Actually the opposite. The quality of that other big hit (Dalle) has gone noticeably down IMO. I feel like they have over trained it, trying to compensate for t

Actually the opposite. The quality of that other big hit (Dalle) has gone noticeably down IMO. I feel like they have over trained it, trying to compensate for the initial shortcomings, and in the process of trying to solve those by throwing more and more data at it the machine has forgotten how to deviate from sterile stock photos and do anything interpretive. For example, I asked it for "Mother of Exiles" (a reference to the inscription at the base of the Statue of Liberty), in return it showed 4 generic stock photos of female characters, sometimes with a visual prop to vaguely insinuate motherhood but certainly nothing inspiring and certainly nothing to do with exiles. Compare this to the early days, when people were throwing natural language statements at it like "power bottom dad for the people" and it was spitting back results that felt like it truly understood what any of those words meant.