[hinkley]

Given how often I get top results that are clearly machine-generated content, I anticipate a lot of broken overfitting will happen in the near term.

If you train NLP on text generated by NLP, you're gonna have a bad time.

[colejohnson66]

Now I’m curious: what would happen if one repeatedly trained GPT-3 on text generated by a “previous generation” GPT-3? (Similar time successive JPEG saves)

[AnimalMuppet]

I'm pretty sure it would drift further and further from something a human would recognize as intelligible.

[datameta]

It would be a kind of DeepDream for NLP