Asymptotics of Language Model Alignment