The thing is though, LLMs do whatever you tune them to do. If you train them on a sycophantic corporate drone butler dataset, you get the average assistant model that's obviously a bad fit for this use case. If you train them on something else, you get whatever you want, even someone that challenges you. I wouldn't be surprised if having some sort of simulated soulmate partner thing that also does the job of an educator and life guide will be the norm in the future.