When device owners purchase equipment with any exceptional amount of unused memory storage, does either iOS or the installed apps have the ability to supplement on device edge models with eg larger parameterized models or additional models for an ensemble solution?
Thanks for the reply, happy to ”talk shop” just fun seeing if it might make these a better product. I think the common rule of thumb is that powerful AI is most primarily attributable to memory scale, eg parameterization, and in that case the bottleneck is less storage memory driven than it is GPU memory in inference, albeit some chips optimize for these in different manners.
What I was trying to get at though is that any sizable storage memory untapped on an edge device is a potential resource to supplement intelligence features not only because it allows you to save bigger models, but also because a longstanding alternative to bigger general models is ensembles of specialty models. Thus anywhere the edge device relies on specialized models, those can be reframed as ensembles of specialized models either aggregated or routed based on a higher tier model, where it can only be expected to help performance to have a larger selection of specialty models to include in such ensembles.
not sure if it remains common to include specialty models as opposed to just relying on some master model on edge device settings, but if you are looking for how untapped storage memory could become a resource to improve features that could be a starting point.
Hey @NicholasTeague -
Your query is a little vague, but an interesting one, so I’ll bite. (in a proportionally generalized manner…lol)
Powerful open-source models exist that can be operated locally, but I don’t think they’re iOS compatible. (At least not directly.) They are typically installed on more powerful machines (ie, the type that would be running MacOS). They’re primarily designed by/for privacy hawks, so tbh, I haven’t heard of anyone coupling them with one of the usual suspects (eg, GPTx).
That said….under the right circumstances and including the following assumptions…it would hypothetically be technically feasible to have them engage with each other.
Assumptions:
Hope that helps. Cheers.
Unused memory allocations to Apple Intelligence?
