This paper explores the use of evolutionary algorithms to automatically discover neural network architectures for image classification tasks. The authors aim to minimize human intervention by employing evolutionary techniques to evolve models from trivial initial conditions, achieving competitive accuracies on the CIFAR-10 and CIFAR-100 datasets. They introduce novel mutation operators that navigate large search spaces and demonstrate the repeatability and variability of the results. The study emphasizes the simplicity and efficiency of the algorithm, which produces fully trained neural networks without post-processing. The results show that it is possible to evolve models with accuracies comparable to those achieved by hand-designed models, given sufficient computational resources. The paper also discusses the impact of meta-parameters and computational costs, highlighting the potential for future improvements in algorithmic and hardware implementations.This paper explores the use of evolutionary algorithms to automatically discover neural network architectures for image classification tasks. The authors aim to minimize human intervention by employing evolutionary techniques to evolve models from trivial initial conditions, achieving competitive accuracies on the CIFAR-10 and CIFAR-100 datasets. They introduce novel mutation operators that navigate large search spaces and demonstrate the repeatability and variability of the results. The study emphasizes the simplicity and efficiency of the algorithm, which produces fully trained neural networks without post-processing. The results show that it is possible to evolve models with accuracies comparable to those achieved by hand-designed models, given sufficient computational resources. The paper also discusses the impact of meta-parameters and computational costs, highlighting the potential for future improvements in algorithmic and hardware implementations.