The last thirty years have seen the development of comprehensive numerical models of the large-scale circulation of the atmosphere, based on physical principles. Such models are quite skillful at describing the evolving weather up to a few days ahead, despite imperfect theory and inadequate observational data. Yet even a hypothetical perfect model, which exactly represented the dynamics of the real atmosphere, and which used data from the best conceivable observing system, could not produce an accurate forecast of indefinitely long range. Any forecast must eventually lose skill because of the intrinsic instability of the atmosphere itself.