But while bouncing balls and rotating shapes are a reasonable test of programming skills, they’re not a very empirical AI benchmark. Even slight variations in the prompt can — and do — yield ...