I think the claim is that Uism^2 gets increasing returns from scale. But also there are a bunch of other increasing-returns theories, and you’re not sure which is right.
And so if you’re maximizing expected value (or something like it), what you end up doing is randomly picking a theory and then giving it the whole universe I think (because increasing returns make this better than splitting the universe according to theories).
But actually even higher EV than that is just to pick the “most likely” (or something like it) theory and give that the universe every time.
But then it’s not just quadratic theories: there are theories with even more increasing returns. (But I’m not sure that this changes the optimal strategy.)