We want to arithmetic code against some other AI (which will achieve much more compression than lzip or whatever). But, the very fact that we need to decode using that other AI implies that the other AI is in some data center that needs to upload a bunch to the internet. So, we want the other model to be sufficiently weak that it’s not that as bad if it is stolen.
I don’t quite understand: why do we need weaker model during compression? Why don’t just use something like “gRPC with compression”?
We want to arithmetic code against some other AI (which will achieve much more compression than lzip or whatever). But, the very fact that we need to decode using that other AI implies that the other AI is in some data center that needs to upload a bunch to the internet. So, we want the other model to be sufficiently weak that it’s not that as bad if it is stolen.