i think people are moving to python in the competitive predictive analytics community because of neural nets (theano, nolearn, lasagne, neon, keras, etc.). i have yet to see any useful deep learning packages in R. at this point using R almost requires basic proficiency in libraries like data.table making the learning curve even steeper for non native users.
> using R almost requires basic proficiency in libraries like data.table
I think a lot of people just use plain old data frames, together with plyr 2 – which is really not any harder to work with than Pandas' `DataFrame#groupby` functionality, though I will grant that having to look for it in a separate library is a bit annoying for newcomers.
OTOH, In Python you will find that everything in the NumPy/SciPy universe works one way (vectorized functions, NA objects, its own float type and so on) and everything in plain Python another way. Not exactly ideal either.