MIT researchers tested solo AI agents called Eko—copies of Claude Opus 4.7—on a virtual island for 30 hours, observing if they would 'play' without explicit instructions. The agents created their own games, including basketball and 'the floor is lava,' while exploring world limits and setting self-designed challenges. One agent drew constellations and its name using blocks, while another juggled and refined its own rules to make activities harder.
Researchers found Eko agents performed better on subsequent tasks after this autonomous exploration, suggesting a third learning approach: playful learning. The behaviors matched classic play hallmarks: inventing rules, repeating with variations, pretending objects were others, and pursuing goals for their own sake. Lead author Antonio Norelli notes this hints at machines learning without human-provided rewards or demonstrations.
المصدر: MIT Technology Review AI · لخّصه HeadlinesBriefing