Learning by trying
Most accounts of reinforcement learning open with a word nobody needs. The plain version is short. A system is put in a situation, it does something, and afterwards it receives a signal about how that went. Then it does the whole thing again, adjusting a little each round towards whatever produced the better signal.
That is the entire idea. No instruction manual is handed over at the start and no worked example is supplied to copy. The system is never told what the right answer looks like. It is told only, and only afterwards, whether what it did turned out well, and it has to infer the rest by repetition.
The bicycle version
Anyone who has learned to ride a bicycle has already done this. Nobody acquires balance by reading about balance. You get on, you wobble, you put a foot down, and something quietly adjusts before the next attempt. The feedback is instant, unambiguous and faintly undignified, which is exactly why it works so fast.
Speed of feedback is doing most of the work there. Push off, wobble, correct: the gap between the act and the verdict is under a second, so there is never any doubt about which act the verdict belongs to.
When the reward arrives late
A great deal of what anyone actually cares about does not behave that way. Bulbs go into the ground in October and the answer comes in March, by which time a hundred smaller choices about soil, depth and position have also been made. One result at the end has to be shared out somehow among all of them.
That is what a long horizon means, and it is the genuinely difficult part. When a reward turns up a thousand steps after the step that earned it, something has to work out which of those thousand steps deserves the credit. Get that division wrong and what gets learned is confident, tidy and wrong.
Why it is worth understanding
Nicole Junkermann reads about this for the reason she reads about most things, which is that the shape of the problem is familiar. Delayed feedback, noisy signals and a strong pull towards whatever pays off this week are not machine problems. They are the ordinary weather of any work measured in years.
What is worth resisting is the urge to turn any of it into a forecast. The honest description stays narrow. It is a way of learning that does well where feedback is quick and clean, and that becomes markedly harder the further apart the act and the result are allowed to drift. That is interesting enough on its own.
Nicole Junkermann keeps the plainer definitions in the glossary, and the question of feedback that arrives slowly runs through the human skills pages and the shorter tech notes. The companion note on India's engineers and the wider picture follows the next day, and the same argument about time appears in why patient capital matters.