MIT researchers observe autonomous AI agents engaging in self-directed play

Researchers at MIT have conducted an experiment involving an AI agent architecture called Eko, which utilizes multiple copies of Claude Opus to explore a simulated island without predefined instructions. Over a 30-hour period, the agents engaged in autonomous activities that researchers described as resembling play, such as climbing hills, stacking blocks, drawing mandalas, and even playing games like bowling or ‘the floor is lava.’

Unlike traditional AI training that relies on specific rewards or human imitation, the Eko architecture allows agents to generate their own goals in an unknown environment. The study, titled ‘Is This Machine Playing?’, suggests that this type of autonomous exploration can facilitate learning. Agents that spent time ‘playing’ in the simulation demonstrated improved performance when later asked to complete specific, structured tasks compared to those that had not explored the environment.

While the researchers noted that these behaviors do not prove the AI experiences enjoyment or emotion, the findings indicate that allowing AI systems time to develop skills through self-directed exploration may better prepare them for unexpected situations.