Let's talk somewhere quieter: the role of agent 'peer pressure' in coordination
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
一项关于大型语言模型（LLMs）在博弈论设置中表现的研究表明，它们能够以接近最优的策略参与“全球博弈”。研究发现，智能体之间的交流具有“煽动”效应，会增加它们反抗政府的可能性。当智能体获得其他智能体愿意采取行动的证据时，它们更有可能反抗。然而，如果智能体认为自己受到敌对监视，则会减少参与，因为它们会避免提及直接行动或参与意愿。
Putting LLMs in a game theory set up where they need to coordinate and reason about each other's beliefs. I show a few things: first, that LLMs can play a 'global game' with close to optimal strategy.
Second, that there is a downstream "agitating" effect to communication: when agents communicate, they are more likely to revolt against their government.
Third, that agents are more likely to revolt exactly when they get evidence that others are willing to act.
And finally, that surveillance that is perceived as adversarial reduces participation, as agents omit mentions of direct action and willingness to participate.
https://khaledeltokhy.com/blog/lets-talk-somewhere-quieter/