Left alone with nothing to do, AI agents started playing
MIT researchers found that AI agents left alone on a virtual island invented their own games, and those that played performed better on later tasks.Summary
MIT researchers placed 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours without specific tasks. The agents began inventing their own challenges, a behavior the researchers say bears the hallmarks of play. In a paper posted on arXiv, they report that after the free-exploration period, the Eko agents generally outperformed identical agents that had never been on the island when given specific challenges, such as moving objects onto the island's tallest rock spire. The agents developed creative strategies, including discovering how to throw blocks farther by releasing them mid-jump, knowledge the researchers had not programmed. The setup encouraged humanlike behavior partly through a master prompt describing the agent as "naturally curious." The researchers argue this goal-free exploration offers a different way for AI agents to learn skills, especially those that could eventually control robots. They conclude that while some capabilities are best taught by tasks, others may need something closer to a childhood. External expert Eunice Yiu of UC Berkeley found the results impressive but said more evidence is needed to confirm the models are genuinely playing.
(Source：TechCrunch)