[thesz]

You are right about "LLM improving themselves" is impossible. In fact, LLM make themselves worse, it is called knowledge collapse [1].

[1] https://arxiv.org/abs/2404.03502

[int_19h]

That paper again.

LLMs have been trained on synthetic outputs for quite a while since then and they do get better.

Turns out there's more to it than that.