# OpenAI Introduces Evolved Policy Gradients for Faster AI Learning

OpenAI has announced Evolved Policy Gradients (EPG), an experimental metalearning technique that could significantly speed up how AI agents learn new tasks.

Unlike traditional machine learning approaches that use fixed loss functions, EPG actually evolves the loss function itself during training. This allows AI agents to adapt more quickly when facing novel situations they haven't explicitly been trained on.

The breakthrough demonstrates impressive generalization capabilities. In testing, agents trained with EPG successfully completed basic tasks that fell outside their original training parameters. For example, an agent could learn to navigate to an object placed on a different side of a room than it had seen during training—a task that would typically stump conventional AI systems.

This metalearning approach represents a step toward more flexible artificial intelligence that can transfer knowledge across different scenarios without extensive retraining. Rather than learning just one specific task, EPG helps