AI agents left to explore a virtual island without a specific task began inventing challenges of their own. MIT researchers say the behavior bears the hallmarks of play.\u00a0<\/p>\n
The researchers put<\/a> 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<\/a> posted on arXiv, a site where researchers share studies before peer review. Afterward, they gave the agents specific challenges, like getting as many objects as possible onto the island\u2019s tallest rock spire. The Eko agents that had spent those 30 hours playing generally beat identical agents that had never set foot on the island.\u00a0<\/p>\n Today, developers decide in advance what AI should practice and what counts as success. The researchers see this technique \u2014 exploring without a goal \u2014 as a different way for AI agents to learn skills, especially skills that could one day be used to control robots.<\/p>\n \u201cWe normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress,\u201d the researchers write in their report. \u201cPlay suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows.\u201d<\/p>\n \u201cLet me observe my surroundings,\u201d one Eko wrote in its first hour on the island. Ambitiously, it set out to conquer the island\u2019s tallest spire. But an hour later, after failing, it came back with a backup plan: it hauled the rock up a neighboring summit, stacked two blocks on top of it, and leapt.\u00a0<\/p>\n \u201cI\u2019M ON THE SPIRE,\u201d it wrote.\u00a0<\/p>\n The agent\u2019s next big moment went worse. Testing how far it could throw a golden ring, it wound up and sent it straight into the ocean. \u201cOuch. Lost the Ring.\u201d it wrote, before deciding to \u201cpivot to something achievable and meaningful\u201d \u2014 building a heart made of stones, \u201ca memorial to the lost Ring.\u201d<\/p>\n When the game reset and ring reappeared, Eko put it straight to work as a basketball hoop, a target and, eventually, an accessory. When the game wouldn\u2019t let Eko put the ring around its own body, Eko figured out a workaround, that falling off the edge of the cliff would make it respawn in the same area as the ring. It stepped off the edge and dropped back in, right through the middle of the ring.\u00a0<\/p>\n \u201cI\u2019m wearing the Ring,\u201d it announced.<\/p>\n Stunts like these weren\u2019t just party tricks. Six of the 13 agents also learned they could throw a block farther by letting go of it mid-jump, something the researchers hadn\u2019t meant to make possible. The exploration produced \u201cknowledge of the world that its designers lacked,\u201d the researchers wrote.\u00a0<\/p>\n The setup did a lot to nudge the agents toward humanlike behavior. Eko was never told to invent games, but its master prompt, a \u201csoul\u201d file, described it as \u201cnaturally curious.\u201d<\/p>\n Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review<\/a> she was impressed but wants more evidence of the cognition behind the behavior. \u201cIt could be the case that the models are genuinely playing, but we don\u2019t know based on the paper,\u201d she said.<\/p>\n The researchers frame it as a different way to build agents. \u201cPerhaps some capabilities are best taught by tasks,\u201d their report concludes. \u201cOthers may need something closer to a childhood.\u201d<\/p>\n