Haniya Rae’s Post

Giant Self-Licking Ice Cream Cone Great paper in Nature Magazine (Open Access! "AI models collapse when trained on recursively generated data" https://lnkd.in/gZpAZUhu) on what Jathan Sadowski called "Habsburg AI": data inbreeding leads to unseemly results (https://lnkd.in/gXMWhkJV: "a system that is so heavily trained on the outputs of other generative AI's that it becomes an inbred mutant, likely with exaggerated, grotesque features. It joins the lineage of Potemkin AI."). Or "Model Autophagy Disorder", MAD (https://lnkd.in/gyNE24KQ). Or "Giant Self-Licking Ice Cream Cone". https://lnkd.in/gXFY_Pkb

You are what you eat... And it applies to AI systems as well. This is a growing concern since many texts and images online are AI-generated and will feed large language models if we are not careful. Therefore, the output of AI tools will be less and less reliable and useful. Just have a look at the images in this article below to see what I mean: https://lnkd.in/eDmWQxuw

AI is like a new toy. Folks are excited and play around with it to find out what it can or can't do. Some think it is the panacea to world problems. Others think it is the end of civilisation as we know it. The reality is that more research is needed to understand the limitations as well as benefits of this new tool. This recent article from researchers at Cambridge shows why it is imperative that organisations invest in human capital with skills such as critical thinking and not be over reliant on large language models which are likely to be based on out of date, limited and bias real world data as well as synthetic data. The old adage stands, garbage in, garbage out. https://lnkd.in/egMCGPAr

As more AI generated data gets mixed with real data then there is a risk to model output accuracy and stability. This paper highlights the rapid collapse of AI models as they consume synthetic data

What is AI-cannablism? It turns out that eating one's own kind causes problems for AI models as well. When researchers gave successive versions of an LLM information produced by previous generations of the AI, a rapid collapse in output quality was observed. With an estimated 90% of internet content expected to be AI-generated by 2026, understanding whether or not content is organic will be crucial. https://lnkd.in/gisC6UQW

AI models fed AI-generated data quickly spew nonsense, in a phenomenon called model collapse, which appears to affect models of all sizes. This is highly consequential research at a time when models are reaching limits of existing--and available--data.

Why you can't train #AI models with data generated from other #LLMs. #ModelCollapse. AI models fed AI-generated data quickly spew nonsense https://lnkd.in/g9j2zyZK

AI models fed AI-generated data quickly spew nonsense, and it makes perfect sense. Remember making copies on a Xerox machine? copy → copy → copy → copy → garbage.

The fact that feeding AI-generated data into AI models results in model collapse is hardly surprising to anyone with experience in Chaos Theory. One of the fastest ways to create a fully chaotic system is to recursively feed data from the output of a system back into itself. You can try this at home by pointing your video camera at your monitor/TV as it displays the video coming from your camera. Alternatively, feeding audio from your computer back through your computer microphone. The screech that you hear is pure chaotic noise. #ai https://lnkd.in/d3EfHhAw