It's weird that there's little to no focus on making AI describe problems coherently for use-cases like this?