Learning through consequences
Follow the loop from observation to action to delayed reward. Connect that idea to robotics, preference training, and the infrastructure needed to generate rollouts.
Understand the loop
Separate the policy that acts from the value function that predicts future return.
Course · Hugging Face · Deep reinforcement learning course ↗
Course + videos · UC Berkeley · CS285 Deep Reinforcement Learning ↗
Connect it to language models
Human preferences and verifiable rewards supply different signals. Look for reward hacking and mismatches between score and usefulness.
Follow the systems cost
Map rollout generation, scoring, learning, and weight refresh. Each stage needs capacity and has its own queues.