AI Models Are Starting to Learn by Asking Themselves Questions

Researchers from Tsinghua University have developed Absolute Zero Reasoner (AZR), an AI system that learns by generating its own coding problems and solving them, rather than copying human examples. The system creates challenging Python problems, attempts to solve them, and uses the results to improve both its problem-creation and problem-solving abilities. This self-teaching approach significantly improved AI reasoning skills and outperformed some models trained on human data, potentially offering a path toward AI systems that surpass human capabilities.