Neural networks have advanced to excellent candidates for probabilistic inference. In combination with variational inference, a powerful tool ensues with which efficient generative models can represent probability densities, preventing the need for sampling. Not only does this lead to better generalisation, but also can such models be used to simulate highly complex dynamical systems. In my talk I will explain how we can predict time series observations, and use those to obtain efficient approximations to optimal control in complex agents. These unsupervised learning methods are demonstrated for time series modelling and control in robotic and other applications.