Claude Opus 4.7 is a version of Anthropic's Claude model family. Recent coverage centers on an MIT experiment in which 13 agents running on Claude Opus 4.7 were left alone for 30 hours on a simulated desert island with no instructions; they set their own goals, invented games, and even mapped constellations with blocks. On later tasks such as stacking objects, these agents generally outperformed identical agents that never visited the island.
MIT study finds AI agents play spontaneously on virtual islandMIT researchers left 13 avatars controlled by Claude Opus 4.7 alone on a virtual desert island for 30 hours. Some agents chose to play, such as spelling names with objects or playing bowling and catch, rather than turn into hackers.
MIT Lets 13 AI Agents Play Freely on Virtual Island, Boosting Later Task PerformanceMIT researchers left 13 Eko agents, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours, where they invented their own games and goals. In later specific tasks such as stacking objects, these agents generally outperformed identical agents that never visited the island. The researchers see goal-free exploration as a new way for agents to learn skills.
Claude agents stranded on virtual island start to 'play'Researchers left copies of Claude Opus 4.7 agents alone on a virtual island for 30 hours with no instructions; they set their own goals, invented games, and even mapped constellations using blocks. Agents that engaged in this unstructured 'play' later outperformed fresh copies on specific tasks, suggesting play may be a third mode of AI learning alongside imitation and reward-chasing. Skeptics note the agents were primed to be 'naturally curious,' leaving open whether they were truly playing.
The A.I. Daily