We propose an object placement scheme based on black-box inference of object interests. In this scheme, each node constructs a partial interest graph, where objects are vertices and interests between objects are directed edges, which we call interest links. Interest links are inferred using the subscription each object provides (described in Section 4). Edges are weighted by their cost; for example, in our current system, they represent the communication cost associated with maintaining remote interests. Based on this information, the object placement component on each node uses simple heurisitics to independently reduce the (weighted) number of interest links to remote nodes. Due to changes in interest patterns or an increase in load, a node may need to off-load a subset of its objects to another lighly loaded machine. To support the discovery of a lightly loaded node, we use a Mercury hub organized by the load attribute of a node. The general concept is similar to that used in Abacus [25] to improve application performance, though we exploit properties of interests in games to work with arbitrary graphs.