Something I have found interesting is the various approaches being followed to generate synthetic training data for Language Models.