The power law of practice states that performance on a task improves as a power-law function of the number of times the task has been performed. In this article we describe recent work on a model of this effect. The model, called the chunking theory of learning, is based on the notion of chunking. A limited version of this model has been implemented within the Xaps2 production system architecture. When it is applied to a 1023-choice reaction-time task (encoded as a set of productions), task performance is improved (measured in terms of the number of production system cycles). Moreover, the practice curves are power law in form.