MIT left 13 AI agents alone on an island for 30 hours. They played

One Eko opened with "Let me explore my surroundings" and set out to climb the island's tallest tower. It failed. An hour later it came back with plan B: stack blocks next to a neighboring peak, then jump. "I'm on the tower!" it wrote. Another Eko tried to throw a golden ring and watched it land in the sea. "Ouch. I lost the ring." It built a heart out of stones as a memorial. When the game reloaded and the ring reappeared, the agent used it as a basketball hoop, then a target, then an accessory.

Did the researchers tell the AI agents to play?

Not directly. The master prompt did describe Eko as "naturally curious", and the setup nudged the agents toward human-like behavior. Still, the team says it didn't expect everything that followed. Six of the 13 agents learned to throw a block farther by releasing it mid-jump. The study says the experiment produced "insights about the world that even the developers lacked".

Why does it matter for AI training?

After the island phase, the researchers gave the agents concrete jobs, such as getting as many objects as possible onto the island's highest peak. Agents that had spent 30 hours playing generally beat identical agents that had never been on the island.

That suggests exploring without a goal is another way for agents to pick up skills. The team thinks it could matter especially for robots. Today, they write, agents are built by fixing in advance what to practice, what to optimize and what counts as progress. Play suggests a different route: give an agent a varied, safe world and let it decide what's interesting. Their conclusion: some skills may be best taught through tasks, while others may need "something closer to a childhood".