[abeppu]

I think a key difference here is that with chess, 'goodness' is defined by winning. With content generation, the training methods point towards some form of comparing the generated thing to some observed data, but the 'goodness' of the content from the perspective of potentially competing with or displacing human creators is "do people like to consume it?"

If one trained using e.g. a tiktok like dataset showing viewer response measurements for each video, and do conditional generation on those response values ("prank video watchers are highly likely to watch the full video"), are we really that far from a system that learns to generate content that attracts and hold eyeballs? Not so long ago there were a lot of concerning trend pieces about how youtube had a network of creators making bizarre, disturbing or transfixing videos being watched entirely by young children. Before that, it was clickbait listicles. "Bad" content that can get eyeballs can still wildly steer what humans create and consume. I'm wondering if in 2 years we'll have an enormous number of short videos that we all agree are "bad" but which are nevertheless constantly watched.