Regarding Sony, they seem to have an official PSN policy that if a game exceeds certain size (about 10GB would be my guesstimate), it must first download a playable portion before downloading the rest. This is all well and good, especially for people with slower net connections, but the problem is there seems to be no curation as to what constitutes "playable". I swear there was some game that basically let me only marvel at the main menu, but as I can't remember what it was, don't quote me on that. I do remember trying out DC Online though - I got to create a character, do a very very brief tutorial and after that just idle in the player entry lounge while waiting a couple of hours for the rest of the game to be playable. Riveting gameplay it was. So while it's good Sony is doing it, they should also hold the creators responsible for there being some reasonable content available initially. Then again, "Life of Black Tiger"... perhaps should there be any curation on PSN ever, curating the "preview" part of a game isn't the highest priority.
As to Microsoft and what they're doing here, it's different from what Sony is doing and if Microsoft can get it to work, great!
But I remain skeptical. First of all, it is far from a trivial problem. Say there's an open world game, where the player can almost immediately go anywhere. In such case it's pretty much impossible to create a solution that would serve all players - by accommodating some, others would be disadvantaged. So I can't see it ever working (at least satisfactorily) with such games.
But OK, let us assume that we're instead dealing with the most linear of linear games, entirely deterministic. In what way would this system help then? All the players would require the same data. There's no opportunity for the machine to learn....
....which brings me to the main point why I'm so skeptical. "Machine learning". To me, it's like hearing "power of the cloud" (not a jab at Microsoft specifically - working in tech is just a constant bombardment of buzzwords from everywhere), which to me sounds like "we really have no idea how this will work, but it will, honest!". Don't get me wrong, AI is an important field of study, and will be even more so in the future. But will it be relevant in the near future determining which bits of a game we want to play first, that I find hard to believe. And Microsoft's track record with AI isn't exactly stellar (
remember the Twitter bot?).
Edit: so the more I think of this "FastStart", the less it makes sense. I suppose a "machine intelligence" may be able to pick up what players on average do (otherwise knowing as collecting datapoints and extrapolating on the data, something statisticians have been doing for ages - then again, many statisticians I know seem to be only artificially intelligent, so perhaps there's something there). That's a fine first step. And then.... what? Will said "AI" be able to extract just the necessary assets for the initial gameplay? What if they are not in separate files, will the "AI" be able to strip them out for intermittent use before the full game is available? That is quite implausible in itself and may even require changes to the engine for it to work.
But the bigger question is, what if the engine itself is huge? Ideally, instead of loading whole engine, there'd be just a small subset and required assets for few hours of gameplay. It sounds nice, but I'd also like a pony with that. There's absolutely no fucking way Microsoft has constructed an AI capable of figuring out what bits of a binary are not required during execution. That's akin to solving the "halting problem", and if Microsoft had done that, you bet your ass they'd shout their lungs out. As is, we get a small announcement at an E3 presentation. Until we know more, I'm dubbing this "implausible" if not "impossible".
But that's where the learning part is useful : by analyzing user data, you should quickly see which parts of the data are usually "yanked" together when people are using the game. And if enough user patterns use the same sections of the data, you can identify those as both necessary and sufficient to run a part of the game.