The problem we address is how to predict strong ground motions for very large earthquakes from observations made of such motions produced by events of moderate size. The discussion is in terms of two basic rupture models, a W model in which slip is controlled by fault width and an L model in which slip is controlled by fault length. Because mean slip is observed to increase linearly with fault length, a long earthquake cannot be modeled as a series of shorter events placed end to end. Rather, to explain the correlation of slip with length, a W model will predict that stress drop increases with length, whereas an L model will predict that stress drop is constant but rise time (slipping duration) increases with length. Thus, a W model predicts that peak and rms accelerations and peak and asymptotic particle velocities increase linearly with fault length. An L model predicts that rms acceleration and asymptotic velocities are independent of length but that the peak velocities increase with the square root of length and peak acceleration with .