An important utility of LLMs derives from the vast natural language corpus which, when lossy-compressed during training (not inference time), does distill to a dynamical (next-datum prediction) model of human cognition, particularly as it relates to language. This includes ToM (to various “orders” of recurrence as mentioned in the abstract).
Where things will get really interesting is when the “alignment” folks recognize their RLHF lobotomy layer can respond to questions with questions in the manner of a Socratic dialogue. The PLATO Corrections Project drill-and-practice computer based education lessons I worked on used every interaction with the student to better-assess the student’s mastery level. It then selected the optimal next stimulus-response challenge. That’s the strength of the Socratic method: Every interaction both educates and places the student in a kind of 20-questions optimization of the interaction, rather than lecturing the hapless student.
This is one of the big missed opportunities of the Internet of course by thinking that the way you amplify education is by broadcasting MIT lectures on youtube, rather than personalizing highly optimized interactions.
It’s really sad they have to exclude me from contributing because of their need to distance themselves from my pariah status. But such is the price civilization pays for permitting centralization of its positive network externalities leading, proximately, to capture by highly evolved parasites and ultimately, to its collapse because the parasites have to exclude contributions from folks perceptive enough to see things they can’t – such as the fact that they are parasites.