Nice work from Microsoft Research, including it being open source, and with open sourcing of the training dataset to come (which I suspect will actually be the most valuable contribution long term)