The doom cycle came round again this week, and I want to resist the easy cynicism about it. A 27-year-old researcher who had only recently moved from OpenAI to Anthropic quit with a public warning that the labs are racing toward self-improving systems nobody can control. Then Dario Amodei published an essay calling for the industry to pace itself, with outside evaluators given employee-level access to models, and within hours Sam Altman said he agreed and Elon Musk posted "Dario is right". Google and Meta made polite noises. Trump said the opposite.
The cynical reading is obvious: regulatory capture, fundraising theatre, the incumbents pulling up the ladder. I think that misses what is actually going on, which is that these people are in a tangled position. There is a Wodehouse line from 1934 where Bertie Wooster reads about the chaps trying to split the atom and notes that they have no idea what happens if they succeed, and would feel pretty silly if the house went up. That is roughly where the labs are. They have no real theory of how these models work or how far they scale. They believe the risk is real, they say so, and then they go straight back to building faster and raising more money, because the alternative is someone else doing it. Both halves are sincere. That is the uncomfortable part.
Useful context for the same week: Anthropic published a report on the misuse it has caught, including a cell in northern Yemen using Claude Code to write guidance software for a rocket, which it then test-fired, apparently unsuccessfully. And OpenAI was so short of compute that it stopped selling its $200 a month plan a week after launching its new model. Marginal cost is real for this technology, which matters for everything that follows.
Then the maths. OpenAI announced that a swarm of its agents had solved a Millennium Prize problem in 88 hours. Two mathematicians who had spent months on the same problem using AI, one of whom works at Anthropic, say OpenAI heard they were close and threw a research team and millions in compute at it to get there first, then pressured them over credit. OpenAI denies it. Twenty-five Fields Medallists signed a letter saying the labs' goals and the field's goals are badly misaligned. I am not qualified to judge the mathematics. But I notice two things. First, if telling people you have a promising approach means someone with a hundred times your compute will finish it before you do, people will stop telling each other things, and open science dies quietly. Second, in pure maths the answer is often the least interesting part; the new ways of thinking you develop on the way are the point, and a machine that brute-forces the result skips exactly that. A cancer researcher who wanted to be the one to find the cure would be putting ego before lives. Mathematicians have a better argument, and they are making it.
Meta launched Muse, a personal agent that reads your email, books your travel and pays with your card, free with limits and then $20 or $100 a month. Nobody knows what the right shape of a consumer agent is, but there is a lot of careful work here, and Meta will point the same growth machine at it that took Threads to half a billion users. I will say more about this when there is more to say.
Numbers worth keeping. Microsoft plans to go from 12 gigawatts of data centres to 38 by 2032, with AI chips going from a sixth of that to a third. Bernie Sanders announced a Ban Artificial Superintelligence Act with 20-year prison terms; it will not pass, and the point is that he thinks it will play, which tells you how toxic the word has become in American politics. Bending Spoons bought Miro for $1.36bn, about 90% below its 2022 valuation, which is the pandemic SaaS bubble, AI eating the category, and plain consolidation all at once. And Mistral raised €3bn, the largest European round ever. It has not been near the frontier leaderboards in a long while and has leaned into "sovereign AI" for enterprises that want to run models themselves. I have mixed feelings about that as a Frenchman, but a billion dollars of recurring revenue is not nothing.
The UK's air traffic control fell over again on Tuesday, more than a thousand flights cancelled, three years after the last time. The system that failed turns filed flight plans into something controllers can use, and replacing it has been discussed for decades. Memo to everyone in the Valley who thinks the world will be rebuilt in eighteen months: it is much more complicated and much slower to change than you imagine.
Apple's new CEO launched the iPhone Duo, a $1,999 folding phone, and did the usual Apple thing of arriving late and claiming to have done it properly, with real attention to the hinge and the crease. It is an interesting object, something between an iPhone and an iPad that asks again what a computer is. It is also the last platform, not the next one.
And a small story I keep thinking about. A 21-year-old Chinese student fed the new Murakami novel through DeepSeek in 2,000-word chunks and had a complete Mandarin translation online four hours after the Japanese release, for under $2. Readers are arguing about whether it is good. It will get better, fast. I do not think we have begun to understand what it means when the wall between the English-language internet and everyone else comes down.
Finally, WordPress. Automattic's board voted to put Matt Mullenweg on leave and he was back in charge within 33 hours, calling it his fifth coup attempt. There is a whole class of lifestyle tech companies, Craigslist, 37signals, this one, that found a niche too deep to dislodge, made a founder very rich, and opted out of the rest of the industry. They tend to come with drama.