Training AI on human-written examples hits a ceiling — the model can only be as good as its teachers. Reinforcement Learning (RL) breaks through this limit with a simple but powerful idea: instead of showing the model the "right" answer, let it try many different answers and learn from the ones that work.

For this loop to work, you need one crucial ingredient: a way to check if the answer is right. A math problem has a correct answer you can verify. A piece of code either passes its tests or it doesn't. But "write a beautiful poem"? That's subjective — there's no automatic checker.

This is a major reason AI coding has improved so dramatically. Code is the perfect RL playground: you write a function, run the test suite, and know instantly whether it works. No human reviewer needed — the computer checks the answer for you, millions of times.

The same principle applies to math (check the answer), logic puzzles (verify the solution), and even games (did you win?). Anywhere a computer can judge the result, RL can generate millions of training examples — far more than any human could write.

But look at the red zone on the spectrum above. "Is this poem beautiful?" "Is this explanation clear to a 10-year-old?" "Is this advice helpful without being harmful?" There's no test suite for these questions. Yet humans intuitively know a good answer when they see one.

This is where RLHF — Reinforcement Learning from Human Feedback — comes in. Instead of an automated checker, you use people as the reward signal.

The key insight: it's much easier to judge which response is better than to write a perfect response yourself. You don't need to be a chef to know which of two dishes tastes better. Thousands of these comparisons train a "reward model" that learns to predict what humans prefer — and that model becomes the automated judge for RL. This is what teaches models to say "I'm not sure" instead of confidently making things up, and to decline dangerous requests while staying helpful for everything else.