Last I checked, feeding the output of an LLM back into its training data leads to a progressively worse LLM. (Note I'm not talking about distillation, which involves training a smaller model, by sacrificing accuracy. I'm referring to an equal or greater number of model parameters)