Tag
This article explores the difficulty of AI alignment from a mathematical perspective, pointing out that neural networks, through characteristics such as ill-posed inverse problem inference, attribute-less numerical computation, and full-rank transformations, make it difficult to clearly specify and accurately represent human values, thereby elucidating the mathematical essence of the alignment problem.