There's an upper bound for the sum total of human text output on the Internet. when large language models run out of training data, we can only train on large language outputs.