Could AI be getting dumber? | TechCrunch Minute

Key Takeaways

[Music] although AI models have made big breakthroughs in the last few years there's now a risk that they could get dumber and even collapse according to a new research paper published in nature the phenomenon that the researchers describe is called Model collapse think of it as a snake eating its own tail or as this illustration shows a process that starts with a bunch of photos of real dogs of different breeds then ends with identical weird smudges that kind of resemble golden retrievers the thing to remember is that for AI models to get smarter they need to get trained on new data ideally with diversity and depth so that's what goes in what comes out is essentially the result of pattern matching giving you the most common or likely output for a given prompt for example if you ask an AI image generator to give you a picture of a dog it probably won't give you a picture of some obscure or Rare Breed instead you'll get a golden retriever or a lab because that was most common in the images it was trained on now consider the fact that there's already a lot of AI generated content on the web and the quantity is only going to go up from here consider also that some companies are so hungry for data that they're actually choosing to train their models on synthetic data that's been generated by AI what results according to researchers is quote a degenerative process whereby over time models forget the true underlying data distribution in other words over time the content going in and the content coming out converge and turn into garbage Golden Retrievers in Golden retriever's out and eventually the model loses track of what a dog even is you can imagine something similar with images of real people or blog posts on pretty much any topic so can this be prevented well there could be different benchmarks the AI companies aim for when they're looking at the sourcing and variety of their training data and there could also be watermarks that identify content as AI generated unfortunately there hasn't really been any standardization in those areas and companies don't necessarily have an in incentive to share this information so in the future large language models might really struggle unless they're trained on data that was crawled from the internet before it was overrun with AI content I'll see you tomorrow [Music]