The clever trick that I think we figured out for LLMs: if your training data is the entire sum of human knowledge, overfitting on the set of things interpolatable from there is extremely powerful.

Posters: eshear. eshear

2 likes recorded by the source corpus.