This paper presents a large-scale evolutionary approach to automatically discovering image classifiers. The authors use evolutionary algorithms to evolve neural network architectures without human intervention, achieving high accuracy on the CIFAR-10 and CIFAR-100 datasets. The method employs simple evolutionary techniques at unprecedented scales, starting from trivial initial conditions and reaching accuracies of 94.6% (95.6% for ensemble) on CIFAR-10 and 77.0% on CIFAR-100. The approach uses novel mutation operators to navigate large search spaces and produces fully trained models without post-processing. The algorithm is simple, with few impactful meta-parameters, and emphasizes repeatability, variability, and computational requirements. The method is parallelized and uses a simplified graph encoding for the DNA, which is transformed into a full neural network for training and evaluation. The algorithm allows for the evolution of complex convolutional neural networks with arbitrary skip connections and numerical parameters. The results show that the evolved models are competitive with hand-designed architectures, and the method is efficient in terms of computational cost. The paper also discusses the importance of weight inheritance and the challenges of escaping local optima in evolutionary processes. The results demonstrate that the evolutionary approach can produce high-accuracy models for image classification tasks, with the potential for further improvements through algorithmic and hardware advancements.This paper presents a large-scale evolutionary approach to automatically discovering image classifiers. The authors use evolutionary algorithms to evolve neural network architectures without human intervention, achieving high accuracy on the CIFAR-10 and CIFAR-100 datasets. The method employs simple evolutionary techniques at unprecedented scales, starting from trivial initial conditions and reaching accuracies of 94.6% (95.6% for ensemble) on CIFAR-10 and 77.0% on CIFAR-100. The approach uses novel mutation operators to navigate large search spaces and produces fully trained models without post-processing. The algorithm is simple, with few impactful meta-parameters, and emphasizes repeatability, variability, and computational requirements. The method is parallelized and uses a simplified graph encoding for the DNA, which is transformed into a full neural network for training and evaluation. The algorithm allows for the evolution of complex convolutional neural networks with arbitrary skip connections and numerical parameters. The results show that the evolved models are competitive with hand-designed architectures, and the method is efficient in terms of computational cost. The paper also discusses the importance of weight inheritance and the challenges of escaping local optima in evolutionary processes. The results demonstrate that the evolutionary approach can produce high-accuracy models for image classification tasks, with the potential for further improvements through algorithmic and hardware advancements.