The last few years has seen dramatic progress in artificial intelligence, particularly in machine learning, most notably in new work in the connectionist tradition, such as deep learning, but also in work on inferring structured generative models from data. Nevertheless, this new work still is limited to relatively narrow and well-defined spaces of hypotheses. In contrast, human beings and human children, in particular, characteristically generate new, uninstructed and unexpected, yet relevant and plausible hypotheses. I will present several studies showing a surprising pattern. Not only can preschoolers learn abstract higher-order principles from data, but younger learners are actually better at inferring unusual or unlikely principles than older learners and adults. I relate this pattern to computational ideas about search and sampling, to evolutionary ideas about human life history, and to neuroscience findings about the negative effects of frontal control on wide exploration. I uggest that children solve these problems through model-building, exploration and social learning. My hypothesis is that the evolution of our distinctively long, protected human childhood allows an early period of broad hypothesis search, exploration and creativity, before the demands of goal-directed action set in. This evolutionary solution to the search problem may have implications for AI solutions.