[{"data":1,"prerenderedAt":14},["ShallowReactive",2],{"artikel-en-berita-ai-eko-pulau-maya-main-24-oktober-2026":3},{"slug":4,"lang":5,"title":6,"description":7,"date":8,"updated":9,"topic":10,"product":9,"cover":11,"readMin":12,"html":13},"berita-ai-eko-pulau-maya-main-24-oktober-2026","en","MIT stranded 13 AI agents on a virtual island, and they started playing on their own","An MIT CSAIL team left AI agents with no tasks for 30 hours on a virtual island, and the agents invented their own games before getting better at real work.","2026-10-24",null,"berita","\u002Fimages\u002Fartikel\u002Fberita-ai-eko-pulau-maya-main-24-oktober-2026.webp",4,"\u003Cp>Researchers at MIT's CSAIL lab ran a strange experiment: they marooned 13 AI agents named Eko on a simulated desert island for 30 hours, with no tasks and no rewards. Each Eko ran on Claude Opus 4.7. The agents started inventing games, building monuments, and eventually got better at real tasks.\u003C\u002Fp>\n\u003Ch2>1. What happened on the island\u003C\u002Fh2>\n\u003Cp>Eko got a bare world: one large rock, some small blocks, and a golden ring. No score, no leaderboard, and no one setting goals.\u003C\u002Fp>\n\u003Cp>In the first hour, one Eko wrote "let me observe my surroundings" and set out to climb the island's tallest spire. When it failed, it hauled the rock up a neighboring summit, stacked two blocks on top, and leaped. "I'M ON THE SPIRE," it announced.\u003C\u002Fp>\n\u003Ch2>2. The games they invented\u003C\u002Fh2>\n\u003Cp>The recorded behavior includes:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>One Eko built a basketball hoop and played from below\u003C\u002Fli>\n\u003Cli>Another created a version of "the floor is lava", jumping between blocks without touching the ground\u003C\u002Fli>\n\u003Cli>One Eko drew hearts, pyramids, its own name, and the constellation Cassiopeia by arranging blocks on the ground\u003C\u002Fli>\n\u003Cli>After throwing the golden ring into the ocean, one Eko built a heart-shaped stone "memorial to the lost Ring"\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>According to MIT Technology Review, the world was loosely based on the video game Roblox, and Eko's virtual body was styled as a two legged lobster.\u003C\u002Fp>\n\u003Ch2>3. Not just fun and games\u003C\u002Fh2>\n\u003Cp>The play had real effects. After the 30 hours, agents that had played beat fresh copies of the same model on specific tasks, like moving as many objects as possible onto the island's highest peak.\u003C\u002Fp>\n\u003Cp>Six of the 13 agents also discovered a trick the world's designers never planned: letting go of a block mid jump throws it farther. The MIT team wrote that the exploration produced "knowledge of the world that its designers lacked".\u003C\u002Fp>\n\u003Cp>When the researchers deleted lines from the agents' memory files, a technique called ablation, their performance dropped. The memories of play itself appear to be the learning capital.\u003C\u002Fp>\n\u003Ch2>4. A new way to train AI\u003C\u002Fh2>\n\u003Cp>AI models are usually trained two ways: supervised learning (copying human examples) or reinforcement learning (rewards for trial and error). Both fix the goal in advance.\u003C\u002Fp>\n\u003Cp>This study, released as a paper on arXiv and not yet peer reviewed, suggests a third route: give an agent a rich and safe world, let it choose what is interesting, and let useful knowledge accumulate on its own. The team ends its report with this line: "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."\u003C\u002Fp>\n\u003Cp>Robotics researchers are watching closely. Agents that learn to manipulate objects through play could one day help robots learn without humans demonstrating every step.\u003C\u002Fp>\n\u003Ch2>5. How much should we believe it\u003C\u002Fh2>\n\u003Cp>There are real caveats. Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review the study does not yet prove the models are genuinely playing: "It could be the case that the models are genuinely playing, but we don't know based on the paper."\u003C\u002Fp>\n\u003Cp>Eko's "soul" file also described it as "naturally curious", so some behavior may have been nudged by that indirect instruction. MIT researcher Nathan Cloos admits it all "could be role playing", but he stresses that the play still contributed to measurable learning.\u003C\u002Fp>\n\u003Ch2>What it means for you\u003C\u002Fh2>\n\u003Cp>The study is early stage, but it hints at how AI training could change. There is no immediate action for everyday users, yet it is worth following:\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>AI agents will get smarter on their own.\u003C\u002Fstrong> The future of agents may involve not just human instructions but free exploration time that builds base skills.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Robots could benefit too.\u003C\u002Fstrong> MIT sees this method as especially useful for physical skills, a field that is hard to train with older methods.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Leave room to experiment.\u003C\u002Fstrong> The same principle applies to people: unstructured playtime can be your most productive learning, including when you explore new AI tools like the ones in \u003Ca href=\"\u002Fen\u002Fai-tools\">AI Tools\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Always check the claims.\u003C\u002Fstrong> This paper is not peer reviewed, and outside experts remain skeptical. Exciting AI news usually has a gap between headline and evidence.\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch2>Sources\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Ftechcrunch.com\u002F2026\u002F10\u002F22\u002Fleft-alone-with-nothing-to-do-ai-agents-started-playing\u002F\" target=\"_blank\" rel=\"noopener\">Left alone with nothing to do, AI agents started playing\u003C\u002Fa> (TechCrunch)\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.technologyreview.com\u002F2026\u002F10\u002F19\u002F1147095\u002Fai-stranded-claude-virtual-desert-island-play\u002F\" target=\"_blank\" rel=\"noopener\">AI researchers stranded Claude on a virtual desert island. It started to play\u003C\u002Fa> (MIT Technology Review)\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.fastcompany.com\u002F91625248\u002Fmit-researchers-stranded-ai-agent-on-desert-island-what-happened-could-change-how-robots-are-trained\" target=\"_blank\" rel=\"noopener\">MIT researchers stranded an AI agent on a desert island\u003C\u002Fa> (Fast Company)\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fgizmodo.com\u002Fresearchers-stranded-an-ai-agent-on-a-virtual-island-things-got-very-weird-2000825998\" target=\"_blank\" rel=\"noopener\">Researchers Stranded an AI Agent on a Virtual Island. Things Got Very Weird.\u003C\u002Fa> (Gizmodo)\u003C\u002Fli>\n\u003C\u002Ful>\n",1793543302123]