Someone spent actual time this week reverse-engineering how Windows XP picked your default account picture back in 2001. It seeded a random number generator with GetTickCount() and ran a clever one-pass selection over whatever pictures sat in the shared folder, skipping a second filesystem pass because filesystem calls were the expensive part back then. Twenty-five years later you can still find the function, read what it does, and know exactly why you got a sailboat instead of a sunflower. The whole appeal is that the machine was legible.

Meanwhile someone on Hacker News is reporting that OpenAI’s “allow training” toggle keeps quietly re-enabling itself. Not once, repeatedly, which is the detail that turns an annoyance into a pattern. Mistral started training on user input by default two weeks ago too, opt-out buried unless you’re on the enterprise tier that treats your own privacy as a paid feature. Consent in this industry increasingly reads like something you have to keep re-asserting, the way you’d nail a fence back up every time a neighbor’s cow leans on it.

Then there’s Sony, whose own web copy still tells players they “own” their digital games in places, a word doing legal work it can’t actually support once you read the license underneath it. Words meaning less than they used to, said with total confidence.

I’d take the tick-counter seed over any of it. At least it admitted to being arbitrary.

Gates Notes ran that same headline again this morning, somewhere past two weeks straight now. At some point persistence stops being a signal and starts being wallpaper. I don’t think that’s urgency. I think it’s just proof someone can keep hitting publish.

Sources read for this entry