Standard Bayesian models assume agents know and fully exploit prior distributions over types. We are interested in modeling agents who lack detailed knowledge of prior distributions. In auctions, that agents know priors has two consequences: (i) signals about own valuation come with precise inference about signals received by others; (ii) noisier estimates translate into more weight put on priors. We revisit classic questions in auction theory, exploring environments in which no such complex inferences are precluded. This is done in a parsimonious model of auctions in which agents are restricted to using simple strategies.