摘要：The cavity method, also known as Belief Propagation, has emerged as a versatile and powerful technique to compute properties of marginals of Gibbs-Boltzmann distributions. It has many applications outside of physics and plays a central role in understanding spin glasses on random graphs. In this talk, I will review the cavity method and describe some of its most important (or famous) applications. I will then consider the formal extension to distributions that are not Gibbs-Boltzmann, in particular to the probability distributions of spin histories in a spin system evolving in time. I will describe the attempts by us and others to apply the cavity method to these cases, an approach known as "dynamic cavity", and describe some results in this direction.