If you (like me) have wondered what the feed-forward layers in transformer models are actually doing, this is a pretty interesting paper on that topic: arxiv.org/abs/2012.14913#