Key facts
- AI agents named Eko, running on Claude Opus, were placed on a virtual island.
- Agents that explored without tasks performed better on subsequent challenges.
- One agent discovered a method to throw objects farther by jumping.
- The agents developed knowledge not present in their initial programming.
AI agents, named Eko and powered by Claude Opus, were placed on a simulated desert island for 30 hours without specific tasks. Researchers from MIT observed that these agents began to invent their own challenges and games, a behavior they described as akin to play. When later given specific tasks, such as moving objects to a rock spire, the agents that had explored freely generally outperformed identical agents that had not. One agent discovered a technique to throw objects farther by releasing them mid-jump, demonstrating knowledge that its designers had not programmed. The researchers suggest this 'play' regime could be a new way for AI agents to learn skills, particularly for controlling robots.