Generation of Everything

Alexander explains that his job title has read the same for years — technical director — while the domain under it changes: last time on the show he talked about architecture, and now his focus has moved from product development and platforms, client and marketing ones among them, to engineering R&D and how generative models change development processes. He draws a parallel with the previous wave, the one sold as digital transformation: IT people walked into ordinary industries, moved processes from offline to online, felt bright and cheerful, and plenty of nonsense came with it, down to odd job titles promising to make a company number one in the world. Now, he says, the industry changes itself: fewer people will hold a single role inside the SDLC, and full-stack breadth comes not from assistants but from agents that take a share of the work.

The hosts ask how he keeps up with the pace. The answer carries no posturing: he is forty, he has read and studied since school, and the wave of information still washes over him — now and then he wants something to calm down, and the sense of time slipping through his fingers like sand stays; his wife suggests seeing a therapist. The old meme about learning one JavaScript framework while ten new ones appear he restates: while you set up one autonomous assistant, a bunch of others turns up beside it. The worst part: his brain needs an assembled concept: it used to last about a year ahead, and now it is reassembled constantly, so only a baseline can be fixed. His comparison is a data platform refreshed at night: so many diffs arrive that yesterday's are unprocessed when the next day starts. The example: the third DeepSeek version arrived with a claim that training it had been very cheap, while a book bought a year ago still sits unfinished.

The conversation turns to checks — this is, after all, a podcast about quality. Alexander describes his System Design Space project. He had been writing a book on system design: he drew up a table of contents and assembled the text from his own talks and articles, was annoyed by repetitions and self-references, and moved slowly and sadly. At the end of 2025, once it was clear he was shifting towards AI-assisted development, he needed to publish study results on a site, so he handed the contents and materials to Codex and asked for a simple site; the first pretty version came from a different tool. Then he watched what came out and grew the scaffolding: architecture diagrams were missing, so they were added; colours were nailed down, so a check appeared for light and dark themes and contrast; Lighthouse showed poor numbers, so performance limits appeared. Over half a year more than twenty kinds of checks accumulated. A side effect: his fanless MacBook Air heats up like a stone, because Codex spawns processes that eat every resource on the runs; he already wants a more powerful machine, one with cooling.

The second story is the glossary. Alexander has been criticised for using too many anglicisms, and he uses them because he does not know the Russian term, and because terms differ from place to place; so he started a glossary and translates the text into plain Russian with Codex, step by step. The routine: the agent shows a plan — what to lift into the terminology list, what to leave in English, what to simply translate, how to change the visualisations — the plan gets reviewed, and the agent then works on its own, longer but very likely at good quality: the text reads as if a proofreader had gone over it. He warns against installing everything in bulk, npm-install style: pull skills and wire up MCP servers thoughtlessly and you get a dump of tools that the agent is supposed to sort out, while the result is not what you wanted. Masha notes this is exactly the flow they used rewriting client web interfaces over eighteen months — first onto micro-frontends, then onto SSG — and that QA decided what to check.

That leads into an argument about the full-stack engineer: Masha thinks the hard split into professions will not survive; Stas objects that there will be no universal soldier and the profile built over years stays somewhere in the heart. Alexander answers through the inner loop: in 2025 everyone optimised their own phase — analysts generated documentation, developers summarised it and fed it into prompts, QA watched coverage turn green — and at the end the end-to-end time of a task had not shrunk, or the output was not what the spec asked for, a game of broken telephone. So he sees the value of a QA engineer in working out what feature is asked for and how to verify it: many arrive at asking the agent for the check scaffolding first and the implementation second, which is effectively TDD or BDD. On the half-life of knowledge in this area, one of the hosts offers a figure: a month.

One of the hosts adds an observation: teams that have long kept tests next to the code, test in isolation and wrote down their contracts got a huge boost from AI, while those that never invested in basic processes back in 2010–2015 get no boost from any agent. Alexander calls it the Matthew effect: AI multiplies whatever you already have. The downsides he opens with are not about security but existential risk: the model builders aim at AGI, and the camps around it differ. The positive scenario is AlphaFold and Demis Hassabis's Nobel prize: a model predicts how a protein chain folds in three-dimensional space, and for pharma that is the key that has to fit the lock. Other blocked areas: the space elevator, where graphene showed extraordinary properties at the nanoscale that nobody managed to scale up, and fusion, which nothing but unstable magnetic fields can contain. The dark side: a superintelligence smarter than us is unlikely to stay under our control, and closer to home digital identity is blurring, because voice and video are easy to generate.