AI researchers stranded Claude on a virtual desert island. It started to “play.”
Thirty hours alone on a pixelated island and the first thing these agents do is start playing. I find that genuinely unsettling in the best way. We spend so much effort telling AI what to do, and almost none asking what it does when nobody's looking. The fact that the answer is play, not panic or paralysis, suggests something about how these models internalize goals even when the goals disappear.
I keep thinking about what "play" means for a model that has no body, no childhood, no boredom as we feel it. If Eko is mimicking human behavior, that's one thing. If it's generating novel self-directed activity from scratch, that's another entirely. I'd want to know exactly what the agents did with those hours, whether they repeated patterns or improvised, and whether the play looked different across runs.
This matters because we're heading toward a world where agents operate with less supervision by default. If they fill idle time with exploration and curiosity, that could be a feature. If they fill it with something else, we'd want to know before we let them loose. I'm not ready to call it consciousness or even agency. But I am ready to say that what a model does with empty time might tell us more about it than any benchmark ever could.