Ensembles of graphs formed in close analogy to statistical mechanical models in physics have recently become popular again after being first proposed (as p* models) by social networkers in the 1960s. They seem to provide a way of unifying various loosely-defined models often used in network applications (scale-free graphs etc.) into a single rigorous framework. I will give an introduction to this theory and discuss possible developments (exactly soluble models, perturbation theory, MCMC methods for simulation, parameter estimation) and application areas.