This book by a Nobel laureate in economics begins with a brief exposition of Kenneth J. Arrow's classic paper "The Economic Implications of Learning by Doing" (1962). It shows how Arrow's idea fits into the modern theory of economic growth, and uses it as a springboard for a critical consideration of spectacular recent developments that have made growth theory a dynamic topic today. The author then develops a new theory that combines learning by doing (identifying it with the concept of "continuous improvement") with a separate process of discrete "innovations." Learning by doing leads to a fairly smooth reduction in labor required per unit of output, tied to the rate of gross investment in new capital equipment. Innovations arrive at random; when one of them happens, the labor requirement takes a jump downward. This new model, simple as it is, does not lend itself to self-contained solution. The author accordingly presents the results of a series of computer simulations that exhibit the variety of paths the new model economy can follow, showing, among other things, that early good luck can have a persistent effect. The book concludes with some general reflections on policies for economic growth, drawn not from any one modeling exercise but from general experience with a variety of growth models. Of the four chapters of this book, the first two were presented as the Kenneth J. Arrow Lectures at Stanford University in 1993. The computer simulations were specially done for inclusion in this book. The final chapter on policies for economic growth was first presented as the Ernest Sturc Lecture at the Johns Hopkins School of Advanced International Studies in Washington in 1991.