Google Now looks like Google's most magical, artificially intelligent product. And in some senses it is, but it is also in some senses Google's most manual product. Each category that Now knows about and makes magical suggestions around is the result of a human making an editorial choice to create that category. Google Now gained cricket scores not because it 'learned' that there was cricket but because someone at Google wrote a recipe for cricket.
Obviously, you can't do this for everything - Google Now cannot scale to the whole internet. It can do a lot, and certainly enough to be a great product, but it can't do everything, not until we have 'true' HAL 9000-style AI, which is a long way off. The underlying domain is not actually infinite, but like Borges's library it might as well be for any manually-edited project - Yahoo's manually-edited directory peaked at 3.2m sites but had become absurd long before that. So there are gaps - you're making something that looks like AI but isn't. Google Now covers the gaps by keeping quiet, where Siri covers the same kinds of gaps by making jokes (that was probably a better product management decision - if you baffle Siri you get a blank stare or a laugh but if you baffle Now you never know about it).
Contrast this with Google Maps, which is also partly a manual product. Hundreds or thousands of people drive Streetview cars down roads and hundreds or thousands of people check the images and edit the maps. There's lots of machine vision too, as well (and crowd-sourcing), but the back-stop is lots of paid employees. The algorithms alone aren't good enough yet. Editing the internet by hand was just too big, but editing maps isn't - it's just very expensive. Google is willing to spend that (and Apple is going to try too).
So, mapping the internet by hand is impossible - you need algorithms. But mapping the world with algorithms alone is also impossible, for now - you need people to look at the data too.
I see Apple Music as sitting somewhere in this second category - manual curation at scale. Here the problem to be solved is that the commodity streaming service's 'search box plus 30m tracks' offers no way to discover anything you might like but haven't already heard of. You need personalised suggestions, from somewhere. But how do you suggest things from 30m tracks? It is in principle pretty easy to get some data about what people like, especially once they start listening (and so stop lying). But how do you know what aligns to that preference data?
Pandora applies an algorithm. Apple Music is trying the Google Maps approach: "doing this at least partly by hand is not an impossibly large problem, just a very big one". Just as Google manually adds metadata to the raw maps (street names, one-way streets, business names etc) Apple manually places tracks into playlists with its own metadata. Then it takes the interest data it has about a user and suggest 10 or 20 or 30 playlists, and, hopefully, there's a good chance that enough of them will be right. (It covers the gap that Siri covers with jokes and Now covers by keeping quiet by overshooting - it's probably OK if 15 are not quite right if 10 are on target.)
How many playlists you need? 5,000? 10,000? With a few hundred editors that's not actually an impossible challenge, just a big one. And a cynic might suggest that really, you only need a few hundred for the vast majority of music listeners, if you do them right.
Conversely, Google Search is itself a manually edited product - it gives you ten results and makes you pick the right one. It's curated by the users. So, perhaps the real point is that today, without HAL 9000, almost everything has both human editors and algorithms - it's just a question of how you combine them most efficiently.