* the snake eating it's own tail This feedback loop, when you extend LLM to the horizon, is my primary point against this approach. When 90% of the new trainin

* the snake eating it's own tail This feedback loop, when you extend LLM to the horizon, is my primary point against this approach. When 90% of the new training data is from content (a worse version of) it previously generated you get a negative feedback loop to zero quality.