Artificial intelligence, now touted as the most promising field in technology, also seems to have hidden crises. Recently, a new study found that feeding back AI-generated content to similar models may lead to a gradual decline in model quality or even collapse. This phenomenon of "self-enophagogy" is called model autophagy by scientists (a bit like self-phagy of cells).
The researchers note that with the tremendous progress made in generating AI algorithms in areas such as images, text, and more, it becomes very tempting to use synthetic data to train models. However, repeating this training process causes the model to become more and more closed, extracting only information from a limited amount of data, and eventually making the model lose diversity and accuracy. In other words, without enough real-world data to feed these models, they are likely to fall into a cycle of self-destruction.
The findings raise concerns about artificial intelligence. At present, the training data of artificial intelligence models mainly comes from the Internet, and a large amount of the data comes from the generated content. With the widespread use of generative models, more and more synthetic content emerges that is not easily discernible. This makes AI companies face significant challenges in ensuring the quality of training datasets and network structure.
Although there is no conclusive evidence of how this phenomenon will affect the real world, we still need to be vigilant about the future of AI. After all, how useful will these models actually be if they can't accept human input? AI may not completely replace humans as we think, but it may make us slaves to content farms, constantly generating content for it and keeping it running.
Thankfully, there are some ways to deal with this, such as adjusting model weights, etc. We hope that while exploring the development of artificial intelligence, we can avoid this situation where the future seems meaningless.
Self-devouring infinite loop
What do you think of the results of this study? How do you think this will affect the development of AI? Everyone is welcome to speak and share your views and opinions! Let's explore this fascinating and thought-provoking topic together!