Israeli Study Explores AI Training with Self-Generated Data, Potentially Addressing Data Scarcity

Researchers from Tel Aviv University and Stanford University have demonstrated that artificial intelligence (AI) models can develop general learning abilities without being trained on real-world data. This preliminary study, published as a preprint on arXiv, suggests a new approach to AI development that could mitigate the increasing scarcity of high-quality human-generated training material. The findings indicate that models exposed to self-generated exercises learned faster when subsequently introduced to real-world data. However, the experiments involved models significantly smaller than current large-scale AI systems, and further research is required to assess the scalability of these results.