Researchers from Tel Aviv University and Stanford University demonstrated that artificial intelligence can develop general predictive abilities without training on real-world information. The study introduced a method where two models collaborate, with one generating computer programs to create training sequences and the other learning to predict them. This approach allowed the system to improve its performance across various domains, including language, images, and DNA sequences. The findings challenge the assumption that AI requires vast amounts of external data to acquire useful skills, suggesting a more efficient path for future model training.
Read the full story at english.tau.ac.il.
