The general observation is of course true, but we're talking about a particular instance, and I think it could be a pretty good fit here. The expectation when

The general observation is of course true, but we're talking about a particular instance, and I think it could be a pretty good fit here. The expectation when you first try modern image or text generators is sth like 'this will replace all artists, programmers, writers and probably most white-collar workers'. Upon closer inspection, it turns out that the image generators fail on details like hands, and the text generators are too unreliable for production-grade code or writing articles. So the result in the short term is that you still need the same people with the same skills babysitting the models' outputs. They've just become a bit more productive. But in the longer run, even plain efficiency gains can have a large impact, and it seems like we might not need any more fundamental breakthroughs to make those tools actually production-ready (at least largely) on their own. The hands in the images, the bugs in the code and the hallucinations in the texts can probably be fixed with clever, but ultimately straightforward, iterations over the existing models.