By analyzing the function of individual neurons, these guys were/are able to gain a deeper understanding of how language models process language, which could lead to improved model architecture and training methods.