Most people here don’t understand what this is saying.
We’ve had “pure” human generated data, verifiably so since LLMs and ImageGen didn’t exist. Any bot generated data was easily filterable due to lack of sophistication.
ChatGPT and SD3 enter the chat, generate nearly indistinguishable data from humans, but with a few errors here and there. These errors while few, are spectacular and make no sense to the training data.
2 years later, the internet is saturated with generated content. The old datasets are like gold now, since none of the new data is verifiably human.
This matters when you’ve played with local machine learning and understand how these machines “think”. If you feed an AI generated set to an AI as training data, it learns the mistakes as well as the data. Every generation it’s like mutations form until eventually it just produces garbage.
Training models on generated sets slowly by surely fail without a human touch. Scale this concept to the net fractionally. When 50% of your dataset is machine generated, 50% of your new model trained on it will begin to deteriorate. Do this long enough and that 50% becomes 60 to 70 and beyond.
Human creativity and thought have yet to be replicated. These models have no human ability to be discerning or sleep to recover errors. They simply learn imperfectly and generate new less perfect data in a digestible form.
Wasn’t there a paper not long time ago that it was possible to generate data with AI as a training set for AI? I was surprised (and the math is to much for me to check out my self) but that seems to solve that problem.
As far as I know, that is mainly used where a better, bigger model generates training data for a more efficient smaller model to bring it a bit closer to its level.
Were there any cases of an already state of the art model using this method to improve itself?
I will search for the paper.
EDIT: can’t find it, dang.
Sorta. This “model collapse” thing is basically an urban legend at this point.
The kernel of truth is this: A model learns stuff. When you use that model to generate training data, it will not output all it has learned. The second generation model will not know as much as the first. If you repeat this process a couple times, you are left with nothing. It’s hard to see how this could become a problem in the real world.
Incest is a good analogy, if you know what the problem with inbreeding is: You lose genetic diversity. Still, breeders use this to get to desired traits and so does nature (genetic bottleneck, founder effect).
Training data for models in general was a big problem when I studied systems biology. Interesting that we finding works around, since it sounded rather fundamental to me. I found your metaphor rather helpful, thanks.
Microsoft’s Phi model was largely trained on synthetic data derived from GPT-4.
I’m to lazy to search for the paper, not sure it was Microsoft, but with my rather basic knowledge of modeling (studied system biology) - it seemed rather crazy and impossible, so I remembered it.