10 Key AI Trends You Can't Afford to Miss: What to Learn and What Not to Waste Time on in the Age of AI
What should you learn and what should you stop wasting time on when artificial intelligence changes every day — and which shifts in work, software and access to strong models must not be missed?
The lead of those who really use artificial intelligence over everyone else is growing catastrophically: a year and a half ago Alexander Volchek's wife was a strong user, and now the gap between them is huge — and strong models are expensive: an hour of a normal task, according to OpenAI employees, costs about thirty dollars. Chasing 'agents' and learning to write prompts is pointless: models such as Astra understand the task and context themselves, and what is valued is the ability to understand a task as a whole. One main system is enough, two at most, and processes should be rebuilt around AI rather than bolting a module onto a CRM or Google Sheets. Any decent model is already smarter than any expert; there will be no familiar programs, sites and search, and OpenAI, Alexander believes, will become an operating system. Dots, the DevDay newcomer, took on the task of sorting school emails but asked nothing — the future belongs to AI that asks the questions itself. Access to strong models is getting more expensive — a five-hundred-dollar ChatGPT Pro plan, expensive API tokens — and change has become daily.
What to watch for
Key takeaways
A year and a half ago Alexander's wife was a strong AI user, and now the gap between them has grown catastrophically. The gap will widen also because good models are limited in access and price: an hour of a normal task, according to OpenAI employees, costs thirty dollars on average.
People and companies lose money on agents, and an agent that uses eight and a half thousand subagents is like counting tokens. 'Agentic' thinking, Alexander believes, keeps people from understanding what artificial intelligence and the system of the future are.
Modern models — Astra in ChatGPT and Codex or Sonnet 5.5 at Anthropic — are built not for people to know how to write prompts but to understand their task and context as fully as possible. What is valued is not a profession or a diploma but the ability to understand a task as a whole.
At a prestigious university in Warsaw, first-year students are taught how to use Windows 10 and take screenshots. It is as strange as teaching children to vacuum when a robot vacuum drives around the house all day.
There are four leaders — OpenAI, Anthropic, Grok and Gemini; everything else is secondary, and Russian models do not even reach the sixth level. One system lets you understand what is really happening, even though the leaders themselves are turning into a zoo.
A module in a CRM, an AI module in a CMS or a site on ChatGPT Sites on top of Google Sheets is an attempt to use an elephant where there were toys. Why have Google Sheets and a site if your own Codex will tell you that a status has changed and make the visualization you need?
A Cybercab with no steering wheel stands in a Tesla store, and such cars drive past Alexander's home every day. His daughter, who has turned sixteen, did learn to drive because the companies do not yet have licenses, but the younger ones, he believes, will not: within four years the licenses will be worked out.
By every test a model is smarter than any human, and no human will beat it, although it does not know your inner details and will not dig a hole in the garden. Waymo in San Francisco drives more safely than any driver because it sees 360 degrees.
Sasha Mashrabov, the creator of Higgsfield, spoke at the conference about a ten-minute video that no one in the world can make — thirty thousand dollars were spent on it. OpenAI, according to Alexander, burned a million or tens of millions of dollars but solved part of the Navier–Stokes problem.
What remains is a dialogue with artificial intelligence through a command line, an interface or voice across different devices, and the intermediate layers will go — from Booking to Google and Yandex search. Thousands of startups have already died on this.
Codex and Dots scan Alexander's letters, tell him about them and sort them by rules, and he stops opening his email. What remains is the address — his name — and which provider serves it matters as little as which courier delivered an Amazon order.
Not a marketplace and not a platform, even though the company has announced what is, by Alexander's count, its fifteenth marketplace: someone will definitely become the operating system, as Google once did with Android.
A service in Dots took on the task of sorting emails from four schools but did not ask who in the family takes care of the children and who reads those emails first. The future, Alexander is sure, belongs to AI that asks the questions itself; cost and safety may stand in the way.
The two-hundred-dollar ChatGPT Pro plan will be made simpler, a five-hundred-dollar plan has appeared, and Astra's work through the API is estimated at roughly thirty dollars an hour. Strong systems are already available only to certain corporations, structures and countries — the US and China.
Models come out every week, the stage explains what is cheaper and what is more expensive, and change has become daily. When AI starts creating hardware itself, 'then it's game over' — and systems like Astra and Fable are still worth living with for months.
What this episode is about
A ToTheMoon episode that is the second part of a series: after ten rules for working with artificial intelligence, Alexander Volchek names ten trends that must not be missed — what will be valued at work, in a profession and in life, what to learn and what not to waste time on. Together with the first part that makes twenty thoughts, and all of them are published on the project's website. It was recorded the day after OpenAI DevDay in San Francisco, and impressions from the conference run through the whole episode — while Alexander asks viewers not to live in hype mode and not to rush after every new release.
The first trend is that the gap between those who really use AI and everyone else will become catastrophic: a year and a half ago Alexander's wife was a strong user, and now the gap between them is huge — and strong models are expensive: an hour of a normal task costs about thirty dollars. The second is not to chase the concept of agents: people lose money on it, and an agent that uses eight and a half thousand subagents is, to Alexander, like counting tokens.
The third trend is that what is valued is not a profession or a diploma but the ability to understand a task as a whole. Alexander calls Prompt Engineering courses an illusion — Astra and Sonnet 5.5 are built to understand the task and context themselves — and teaching Windows 10 at a prestigious university in 2026 an absurdity, like teaching children to vacuum while a robot vacuum drives around the house.
The fourth is one main system, two at most. The leaders, in Alexander's view, are four — OpenAI, Anthropic, Grok and Gemini — while Russian models do not even reach the sixth level. The leaders themselves are turning into a zoo, and the versions of ChatGPT and Codex are ever harder to figure out; he himself has two systems — Codex and Claude Code.
The fifth is to rebuild processes around AI rather than bolt AI onto the old: a module in a CRM, an AI module in a CMS or a site on ChatGPT Sites on top of Google Sheets is an elephant where there used to be toys. Interfaces like the iPhone will change, and children, Alexander believes, no longer need to learn to drive: a Cybercab with no steering wheel stands in the Tesla store nearby.
The sixth is that models are already smarter than any expert. Waymo drives more safely than a human, the Higgsfield team made a video no one else can make for thirty thousand dollars, and OpenAI solved part of the Navier–Stokes problem. AI can make mistakes, but there is no need to argue with those who call it dumb.
The seventh is that there will be no familiar search, browsers, sites and programs: what remains is a dialogue with AI across different devices, and the intermediate layers — from Booking to Google search — will go. Codex and Dots already go through Alexander's email for him, and he bets that OpenAI will become an operating system, as Google once did with Android.
The eighth is that real AI will ask the questions itself and remember everything: Dots took on the task of sorting emails from four schools but did not ask who in the family takes care of the children. The ninth is that access to strong AI will become expensive: the two-hundred-dollar ChatGPT Pro plan will be made simpler, a five-hundred-dollar plan has appeared, and Astra's work through the API costs about thirty dollars an hour. The tenth is that change has become daily, and the companies themselves no longer understand what is happening.
The second part turns the rules of the first into a forecast: value shifts away from skills like prompts and from the fashion for 'agents' toward understanding a task as a whole, one deeply mastered system and processes built around AI. The episode's boldest claims — that OpenAI will become an operating system and that familiar programs, sites and email interfaces will disappear — Alexander tests on his own email and on Dots, just unveiled at DevDay. The most practical one is to work with strong models now, while they are still accessible to ordinary people.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 86 segments: 86 identified, 0 mixed, 0 probable, and 0 unresolved.
Read transcript on a separate page
Loading…