Agreed. NLTK's value is not that it should be your go-to source for NLP algorithms -- when there's a specific task you need to accomplish, there'll be a specific solution by now that works better than NLTK.

NLTK's value is that it shows you how to write NLP algorithms, and gives you an understandable starting point when you need to do something that nobody has implemented yet.

agreed: I also learned a lot from nltk too, esp. freqdist/probdist as well as the classifiers in nltk and sklearn