Showing a model examples and nudging its numbers each time it gets one wrong.
Training is how a model gets good. It tries an example, measures how far its answer is from the right one, and shifts its internal numbers (parameters) a tiny bit to shrink that gap. Repeat this an enormous number of times and answers that started as nonsense become sensible.
The method that works backward from the error to figure out which numbers to change, and by how much, is called backpropagation. Training a big model can take thousands of graphics chips or more, running for weeks or months. Training happens when the model is made; using the finished model is called inference.
Think of it like this
It's like doing a set of math problems, marking them, and fixing your method on the ones you missed. The difference: an AI never has an "aha!" moment. It just moves its numbers a little in the right direction, over and over.
For example
Hide the next word in a sentence, let the model guess, adjust if it's wrong, repeat endlessly: that's how a large language model is trained.
Common mix-ups
Mix-upDuring training, AI has an "aha!" moment.
ActuallyDuring training, researchers watch a number called the "loss", which measures how wrong the model is, on a graph. The line usually wobbles up and down while gradually going down — less a moment of insight than the trace of countless tiny corrections.
Mix-upAI trains on the spot every time I use it.
ActuallyThe numbers inside the model don't change the moment you ask something. But some services store chats and later use them to train their next model, so check whether the settings let you turn that off.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- If it only fixes a little at a time, why does it need so many rounds?
- Activity
- Imitate training by throwing a paper ball into a box. After each throw, measure how far you missed, and next time change your aim or force by only half of the miss. Record ten throws in a table and draw a graph of the misses shrinking. Then connect it to the 1986 backpropagation record (E0010): a method for working out which numbers to change, and by how much.
Records about this
Source-checked moments from AI history.
Showed that hidden layers learn useful internal representations for a task on their own, making the training of multilayer neural networks practical