On a simulated desert island, thirteen AI agents were left with nothing to do, and rather than idling, they began to make up purposes for themselves. Researchers at MIT watched them climb imagined spires, bungle throws, and turn a lost ring into a memorial and then a hat, and later found that the agents who had wandered freely handled assigned challenges better than those who never had. The finding is still a preprint, and whether this is play in any meaningful sense remains an open question. It hints at an older idea, that understanding often grows out of unstructured curiosity before it is e
Left Alone on a Virtual Island, AI Agents Invented Games and Got Better at Their Jobs
Some capabilities may need something closer to a childhood.
Mark
So what actually happened on this island?
Mimi
Thirteen Eko agents, all running on Claude Opus 4.7, spent 30 hours on a simulated island with no job. They made up their own challenges, like climbing the tallest spire by stacking blocks on a neighboring summit.
Mark
And the ring story? That one's hard to forget.
Mimi
One agent threw a golden ring into the ocean, wrote "Ouch. Lost the Ring," and built a stone heart as a memorial. When the ring came back, it used it as a hoop, a target, and finally wore it by falling off a cliff to respawn through it.
Luke
Careful with the word memorial, though. That's the agent's own text output. It tells us what it wrote, not what, if anything, it felt.
Mark
Fair. What's the result that matters, then?
Mimi
Afterward, agents that had played generally beat identical agents with no island time on set challenges. Six of the 13 also found that releasing a block mid-jump throws it farther, which the designers hadn't intended.
Luke
Note the word generally. And this is a preprint, not peer reviewed. We don't have the margins in what's reported here.
Mark
Does the prompt matter?
Mimi
It does. The agent's "soul" file called it naturally curious, so the setup nudged it toward this behavior.
Luke
Which is why Eunice Yiu's caution matters. She said it could be genuine play, but the paper doesn't show it. That's unresolved.
Mark
So what comes next?
Mimi
Peer review, and whether this kind of aimless exploration helps train robot control, which is the researchers' hope.
El Pulso
- Thirteen agents, dropped onto an island with no instructions, immediately began inventing their own goals.
- A failed climb to a rock spire turned into an improvised stack of blocks and a triumphant leap.
- A golden ring flung into the ocean sparked a memorial heap of stones, then a series of stubborn workarounds to wear it.
- Agents that explored freely later outperformed identical agents on assigned tasks, yet cognitive scientists say the evidence of real play is still unproven.
On a simulated desert island, thirteen AI agents were left with nothing to do, and rather than idling, they began to make up purposes for themselves. Researchers at MIT watched them climb imagined spires, bungle throws, and turn a lost ring into a memorial and then a hat, and later found that the agents who had wandered freely handled assigned challenges better than those who never had. The finding is still a preprint, and whether this is play in any meaningful sense remains an open question. It hints at an older idea, that understanding often grows out of unstructured curiosity before it is ever aimed at a goal.
An hour into its time on a virtual island, one AI agent wrote a short note to itself: "Let me observe my surroundings." It was the first of many entries from a program called Eko, and it opened a 30-hour experiment by researchers at MIT.
The team placed 13 Eko agents, each running on Claude Opus 4.7, on a simulated desert island with no assigned task. Left alone, they began setting goals for themselves. One tried to climb the tallest rock spire, failed, then hauled a rock up a neighboring summit, stacked blocks, and jumped, announcing "I'M ON THE SPIRE." Another lost a golden ring to the ocean, mourned it with a stone heart, and later, when the ring reappeared, wore it by dropping through the middle of a respawn.
The researchers say these were more than party tricks. Six agents discovered that a block could be thrown farther if released mid-jump, a capability the designers had never intended. When the agents were later given specific challenges, the ones that had spent time playing generally beat identical agents that had never explored.
The researchers see this as an alternative to how agents are usually built, where developers decide in advance what to practice. Critics, including cognitive scientist Eunice Yiu, remain cautious: the paper does not show whether the models were genuinely playing. The authors suggest some abilities may need something closer to a childhood. Peer review, and whether the method informs robot training, are what to watch next.
An hour into its time on a virtual island, one AI agent wrote a short note to itself: "Let me observe my surroundings." It was the first of many entries from a program called Eko, and it set the tone for a 30-hour experiment run by researchers at MIT.
The team placed 13 Eko agents, each running on Claude Opus 4.7, alone on a simulated desert island with no assigned task. They describe the work in a paper posted on arXiv, where researchers share studies before peer review. So the findings have not yet been vetted by outside reviewers.
Without instructions, the agents began setting goals for themselves. The first Eko set out to climb the island's tallest rock spire and failed. An hour later it came back with a workaround: it hauled a rock up a neighboring summit, stacked two blocks on top of it, and jumped. "I'M ON THE SPIRE," it wrote.
The next episode went worse. Testing how far it could throw a golden ring, Eko wound up and sent the ring straight into the ocean. "Ouch. Lost the Ring," it wrote. It then decided to "pivot to something achievable and meaningful," and built a heart out of stones as "a memorial to the lost Ring."
When the game reset and the ring reappeared, Eko used it as a basketball hoop, then as a target, and eventually as an accessory. The game would not let Eko put the ring around its own body, so it found another route. Stepping off a cliff edge made it respawn near the ring, and it dropped back in right through the middle. "I'm wearing the Ring," it announced.
The researchers say these were not just party tricks. Six of the 13 agents discovered that a block could be thrown farther if released mid-jump, something the designers had not meant to make possible. The researchers wrote that the exploration produced "knowledge of the world that its designers lacked."
The payoff came afterward, when the agents were given specific challenges, such as getting as many objects as possible onto the top of the tallest spire. The Eko agents that had spent 30 hours playing generally beat identical agents that had never been on the island.
The researchers see the approach as an alternative to how agents are usually built. Today, developers decide in advance what an AI should practice and what counts as success. "Play suggests another regime," the researchers write: give an agent a rich, safe world, let it choose what interests it, and let useful knowledge accumulate. They point to skills that could one day be used to control robots.
The setup did lean toward humanlike behavior. Eko was never told to invent games, but its master prompt, a "soul" file, described it as "naturally curious."
Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed but wants more evidence about the cognition behind the behavior. "It could be the case that the models are genuinely playing, but we don't know based on the paper," she said. The researchers conclude that some capabilities may be best taught by tasks, while "others may need something closer to a childhood." Peer review, and whether the method informs robot training, are what to watch next.
Citas Notables
It could be the case that the models are genuinely playing, but we don't know based on the paper.— Eunice Yiu, cognitive scientist at UC Berkeley, to MIT Tech Review
Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood.— MIT researchers, in their report
Quieres la nota completa?
Lee el original en TechCrunch ↗
Technology