* Notice in LLMs, we learn by written language FIRST. We expect, with no "world model" built up at all, the LLM to just get what we are talking about.
* Notice in humans, we learn by written language last.
* First we learn by 5 senses. We touch we feel we crawl we put things in our mouths. We hear. We smell. This gives us a strong world model. But required parenting and a lot of tolerance for “foolishness” for years. It’s an on policy learning algorithm for sure. Piaget calls this the sensorimotor learning phase. The objective function at this phase is entropy reduction in our T+1 prediction. We seek a high or the highest entropy s+a prediction in our world model and then execute that action to and learn with that greedy example. This is curiosity. Another thing to note is the brain is very sparse at this time. The brain is in its earliest development point so neurons are still being added during this phase.
* Then we imitate. We play make believe. We play doctor or nurse. Piaget calls this the pre-operational phase (2-7). Girls play mama, and put their dolls night night. Boys hammer nails into wood if that’s what they see their dads doing. We go up the abstractness ladder to “exploit” our newly acquired world model in an egocentric manner to get better at mimicking the behavior we see. We are not putting things in our mouth as much anymore. We aren’t randomly swatting at things. We have a “goal” to imitate as best we can, and just nudge our way there. Not worrying about hitting zero loss by any means, keeping “temperature” high, we are ok if it isn’t perfect.
* Then we continue up the levels of abstraction from merely mimicking to mimicking the goals of the teacher. 7-11. Piaget calls this concrete operational stage.
* Then finally we abstract to the final stage. We mimick the act of autonomy itself. Formal operation. 11-18. We aim to do what we want. Not what someone else does or has done.
* Then we chastise the LLM that it doesnt get key concepts like "drying k shirts" at the same time doesnt take k*# shirts time. Silly. We shouldn't test a fish for how well it can climb trees.. it wasn't designed for that.
* If we want an LLM to be human-like, we should design it to be human like from the ground up. It should start with ONLY sensorimotor training.. allow it to "play" in real-world or a sim.. allow it to hear, feel, touch, smell, and taste. Then give it a "mother" to imitate and "love". Then teach it speach and how to talk. Then teach it written language dead last vs. first.