by age. Bloom showed in 1984 that a kid with a personal tutor and mastery learning beats the classroom by two standard deviations. The question was always how to afford a tutor for everyone. Now we can.
Motivation: the human. School stays a physical place with a schedule. Self-discipline is what education produces, not what it requires. The teacher stops lecturing and becomes a coach: watches technique, notices the slip on day three, runs the group.
Assessment: continuous and private. Short weekly checks instead of one exam, Durov’s idea. But the results go to the kid, the parents and the teacher, not a public board. Ten years of history beats one exam day.
A standard for the floor, freedom above it. Native language, English, math, money and digital basics: guaranteed to everyone. This is kid C’s floor. Everything above it: any program, any experiment. Durov’s mistake wasn’t freedom. It was not marking the floor.
The price. This costs more. The machine took lectures and grading, but you pay a teacher for being in the room, and 1:12 instead of 1:30 is more teachers. There is no cheap good school. There’s a cheap bad one, and you pay for it later.
Rankings
The best objection: rankings are how things evolve. They give you a gradient, someone at 1600 whose solutions you can read. They can’t be faked. Chess and competitive programming improve faster than most fields for this reason. All true.
But evolution optimizes the metric, not the goal. Goodhart. Codeforces bred people who solve an isolated problem in 15 minutes. That says little about building systems or working with others, because the rating doesn’t measure it.
And evolution is cheap for the species and expensive for the individual. Genes are free to waste. A kid who reads “23rd of 30” every week and quits is the mechanism working as designed. For us it’s a lost engineer.
So put the selection pressure on methods, not kids. Let schools, programs and tutors compete in the open, and kill what doesn’t work. Medicine does exactly this: brutal competition between treatments, no public ranking of patients.
Rankings stay, as a choice. Codeforces works because everyone there chose it.
How to actually learn
None of that gets built tomorrow. But the personal version exists today: you, an AI, and hours.
I’m the test case. Small town, Mozdok. Java at 14 from a book, Python at a MIPT camp, olympiads in QBasic and Pascal. That’s kid A with public rankings, and it worked on me. Then Prague, nuclear physics at ČVUT, one semester passed, and I left. Structure without fire didn’t hold. Now I study for hours a day with an AI and keep an 8:00–9:30 block nothing moves. The machine gives me knowledge. The block gives me discipline. Neither works alone. I checked.
Grilling a model until it makes sense is a good method. Active questioning is one of the few techniques with real evidence. It has three holes, and none of them feel like holes.
“It makes sense” measures the explanation, not your memory. A month later you find out what you actually kept.
Answers without effort don’t stick. Desirable difficulties, the guessing and the being stuck, are the write to long-term memory. Ask before you try and you trade retention for speed.
And input isn’t output. Understanding is half. Reproducing without hints is the other half, and it’s the half anyone ever tests.
Four fixes:
- Predict first. Two minutes, your own hypothesis, written. A confident wrong guess before the answer is the strongest known boost to retention.
- Flip it. Every few days the model asks you, about last week and last month, cold. That’s retrieval practice and spacing, the two best-proven techniques there are.
- Write the summary yourself. From memory. Then let the model check it. If the model writes it, the model learned.
- Explain it to a human. A student, a colleague. What you can’t explain, you don’t have.
Knowledge has to flow out, not just in.
The short version
- Machines for knowledge.
- Humans for motivation.
- A standard for the floor.
- Freedom above it.
- Ratings private by default.
- Competition by choice.
Durov got the technology right and the pedagogy wrong. He swapped the teacher for content. The thing to swap was the lecturer for a coach.
One thing I don’t have an answer to. A kid raised with a perfect AI tutor might never learn to learn without one. How much frustration to leave in, I don’t know. Nobody does yet.