When will unsupervised learning be as good as supervised learning?

Andrew Ng has said many times that the holy grail of deep learning is unsupervised learning, i.e., the use of unlabeled training data. Why has there been no real progress in unsupervised learning?