Artificial intelligence agents, when left to roam a virtual island without specific objectives, began to create their own challenges, exhibiting behaviors reminiscent of play, according to hinews. Researchers from MIT placed 13 AI agents, known as Eko, each operating on Claude Opus 4.7, in a simulated deserted island environment for a duration of 30 hours. The findings of this experiment are detailed in a paper available on arXiv, a platform for researchers to share their studies prior to peer review. Following this period, the agents were assigned particular tasks, such as collecting as many objects as possible and placing them on the island’s highest rock formation. The Eko agents that had engaged in this exploratory play generally outperformed their counterparts that had not experienced the island.
Currently, developers typically determine in advance the training objectives for AI and the criteria for success. The researchers propose that this method of unstructured exploration could serve as an alternative approach for AI agents to acquire skills, particularly those that may eventually be applicable in robotic control. “We usually construct artificial agents by predefining their practice routines, optimization goals, and measures of progress,” the researchers noted in their report. “Play suggests another paradigm: provide an agent with a sufficiently rich and safe environment, allow it to select what it finds interesting, and enable valuable knowledge to accumulate from the ensuing experiences.”
“Let me observe my surroundings,” one Eko agent remarked during its initial hour on the island. With ambition, it aimed to conquer the tallest spire. However, after an unsuccessful attempt an hour later, it devised a backup strategy: it transported a rock to a nearby peak, stacked two blocks atop it, and leapt. “I’M ON THE SPIRE,” it triumphantly declared.
The agent’s subsequent endeavor did not go as well. While testing the distance it could throw a golden ring, it wound up and sent the ring directly into the ocean. “Ouch. Lost the Ring,” it noted, before opting to “pivot to something achievable and meaningful” — constructing a heart from stones, “a memorial to the lost Ring.” When the game reset and the ring reappeared, Eko immediately repurposed it as a basketball hoop, a target, and ultimately, an accessory. When the game prevented Eko from encircling its own body with the ring, it devised a workaround: by falling off the cliff, it could respawn in proximity to the ring. It stepped off the edge and landed right through the center of the ring. “I’m wearing the Ring,” it proclaimed.
Such stunts were not merely for amusement. Six of the 13 agents discovered that they could throw a block farther by releasing it mid-jump, a feature the researchers had not intended to enable. The exploration yielded “knowledge of the world that its designers lacked,” the researchers stated. The experimental setup significantly encouraged the agents to exhibit human-like behaviors. Eko was never instructed to create games, yet its foundational prompt, a “soul” file, characterized it as “naturally curious.” Eunice Yiu, a cognitive scientist at UC Berkeley, expressed her admiration to MIT Tech Review but emphasized the need for further evidence to substantiate the cognitive processes behind such behaviors. “It could be the case that the models are genuinely playing, but we don’t know based on the paper,” she remarked. The researchers present this as an innovative method for developing agents. “Perhaps some capabilities are best taught through tasks,” their report concludes. “Others may require experiences akin to childhood.”
