What Remains Scarce When Code Becomes Cheap?

Sergey read maths and mechanics at Rostov State University, tried design, noticed he used rulers and guides far more than strokes, and moved into markup; XSLT brought him to Yandex, one of the few companies then using it. He came from a four-person web studio in Rostov to a company of roughly 250, spent his probation on Yandex Money, then asked to move onto Yarushka, Yandex's incarnation of blogging with microblogging about changing moods. It closed after landing in a valley of death: even old-school machine learning for filtering did not exist yet, quality rested on manual assessor labour, and an open window for user-generated content filled with the worst of it. The questions-and-answers service ended the same way: the morning after launch Sergey asked what to do if you had swallowed glass from a shot glass broken in a bar, and got some answers — on such a service, he says, you may be told to treat yourself with electricity. Who answers for advice like that has, in his words, no clear answer even now.

Sergey calls BEM a shard of Yarushka, a project the team — managers, backend engineers and Vitaly Kharisov on the interface side — over-engineered gloriously, in his own phrase. His favourite example: the settings page carried a preview that repainted while you dragged sliders and the colour picker, built from the same blocks as the full-size page; hence the conclusion that marking up against element IDs is bad and a block holds all the technologies it needs. A build step was needed too, to which colleagues replied: is building the frontend C++ or something? The shared blocks grew into a project codenamed Lego, the common library for Yandex services. The methodology spread through free Yandex Subbotnik conferences, meetups and trips around the country, yet the West got a reflection of reflections: their English was not good enough for the stage. Friction was higher inside than outside, where the label of a Yandex technology sufficed: no prophet in his own land, as Alexander puts it.

The matrix has two axes: product — image search, video search, the marketplace — and technology: interface developer, backend engineer, QA. Making product the primary hierarchy Sergey considers deeply mistaken: products move, technology is more fundamental, and a manager should be a technical mentor, with the rest unpacking from a report's grade growth. A product manager with two frontend developers can neither mentor them nor run a proper review, which gets more accurate the wider the comparison panel of people of the same specialisation and grade; some reports stayed with him over ten years. The first version of the matrix, a pool of workers by specialisation, worked like a blender, so the answer is a hybrid: months-long domain attachment, around 20% rotation per half-year at the review boundary, quarterly demo days prepared by the team itself, and personal plans along the product, technology and personal-growth axes. Alexander recalls a head of a JavaScript department with about a hundred reports, seen mainly at bonus announcements; Sergey's norm is four to nine or ten reports and a one-to-one every week or two.

The Staff+ track Sergey explains through the institution of directors: personal data, security, open source, education and accessibility must be done in one sweep for the whole company, yet cannot be delegated to a separate unit whose people would lose contact with reality and stop being practising programmers. He runs the summer schools and the school of interface development himself. The main pyramid is still built by product, with business units and P&L, while the directorates stitch through it perpendicularly: responsibility is broad, handles few, and for several years he went around without a badge doing what is known internally as trolling, pestering and whining, hearing that he was already the fifth director arriving with his own tax. A system crediting such contribution at review, he admits, still does not exist. Alexander recognises his own past year as a high-grade individual contributor responsible for AI in development, where instead of the frontier one ends up dragging the lagging tail towards basic practices.

Sergey calls himself an AI enthusiast and compares the current lift to the moment he first grasped what computers were: if the tool is not helping you, most likely you have not found the way. The value lies in a meta-property — just as a programmer builds his own programming tools, so here; when Node.js appeared they were delighted to write their own command-line tooling. The field is empirical, the way CSS across browsers once was: intuition accumulates through practice rather than deduction from rules. Alexander describes the sag in the middle: money is poured from above and results are expected, engineers below find it interesting, and the middle layer looks like people given electricity while running on steam engines, asking where to plug it in. Sergey's answer is not to ask AI for the same five tickets: that is fitting an electric motor to a mill wheel, while electricity changes the balance of forces itself. You have to climb to the meta-level and shepherd the next tier of agents, who respond faster than people and suffer no shock of change.

Education runs into the same shift. What matters most, in Sergey's view, is the meta-skill of learning an unfamiliar field quickly: by graduation React will have been replaced, and classical education is useful if only because it creates cognitive load. His metaphor is the gym: humanity does not need you to lift 250 kilograms, what matters is that your body changes on the way, and the brain has to sweat the same way. Search engines broke the habit of memorising constants; AI unloads basic cognitive functions so we can ask the next questions. Then comes the two-sigma effect: a personal tutor beats mass teaching by a statistically large margin and AI gives everyone one, which is why Practicum is moving towards an AI-native format while admitting that today's courses, with their brittle autotests, are written for everyone at once. Motivation, he answers with a borrowed joke, cannot be instilled: the share of the curious is near a universal constant. Content costs almost nothing; people buy focus and willpower. They close on his rollercoaster metaphor for AI, between how brilliant and how stupid it is, and agree to return in a year.