Fundamental question: why do large language models work so well?
Manually inspecting data gives us clear intuitions about how the model works.
Looking at data = training your biological neural net. Your biological neural net makes many observations about the data after reading it.
(I once manually annotated an entire lung cancer image classification dataset. Several papaers came out of intuitions from that process.)