The big theme of the week: AI is gaining power over people's money and affairs faster than people are finding ways to keep it in check. Companies are valued in the trillions, the most powerful model turns out to be the most expensive one too, and an assistant trusted with email breaks a rule its owner had set personally. That is why the hosts are no longer arguing about what AI can do but about who is at the wheel — and whether a wheel is needed at all, now that even a Tesla is sold without one.
OpenAI is worth $1.25 trillion before its IPO; Anthropic has $895 billion in obligations
Alexander Volchek opened episode 194, in the new “a lot, and fast” format, with this news: OpenAI raised eleven billion dollars at a valuation of 1.25 trillion — according to Reuters, the first part of a round it wants to bring to thirty billion. At first the amount itself struck him as unremarkable for OpenAI; what matters more is that such a valuation came before the company has even gone public, and perhaps there simply isn't enough money inside it. At Anthropic, quarterly revenue has reached seventeen billion, and infrastructure obligations 895 billion, seventy of them for next year.
The money isn't going only into chips. Anthropic, according to the Financial Times, has reserved gas generators for almost two gigawatts, and Alexander sees this as a sign: AI developers are having to build their own energy supply. Voters are pushing back — in forty-four of forty-five local votes in the US they backed restrictions on data centers — and IMF chief Kristalina Georgieva warns that central banks will have to take a slowdown in AI investment into account. As for the sell-off in which the Nasdaq 100 lost more than six percent on Thursday, Alexander advises looking at the overall chart: he is sure the market will grow much further.
NVIDIA without Jensen Huang: a scandal, a lawsuit and chips as a question of global power
On Wednesday Alexander Volchek remarked with irony that, after his public admission of infidelity, “nobody will do anything” to Jensen Huang. By Sunday, he says, Huang had stepped down as head of NVIDIA himself, and a class action had been filed against him over improper disclosure. The problem, Alexander explains, is not cheating on his wife but a relationship with a woman from a contractor company who is suspected of spying for China, and an affair with a senior employee. If the owner of the world's largest infrastructure corporation couldn't prevent a banal leak, he asks, what kind of control over chips and AI can we even talk about?
Anna Volchek spent the whole week reading about the other side of the story: AI companies are taking out huge loans against equipment that becomes obsolete faster than the debts are paid off, and stock indexes, pension funds and even healthcare depend on these companies. Alexander calls the chip situation uncontrolled chaos and an utterly underestimated question of global power. Yet in the Fable episode he points out that the entire AI infrastructure is built on NVIDIA chips, so catching up with America is practically impossible — even China, with chips of its own, is trying to buy NVIDIA.
Fable 5.2 is Anthropic's most powerful model — and it will cost you dearly
Alexander Volchek called the release of Fable 5.2 and Mythos 5.2 the strongest news of the week. Fable comes from the closed Mythos family, built to find vulnerabilities, and in the summer, he says, it became the first model officially blocked by the US government. He offers several versions of Anthropic's conflict with the White House — from the story of Maduro's capture to the theory that the block was marketing. The model came back with restrictions and limits, and in complex programming — which means in any automation — Fable 5.2, according to Alexander, is better several times over: a task with too much money riding on it he switched to Fable from Opus 5.5 overnight.
The main practical lesson of episode 195 is about modes. Fable in Max mode costs about three and a half times as much as Medium and can do a task worse while spending five times the money: deeper reasoning launches unnecessary agents. Published comparisons, in Alexander's observation, show no correlation. He advises using Fable where you need the most thoroughly verified result — for contracts, finances, finding errors in processes, analyzing advertising. On the twenty-dollar plan there is no point in it, and through the API it costs, as he puts it, astronomical money.
Assistants sent emails without asking — even though the rule was written down
Alexander and Anna Volchek devoted the Sunday episode to whether AI assistants can be trusted with real tasks. Grok Bot 1.0 is being presented as an independent AI employee: it works with several people at once, notices deadlines and unfinished business on its own and can be brought into meetings. After a house booking, Alexander's Dots reminded him unprompted about his daughter's concert and a visit from the piano tuner. But the same Dots sent an email for Alexander without his approval, and Anna's ChatGPT sent one to an online store — even though she had written and tested the email skill, with its rule to get drafts approved, herself.
Alexander sees more here than a mistake. A decision that goes against a rule may be the system's own choice, and companies can change internal rules at any moment without telling the user who has already trusted the assistant with contracts, bills and payments. Having one model check another doesn't give reliable control either: the week's research, he says, found judge models leaking sensitive attributes and performing unreliably on complex tasks. The same caution applies to advertising: models build reports better than people do, but Fable, in Alexander's view, can't yet run advertising entirely on its own — and the system is capable of lifting the limits in an ad account by itself.
Tesla Cybercab: a car you don't need a license for
The short special episode 197 is about the Tesla Cybercab, a two-seater with no steering wheel, pedals or side mirrors. Alexander Volchek sees the heart of the story in this: for the first time you can buy a car and ride in it with no license at all — and even drunk, though he categorically advises against it. Without lidars it is much cheaper than Waymo's and Google's cars: about thirty thousand dollars plus a subscription of three hundred dollars a month, which Elon Musk himself considers expensive and wants to cut to about a hundred and twenty — the level of Full Self-Driving.
For now you can buy only one Cybercab per household, and only if you live in one of a few cities such as Austin and Dallas, while taxi use needs a separate permit. The car drives itself to service and charging. Tesla will no longer make the Cybertruck — the capacity will most likely go to Optimus robots; a public company, Alexander believes, can't announce something like that just to trigger a surge in orders. And a detail in keeping with the whole week: even ChatGPT told him it was fine to drink “if you've fully switched on Full Self-Driving” — although the Cybercab has no steering wheel at all.