A knowledge representation system provides an important service to the rest of a knowledge-based system: it computes automatically a set of inferences over the beliefs encoded within it. Given that the knowledge-based system relies on these inferences in the midst of its operation (i.e., its diagnosis, planning, or whatever), their computational tractability is an important concern. Here we present evidence as to how the cost of computing one kind of inference is directly related to the expressiveness of the representation language. As it turns out, this cost is perilously sensitive to small changes in the representation language. Even a seemingly simple frame-based description language can pose intractable computational obstacles.