This is something like the holy grail for me. I am a machine learning research, using the python/scipy/theano stack. One problem performance wise is that I can

This is something like the holy grail for me. I am a machine learning research, using the python/scipy/theano stack. One problem performance wise is that I can get my models reasonably fast with Theano--but as soon as I need a python for loop (e.g. by using Theano.scan) performance dies. As soon as you have a couple of nested (e.g. due to a Jacobian matrix instead of vector) performance is virtually gone again. I'd love to use numpypy, and the way I see it not much is missing. Keep up the good work! I wish I cold fund you! :/