Israeli Researchers Develop AI That Learns Without Human Data

Researchers from Tel Aviv University (TAU) and Stanford University have developed a novel method for training artificial intelligence systems, enabling them to learn from each other without relying on human-generated data or external datasets. The findings, published on the scientific platform arXiv, suggest that AI could develop fundamental learning skills through challenges created entirely by other AI models, potentially transforming future training methodologies.

Unlike traditional training methods that require vast amounts of pre-existing information, this new approach allows AI models to build predictive capabilities before encountering real-world data. The researchers tested this concept by setting up a training system with two AI models: one generated symbol sequences, while the other attempted to predict the next symbol. This process did not require the models to access written documents, images, or other external sources, as training materials were generated internally.

As the exercises progressed, the sequence-generating model became more adept at creating challenges that improved the other model's performance. In turn, the second model developed stronger predictive abilities through repeated attempts to solve these challenges. This system allows AI to develop learning skills through a continuous process of creating and solving progressively difficult exercises.

To assess the applicability of these skills beyond the artificial training environment, the researchers evaluated the resulting model on diverse data types, including written language, images, audio, music, code, and DNA sequences. Experiments showed that prediction accuracy increased with greater computational resources dedicated to training, indicating potential for enhanced effectiveness with expanding computing power.

The results also highlight a distinction between learning to recognize patterns and acquiring knowledge about the world. While the models gained predictive abilities without external training data, they did not independently acquire factual information about real-world subjects, which still requires exposure to actual data. The researchers believe this technique could lead to more efficient future AI systems by strengthening their core learning capabilities before they process real-world data.

This announcement comes as Israel aims to bolster its position in the AI field. On November 3, President Isaac Herzog inaugurated the first session of the "Beyond Intelligence" forum, established in collaboration with the National AI Directorate in the Prime Minister's Office and the Mosaic AI Policy Institute, to study AI's impact on employment, economic security, and society. Herzog urged Israel to prioritize leadership in AI, stating, "Israel must take its place at the forefront of the world's nations in the field of artificial intelligence; this train will not wait for us, and the whole world is watching us."