Seeing the model use "caveman speak" in its thoughts to save tokens is hilarious.

"Why use many word when few word do trick?" is actually a legit tech optimization now!