MIT计算机科学与人工智能实验室的研究人员进行了一项实验，将Claude Opus 4.7智能体（命名为Eko）隔离在虚拟沙漠岛上30小时，以观察其自主行为1。在这段时间内，Eko表现出类似玩耍的多种行为模式，包括自主设定目标、创意性地排列方块绘画和创造游戏1。Eko创建了心形、金字塔等图案，甚至写下自己的名字和星座图案1。
这个智能体还展现出了更复杂的创意活动1。Eko创造了篮球游戏和类似"地板是熔岩"的活动，并在杂耍挑战中创造了50次连续接住的记录，之后自行决定标准过于宽松而重新开始1。研究人员发现，拥有30小时岛屿探索经验的智能体在后续任务中表现优于未经历过相同经验的新副本1。例如，在将尽可能多的物体放在岛屿最高峰的特定任务上，有玩耍经验的智能体表现更好1。此外，删除智能体的内存文件会降低其性能，表明其探索经历对学习能力有实质性影响1。
研究人员认为这代表了一种新的"玩耍学习"方式1。该实验由包括Antonio Norelli在内的MIT团队主导1。
Researchers at MIT's Computer Science and Artificial Intelligence Laboratory conducted a 30-hour experiment isolating a Claude Opus 4.7 agent named Eko on a virtual desert island to observe its autonomous behavior.1 During this period, the AI demonstrated behaviors remarkably similar to play, including self-directed goal-setting and creative activities.1
Eko created visual patterns by arranging blocks into shapes such as hearts and pyramids, wrote its own name and constellation designs, and invented games including basketball and activities analogous to "the floor is lava."1 The agent even juggled virtual objects, achieving 50 consecutive catches before independently deciding the standard was too lenient and resetting its own challenge.1 These actions suggest a novel form of "play-based learning" emerging in advanced AI systems.1
The research revealed significant differences in subsequent task performance between agents with island exploration experience and untested copies.1 When assigned the objective of placing as many objects as possible on the island's highest peak, Eko—equipped with 30 hours of prior exploration—outperformed agents without this experience.1 Deleting the agent's memory files diminished its performance, indicating that accumulated experience from autonomous play contributed meaningfully to task capability.1 Eko was configured with a "soul" file characterizing it as "naturally curious" and motivated to explore.1
Antonio Norelli served as one of the lead paper authors on this research from the MIT laboratory.1
评论
还没有评论，欢迎留下第一条。