> We got skidillions of data, but that alone isn't going to tell you
> where else we should be looking for more, or what different kinds,
> etc. etc. Plus it's nice to be able to winnow down your options
> because it's usually expensive to get.
I think that's the whole point: it's a different way of approaching the problem, and given that many 'scientific discoveries' were essentially mistakes, why not? If we can dispense with your "Don't do it because"s (i.e., you can always say that something doesn't tell you where to look for me -- so keep looking! ... and well, what if it's not as expensive as you think? And on ...), then I don't see it as being a particularly _bad_ course of action. You certainly can't stop "old school" science, but there's plenty of room for brute-force attacks on problems that haven't gotten much attention recently with this kind of thing.
Ok, it's not 'Science' and it's not 'Theory' ... but is it useful?
Steven Wright has a joke about 'I have a map of the world, where 1" = 1"' and that's essentially the Google Ahah! Moment: why subsample when you have essentially "all" the data?
We ran into this in the late 80s with stock market data; you still have macro modelling to deal with, but micro modelling is a thing of the past.
/jordan