USA vs China: Two Machines of Destiny Are Building AI to Run the World

The encyclical places AI in the tradition of the Church's social doctrine: human dignity, power, the concentration of control. The host quotes the words about a "new Tower of Babel" and about AI needing to be "freed from the logics of domination, exclusion or death" — and answers that this can no longer be stopped and such appeals do not affect what is happening. Tatiana Tsvetkova adds that setting unattainable ideals is the point of religion, and the truth lies somewhere between apocalypse and idyll.

This week Sam Altman said he had thought more entry-level office jobs would have been eliminated by now, and that he cannot imagine handing the human part of interaction to AI. The host calls it nonsense: OpenAI has won its case against Elon Musk and is preparing for an IPO, so the right rhetoric is needed. The same chaos reigns at state level — Trump's bill on giving politicians access to models ninety days before release was reversed this very week.

Daniel Kokotajlo, a former OpenAI researcher, says the people who own AI companies and the governments of their countries will never stop short of deploying AI in their own interests. The host adds that OpenAI will block countries not over sanctions or GDPR but over profit, because trillion-dollar companies are companies of the country. There are two leaders, the US and China; Europe, given two years by the head of Mistral, has in the host's view already lost.

In the US the center of power is private AI companies, clouds, chips, capital and the Pentagon; in China it is the state, the party, industry, data and control. The US logic is to invent, deploy and monetize faster, China's to build the state, production and strategy faster; in the military sphere AI is an accelerator of command for one and part of a mobilization system for the other. The US weakness is the gap between corporate interest and public control, China's the rigidity and self-deception of centralized data — from which the host draws the theme of false correlations, from a Whoop band praising coffee at night to laws passed on AI's false conclusions.

Kokotajlo's main phrase: the danger is not evil AI but AI that thinks faster than a person, plans deeper, knows society better and creates texts, videos and ideologies of fear and hope. A person loses not a job but the inner position of a subject: he no longer asks what is true but what the system considers optimal. It is a transition from wisdom to optimization, and two "machines of destiny" — the US and China — are building systems to run the world.

In a study reported by Tech Explorer, in one version of the experiment the old GPT-4.5 model was judged human 73% of the time. An essay the host's nephew wrote himself to apply to Stanford was flagged by a machine as robot-written, while texts written entirely by AI now pass as human. The model itself puts its probability of detection at 12%, but for text, in the host's view, that is not a probability but an impossibility of telling.

Alexander Mashrabov's Higgsfield showed a full-length film of about ninety minutes made by AI at Cannes, with a script from a serious director and screenwriters. To the host it "makes no difference" who wrote a text — what matters is whether it is genuine in meaning and logic and free of manipulation; handmade at fairs he regards as the same kind of convention. Tatiana replies that manipulation should be taken as a given, AI already knows what she will like, and she would buy a two-thousand-dollar pendant only with a story behind it — a thing made by a machine has no soul.

An offer from a man who built a hundred agents does not interest the host: knowing the menu items, ten thousand opened chats and a configured API are not value. Implementing automation since 2002–2003, he has watched people retrain into different work, and he believes the winner will be whoever genuinely wants to understand what the client and the company need and which metrics to hit. Of a thousand companies promising to improve a sales department, one will offer something worthwhile, and those who teach "earning with OpenAI or on Amazon" are frauds.

Some speak of mass displacement, others call it heated management rhetoric; a chief operating officer who "sees no effectiveness in AI" is, to the host, just as much a populist. Jensen Huang reminds that generative AI arrived in real productivity only recently, so it is strange to blame it for layoffs that began earlier — "the facts are being fitted," Tanya sums up. The host is sure work will change and many will disappear from their roles, with the fanatics of their own business most vulnerable — ready to go into the red rather than admit the business is gone.

Running a company in Russia from 2020, the host built IT programs for hundreds of thousands of students with Ilnar Shafigullin and kept asking how anyone becomes a programmer in six months: at the first meeting none of twenty directors knew DevOps. The hype of "two thousand dollars a month in ninety days" has to be balanced against mathematics, algorithmic thinking and a broad outlook. Tatiana backs classical education: five years of university build connections in the head, after which even a course on launching an Instagram reads structurally.

This week's series of studies reminded that LLM judges are unreliable when evaluating complex research agents and leak sensitive attributes. The notion has appeared of an answer that is correct but useless to you: Sundar Pichai admitted Google's AI search can interpret a question too confidently, while AI Mode has passed a billion weekly users. When one model evaluates another, it is the same mistake as a person judging another without understanding who is in front of him — like ChatGPT recommending the host three events for a Saturday.

Having loaded a large core written in Codex into Claude, the host got, on one project, a pile of architecture errors and a proposal: "Since Codex created such a mess, I will build everything first and then Codex fixes the small things" — Anthropic pushing its line. A second chat, on two other repositories, suggested continuing in Codex and accepted his refusal without that narrative. Claude wrote that "traces of Codex" are visible at once, and the host concludes: think not in agents but in software created by AI, and code will soon stop concerning us — just as Tanya does not care about the code in Figma or Archicad.