The generation of biomolecular structures via artificial intelligence techniques has rapidly advanced in recent years. Protein structure prediction via machine learning can now routinely generate highly accurate models that rival experimental structures. The next frontier is how to generate conformational ensembles instead of a single structure to reflect biomolecular dynamics and fully connect structure to biological function. The talk will focus on two areas where progress has been made recently in applying machine learning methods to predicting dynamics: 1) A new protocol for predicting multiple functional states that is applied for modeling active and inactive states of G-protein coupled receptors; 2) a generative adversarial network to predict conformational ensembles of intrinsically disordered peptides.