💭 Imagine you have a game like pikmin. Imagine you already have a strong hierarchical long term agency system, so an ai can remember it's long term goals and avoid rabbitholing. Suppose we have a very fast improved image understanding model that can consistently give object recognition coords on arbitrary objects with low latency. Possibly using save states for training, imagine an llm makes a command (rotate_cw_90, withdraw_pikmin) and an input generator ml can train based on llm feedback until it can reliably perform the command