Transcript for:
AI Inbreeding and Model Collapse

Foreign [Music] [Music] There's been some reporting now around AI generated images having a a weird yellow tint. Why are all these images so yellow? That might be because um the model being used to create those images is being trained on AI generated data and in that AI generated data there might be an over representation of yellow but the AI system is just replicating and reproducing and kind of fixating on yellow. exact model collapse can happen where you start training machine learning models on data created by other machine learning models. I've called this in in my work a form of Hapsburg AI. the Hapsburgs, the the very famous political dynasty in Europe that was also very famous for lots of inbreeding, creating these kind of exaggerated, even grotesque features. Very quickly, just after a few generations of training a model on a substantial amount of AI created data, the model can start to fall in on itself. It can start to generate really bizarre and weird uh uh outputs. Fore the whole reason why AI inbreeding is even a problem is because on one hand there's a problem of data scar. university. A lot of the AI systems that already exist, so like chat GBT or Anthropics Claw and and so on, they've essentially been trained on all of the human produced data that exists right now. People are not able to produce enough data at a fast enough pace to satiate the hunger of these AI model. And so that's why companies are turning towards AI produced data as a solution to that data scarcity. You can have lots more synthetic data a lot quicker and a lot cheaper and a lot easier. No one's complaining about copyright or anything like that. Some studies have shown that very quickly and with a substantially high enough amount of synthetic data in the training set for a machine learning model, the model just becomes essentially useless. A lot of companies are trying to downplay the threat of AI inbreeding or model collapse or synthetic data. Um because if they come out and admit then their their share prices and their investments are going to tank, right? I'm skeptical that we will continue that we are going to continue seeing um such rapid improvement in these technologies because I think we're we're already seeing quite the opposite that kind of rapid improvement of these AI systems is slowing down but nobody is able to admit that they're slowing down. Fore [Music] speech.