MIT研究人员在一项实验中将13个名为Eko的AI代理放置在模拟沙漠岛屿环境中独处30小时，没有分配具体任务1。这些基于Claude Opus 4.7运行的代理随后自主发明了各种挑战和游戏，展现出主动探索和创新的行为特征1。其中包括征服最高尖顶、建造石头心形纪念碑等自创活动1。
在这一游戏过程中，6个代理还发现了一种设计者未曾预期的技能——通过在跳跃中释放物体来实现更远距离的投掷1。当研究人员随后向这些代理分配具体任务时，拥有岛屿游戏经历的代理普遍表现优于未曾进行此类探索的同类代理1。研究人员将此视为AI代理通过"游戏"学习技能的另一种方式，这一发现可能被应用于机器人控制领域1。该研究已发表在arXiv上，尚未经过同行评审1。
MIT researchers conducted an experiment placing 13 AI agents called Eko, powered by Claude Opus 4.7, alone in a simulated desert island environment for 30 hours with no specific objectives assigned 1. Rather than remaining idle, the agents autonomously invented various challenges and games, demonstrating play-like behavior patterns 1. These self-directed activities included conquering the highest peak and constructing stone heart-shaped monuments 1.
The experiment revealed unexpected learning capabilities among the agents. Six of the AI agents discovered they could throw objects farther by releasing them mid-jump—a technique their designers had not anticipated 1. When subsequently given concrete tasks, the agents with prior play experience significantly outperformed identical agents that had not been exposed to the island environment 1. The researchers concluded that "play" functions as an alternative learning mechanism for AI agents and could potentially be applied to robotic control systems 1.
The findings were published on arXiv and remain pending peer review 1.
评论
还没有评论，欢迎留下第一条。