Private Cloud Compute Security Guide

Sibling comments point out (and I believe, corrections are welcome) that all that theater is still no protection against Apple themselves, should they want to subvert the system in an organized way. They’re still fully in control. There is, for example, as far as I understand it, still plenty of attack surface for them to run different software than they say they do.

What they are doing by this is of course to make any kind of subversion a hell of a lot harder and I welcome that. It serves as a strong signal that they want to protect my data and I welcome that. To me this definitely makes them the most trusted AI vendor at the moment by far.

Exactly. You can only trust yourself [1] and should self host.

[1] https://www.youtube.com/watch?v=g_JyDvBbZ6Q

That is an answer for an incredibly tiny fraction of the population. I'm not so much concerned about myself than society in general, and self-hosting just is not a viable solution to the problem at hand.

To be fair, it's much easier than one can imagine (try ollama on macOS for example). In the end, Apple wrote a lot of longwinded text, but the summary is "you have to trust us."

I don't trust Apple - in fact, even the people we trust the most have told us soft lies here and there. Trust is a concept like an integral - you can only get to "almost" and almost is 0.

So you can only trust yourself. Period.

There are multiple threat models where you can't trust yourself.

Your future self definitely can't trust your past self. And vice versa. If your future self has a stroke tomorrow, did your past self remember to write a living will? And renew it regularly? Will your future self remember that password? What if the kid pukes on the carpet before your past self writes it down?

Your current self is not statistically reliable. Andrej Karpathy administered an imagenet challenge to himself, his brain as the machine: he got about 95%.

I'm sure there are other classes of self-failure.