It's not collapse, it's monoculture. It can be sustained but it becomes stagnant. The introduction of bias in training data seems to be even worse. Ironically,

It's not collapse, it's monoculture. It can be sustained but it becomes stagnant. The introduction of bias in training data seems to be even worse. Ironically, it seems that most models nowadays do it. Either by contextual annotation or patching. It's a mistake.