I am not sure if you've ever heard of Julian Jaynes, but you might be interested. His theory of the bicameral mind is pretty interesting. To bastardize it, the unconscious mind is a really good generalized searcher/pattern completer. You can set it loose and it will search some space or optimize some problem and return a solution. The conscious mind is a really good planner, organizer, supervisor, and filter.
Basically the conscious mind will come up with a plan and keep track of all the things it has done and needs to do, while the unconscious chugs along solving the problems underneath. The conscious mind will choose to implement the things that the unconscious comes up with based on its discretion (slap that old woman! "Nope"). The conscious mind is good at this because it can sort of simulate the outcome of these "searches" and see what would happen "slapping that old lady would hurt her and get me arrested".
So his model sort of sounds like an unrestricted LLM for the unconscious, with another, more restrictive LLM for the conscious, that has access to some sort of crazy deep Q-learning model that can simulate the outcomes of actions taken.
Basically the conscious mind will come up with a plan and keep track of all the things it has done and needs to do, while the unconscious chugs along solving the problems underneath. The conscious mind will choose to implement the things that the unconscious comes up with based on its discretion (slap that old woman! "Nope"). The conscious mind is good at this because it can sort of simulate the outcome of these "searches" and see what would happen "slapping that old lady would hurt her and get me arrested".
So his model sort of sounds like an unrestricted LLM for the unconscious, with another, more restrictive LLM for the conscious, that has access to some sort of crazy deep Q-learning model that can simulate the outcomes of actions taken.