The University of California studied the problem of "task contamination", where the parameters of a large model freeze after training and people keep asking new tasks, i.e., the input distribution keeps changing. Over time, people start asking more new questions, and the AI doesn't perform as well.
New explanation for ChatGPT getting stupid: it's not the same as it was in training
Previous: “人形机器人第一股”诞生！优必选上市首日市值逼近390亿港元