As models improve, their representations converge. This even happens across modality for vision and text.