Access to the latest foundation models, which can’t be run locally. AI feels like it’s in this really weird place where the latest Claude model sets expectations that can’t be matched by an on-device model.

Even Apple Intelligence is getting a lot of negative feedback by the review crowd due to its limitations like being unable to summarize a very large document (which is pretty much the point of such a feature).

The problem is that AI has few well-defined use cases and a mountain of expectations, and this really shows in the execution by these companies. It’s hard to build good products when the requirements are “we don’t really know”

Looking at the logs for Apple Intelligence on my iPhone 15 Pro Max, almost everything is actually run remotely