I improved the model based on Simplicity Theory by giving certain agents the role of W-machines (generating the world) and another the role of O-machine (observing and describing the world). The results are better according to this initial test. But I can't go much further at the moment with my limited resources. I urgently need GPU, tokens and also some $ for my time, for debugging, fine tuning, and to conduct larger and more meaningful runs. Have a great weekend!