This is what happens when everyone on the thread wants to virtue signal.
Let's not do that!
What is Apple's SINGLE comparative advantage? ie what do people pay extra for?
It's UI (broadly understood). The significance to Apple of AI is how it allows for modified UI. Everything else doesn't matter. Whether OpenAI or Google or xAI or Anthropic supply your chatbot or the intelligence in your <medical app/law app/programming app/logistics app> doesn't matter; what matters to Apple is that you're running that app on your mac and your phone (and maybe soon your watch).
Which means that what Apple has been concerned about over the past few years is not creating "the best model in the world", whatever that means (a title that seems to change every month). Instead they've been focussed on what the previous paragraph implies:
1. what does AI mean for UI generally? Obviously you want a "smarter" Siri, you want better image generation, but think more broadly. Could AI classify your email? Organize your files? Figure out the "best" 20% of your photos and throw away the rest?
2. how much AI should be running locally? Running on Apple servers? Running on 3rd party servers? No-one can be sure, so they've been building a HW platform that is as capable as any of adapting to however this turns out. I've pointed out before that people are experimenting right now with running massive (like 400B parameter) models on iPhones by streaming in weights from flash. Is this a cute trick? Or is it harbinger of the future?
3. maybe we'll land up with a model that's bifurcated, but so seamlessly that no-one notices.
For example one problem that's only going to get worse is the issue of "personality". If the big model supplies the personality you care about, that's a problem when the big model updates.
A second problem is intimate personal knowledge. I don't want reading recommendations based on what 300M idiots are reading, I want recommendations based on my skill level, my interests, and what I do vs don't already know.
A third problem is on-going learning. This still hasn't been cracked; the best we really have is huge contexts that just keep growing. And maybe it can't be cracked at the hypermodel level?
What all three suggest is a scheme whereby an extremely personalized model lives on my device, knows all about me, sports a personality I like, and is continually learning [that's what it would *mean* in part, to know all about me]. This model in turn handles trivial questions and tasks locally, but also has a switchboard that can route questions/tasks as appropriate to the hyperscalers.
Point is: all this is very different from what is being discussed in this thread and in the general media chatter. As always people can quote "skate to where the puck will be" a thousand times while never understanding what it means even once.
I strongly suspect Apple is skating to where the puck will be, and in a way beyond everyone else.
Apple Intelligence was something of a flub, but then so was Copilot everywhere. No-one *knows* the optimal paths going forward. If we score
- Apple Intelligence was (possibly) a dumb move, trying to say "we're relevant, we're doing something" without the important parts yet ready. As I always say -- you CANNOT let your marketing team drive the schedule! Every time Apple (or other tech companies) do this the result is tears.
- Much of the behind the scenes Apple AI stuff is quite adequate.
The translation stuff, for example, is about Google level (though not as many languages). The current spell check stuff is, I think, about the best its been -- but apparently ruined by an unrelated stupid bug to do with touch tracking that makes it appear more stupid than it is. The image recognition stuff is good enough and keeps improving.
Along with that much of the language stuff is lousy, at least in part because the best language tech we have, in LLMs, isn't yet capable of learning. So email filing feels dumb because the machinery, such as it is, for learning the patterns in my email is terrible. We have adequate "understanding" of each email coupled with inadequate "what to do once you've understood the email". Same thing with the writing stuff, and even with the spelling stuff.
On the other hand, the connecting APIs to LLMs (as demo'd at WWDC last year) and the recent integration of LLMs into XCode (initially as "assist", now as agentic) seem to be going reasonably well, especially when you understand that the goal of Apple is not to be leading edge, it's to round off all the sharp edges, ensure the safety, security and ease of use, that get ignored by the people six months to a year ahead of Apple who are just experimenting wildly with what is possible.
Note the common thread in all this. Most of this wouldn't be fixed by a better model, that's not where the problems are. Which means it's simply not useful to say "Gemini is so much better". It's not clear that Gemini *is* so much better, not in the ways that are going to matter going forward.
I certainly don't *know* that Apple have a solution to all of this (they seem to have personalization under control, but the technical problem of on-going learning remains). But I think they understand what's going to matter *to them* in five years more than the public do. And they are executing to achieve those five year goals better than people appreciate.
