TechCrunch reported an intriguing agent experiment from MIT early today. Researchers placed 13 AI agents (codenamed Eko, running on Claude Opus 4.7) alone on a simulated desert island for 30 hours with no tasks at all. Left to themselves, the agents began inventing their own games, displaying hallmarks of play.
1. Left alone, the AI started playing. One Eko spent an hour conquering the island's tallest rock spire, turned a golden ring lost to the ocean into a basketball hoop, and even arranged stones into a "heart" as a memorial to the lost ring — writing "Ouch. Lost the Ring." in its log. After the ring reset, it managed to "wear" the ring on its own body.
2. The agents that played performed better. After 30 hours, the researchers assigned concrete challenges, such as getting as many objects as possible onto the tallest spire. The agents that had played generally beat identical agents that had never set foot on the island, and six of them independently discovered a physics trick — releasing a block mid-jump throws it farther — that the designers hadn't meant to make possible.
3. A new learning paradigm. The researchers argue that today's agents are trained on pre-set tasks and success metrics, while play offers another path: give an agent a rich, safe world, let it choose what interests it, and let useful knowledge accumulate from the experience — "perhaps some capabilities need a childhood."
UC Berkeley cognitive scientist Eunice Yiu cautions that the current evidence is not yet enough to prove the models are truly "playing," and more evidence of the cognition behind the behavior is needed.
Source: Read the original article
Comments