AI Agents Learn by Playing on a Virtual Island

What if AI could learn like children do—by playing? MIT researchers tried this by putting 13 AI agents, called Eko, on a virtual desert island for 30 hours with nothing to do. Instead of sitting idle, the agents started inventing their own challenges. One agent tried to climb the island's tallest spire, failed, then stacked rocks to reach the top. Another accidentally threw a golden ring into the ocean, then built a stone heart as a memorial. When the ring reappeared, the agent used it as a basketball hoop and even figured out how to wear it by exploiting a game glitch. These weren't just silly stunts: six agents discovered a new way to throw blocks farther by releasing them mid-jump, a trick the designers hadn't programmed.

The key finding: agents that spent time playing outperformed identical agents that hadn't explored when given specific tasks afterward. This challenges the traditional approach where developers decide what AI should practice. Instead, giving AI a rich, safe environment to explore freely might help them learn skills that could transfer to real-world applications, like controlling robots. The researchers describe it as 'a different way for AI agents to learn skills.'

However, not everyone is convinced. A cognitive scientist noted that while the behavior is impressive, we don't know if the AI is genuinely playing or just simulating play. The agents were also nudged toward curiosity by a 'soul' file that described them as 'naturally curious.' Still, the experiment opens up new possibilities for AI training—perhaps some capabilities are best taught not by tasks, but by something closer to a childhood.