Models can already process and understand images we give them, but I wish they'd interleave parts of those images in their responses to better their answers.