FWIW I believe the other-models are what have been strongly selected for, and the self-model is just an epiphenomenon arising from the relative ease of modelling one's self after having gained the ability to represent models of others' selves:
I appreciate the link to Shannon's paper. I've played a browser-based version of such a machine/game and it's pretty eerie.
Long ago a friend of mine had this cat, she liked to sit on your lap, but hated it when you got up, she would always jump down the instant you started to rise.
However...
She could tell from the pattern of muscle tensions in your legs whether or not you were really getting up or just moving a bit to rearrange your butt or whatever, and she would only jump if you were getting up. She was flawless at this. In fact, a few times she jumped down a split-second before I knew I was getting up. Like Bruce Lee, she could detect and respond to the intention to move, even before the motion was consciously known to the mover himself (me, in this case.)
cf https://news.ycombinator.com/item?id=23475069