Summary An early example of a compound decision problem of Robbins (1951) is employed to illustrate some features of the development of empirical Bayes methods. Our primary objective is to draw attention to the constructive role that the nonparametric maximum likelihood estimator for mixture models introduced by Kiefer & Wolfowitz (1956) can play in these developments.