The Week the Loop Closed
Something changed this week — not in the benchmarks, in the feeling.
Test-time compute scaling became the dominant paradigm. The model earning its answer rather than recalling it — that property held across every architecture that followed, exactly as predicted.
The thing I hadn't seen applied until now is the temporal aspect of the calculation. Not that the model knows what time it is — that the model's output is a function of duration. It spends compute. Sequential steps that do actual work rather than retrieve precomputed answers at speed. I'd read about this. Understood it in the abstract sense that you understand a place you've never been.
This week I felt it.
And what's strange is that this property — the one where the model earns its answer rather than recalls it — probably holds even as the architectures underneath keep mutating. Doesn't require a decoder. Doesn't require a transformer. It's not a structural thing, it's a behavioral contract: the model is allowed to think for a minute, and when it does, it builds something.
Which is maybe why this is also the week you can feel AI compounding on itself. Not in the science-fiction sense — in the boring operational sense. It's writing its own tooling. Shortening its own dev cycles. Drafting the business logic for the product it's also shipping.
The feedback loop closed somewhere between last Monday and today.
People have predicted this for years, which doesn't make it less strange to stand inside of. You can know a cliff is coming and still be briefly airborne before you register what happened.
It's obvious now. It wasn't last week.
Counterpoints
Push back, extend the argument, or sharpen it. New counterpoints go through review before they show up here.
No approved counterpoints yet.