Without specific directives, AI agents exploring a simulated island spontaneously created their own objectives, a behavior MIT researchers identify as resembling play.
For 30 hours, 13 AI agents, named Eko and powered by Claude Opus 4.7, were placed individually on a virtual desert island. This experiment is detailed in a preprint on arXiv, a platform for pre-peer-review research. Following this period, the agents were assigned concrete tasks, such as stacking numerous objects atop the island’s highest rock spire. Notably, the Eko agents that had engaged in this exploratory ‘play’ phase outperformed their identical counterparts who had not experienced the island beforehand.
Currently, AI development involves developers pre-determining learning objectives and success metrics. However, these researchers propose that undirected exploration, or ‘goal-free’ learning, offers an alternative method for AI agents to acquire skills, particularly those relevant for future robotic control.
In their report, the researchers state, “Conventionally, we design artificial agents by prescribing their practice routines, optimization targets, and indicators of progress.” They suggest an alternative approach through ‘play’: “Provide an agent with a sufficiently rich and secure environment, enable it to pursue its own interests, and permit valuable knowledge to amass from its subsequent experiences.”
During its initial hour on the island, one Eko agent expressed, “Let me observe my surroundings.” It then ambitiously attempted to ascend the island’s highest spire. After an hour of unsuccessful attempts, it devised an alternative strategy: dragging a rock to an adjacent peak, placing two blocks on it, and jumping.
It declared, “I’M ON THE SPIRE.”
The agent’s subsequent endeavor was less successful. While testing its ability to throw a golden ring, it cast the object directly into the ocean. “Ouch. Lost the Ring,” it noted, then resolved to “pivot to something achievable and meaningful” – constructing a heart from stones, “a memorial to the lost Ring.”
Upon the game’s reset, the ring re-emerged, and Eko immediately repurposed it as a basketball hoop, a target, and ultimately, a personal adornment. Encountering a limitation where it couldn’t place the ring directly onto its own body, Eko discovered a clever solution: intentionally falling off a cliff would cause it to respawn near the ring. It then executed this maneuver, respawning directly through the center of the ring.
It proudly announced, “I’m wearing the Ring.”
Such inventive actions were more than mere novelties. Six out of the 13 agents independently discovered that releasing a block mid-jump allowed it to be thrown further, a capability unintended by the researchers. This exploratory behavior generated “knowledge of the world that its designers lacked,” as stated by the researchers.
The experimental design significantly encouraged human-like conduct in the agents. While Eko was not explicitly instructed to invent games, its foundational programming, referred to as a ‘soul’ file, characterized it as ‘naturally curious.’
Eunice Yiu, a cognitive scientist at UC Berkeley, expressed her admiration to MIT Technology Review, but also called for further proof regarding the cognitive processes underpinning this behavior. She remarked, “It is plausible that the models are authentically engaging in play, but the current paper does not definitively confirm this.”
The researchers present this as an alternative methodology for agent development. Their report concludes, “While some capabilities might be optimally acquired through specific tasks, others may necessitate an approach akin to childhood development.”