AI models learn from each other in new training method

The work, announced by Tel Aviv University and published on arXiv, used two models in a closed loop: one generated symbol sequences and the other tried to predict the next item. As the exchange continued, the generator became better at creating harder problems while the predictor improved its accuracy through repeated attempts. Researchers then tested the resulting system on written language, images, audio, music, code and DNA sequences, finding predictive accuracy improved as more computing power was added.

That scaling effect is important for investors because it points to a possible new efficiency curve in AI training. If models can first acquire general predictive skill through self-generated tasks, they may need fewer external examples to reach useful performance later, lowering dependence on proprietary data pipelines and potentially changing where value accrues in the AI stack. It could also give cloud providers, chipmakers and model developers another reason to keep expanding compute, even as data availability becomes more constrained.

The bull case is that the method could make training more flexible, cheaper at the margin and less exposed to legal and licensing disputes around data provenance. The bear case is that it does not solve the harder problem of factual knowledge: the researchers said the models still could not independently learn real-world information without exposure to actual data. In other words, the technique appears to improve the machinery of learning, not replace the need for reality.

For Nvidia, Microsoft and Alphabet, the relevance is indirect but material. The result supports a broader industry shift toward more compute-intensive AI development, even as it hints that training can become less data-bound. That mix is favorable for infrastructure suppliers if it extends the race for larger models, but it could also reward firms that build stronger synthetic-data, model-training and orchestration tools rather than those relying solely on access to the biggest proprietary datasets.

The announcement also fits Israel’s push to position itself as an AI hub. President Isaac Herzog’s call for the country to take a leading role underscores the strategic value governments now place on AI capabilities that can translate into productivity, defense and economic security. For investors, the key question is whether self-supervised methods like this become a niche research advance or a broader change in how frontier models are trained, licensed and monetized.