This paper proposes a Meta-learning and Markov Chain Monte Carlo (MCMC) based approach to solve the blind single image super-resolution (SISR) problem. The approach aims to learn kernel priors from organized randomness, avoiding the need for handcrafted or pre-trained priors. The method consists of two main phases: the Markov Chain Monte Carlo kernel approximation (MCKA) phase and the meta-learning based alternating optimization (MLAO) phase. In the MCKA phase, MCMC simulations on random Gaussian distributions are used to generate random kernel priors, which are then optimized using a lightweight network. This process provides a rational blur kernel and introduces Langevin dynamics into the SISR optimization, preventing the optimization from getting stuck in local optima. The MLAO phase refines the blur kernel and restores the high-resolution (HR) image, alternating between kernel estimation and image restoration. A meta-learning-based adaptive strategy is used to optimize the non-convex and ill-posed blind SISR problem, ensuring better convergence performance. The proposed approach is evaluated on synthesis and real-world datasets, demonstrating superior performance and generalization compared to state-of-the-art methods. The code is available at <https://github.com/XYLGroup/MLMC>.This paper proposes a Meta-learning and Markov Chain Monte Carlo (MCMC) based approach to solve the blind single image super-resolution (SISR) problem. The approach aims to learn kernel priors from organized randomness, avoiding the need for handcrafted or pre-trained priors. The method consists of two main phases: the Markov Chain Monte Carlo kernel approximation (MCKA) phase and the meta-learning based alternating optimization (MLAO) phase. In the MCKA phase, MCMC simulations on random Gaussian distributions are used to generate random kernel priors, which are then optimized using a lightweight network. This process provides a rational blur kernel and introduces Langevin dynamics into the SISR optimization, preventing the optimization from getting stuck in local optima. The MLAO phase refines the blur kernel and restores the high-resolution (HR) image, alternating between kernel estimation and image restoration. A meta-learning-based adaptive strategy is used to optimize the non-convex and ill-posed blind SISR problem, ensuring better convergence performance. The proposed approach is evaluated on synthesis and real-world datasets, demonstrating superior performance and generalization compared to state-of-the-art methods. The code is available at <https://github.com/XYLGroup/MLMC>.