Gradient Norm Aware Minimization Seeks First-Order Flatness and Improves Generalization | Read Paper on Bytez