Conncetionist learning schemes have hitherto seemed far removed from cognitive skills such as reasoning, planning, and the formation of internal models. In this article we investigate what sort of world models a conncetionist system might learn and how it might do so. A learning scheme is presented that forms such models based on obserrved stimulus-stimuluus relationships. The basis of the scheme is a recurrently connected network of simple, neuron-like processing elements. The net produces a set of predictions of future stimuli based on the current stimuli, where these predictions are based on a model andn involve multiple-step chains of predictions . Results are presented from computer simulatinos of the scheme connected ot a simple world consisting of a stochasitc maze (Markov process). By wandering around the maze the network learns its construction. When reinforcement is subsequently introduced, the solution to the maze is learned much more quickly than it is without the exploration period. The form and logic of the experiment is the same as that of the latent learning experiments of animal learning research