I just had a thought about Pinterest's plan to feed everyone’s pinned images into an ML engine: lots of the content on Pinterest is now synthetic (“AI”-generated) imagery. In general, training models on the output of "AI” systems is a really bad idea; it noticeably degrades the model, a phenomenon called model collapse.
Will they only train on images pinned before slop became ubiquitous? Or will they try to filter the synthetic images somehow?
https://mastodon.social/@angusm/114143550897370005