The article presents a clear and simplified definition of reinforcement learning (RL), tying it to psychological principles of learning. Lior Fox, a cognitive scientist, frames RL as an iterative process: receiving feedback on performance and adjusting actions based on that feedback. This definition is applied to both human learning and computer science, where a program interacts with an evaluation environment, produces responses, and is scored on those responses. The goal is to optimize performance over time, aiming for the highest possible average score through repeated iterations. The author highlights Reformist Reinforcement Learning, which utilizes a generative model to create responses. In this approach, the computer generates random outputs based on input from the evaluation environment, scores them, and updates its responses using only data associated with positive scores. This method simplifies implementation, especially in large language models, by allowing for rapid adjustments between pretraining and posttraining processes. The article also addresses the role of Markov Decision Processes (MDPs) in traditional AI education. While MDPs have been central in classical reinforcement learning, the author argues that in Reformist RL, they are less fundamental. MDPs are seen as a specific framework within a broader set of RL problems, where scoring and evaluation dynamics are crucial. The author’s concise overview reduces complex RL concepts to a digestible format while maintaining a connection to broader machine learning principles.