I played many hours with AI Dungeon's GPT-3 based "Dungeon" model, instructing it to write sci-fi fan-fics. Its output never ceased to amaze me. It's junk by the standard of high literature, obviously. Yet, when you consider the number of bad fan-fics in existence, I'd say GPT-3 probably ranks high in a relative comparison, at least two order-of-magnitude better than my creativity (since its training data already included a huge collections of fictions, including web fan-fics and numerous text adventure games, and the job of GPT-3 is mixing and rewriting them, it shouldn't be a surprise).
I'm never a fan of the "AI" hype, but GPT-3 decidedly changed my mind. If I can be attracted to read, for hours, the meaningless stories generated by a mindless language model in an experimental stage - still, with a quality arguably outperforms many unskilled humans, I'd say I'm convinced that there are lots of promising applications of machine learning (for better or worse), even if it's only capable of feeding low-quality content to the people. 1984 got it right.
No reason the AI composing tools have to be limited to pop music. All the most "authentic" types of music could be assisted by AI just the same way. I would have thought it would be particularly useful for things like scoring film and TV.
Scoring is tricky because the music has to be synced to the action. Once you've done all that feature engineering, it's just a small bit of work for the composer to score it. The AI bit is mostly useful in extrapolating detail from the melody to a MIDI orchestra.
