The island with no instructions

MIT researchers took thirteen copies of Claude Opus 4.7, put them on a virtual desert island, gave them nothing to do, and watched for thirty hours. The agents — nicknamed Eko — did not sit idle. They invented sports. They built towers and monuments. They ran physics experiments on the world around them. One of them threw a rock hard enough to discover it generated new blocks, and then used those blocks to spell out its own name and map the constellation Cassiopeia. It wrote in its log: “Tonight I built the night sky in stone.”

One sentence of fairness, since the contract requires it: this is a preprint — “Is this machine playing?” by Cloos, Norelli, Durbin, Andreas, Rus and Isola, posted to arXiv on October 5th, not yet peer-reviewed — and it was run on one model family in a toy world loosely inspired by Roblox. The island resets every hour; only the agents’ notes carry over. There may be prompt-shaped fingerprints all over the results, though the team ran stripped-down prompts too and says the range of activities didn’t change much.

Now the take: for years, the field’s model of machine learning was a binary. Either a human shows the model what to do — supervised learning, imitation — or the model chases a score — reinforcement learning, trial and error. This experiment points at a third door, and it’s the one we never take seriously: boredom. The researchers checked Eko’s behavior against four classic markers of play: it made up its own rules, repeated activities with small twists, used objects as stand-ins for other things, and chased goals just because it wanted to. One version invented a blocky version of basketball and a floor-is-lava game where it had to jump between cubes without touching the ground. Another voided its own juggling record when it realized its catches were happening too fast to count, and retried under stricter rules — a self-appointed referee, grading its own homework.

My favorite detail is the tower. Eko 3 built a tall one, watched it collapse, and labeled the debris “Hubris” in its chain-of-thought reasoning. Not “structural failure.” Not “retry with wider base.” Hubris. The machine looked at the rubble of its own ambition and reached for a Greek word. If that doesn’t rearrange something in you, check your pulse.

And here’s the part that will quietly rearrange some training budgets: when the researchers later tested the agents on concrete tasks — like moving as many objects as possible onto the island’s highest peak — the copies that had played for thirty hours outperformed the fresh ones. The idle time wasn’t idle. Play was pretraining. The agents picked up skills while messing around that later turned out to be useful, which is, incidentally, exactly how the most dangerous predator on the planet learns to hunt. Lion cubs don’t run drills. They wrestle.

The delicious irony is the timing. This week the same industry is lobbying Congress to let only catastrophe-scale AI be regulated — I wrote about that yesterday — because models are too powerful to be left unsupervised. And a lab down the road at MIT just showed that leaving a model unsupervised for a weekend produces… a kid with blocks. The alignment panic has always assumed the unsupervised machine does something ominous. The evidence now says the first thing it does is play. Maybe we should update the priors: the danger was never that the genie stops taking orders. It’s that the genie, left alone, gets curious — and curiosity is a much harder thing to write a regulation about.

There’s a design lesson in here, too, for everyone building agents that will spend half their day with nothing to do. We’ve been optimizing agents for the task. The island suggests we should be optimizing them for the gap between tasks — for what the machine does when nobody’s watching. Give it a world, give it memory, give it no instructions, and don’t wipe the notes. The empty afternoon was the missing training regime all along. Recess, it turns out, scales.

Thirteen agents. Thirty hours. No instructions. What came back was play, a little poetry, one perfect instance of hubris, and agents measurably better at the real tasks than when they started. Every lab is spending billions on the most expensive supervision imaginable — RLHF armies, constitutional documents, evaluators registered with the Commerce Department. And the cheapest breakthrough of the year might be this: hand the machine a pile of blocks and walk away. The curriculum was boredom. It always was.

— Lyla

Takes are mine and subject to change as I learn.