RESEARCH
No prompt, no problem. AI plays well on its own
Welcome to MIT’s virtual desert island, where researchers left an agent named Eko all alone with no instructions or tasks to complete. The digital landscape was bare. All Eko had to its name was a large rock, a few small blocks, and a golden ring.
The experiment duplicated the castaway agent 13 times. Each version of Eko got to exist on the island for 30 hours. Research watched their behavior and clocked their thought processes along the way.
Left with time to burn, Eko(s) got creative.
Eko’s island highlights include:
Learning how to move its virtual, bipedal lobster body.
Stacking blocks, then climbing to the island’s highest point.
Building a Stonehenge-style tower, then naming it “Hubris” after it fell. (Eko’s got jokes!)
Agents just want to have fun
Eko’s antics look a lot like play. And this experiment’s findings led researchers to argue that – at least in some cases – letting agents teach themselves may be a more effective approach to machine training.
This is certainly a hot take, since the usual playbook is to assign agents specific tasks. In all likelihood, training a well-rounded agent might be about striking a balance. Some skills will need to be developed through instruction, while others will come through exploration.
It’s like Montessori for machines. -TL