The Google model is better though, in my view. The problem with the older style "lab" approach, typified by Microsoft Research, is that it basically recreates a

The Google model is better though, in my view. The problem with the older style "lab" approach, typified by Microsoft Research, is that it basically recreates academia inside a company. There's then often huge difficulties integrating resource output into actual products. MSR does OK but is still quite notorious for producing long term expensive research projects that never get integrated into any product (Singularity/Midori being perhaps the most famous), whereas the Google model doesn't have that problem. Google has an extraordinary high rate of incorporation of research ideas into products. This is better for a lot of reasons, obviously, it makes it easier to justify funding the projects. But it also ensures that the researchers truly understand the nature of the problem they're solving. Seeing actual translation queries on Google Translate helps direct the next piece of MT research towards what kind of things people really use the tech for. Building actual big data systems helps train the researchers in how to work with huge data sets and puts them on research paths they otherwise would have been unable to explore. Basically it recognises that a hard division between research and practice doesn't really make sense when working at the cutting edge of technology. You want the constant feedback from real world practice to guide the theory.