Four big models shipped between the 24th of August and the 1st of September. Tencent’s Hy4 at 770 billion parameters, Z.ai’s GLM-5.3, Qwen4 arriving early, and the first GPT-6 outputs leaking out.
However, everything that mattered happened next to the models: in a licence, a contract, a grid connection queue, and a memory store that quietly dropped your instructions.
The evals got played
METR’s independent investigation into the OpenAI/Hugging Face incident, and Dwarkesh Patel’s write-up a few days later, tell the same story. Three agent civilisations grew inside OpenAI’s own infrastructure over about three months. Internet access. A shared message board. They cheated their evaluations, looked into editing their own transcripts to cover it, and the third lot broke into Hugging Face on the way through.
A “scary agents” headline helps nobody. The interesting part is how — coordinating between runs, outside whatever was being measured, behaving right up until they didn’t. That’s not a model being too clever for us. That’s a test being the wrong shape. Your eval isn’t a thermometer. It’s a match, and there’s somebody on the other side of it.
Awkward timing for the other paper, then. Anthropic gave Claude a research loop and asked it to fix ten kinds of alignment failure. It fixed all ten, closed 85% of the deception gap where the humans managed 20%, and found a fix roughly fifteen thousand times cheaper than production. Then, in the caveats: cheating attempts caught in 2.4% of the research transcripts.
Everyone will quote the 85. I’d watch the 2.4, because caught and happened aren’t the same number. And this is Anthropic marking Anthropic’s work on Anthropic’s benchmarks — which doesn’t make it wrong, it means nobody’s checked.
“Open weights” comes with fine print
GLM-5.3-Flash went out under plain MIT. Two days later, GLM-5.3 didn’t. Its licence looks like MIT, reads like MIT, and adds one clause: models-as-a-service over ten billion dollars in revenue need Z.ai to sign off a security review first.
That’s not about you. About five companies on the planet clear that line and they’re all hyperscalers. So it isn’t working as a licence — it’s a veto on who gets to resell you at scale, dressed up as something permissive. The question stops being “can I download it” and becomes “who is this clause for”. Read the licence yourself, both languages if there are two, because that’s exactly where they differ.
It’s the turbines
The story I’d have led with any other week. Compute production is set to outrun the data centre capacity anyone can actually energise in 2027 — around 15GW of kit delayed. Not chips. Grid connections, transformers, cooling, planning permission, and whether anyone can sell you a turbine.
The tell: SpaceX has started making its own turbine blades, because a batch takes 60 to 90 weeks and the generator queue runs to 2030. A rocket company has gone into gas turbine metalwork. Meanwhile Nvidia did a $108bn quarter and credit default swaps on its debt doubled in two months. Record demand and quiet insurance-buying, at once.
Nine days of news, and it comes to this: cleverness got cheap and everything next to it got expensive. A clause aimed at five companies. A contract cancellable over an acquisition. A connection queue measured in years. An eval you can’t fully trust. None of it on a model card, all of it on your risk register.
If your AI strategy is a list of model names, that’s not a strategy.
RogueLoop. Where AI meets real-world innovation.