OpenAI shipped GPT-6 Astra, which puts it back on top of the benchmark tables for now, and Greg Brockman closed the press briefing with "Welcome to the AGI era". OpenAI's own charter defines AGI as systems that outperform humans at most economically valuable work, and the one benchmark OpenAI built to measure exactly that, GDPval, was missing from the launch materials. Make of that what you will.
My problem with the AGI claim is not that the model is unimpressive. It is that AGI started life as a thought experiment: a calculator beats you at arithmetic and a database beats you at memory, so surely we can build something that is as capable as a person in the sense any normal person would recognise. What we have instead is a system that does a growing list of things better than people, or just faster, which is not the same, and passes a growing list of tests. That tells you it can do pieces of what people do. It does not tell you it is what people are. The technology might scale to that, it might not, and nobody knows, but we are plainly not there. There is an old theologians' joke that the Jewish problem is you wait and wait for the Messiah and he never comes, while the Christian problem is he came and nothing changed. The AGI discourse has both problems at once. None of which means this is a scam. Anyone still saying it does not work is not paying attention.
The release I think matters more in practice is Meta's, because its new Muse Spark model is now being named alongside OpenAI, Google and Anthropic as a peer. A year ago Meta and xAI had both fallen off the leaderboards; billions of dollars of acquisitions later, they are back. Which is the point. Everyone is building roughly the same thing, roughly the same way, and getting roughly the same results, with no network effect to lock anyone in. That is the path to a commodity, not to pricing power.
Follow the money anyway. OpenAI's ad business hit a $1bn run rate, about 200 days after launch, which is impressive for a new line and still implies only a few hundred million actually booked this year against a $2.5bn target. Nvidia confirmed it is buying Hugging Face for $12.9bn, its largest outright acquisition. And the Journal reported that Anthropic signed a $35bn compute deal with Lambda, where Nvidia backs Lambda, supplies the chips, and holds the lease on the Texas data centre itself. Nvidia has become the industry's central bank: it uses its balance sheet to create as much competition and fragmentation as possible among its customers and suppliers, which pumps up demand while cementing its own monopoly at its own layer. The Information added up everything Anthropic has committed to in eleven months and got $517bn and 14.8 gigawatts. None of those numbers have a timeline attached; they are neither annual nor one-off, which makes them hard to think about. Leverage is wonderful in a rising market.
Google, meanwhile, won the remedies round in its ad-tech case: the judge rejected the break-up of its ad exchange and accepted behavioural fixes instead, a year after the search case ended the same way. The decade-long push to use antitrust against big tech is now more or less over in the US, and the EU version has produced rules that annoy everyone and change little. The people who shouted loudest for it never had a coherent theory of harm, a legal approach that matched what the law says, or a model of how these products work that did not make engineers laugh. The deeper problem is structural. Platforms shift every fifteen years or so. It takes five to see who won, five more to see the abuse, and five to get a case through. By the time the judge hands down remedies the next platform has made the whole thing look quaint.
Shorter notes. New York banned student-facing AI through eighth grade for a year, and Los Angeles blocked it on all school devices. I learnt arithmetic without a calculator and then used one to go faster at the hard stuff; that is probably the model. Fei-Fei Li's World Labs shipped Atlas, a model that builds 3D scenes from a few photos; OpenAI gets similar results by having its model drive Blender and Unreal as tools, and which approach generalises is a live question. The FTC and 22 states sued Amazon over its ad auctions, claiming it told advertisers it ran a second-price auction while quietly inserting its own reserve price to push bids up. Amazon says reserve prices are standard and nobody paid more than they bid. Either someone at Amazon or someone at the FTC has made a very large mistake. The EU classed ChatGPT as a search engine for the purposes of its platform rules, which tells you how regulators think about new things: as the old thing with a new name. And someone breached an identity verification firm used by Hertz and others and is selling 153 million North American driving licences on the dark web, including scans with the infrared and UV layers used to check that a licence is real.
Two human stories. The New York Times found that Nairobi once had something like 40,000 people writing essays for Western university students, $40 to $70 a paper, and the industry evaporated within two years of ChatGPT. The only work left is "humanising" AI output to get past AI detectors. And Semafor traced a network that paid gig actors $26 a video to read AI-written political scripts to camera, because a real face defeats every synthetic-media detector YouTube has. Tens of millions of views. YouTube pulled the channels, the incentive is untouched.
Remember GoPro? Thirteen billion dollars at the peak. It sold for $285m to a photonics company that wants its optics for defence and data centres. People used to call it a network. There was no software, Chinese makers could do the hardware just as well, and most people did not have the use case. I remember a sell-side note where the analysts all wore one for a week and concluded that either they led uniquely boring lives or nobody needed this thing.