On the Alignment of Large Language Models with Global Human Opinion