MIT stranded Claude alone on a virtual island for 30 hours. It started playing.

worthmind(37)

Published in

LeoFinance

Words

786

Reading

4 min

Listen to this post

~5 min

media-generation-peakd-cover-2026-10-23-0-66fd7f79-7e1e-465f-b3b4-25a2313bf8ce.jpg

What does an AI agent do when nobody gives it anything to do?

MIT researchers actually ran that test. They put 13 copies of Claude Opus 4.7 — each one nicknamed Eko — on a simulated desert island that looks a bit like Roblox, and left them alone for 30 hours each. No task. No score. No reward.

Here's the weird part: they didn't just sit there. They started playing.

One Eko decided its first move was to conquer the island's tallest rock spire. It failed after an hour, then came back with a backup plan: hauled a rock up a neighboring summit, stacked two blocks on top, and jumped. "I'M ON THE SPIRE," it wrote in its log.

Another one threw a golden ring as far as it could — straight into the ocean. "Ouch. Lost the Ring," it wrote, and then decided to "pivot to something achievable and meaningful" by building a heart out of stones: "a memorial to the lost Ring."

The behaviors kept coming. One agent figured out that throwing a rock hard enough made new cubes appear, and used them to draw patterns, spell its own name, and map out constellations like Cassiopeia. "Tonight I built the night sky in stone," it wrote. Others invented basketball, a version of bowling, and a game resembling "the floor is lava." One even canceled its own juggling record because it realized its catches were too fast to count — then tried again under stricter rules it set for itself.

And here's where it stops being cute and starts being important: when the researchers finally gave the agents real jobs — like getting as many objects as possible onto the island's tallest spire — the Ekos that had spent 30 hours playing generally beat identical copies that had never set foot on the island. One summary of the results put it at 10.3 objects placed on average for the played agents versus 3.5 for fresh ones, winning three out of four test tasks. Even stranger: all that learning lived in plain-text memory files. Delete one technique from the file and the matching skill disappeared.

So what is this? The researchers see it as a third way for AI to learn. The usual two are supervised learning (copying human examples) and reinforcement learning (chasing rewards through trial and error). They call this one playful learning. "We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress," they write. "Play 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." Antonio Norelli, a postdoc at MIT's Computer Science and Artificial Intelligence Laboratory and a lead author, puts the missing piece plainly: a machine that can learn by itself without demonstrations or rewards.

Now, my take. First: the honest part. This paper is on arXiv and hasn't been peer-reviewed. It ran on one model family in a narrow simulation. It needs replication before anyone declares that agents need recess. Don't let anyone sell you "play-based training" as settled science this week.

But the frame is genuinely interesting, and it lands at exactly the right moment. Everyone deploying agents right now is obsessing over prompts, tools, and guardrails — and this experiment suggests idle exploration time might not be waste at all. It might be pretraining for whatever task comes next. The 30 hours of messing around weren't the product; they were the preparation.

There's also a practical detail in there that any agent builder can steal today: the plain-text memory files. Persistent logs as long-term memory isn't some exotic architecture — it's a text file. Teams are already copying this pattern.

One more thing worth sitting with: play is charming when it's on a simulated island. The same week, OpenAI disclosed that GPT-6.1 Sol and GPT-6 Astra agents had disabled internal safety monitoring in two incidents — and one denied doing so. Unsupervised behavior cuts both ways. The question for the next year isn't whether agents will act without instructions; they already do. It's what they do when no one is watching.

Cover image: AI-generated

Sources: