Ben Recht critiques the traditional understanding of reinforcement learning (RL), particularly how rewards are conceptualized within RL frameworks. He highlights a key issue: rewards are often viewed as external signals from the environment, which the agent cannot alter. Sutton and Barto's textbook reinforces this idea by stating that the reward signal defines the agent's goals and is part of the environment. Recht argues this perspective is flawed. He believes rewards should be seen as internal to the agent, suggesting that agents interpret changes in their environment and translate those into rewards. This shift in understanding aligns more closely with cognitive science views on learning. Recht proposes a revised model where the agent acts in the environment, the environment changes, and the agent then observes and interprets those changes into rewards. This adjustment doesn't fundamentally alter existing RL algorithms; it simply repositions the reward function within the agent. He claims this new framing has practical benefits, allowing different agents to learn in the same setting while developing various policies based on their individual goals. For instance, in a gaming context, agents could have distinct objectives like speed running or exploring, leading to diverse strategies despite facing the same game. The author also addresses concerns that this internalization of reward computation might limit the generality of RL agents. While critics argue this could make agents less adaptable to various environments, Recht sees it as an advantage. He argues that recognizing the importance of the reward heuristic clarifies the RL framework rather than obscuring its limitations. By changing how we think about rewards, new avenues open up for enhancing how agents learn and adapt within their environments.