Observing Micromotives and Macrobehavior of Large Language Models