I think it's better to think of the hutter prize as a challenge to see how close people can get to efficiently approximating AIXI without the exponential cost of AIXItl. The fact that winners don't generalize well is just that they're not great approximations, because the behavior is different in the limit.