Impy

I think there's this mismatch between what's in the raw datasets, and what's actually being trained on that laymen don't understand. Because when you talk to people, and ask "Well, why are you lobbying for the massive rights of conglomerates?" They usually say something like, "No, I'm defending the little guy." The problem is that they're really not. Little guys don't score high enough on the aesthetic scoring to be trained at all in 98% of cases. What's bothersome, I think, is that when those guys who were associated with RunwayML put up the "Am I being trained?" website, they could have put the whole argument to bed. Instead they offered a search index with raw metadata and no context at all, and arbitrary image reverse engineering and detection tools that they themselves even admitted were unreliable. Sure, very cool experiment. It's just that the timing was awful for the product they delivered. These are people who know what the metadata means, and how it's used. And it's hard to say if the whole thing was an intentional act of fraud, cynical misinformation, or lazy application design. I lean towards the latter, because data scientists are some of the worst engineers for anything other than model weights that you'll run into. It's like I was saying earlier about how disciplines don't always transfer between engineering fields. But the net result is an angry mob that's angry about the wrong thing, because whether naively or not, they've been totally mislead by people who should fucking know better.