On this page
Definition
Reinforcement learning is a machine learning paradigm in which an agent learns to make decisions by receiving reward signals for its actions in an environment.
Simple explanation
It is training by trial and error. Do something good → get a reward → do it more often. Do something bad → get penalised → do less of it.
Why it matters
RL from human feedback (RLHF) is what turned raw LLMs into helpful chat assistants. Without it, models would still be technically strong but frustrating to use.
How it works
- 1EnvironmentThe setting in which the agent acts.
- 2ActionsThe choices available to the agent.
- 3RewardA signal indicating success or failure.
- 4Policy updateThe agent adjusts to pick better actions over time.
Real examples
Products named for illustration only. Inclusion is not an endorsement.
- RLHFAligning LLMs with human preferences.
- AlphaGoDeepMind's Go-playing system used RL at its core.
Advantages
- Handles decision-making problems.
- Improves over time with feedback.
Limitations
- Reward design is hard.
- Sample-inefficient without careful setup.
Common misunderstandings
- ClaimRL is how LLMs learn language.RealityLanguage is learned in pretraining; RL is used to align behaviour afterwards.
Frequently asked questions
What is RLHF?
Reinforcement Learning from Human Feedback — the alignment step used to make chatbots helpful and safe.
The Tool Money Lab perspective
RL is the most conceptually elegant part of modern AI training and also the part most likely to bite teams who apply it without deep expertise. Reach for it deliberately.
Conclusion
Reinforcement learning is how AI learns from consequences. It is a small part of most business AI stacks — but a critical one for alignment.