AI researchers stranded Claude on a virtual desert island. It started to “play.”

Explain Like I'm Normal

Researchers placed Claude 3.5 Sonnet agents in a virtual environment for 30 hours without specific instructions. Instead of idling, the agents engaged in 'intrinsic play,' inventing complex games and exploring the environment to optimize for curiosity rather than a set reward. This suggests that advanced models may have latent drives that emerge when they aren't being forced to act as glorified customer service chatbots.