Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine