Tel Aviv and Stanford researchers develop self-teaching AI method

Researchers from Tel Aviv University and Stanford University have developed a new artificial intelligence training method that allows AI systems to teach one another. This approach enables models to improve their predictive abilities without relying on human-generated content or external datasets.

The study, published on the arXiv scientific platform, utilized two AI models: one tasked with generating sequences of symbols and another attempting to predict the next symbol in the sequence. As the process continued, the first model created increasingly sophisticated challenges to improve the second model's performance. This allowed the system to develop fundamental learning skills through a continuous cycle of generating and solving exercises.

To test the applicability of these skills, researchers evaluated the model against various data types, including written language, images, audio, music, computer code, and DNA sequences. While the researchers noted that this method does not allow AI to learn specific facts about the real world without external exposure, it could significantly improve how future systems learn more efficiently when they eventually encounter real-world data.