Jews in the News / Science & Medicine
Science & Medicine

Israeli Researchers Show AI Can Learn General Skills From Self-Generated Data

A Tel Aviv University and Stanford team found that AI models can gain broad predictive abilities using only sequences they generate themselves, not real-world data.

Researchers at Tel Aviv University, working with colleagues at Stanford University, have shown that artificial intelligence systems can acquire broad predictive skills without ever training on real-world data such as text, images, or recordings. The study, led by Dr. Yoav Levine of Tel Aviv University's Blavatnik School of Computer Science and conducted as part of Dr. Kfir Dolev's postdoctoral research, challenges the long-standing assumption that AI must first absorb enormous amounts of real-world material to learn anything useful.

The team built a training method in which two small models learn together from a blank slate. One model writes short computer programs that generate sequences of symbols, acting as a teacher that designs exercises. The other model tries to predict the next symbol in each sequence. Feedback lets the teacher model gradually craft challenges that match the learner's growing abilities, so the two improve together without any outside input.

After this self-generated training, the researchers tested the learner model on real material it had never seen, including text, images, sound, music, computer code and DNA sequences. Its predictions improved consistently as more computing power went into the self-play process, and it could pick up new rules from examples shown during testing alone. The researchers also noticed that the programs generating training data began producing recognizable mathematical structures, including the Fibonacci sequence, suggesting the system was discovering useful patterns on its own.

The team cautions that self-generated training cannot teach a model facts about the real world; it only seems to build general predictive ability that can later make learning from real data faster and more efficient.

Why it matters

If this approach scales up, it could reduce AI developers' reliance on scraping or licensing vast amounts of real-world text, images and other content, easing some cost and privacy pressures tied to data collection. It also points toward more efficient training methods that could speed future AI research and applications.

Sources

Summarised in our own words from the reporting above.

Tikkun olam, today. This is part of tikkun olam, the Jewish idea of repairing the world, and every one of us can add to it.