Sophisticated folks aren't doing simplistic/stupid decoding.
Gotta go beyond LLMs 101 to see what's actually happening. Even in training folks are building models which predict several tokens ahead.