[currymj]

The odd thing is, the motivation for the research was not really to build this stuff as a product.

If you're going to do machine learning dealing with images, music, or text, a completely natural thing one might want to do for many reasons -- to solve "real problems" -- is to try to approximate the probability distribution of the data.

Having done this, it was seen as a kind of strange curiosity that you can actually sample from this probability distribution, which amounts to generating new images or new songs or new text.

It was only recently and somewhat unexpectedly that these generations actually got good enough to be interesting to people in themselves, and then people started trying to turn them into products.