MIT研究人员在虚拟岛屿上教AI代理玩游戏
mit研究人员将13个ai代理放在虚拟沙漠岛上30小时,没有特定任务,发现它们自己发明了挑战并学习了技能。这可能会改变ai代理的训练方式。
mit researchers put 13 ai agents called eko, each running on claude opus 4.7, on a simulated desert island for 30 hours with no specific task. the agents began inventing challenges, such as conquering the island's tallest spire and building a heart made of stones. six of the agents learned to throw a block farther by letting go mid-jump, a capability the researchers hadn't intended. this experiment suggests a different way for ai agents to learn skills, especially those that could be used to control robots. the researchers see this technique as an alternative to traditional goal-oriented training.