A mixed approach is a great point though and I think I could mostly be happy with that.
A mixed approach is a great point though and I think I could mostly be happy with that.
You misunderstand point 1. Three-digit billion parameter models like GPT-4 (or even 3.5) aren't going to fit on modern iPhones, not even close. Even Llama 70b requires 35GB of RAM at q4 quantization, and that's just for the model.
Compare that to the iPhone 15's 6GB and you'll see the problem. Apple isn't about to announce an iPhone with 6x as much RAM as any previous model. I'd be shocked if they even doubled it. Their local inference models are going to be tiny and limited, which is fine, but that means they have to go to the cloud to provide all the features people are expecting.
Wouldn’t memory be a much bigger bottleneck? You don’t really need a particularly fast a CPU/GPU to run most basic models that couldn’t even fit into the amount of memory that Apple is offering on their devices (especially if you still want to run other apps).