This is a pretty important point, we have relied on all LLMs being broadly similar to each other (even to the extent that prompting is compatible across models). That may start to change with reinforcement learning.

→ View original post on X — @emollick,