“More stuff that sounds like the Sex Pistols is a much easier problem than inventing the Sex Pistols. Would it know this isn't in the training data, but people would like it? Can you build a model of taste that can evaluate something new, or can you only evaluate how well it matches the things we already know people like?”

Yes. Do we want something that isn't in the training data? And this is, you could argue, the sort of the essence of creativity is something that is different but good. And the theoretical problem there is that you can very easily imagine a model making more generic pop music or more generic images like that. More difficult or certainly a very different level of abstraction would be for a model to say, hey, we've all been listening to disco and Progrock and ABBA and it's time and the economy's turned down and it's time for something completely different. So here are the sex pistols. And again, more stuff that sounds like the sex pistols is a much easier problem than inventing the sex pistols. Would it know this isn't in the training data, but people would like it? Can you build? Build a model of taste that can evaluate something new, or can you only evaluate how well it matches the things we already know people like? This is, of course, I often go back to thinking about AlphaGo, that AlphaGo could invent a new move, but it could test it. This is the validation point. It could make a new move, a new…