I tried briefly and was awfully disappointed. For starters, the documentation was crap. Followed closely by the fact that I wanted to take the norm of a vector and had to try 3 different norm functions before one worked. Further, I had a sparse matrix and all the functions -- of course, this was mentioned nowhere -- magically desparsified my vectors. Apparently if you want to operate on sparse matrices/vectors, you need to pull in an entirely separate import and potentially change the name of all functions.
I was left completely underwhelmed.
Not to mention that installing matlab was a pain, but installing scipy was an order of magnitude worse. ports failed to work, the binary distributions wanted much older python than I had, etc.
It would be awesome to see them seriously take on matlab, but I have to conclude they have a very long way to go.
If their gpu implementation is anything like their sparse vector/matrix implementation, matlab will remain the language of choice for a long time.