Many optimization problems have to be solved repeatedly. We consider a model in which the linear objective function may be changing dramatically over time and may not be known in advance. We propose "follow the leader" algorithms that perform with epsilon of the optimal static solution. As an example, in the classic diet problem, you have to choose a set of food to buy that minimizes cost subject to nutritional constraints. We consider a repeated model where the nutritional constraints remain the same, but prices of the food are not known in advance. Each day, only after you have made out your shopping list, the prices of all ingredients are revealed. Our algorithms have the property that they perform almost as well as the best fixed menu, if you had to choose a single diet to use the whole time, where the best is chosen in hindsight.