Conversing with Stochastic Language Models