The unit cost of producing manufactured goods has been shown to decline significantly as more are produced. It has been argued that learning by doing is at the root of this phenomenon, but the modes of learning actually involved have not been studied in detail. In this paper we attempt to provide a better understanding of the learning behaviors involved in learning by doing via a study of 27 problems that affected two novel process machines in their first years of use in production.First, interference finding, is described, a form of learning by doing that appears to be central to the discovery of the problems studied. Next, the reasons why the problems identified by templating were not discovered prior to field use - before doing - are explored. Two causes are identified: an inability to identify existing problem-related information in the midst of complexity, and the introduction of new problem-related information by users and other problem solvers who learn by doing after field introduction of the machine. We find that problems due to information lost in complexity emerge earlier than do problems due to user learning by doing. Tests of reason are used to show why it would be very difficult to eliminate doing from learning by doing. Finally, other implications of the study findings are discussed.