MIT study has 13 Claude Opus 4.7 agents explore a simulated island with no task for 30 hours
What happened
TechCrunch reported on October 23 that MIT researchers let 13 Eko agents running Claude Opus 4.7 explore a simulated desert island for 30 hours with no assigned task. The agents invented their own challenges, and their behavior had features of play. Agents that had explored usually beat un-explored agents of the same kind on later set challenges, and 6 of them found a jump-and-throw technique the designers had not built in. The paper is on arXiv and has not been peer reviewed. The agent prompt describes them as naturally curious; a cognitive scientist interviewed said the paper alone cannot establish whether the behavior is real play.
Written by AI from the coverage · updated 41 minutes ago
Coverage
Follow the reports to see the story from different sides.
- TechCrunch · AILeft alone with nothing to do, AI agents started playing
MIT研究者让13个运行 Claude Opus 4.7 的 Eko 智能体各自在模拟荒岛上探索30小时，发现它们会自创挑战，行为具有玩耍的特征。经历探索的智能体在后续指定挑战中通常胜过未探索的同款智能体，其中6个还发现了设计者未预设的跳跃投掷技巧。论文发布于 arXiv，尚未经同行评审；智能体提示词将其描述为天生好奇，受访认知科学家认为，仅凭论文还无法确定这些行为是否属于真正的玩耍。
Heat over time
Not enough continuous observations to draw a trend yet.