Cosmology, at the present day, works with static catalogs (of, say, galaxies) and point estimates of fundamental physical quantities (of, for example, two-point functions, on which most cosmological science is based). Next-generation cosmological experiments are looking for ever more subtle effects in larger and larger data sets; they will not provide all the cosmological information of which they are capable unless we can develop probabilistic approaches that transmit as best as possible all the information in the raw data to the quantities of interest (for example, the spectrum of primordial density fluctuations or the mass of the dark-matter particle). I will show some baby steps towards creating a probabilistic generative model of the full system, with cosmological parameters at the top and raw image pixels at the bottom. Parts of this work have been performed in collaboration with Fergus (NYU), Lang (Princeton), Marshall (Oxford), and Murray (Edinburgh).