Israeli researchers develop method letting AI teach itself without human-generated data

Researchers at Tel Aviv University and Stanford University have developed a new AI training method that allows AI systems to teach one another without relying on human-generated content or external datasets. The method uses two AI models to generate and solve increasingly difficult exercises, improving predictive abilities. The findings were published on arXiv.

Researchers at Tel Aviv University and Stanford University have developed a new artificial intelligence training method that allows AI systems to teach one another without relying on human-generated content or external datasets, TAU announced Wednesday. The findings, published on arXiv, suggest AI may develop fundamental learning skills through challenges created entirely by other AI models. The researchers tested the concept by creating a training system involving two AI models: one generated sequences of symbols, while the other attempted to predict the next symbol. The training material was generated within the system itself. As the exercises progressed, the generating model became more sophisticated at creating challenges, and the predicting model developed stronger abilities. The resulting model was evaluated against written language, images, audio, music, computer code, and DNA sequences. Predictive accuracy increased with additional computing resources. The method does not allow models to acquire factual knowledge about real-world subjects, which still requires exposure to actual data. The announcement comes as Israel seeks to strengthen its position in AI; on November 3, President Isaac Herzog convened the inaugural session of the "Beyond Intelligence" forum, urging Israel to make AI leadership a strategic priority.