Some really interesting comments below, but the question is still open and requires investigation. I hope a few students pick it up. I liked the discussions on context length generalisation, the fact that we typically train these models as bandits (even when we do RL, which is

→ View original post on X — @nandodf,