The RLHF is happening too late i think. I think the reinforcement learning needs to be during the initial next token prodiction. On that note we need something

The RLHF is happening too late i think. I think the reinforcement learning needs to be during the initial next token prodiction. On that note we need something to represent a complex world state than just language.