Today's AI models start with a first, broad stage of training, when they learn from a huge collection of text, images or video made by people. In a paper posted on Sept. 24, 2026, researchers from Stanford University, Tel Aviv University and elsewhere tried training without that human-made data. They argue that if a model can learn to make the training data most useful for its own improvement, that first stage would no longer be held back by collections gathered by people and would be limited mainly by computing power. They call their work an initial proof of concept.
The method, a form of self-play, pairs two models that start from random settings. The first, a generator, writes programs for a bare-bones universal computer, a machine that can in principle run any program, and running those programs produces strings of data. The second, a learner, is trained to predict that data one piece at a time. The generator learns by trial and error, rewarded for programs whose output sits at the edge of what the learner can still learn, so the lessons get harder as the learner improves. Neither model ever trains on real text, images or other real-world data. That sets the method apart from self-play work such as Absolute Zero, which sharpens the reasoning of a model that has already been trained, and from G-Zero, a separate paper on self-play for open-ended generation.
After this self-play alone, the researchers report, the learner got better at predicting real-world data it had never seen, including text and music, and the gains followed a steady, predictable curve as they spent more computing power. The learner also picked up simple new tasks, such as reversing a string of characters or recalling which item was paired with which, just from examples shown to it, with no further training. Once it had seen enough examples, it got those tasks almost always right.
The generator also found familiar math patterns on its own. By round 512, it was writing programs that produce Fibonacci-like sequences. Without that learning, the researchers estimate, such programs would not be expected to appear until after more than 53,000 rounds.
The models are small, under 25 million parameters, the numbers a model tunes as it learns. The authors say this kind of training is not a replacement for real data, which a model still needs to learn facts about the actual world; they see it instead as a way to isolate universal structure. They did use real text and DNA data to choose some training settings, but they say that data never went into the training itself.
The paper's authors are Aditya Cowsik, Kfir Dolev, Michael Y. Li, G. Bruno De Luca, Nourya Cohen, Noah D. Goodman and Yoav Levine, from independent research, Tel Aviv University, Stanford University and LAPTh, USMB. They had not released the trained models as of Sept. 25, 2026. Whether the method keeps working at much larger scale, once real data is scarce or expensive, has not been tested.