The conversation revolves around the possibility of interesting hardware engineering tricks to scale large scale distributed compute for machine learning, especially in the context of self-supervised learning.