I think Kaggle competitions are probably underrated as far as educational value (i.e. there is a lot of "Kaggle is not data science" sentiment). The forums are full of good stuff to learn regarding practical machine learning. It's important to frame that knowledge in the context of those competitions, but a lot of it translates well to "real world" problems. You probably don't want to use a blend of 30 very different models in production, but a lot of the tuning strategies, etc are very useful.
