This idea that LLMs will get progressively dumber because of a training feedback loop is silly. It's also a straw man. The concern isn't so much that that the

This idea that LLMs will get progressively dumber because of a training feedback loop is silly. It's also a straw man. The concern isn't so much that that theses systems will move backwards, but rather that they will start saturating, and gradually degrade in terms of ROI (measured in terms of both cost/energy and the complexity of building of dealing with systems to deal with all poisoned data). The commenter above was stretching a bit when they said "collapse", but I think they got the basic gist of the issue right. Which is basically: diminishing returns.