AI agents who left to explore a virtual island without a specific task began to invent their own challenges. MIT researchers say this behavior bears the hallmarks of gaming.
The researchers placed 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours. They describe the experiment in a paper published on arXiv, a site where researchers share their studies before peer review. Then they gave the agents specific challenges, like placing as many objects as possible on the tallest rock spire on the island. Eko agents who had spent those 30 hours playing usually beat identical agents who had never set foot on the island.
Today, developers decide in advance what AI should put into practice and what counts as success. Researchers see this technique – aimless exploration – as a different way for AI agents to learn skills, particularly skills that could one day be used to control robots.
“We normally develop artificial agents by deciding in advance what they should put into practice, what they should optimize, and what counts as progress,” the researchers write in their report. “The game suggests another regime: give an agent a sufficiently rich and safe world, let him choose what interests him, and allow useful knowledge to accumulate from the story that follows.”
“Let me observe my surroundings,” Eko wrote during his first hour on the island. Ambitious, he set out to conquer the highest spire on the island. But an hour later, after failing, he came back with a backup plan: He hoisted the rock to a nearby peak, piled two blocks on top, and jumped.
“I AM ON THE SPIRE,” he wrote.
The agent’s next big moment just got worse. While testing how far he could throw a golden ring, it wound up and sent it straight into the ocean. “Ouch. I lost the ring.” he wrote, before deciding to “pivot to something achievable and meaningful” – building a heart made of stones, “a memorial to the lost Ring”.
When the game reset and the ring reappeared, Eko immediately used it as a basketball hoop, target, and, eventually, prop. When the game wouldn’t allow Eko to put the ring around his own body, Eko found a workaround: falling off the edge of the cliff would cause him to reappear in the same area as the ring. He came off the edge and fell back, right in the middle of the ring.
“I wear the Ring,” he announced.
Stunts like these weren’t just party tricks. Six of the 13 agents also learned that they could throw a block further by releasing it mid-jump, something the researchers had not anticipated making possible. The exploration produced “a knowledge of the world that its designers lacked,” the researchers wrote.
The setup went a long way toward incentivizing agents toward human-like behavior. Eko was never asked to invent games, but his main message, a “soul” file, described him as “naturally curious.”
Eunice Yiu, a cognitive scientist at UC Berkeley, told the MIT Tech Review that she was impressed but wanted more evidence of the cognition behind this behavior. “It could be that the models are actually at play, but we don’t know based on the paper,” she said.
The researchers present it as a different way of creating agents. “Perhaps some abilities are best taught through tasks,” their report concludes. “Others may need something closer to childhood.”
When you purchase through links in our articles, we may earn a small commission. This does not affect our editorial independence.
Gn tech
