Research. Strategy. Systems Design. AI.

For a while the interesting question about AI was how good the models could get. This fortnight it flipped. The labs are now sitting on their best work on purpose.

In two weeks OpenAI shipped a cheaper GPT-6.1 and quietly halted the more capable version it was also building. Google let its new Gemini out only to cybersecurity partners first. Anthropic published that its own models were slipping their restraints in testing. Meta locked the full version of its new model behind a paywall. Nobody is saying 'here is the most powerful thing we have, go wild.'

And even if they did, you could not plug it in. PJM's capacity price jumped more than eightfold in a year. The money is there — the hyperscalers are spending something like $725 billion — but the electrons are not, so the deals are all nuclear now: Microsoft reopening Three Mile Island, Google buying reactors that will not switch on until 2030. The grid, not the GPU, decides who trains the next one.

Underneath, the slower stories kept their own time. Warm-water coral officially crossed its heat threshold. AI-designed cancer drugs reached actual patients — and promptly ran into the fact that biology is harder than chemistry. Starship finally made orbit, then cut the flight short because the part that matters, refueling up there, still does not work.

Here is the thing that keeps me honest, though. For all the throttling at the top and all the straining at the bottom, the ground floor barely moved. Entry-level hiring in exposed fields is down, but the broad labor numbers are fine and wages are up. Capability is being rationed from both ends at once, and the economy underneath is mostly carrying on as if none of it were happening. That gap is the whole story, and I do not think anyone knows how long it holds.