> I guess it really shows that we have no way to predict the future. IDK, I think that in the AI circles the idea that this would happen was clear long before

> I guess it really shows that we have no way to predict the future. IDK, I think that in the AI circles the idea that this would happen was clear long before it actually became possible. I recall a university workshop probably 20 years ago with a discussion about generative art of various forms (music, visual, etc) - as some genres of computer-generated art were a thing even back the - and the main conclusion about genres of "mainstream" art (which were far beyond our reach at the time) was that the difficult part is "judging taste"; in essence, if we got a reasonably accurate way of evaluating that one song/picture/sculpture/whatever is "better" than the other, then the rest is "just" relatively straightforward engineering. So the consensus prediction already two decades ago was that it's going to happen. And while we don't have that "magic taste function", what we have made is an "indistinguishability" function, which can fulfil a similar role, and is effectively what the "photographer" in this story has used - if a model generates items that are (almost) indistinguishable from "good portrait photography art", however you define that, then it can make "art" that is as good as the rest of it. It can't really go beyond that, because if it accidentally generated something better then the model wouldn't be able to recognize that it's better, only that it's different, and discard that option, so improving on the status quo would require curation from someone with "good taste", but it apparently can beat mediocre photographers by matching the results of good ones.