A task is k-replica hard if measurements on at most k copies at a time need exponentially many queries, and is r-replica easy if allowing joint measurements on r copies permits efficient learning. For each k>1, are there tasks that are k-replica hard but (k+1)-replica easy, or (k+O(1))-replica easy?