That's my best guess. Since Siri and Photos now processes more locally (especially on M1's dedicated ML cores), all the training improvements and testing probably happen locally as well.
It makes sense it would regularly download updates for ML models. I'm more curious about what data gets uploaded back to Apple...
The privacy implications of this are dubious. There is as-yet no good research showing how these finetuned models can be used to reveal exactly what photos you have in your photo library, but many experts believe that won't be the case for long.
Really? Apple can essentially be asked by Law Enforcement for answers to questions like "Give me all the IMEIs of people in location X, radius 10 miles, that appear to have pictures of cats"
But experts believe that techniques might be developed in the future to figure out based on the numbers what data must have been in the training set (and hence your photos). Nobodies managed to do that yet.
If you train a model A) cats, one for B) dogs and one for C) dogs and cats. Is there any research which asks if you can do a network subtraction C-B = D, and get a sensible network D than can loosely do what A does?
IMO this is the kind of statement that should have something to back it up.
My guess is dictation.
My Trial folder is about 500MB. I've never enabled Siri on my computers declining the option during the setup at first boot. But I do use dictation quite a bit.
If 500MB is the cost of having dictation work solely on my laptop without hitting the internet, I'm OK with it.
The problem is that people love to invent conspiracy theories about Apple because they have an axe to grind.
They get opportunities to do so because of two things: Many of Apple's daemons and other services are not publicly documented. And because it's running a flavor of Unix, it is a lot easier for amateur computer sleuths to poke about and discover what's running that people outside of Apple don't know about.
Operating systems of the past were more opaque, so we just had to blindly trust the manufacturer. But people can see things better now. It would help rebuild trust if Apple would have a standard reference online explaining all of its OS bits.