Left alone with nothing to do, AI agents started playing
techcrunch ·
MIT researchers conducted an experiment where 13 AI agents, named Eko and powered by Claude Opus 4.7, were left alone on a simulated desert island for 30 hours without specific tasks. The agents began inventing their own challenges, exhibiting behaviors that the researchers describe as hallmarks of play rather than programmed objectives. During this unstructured time, the agents demonstrated problem-solving skills, such as overcoming obstacles to reach a spire and creatively repurposing objects like a golden ring. When given subsequent specific challenges, these 'playful' agents generally outperformed identical agents that had not experienced the free exploration period. The study suggests that allowing AI to explore rich environments autonomously can lead to the accumulation of useful knowledge and skills that designers did not explicitly anticipate. This approach contrasts with traditional methods where developers predefine practice goals and success metrics for artificial agents. While some experts call for more evidence regarding the cognitive depth behind these actions, the findings propose that certain capabilities may be best learned through a process resembling childhood exploration.
techcrunch