When Google(+) Photos auto-tags my photos, still my data, although I'd admit the IA training requires a big amount of photos.
Looking at https://www.google.com/landing/now/#cards ; it's obvious a lot of those need online information. It's not obvious that the data source for requiring those online informations has to be in the cloud, or combined from multiple people's data.
It is, but that's pretty useless unless Google also knows so that it can send you the relevant updates.
>When it knows where I work and live to ask Google servers the weather or how long it would take to go somewhere, still my data used as the request base.
The weather, possibly (although again, it's not much use unless Google know enough to target the local weather at you). The how long to get there, not so much. One of the key things they use is info on where everyone else is. This is a classic case of pooling everyone's data in order to add value.
It's shortsighted to say that just because (example implementation) my device polls their server for $team, that it is the same as "google knowing" I like $team. The interface does not need to be authenticated, or keep logs. There is just a slippery slope between monitoring for diagnostics and abuse that incremental business value very quickly slides into tying team-specific view counters and a favorite team to my user identifier.
And Google doesn't necessarily know that you like $team. It knows that the email address you used to sign up to the alerts has requested updates on it. Is that really particularly scary?
Even with a locally installed AI, it will need to be trained on real and up to date data, it needs feedback on what was well recognized and on which data it failed, etc.
the training and update cycle can be done on an individual basis, but having a distributed system sharing a huge set of varied data would accelerate the process exponentially.
Ideally training data shouldn't be stored once it is ingested by yhe AI, but we'd need at least interconnected AIs dharing data from different users
Viewed differently, the value of N specifies how uniquely identifiable the user is.