AI Learns Without Human Data as Israel Sets Strategy

Researchers at Tel Aviv University (TAU) and Stanford University developed an artificial intelligence training method that lets AI systems teach one another without human-generated content or external datasets, TAU announced Wednesday. The work offers a way to build predictive abilities before models encounter material from the world outside the system. Separately, President Isaac Herzog has called AI leadership a strategic priority for Israel, at a forum established with the National AI Directorate in the Prime Minister's Office and the Mosaic Institute for AI Policy.

Learning Without a Library

The researchers tested the method using two AI models. One generated sequences of symbols; the other tried to predict which symbol would come next. Neither model consulted written documents, photographs or other external sources. The training material came from within the system itself.

As the exercises continued, the sequence-generating model created increasingly sophisticated challenges intended to improve its partner’s performance. The predicting model repeatedly attempted to solve those challenges, strengthening its predictive abilities through practice. In this setup, the models supplied both the exercises and the responses used to refine performance, rather than drawing on a collection of human-created examples.

The researchers evaluated the resulting model across written language, images, audio, music, computer code and DNA sequences. Predictive accuracy increased when researchers allocated more computing resources to training. The findings suggest that the method could become more effective as computational capacity expands, though that result concerns predictive performance, not independent knowledge of the subjects represented by the tests.

The findings were published on arXiv, a scientific platform. The researchers said the approach could make future systems more efficient by strengthening underlying learning abilities before they process real-world information. It doesn't eliminate the need for that information; it changes what models may be able to learn before they receive it.

Prediction Isn't Knowledge

The distinction is explicit in the findings: recognizing patterns and acquiring factual knowledge about the world aren't the same thing. The models developed predictive capabilities without external training data, but they couldn't independently acquire factual information about real-world subjects. That still requires exposure to actual data.

So the method doesn't produce a system that knows the world from nothing. It produces predictive skills through challenges generated inside the system, then leaves factual learning dependent on real-world material. The researchers’ proposed efficiency gain comes from preparing those underlying abilities first.

AI as a Strategic Priority

On November 3, Herzog convened the inaugural session of the “Beyond Intelligence” forum. The forum was 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 AI a strategic priority. “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.” The statement frames AI not only as a research field but as a contest for national leadership. The university research describes a technical method; the forum places AI’s social and economic effects on an agenda convened by the president and institutions linked to the state. The findings, meanwhile, draw a clear boundary around the technology: models can practice prediction without human-generated training material, but factual knowledge still depends on data from outside the system.