It doesn't need any world model. What I'm seeing here is a repeat of a very old discussion between empiricism and rationalism. Having "a world model" is simply a tool, a typical human simplification heuristic to compensate for the limitation of our brains to uphold vast amounts of complexity simultaneously. As with all approximations of reality, models are by design very contextual and limited in their bandwidth of legitimacy. Very useful and practical but also extremely limiting straight jackets. I find it far more interesting having access to an ever-expanding amount of raw (empirical) data and then the freedom to tailor extract whatever operational model is needed and the most useful at that specific problem level.