The degree to which just scaling up compute (read: money) in language models just solves problems that architectural tweaks can't... is truly annoying sometimes.