What does more data get you beyond accuracy? I think it opens up certain model classes -- like online regression -- which have proveably low error rates with lots of data, but my argument is that you don't need "the entire web," as another commenter suggests, to be good enough at, say, speech recognition. I could definitely be mistaken though...
I agree that quantitative research must be a core competency -- I'm a ML engineer at a company that's heavily invested in its research team -- and it is most certainly not one of Apple's focii. What's stopping Apple from building that competency by acquihiring the talent, though? This is no different from what Google has done over the years...
At some level, it all boils down to increasing accuracy, but the point I was making was just that doing that seems right not to be best accomplished with loads of data. If you look at the deep learning work that's been big lately, you have models with millions of free parameters. By necessity, you need a lot of data in order to constrain a model that big.
Even speech recognition has gotten a big boost recently from taking a "simple model with massive data" approach.
I'm not convinced that these approaches are sufficient to give you some sort of human-level AI. I'm pessimistic on the timeframes for that in general. And I'm sure there are areas where they fail, and maybe someone else comes along with a better idea, but Apple's not working on that either.
Certainly, they could acquire their way to competency, and I'm certainly not going to chime in with the proverbial "Apple is doomed...DOOOOOOOOMED". The only thing really stopping them is interest. But it takes a while to ramp up from getting results from a research team into making those results into a product.
I agree that quantitative research must be a core competency -- I'm a ML engineer at a company that's heavily invested in its research team -- and it is most certainly not one of Apple's focii. What's stopping Apple from building that competency by acquihiring the talent, though? This is no different from what Google has done over the years...