RL course Part 13 shows how AI labs use reinforcement learning in production

Daily Dose of DS released Part 13 of its reinforcement learning course, the final chapter in a 13-part series that walks readers through the complete mechanics of RL. The course spans foundational concepts—bandits, value functions, Bellman equations, policy gradients—and progresses through advanced topics including PPO, GRPO, judges, trajectories, and environments. Part 13 serves as a capstone, moving from theory to practice by examining four concrete case studies of how AI labs and companies actually deploy RL in production.

The chapter's pedagogical design centers on a practical skill: reading a company's technical blog post about their training pipeline and immediately identifying the environment, reward source, trajectory structure, and likely failure modes—using the vocabulary built across twelve prior chapters. That vocabulary includes concepts like reward hacking (gaming the reward signal), credit assignment (determining which actions led to outcomes), environment design (building the learning space), and the tension between learned rewards and verifiable rewards (signals that can be proven correct versus those that optimize empirically).

The four case studies illustrate different applications of RL. Cursor runs a real-time RL loop that ships improved checkpoints every five hours, showing continuous model refinement in practice. Frontier labs have turned verifiable rewards into reasoning and agentic capabilities, demonstrating RL's role in enabling autonomous reasoning. Scale AI reports enterprise results where a 4B parameter model beat GPT-5 on domain-specific tasks—using RL to optimize a smaller model for a narrow domain. And an emerging market has formed where RL environments themselves are built, priced, and sold as products, suggesting RL infrastructure is becoming a commodity. Together, these examples show that RL is not a single pattern but a set of tools adapted to different industrial problems—from real-time model improvement to specialized domain optimization to agent reasoning.

This capstone chapter represents the culmination of a 13-part series designed to build a working vocabulary of reinforcement learning from foundational concepts (bandits, value functions, Bellman equations) through advanced techniques (PPO, GRPO, RLHF, verifiable rewards). The progression is pedagogical: readers move from learning individual RL mechanics across twelve chapters to applying that knowledge to real-world case studies in the final chapter. The body emphasizes a practical test of understanding—the ability to read a company's technical blog post and immediately map it onto RL concepts learned in the series—which indicates the course's goal is not theoretical mastery but industrial literacy.

The four case studies were chosen to illustrate RL's diversity in production. Cursor's five-hour checkpoint cycle represents continuous learning loops, frontier labs' use of verifiable rewards shows how RL enables reasoning and agency, Scale AI's result demonstrates RL's potential for domain-specific optimization, and the RL-environments-as-products market highlights a new business model. Together, they show that RL is not a single application pattern but a toolkit adapted to different industrial problems—from real-time model improvement to agent reasoning to cost-optimized inference.