Not to bring LLMs into everything but I think they make using boring stuff more likely. Since training data contains a lot of it and less of the shiny new stuff.