This is a cool way to conceptualize it. When conversing with humans, you can explore various parts of the semantic space, but the process also involves a range

This is a cool way to conceptualize it. When conversing with humans, you can explore various parts of the semantic space, but the process also involves a range of top-down and/or bottom-up mechanisms that facilitate a kind of 'hidden navigation'. LLMs, on the other hand, lack this capacity but with the right prompts, one can simulate it by knowing how to steer, which is sort of what prompt engineering feels like to me.