In this talk, I will present some recent results on model selection and model combining. In the direction of model selection, we will examine the consistency property of cross validation for comparing two general regression procedures. When there is much uncertainty in a model/procedure selection process, combining the candidates suitably can result in a much improved performance in estimation/prediction. In addition to presenting some theoretical and empirical results on an information-theoretic approach to combining models, we will discuss a fundamental conflict between finding the true model and estimating the underlying unknown function that cannot be overcome by neither model selection nor model averaging.