@lain @sun but according to the paper even non language models trained on reality derived data converges towards a point but thats not that surprising since universal phenomena like time (audio, video) and geometry (audio, visual) exist and all data is already tuned for human perception and describes stuff from this planet