MIT researchers release findings from 30-hour experiments where multiple AI agents performed goal-free, play-like exploration and demonstrated improved carryover to later tasks
Research publication Provisional 78% confidence first seen
MIT researchers report results from experiments in which they left 13 AI agents (Eko) running for about 30 hours in simulated environments with minimal or no assigned goals. They found that the agents generated their own exploratory challenges and that the resulting skills generally carried over to later tasks compared with agents that did not run the open-ended exploration.