Thirteen agents, an empty island, and not one instruction.

MIT researchers put thirteen AI agents on a virtual desert island and gave them no instructions whatsoever. The agents did not sit still. They made up games — and the ones who played came back better at working.

Give a person a free afternoon and they will invent something to do. Researchers at MIT wondered whether a machine would, so they built a desert island, dropped thirteen agents onto it, and pointedly declined to tell them what to do.

They made up games

The agents invented a version of basketball. Then they invented the floor is lava — hopping between blocks, refusing to touch the ground, for no reason anybody had given them.

Nobody wrote a rule saying the ground was dangerous. They made that up, and then they obeyed it.

And playing made them better at working

Afterwards the researchers set real tasks — get this object onto the tallest rock spire. Agents who had spent thirty hours mucking about on the island generally did better than ones dropped straight into the job.

Which is a finding any parent could have told them, and which is nonetheless strange to watch happen in software.

What we are not saying

That the machines were enjoying themselves. Cognitive scientists quoted on the work are blunt that behaviour which looks like play is not evidence of the thing play is in us. Some write-ups add tidy figures — fifteen per cent faster, twelve per cent fewer errors — that we could not find in the paper, so we are leaving them out.

Sources also disagree on which model was behind the agents. Read the paper before repeating the numbers.

Why it is worth your attention

Not because the machines are becoming children. Because idle time turned out to be useful, and almost nobody is giving these systems any. We build them to be busy. This suggests something is lost in that.