This article discusses a dilemma in teacher education. On the one hand, there is evidence that changes in teachers' behavior can be brought about through the use of relatively simple training procedures (e.g., minicourses, microteaching). These procedures typically focus on directly observable, low-inference categories, such as wait-time and question patterns. On the other hand, studies of effective instruction have shown that good teaching cannot be identified solely with these low-inference, discrete, and trainable behaviors. Rather, good teaching is a more complex and abstract phenomenon that has to do more with inferred qualities and abilities, such as classroom management, lesson structuring, and an “active teaching” style. It is not possible to train teachers directly in these high-inference categories. A solution to this dilemma is discussed by considering approaches to teacher education which attempt to bring about the gradual development and understanding of the process of effective teaching.