What is it, what should it be capable of, where will the training data come from, who's working on it