Breakthrough by Israeli scientists: AI on the verge of cardinal changes

Researchers from Tel Aviv and Stanford Universities have demonstrated that AI can acquire predictive skills without training on real-world data. Two models interacted: one generated sequences of symbols, the other learned to predict them. After this preparation, the model improved its results on data from various fields—from language to genetics.

Researchers from Tel Aviv and Stanford Universities have shown that artificial intelligence can acquire general predictive skills without training on texts, photographs, recordings, or other real-world data. The model was trained exclusively on artificially created tasks and then improved its results when tested on data from various fields—from language and music to genetics. The method is called Zero-Data Pretraining. In the experiment, two models started from scratch and learned through interaction: the first created short computer programs that generated sequences of symbols, and the second tried to predict the next symbol. After this preparation, the trained model performed better at predicting on data that was not in its training tasks. The work was conducted under the supervision of Dr. Yoav Levine. The study offers a preliminary answer to the question of whether intelligence has a universal component that can be developed without external data. Levine emphasizes the limited scale of the experiment.