Human Data to Self-Training
Human Data to Self-Training The Bootstrap Problems
Recursive self-creation in modern AI training For most of the deep learning era, the recipe was simple. Find data made by humans, show it to a model, adjust the weights. The model was a mirror. Whatever the internet contained, it learned to reproduce. Progress came from scaling three things at once, and the third one had a floor under it. There was only so much text on Earth. That floor is why the field turned to something stranger. Instead of feeding models more human output, we began feeding them their own. A model generates candidate solutions, some filter decides which ones are good, and the survivors become the next round of training data. The system becomes both the student and the textbook. This is recursive self-creation, and it now sits at the center of how frontier models are built. The bootstrap problem At first glance the idea should not work. If a model already knows how to produce good training examples, it should already be good. Learning from yourself sounds like lifting yourself by your own belt. The resolution is that generating and judging are not the same skill. It is far easier to recognize a correct proof than to produce one. It is easier to check that code compiles and passes tests than to write it. This asymmetry, sometimes called the generator-verifier gap, is the engine that makes the whole loop turn. Here is the mechanic in its simplest form. Ask a model the same math question a hundred times with some randomness in the sampling. Most answers will be wrong. A handful will be right. If you can tell which is which, and for math you often can because you have the final answer, you now possess a hundred attempts and a reliable filter. Keep the winners. Train on them. The model's average behavior shifts toward what its best behavior used to be. Nothing was created from nothing. The capability was already latent in the distribution, showing up occasionally by luck. The loop simply harvests luck and makes it habit.