>I mostly agree, which is why I question their strategy of even "competing" at all.
If it makes their camera "smarter", it's a win. If they can make Siri do something more than "start a timer", then it's a win. If they can have images translate text more accurately, it's a win. There's a lot of things that an on device AI could help users without having to do all of the power hungry creation of a model or the fine tuning. They can do that in the mothership, and just push models on their device.
Not everyone needs to do AI the way you are trying to do it
I think that's a mistaken way of viewing it. Apple's failure in the gaming space is entirely a matter of policy; you look over at the Steam Deck and Valve is running Microsoft IP without paying for their runtime. Some people really do get their cake and eat it too.
Any of the aforementioned libraries could make their camera smarter or improve Siri/OCR marginally. The fact that Apple wasted their time reinventing the wheel is what bothers me, they're making a mistake by assuming that their internal library will inherently appeal to developers and compete with the SOTA.
The reason why I criticize them is because I legitimately believe Apple is one of the few companies capable of shipping hardware that competes with Nvidia. Apple is their only competitor at TSMC, it's entirely a battle of engineering wits between the two of them right now. Apple is going nowhere fast with CoreML and Accelerate framework, but they could absolutely cut Nvidia off at the pass by investing in the unified infrastructure that Intel and AMD refuse to. It also wouldn't harm the customer experience, leverages third-party contributions to advance their progress, and frees up resources to work on more important things. Just sayin'.
If it makes their camera "smarter", it's a win. If they can make Siri do something more than "start a timer", then it's a win. If they can have images translate text more accurately, it's a win. There's a lot of things that an on device AI could help users without having to do all of the power hungry creation of a model or the fine tuning. They can do that in the mothership, and just push models on their device.
Not everyone needs to do AI the way you are trying to do it