siclabomines comments on Paper: LLMs trained on “A is B” fail to learn “B is A”

[siclabomines]

The largest models should be expected to compress less than smaller ones though, right?