[feed-forward sublayers] … these layers are the mechanism used to gradually learn a general, increasingly abstract understanding of the entire text being processed.

in my opinion, this is the part that people who hates LLMs (large language models) chooses to ignore.