AI models can now learn from better AI models during training, not just after

Researchers found a way to feed a model's output back into its own training process, instead of waiting until training is done to improve it. This means models can learn safety, accuracy, and reasoning patterns earlier and more thoroughly than the traditional approach.

Training large language models has always followed the same recipe: dump raw text into the model, then spend months afterward trying to fix what went wrong. This paper shows that fixing can start earlier, while the model is still learning its basic patterns. The practical effect is that you can build better models without necessarily training longer or using more data. The catch is obvious: it only works if you already have a good model to learn from, which means the winners are organizations that can afford multiple training runs.