I distinguish two roles for a fitness concept in the context of explaining cumulative adaptive evolution: fitness as a predictor of gene frequency change, and fitness as a criterion for phenotypic improvement. Critics of inclusive fitness argue, correctly, that it is not an ideal fitness concept for the purpose of predicting gene-frequency change, since it relies on assumptions about the causal structure of social interaction that are unlikely to be exactly true in real populations, and that hold as approximations only given a specific type of weak selection. However, Hamilton took this type of weak selection, on independent grounds, to be responsible for cumulative assembly of complex adaptations. In this special context, I argue that inclusive fitness is distinctively valuable as a criterion for improvement and a standard for optimality. Yet to call inclusive fitness a criterion for improvement and a standard for optimality is not to make any claim about the frequency with which inclusive fitness optimization actually occurs in nature. This is an empirical question that cannot be settled by theory alone. I close with some reflections on the place of inclusive fitness in the long running clash between ‘causalist’ and ‘statisticalist’ conceptions of fitness.