When the AI agents are placed on a virtual island with no specific tasks to perform, they begin to invent their own challenges spontaneously. Researchers from MIT indicate that this behavior has characteristics of "playing."
Researchers left 13 Eko agents, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours. They described this experiment in a paper published in arXiv. arXiv is a website where researchers share their research findings before peer review. Subsequently, the researchers assigned specific challenges to these agents, such as moving as many objects as possible to the highest rock pillar on the island. Those Eko agents that had experienced 30 hours of "playtime" overall outperformed their counterparts who had never set foot on the island.
Researchers point out that nowadays, developers typically decide in advance what AI should practice and what constitutes success. They believe that this approach of "goalless exploration" could be another path for AI to acquire learning skills, especially those skills that may be used to control robots in the future.
"We usually develop artificial intelligence agents by predeciding what they should practice, what should be optimized, and what constitutes progress," the researchers wrote in their report. "Playing suggests another mechanism: providing the agent with a rich and safe world to choose from on its own, allowing useful knowledge to accumulate from the subsequent events that unfold."
"Let me observe my surroundings," wrote a Eko in the first hour after arriving on the island. It ambitiously planned to conquer the island's tallest spire. But an hour later, after failing, it returned with a backup plan: it moved a rock to a nearby mountain top, stacked two blocks on top of it, and then jumped onto it.
"I'm on the spire," it wrote.
Things didn’t go as smoothly for this proxy from then on. It tried to test how far it could throw a golden ring, but due to using too much force, it ended up throwing the ring into the sea. “Oh no. I lost the ring,” it wrote. Afterwards, it decided to “turn to something more achievable and meaningful” – to build a heart out of stones, “as a monument to the lost ring.”
When the game was reset and the golden ring reappeared, Eko immediately turned it into a basketball hoop and then into a target, and eventually used it as a decoration. Later, when the game no longer allowed Eko to put the golden ring around themselves, Eko came up with a workaround: by jumping off the edge of a cliff, they could be reborn in the same area where the golden ring was located. So they jumped off the edge and landed back down, just passing through the center of the golden ring.
“I’ve put on the ring,” it announced.
Such actions are not just party tricks. Six out of the 13 agents also learned that releasing their grip halfway through a jump allowed them to throw the blocks farther, something that the researchers had not intended for them to do. The researchers wrote that this exploration resulted in “knowledge about the world that its designers lacked.”
To a large extent, this setting encourages these agents to exhibit behavior similar to that of humans. Eko has never been tasked with inventing games, but its primary prompt word, a "soul" file, describes it as "innately curious."
Cognitive scientist Eunice Yiu from the University of California, Berkeley, told MIT Tech Review that she was impressed by these performances, but she hopes to see more evidence regarding the cognitive mechanisms behind such behavior. She said, “Perhaps these models are indeed playing, but based on this single paper alone, we still don’t know for sure.”
Researchers see this as another way to build agents. The report concludes, "Perhaps some abilities are best taught through tasks. Others may require a approach that is closer to that of childhood."
