What happens when generated data of one LLM becomes training data of another LLM? Turns out that models start forgetting the real distribution and as the proc

What happens when generated data of one LLM becomes training data of another LLM? Turns out that models start forgetting the real distribution and as the process repeats models develop dementia.