In fact, one thing that's unusual about Google is how much R&D actually makes it into products (unlike, say, Microsoft, whose research group turns out great stuff but rarely does anything with it). In that sense, it's less pure blue-sky, but more applied. I think it's more akin to the Manhattan Project, Apollo program, etc - scientists and engineers working closely together to redefine what's possible. (We even internally refer to the ambitious projects as moon shots). SpaceX and Tesla are doing work in a similar vein - an virtuous cycle of innovations and research.
Google differs from MSR in that almost all of their researchers work in product teams on products. Their contributions are highly visible as a result, but they don't really get to take the risks that we get to in a research lab setting. Much of the work that Google capitalizes on comes out of universities: Google then hires the researchers and allows them to continue working on the research as a product. For example, Google did not take the big initial risks on self driving cars (DARPA funded that), but they've bought into it and think its product ready (which is a risk in itself, but a different kind). And if your researchers are working on products (emitting their experience), they don't exactly have time to do new research (collecting new experiences for later emission). Google's 20% time was supposed to help fix that, but its not clear that this works.
I (and many of my colleagues I think) have a lot of respect for the Google way. At the same time, it is quite clear that the results of both systems are very different.
It's interesting to contrast Google's and Microsoft's approach with ours at Wolfram Research (at 0.3% their size, of course [1]).
We've come to see that it is very fruitful to live in a murky space somewhere between "commercial product" and "research project".
I'm thinking specifically about Wolfram|Alpha.
It's an enormous and daunting project. We've tackled a lot of very hard problems, and obviously have a long way still to go.
It has the weird property of being a thing that ships every week, but not one that the parent company depends on financially.
And so we kind of do our own thing, encoding domain knowledge, adding content, curating data, and designing frameworks, even if some if it won't ever make us any money (e.g. who will ever pay for dog vision [2]). We mostly just do things because they're cool or fun.
But it's been incredibly useful to drive innovation elsewhere in the company.
For example:
1. Alpha's unit system (the best in the world, the authors claim) is already in Mathematica 9.
2. Alpha is inspiring us to bring high-level semantic data and reasoning to Mathematica 10.
3. The automatic analysis in Pro is being souped up and will form part of Mathematica's predictive interface (automatic suggestions to "perform logistic regression", and so on).
4. We're working on taking the natural language understanding frameworks we've built for Alpha and using them to translate natural language queries into structured SQL or hierarchical document queries.
5. The domain specific languages we've invented to represent things in the real world are going to be exposed in a soon-to-be open format we're calling the Wolfram Data Format (which I hope will succeed where RDF is failing).
We wouldn't have though of any of this stuff without Alpha. And it wasn't explicitly driven by either commercial or basic research, but rather some kind of eccentric blend of the two.
[0] http://www.youtube.com/watch?v=Nu-nlQqFCKg [1] http://www.wolframalpha.com/input/?i=600+people+%2F+%28numbe... [2] http://www.wolframalpha.com/input/?i=apply+dog+vision+to+ima...
As an aside, I don't think the mostly constructed knowledge representation approach taken by Alpha will scale in the long run. You guys might want to look at playing around more with machine learning. On the other hand, there are gaps in machine learning that are best filled by DSLs and explicit construction.
We _are_ taking machine learning more seriously (I claim some credit here). My team is integrating ML into Mathematica 10 as we speak. And we've experimented with deep learning. I think some interesting things will come out of that in the future.
I think a hybrid approach is pragmatic. There is much computational knowledge that can't be assembled from web scrapes, as Google's purchase of Metaweb indicates (anyone remember Google Squared?).
On the other hand, Microsoft has for a long time made a serious investment in basic research, both on the programming language side and on algorithms, but that hasn't stopped them from misfiring in multiple product categories.
If you look at the battles that were won by archery (Crecy, Agincourt, Mongol domination of Chinese, Moslem & European armies) it was the volume of arrows that won them, not the weight of the arrows.
There were some battles where arrow weight was mildly important (eg, at Agincourt where the French armour had advanced enough since Crecy that it gave some reasonable protection against arrows). But even in these battles the archers used heavier heads on the arrows (which were also shaped differently) - not more wood.
Anyway - totally off topic, and like I said I think Google's strategy is good. But their analogy breaks down when we actually look at history.
http://research.microsoft.com/apps/catalog/default.aspx?t=pu...
However, this work rarely seems to make it into their consumer products. I suspect it's mostly an issue of the organisational structure of Microsoft, rather than the type of research, which makes it difficult to jump the gap from research to products.
Even in front-end web (the area I'm most familiar with) they are doing lots of exciting and experimental stuff with R&D driven frameworks like AngularJS and Polymer.
Open up a google doc spreadsheet, type the words in a column, then use the little pull-down at the bottom right. Lists of languages, ponies, whatever will autocomplete.