On the Principles behind Opinion Dynamics in Multi-Agent Systems of Large Language Models