Fair point, GavinB. If you defined "efficiency" in terms of achieving the most benefit from an eyeball using the fewest computations, then I'm not sure whether a Bayesian superintelligence could approach "efficiency". Maybe the theoretically optimal program to run on the eyeball's input, would itself require exponentially vast brainpower to calculate!
But I would concede a much higher chance that a Bayesian superintelligence could get bored at exactly the right time, than that it could simulate all possible universes.
Do you think it would be a "good" thing if the average intelligence level was bolstered by 40 points? Can you think of any ill side effects from this possibility?
Information and calculation beyond a certain point may simply becoming "boring."