[krustyvonklown]

Personally, I find it rather humorous that we've moved from the fear that AI generated output would corrupt training to the idea that it is essential to training. Reality itself has not just a left bias but a bias to fundamentals. Bootstrap from fundamentals without introducing arbitrary error and you have the superior system; it just may not be highly compatible with a trash ecosystem.

[dminik]

I mean, I'm not sure that's the correct read on this.

If you want an Opus class model, it makes sense that you would train on what Opus outputs. But, if you want something better than Opus, training on the same data that Opus was trained on with the same architecture will only result in an Opus class model. Then, if your dataset also contains Opus outputs, many of which are wrong, then it makes sense that the model would have reduced performance.

All this to say that I don't think there's such a thing as a "Model Collapse," but there likely is a "Model Stagnation."

[krustyvonklown]

A model trained on all the data X was trained on should be improved to the extent that X is already out of date. A model trained on X itself has all the errors of X and all of it's own. Society itself seems to show that model collapse is entirely possible today and was presumably a problem in the past given the significance placed on citation and going to original sources that predates obsession with credit.