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

Researchers at Tel Aviv University (TAU) and Stanford University have developed a new artificial intelligence training method that allows AI systems to teach one another and improve their predictive abilities without relying on human-generated content or external datasets, TAU announced Wednesday.  

For more stories from The Media Line go to themedialine.org

The findings, published on the scientific platform arXiv, suggest that artificial intelligence may be capable of developing fundamental learning skills through challenges created entirely by other AI models, potentially changing how future systems are trained.

Unlike conventional AI training, which depends on enormous quantities of existing information, the new approach allows models to build predictive capabilities before encountering real-world material.

Researchers used AI models to generate, predict sequences

The researchers tested the concept by creating a training system involving two AI models. One was tasked with generating sequences of symbols, while the other attempted to predict which symbol would appear next. 

The process did not require the models to consult written documents, photographs, or other external sources. Instead, the training material was generated within the system itself.

As the exercises progressed, the model responsible for producing the sequences became increasingly sophisticated at creating challenges that would improve its partner's performance. The second model, in turn, developed stronger predictive abilities by repeatedly attempting to solve those problems.

The approach effectively allows an AI system to develop learning skills through a continuous process of generating and solving increasingly difficult exercises. 

To determine whether those skills could be applied beyond the artificial training environment, the researchers evaluated the resulting model against several types of information, including written language, images, audio, music, computer code, and DNA sequences.

The experiments demonstrated that predictive accuracy increased as additional computing resources were allocated to training, suggesting that the method could become more effective as computational capacity expands.

The findings also establish an important distinction between learning how to recognize patterns and acquiring knowledge about the world.

Researchers could not acquire information about real-world subjects

Although the models developed predictive capabilities without external training data, they could not independently acquire factual information about real-world subjects. Such knowledge still requires exposure to actual data.

Researchers believe the technique could nevertheless make future AI systems more efficient by strengthening their underlying learning abilities before they begin processing real-world information.

The announcement comes as Israel seeks to strengthen its position in artificial intelligence. On November 3, President Isaac Herzog convened the inaugural session of the “Beyond Intelligence” forum, established with the National AI Directorate in the Prime Minister’s Office and the Mosaic Institute for AI Policy to examine AI's impact on employment, economic security and society.  

Herzog urged Israel to make leadership in the field a strategic priority, declaring: “Israel must take its place at the forefront of artificial intelligence. This train will not wait for us, and the whole world is looking our way.”