This paper explores first-order meta-learning algorithms, focusing on methods that use only first-order derivatives for updates. It introduces Reptile, a new algorithm that repeatedly samples tasks, trains on them, and adjusts the initialization towards the trained weights. The paper also discusses first-order MAML (FOMAML), which approximates MAML by ignoring second-order derivatives. Both algorithms are analyzed theoretically and empirically, showing their effectiveness on benchmarks like few-shot classification. The authors provide insights into implementation practices and discuss the potential of first-order meta-learning algorithms, particularly in scenarios where higher-order gradients are cumbersome to compute. The paper includes a case study on 1D sine wave regression and detailed experimental results on datasets such as Mini-ImageNet and Omniglot, demonstrating the performance and robustness of Reptile and FOMAML.This paper explores first-order meta-learning algorithms, focusing on methods that use only first-order derivatives for updates. It introduces Reptile, a new algorithm that repeatedly samples tasks, trains on them, and adjusts the initialization towards the trained weights. The paper also discusses first-order MAML (FOMAML), which approximates MAML by ignoring second-order derivatives. Both algorithms are analyzed theoretically and empirically, showing their effectiveness on benchmarks like few-shot classification. The authors provide insights into implementation practices and discuss the potential of first-order meta-learning algorithms, particularly in scenarios where higher-order gradients are cumbersome to compute. The paper includes a case study on 1D sine wave regression and detailed experimental results on datasets such as Mini-ImageNet and Omniglot, demonstrating the performance and robustness of Reptile and FOMAML.