AI is basically just a fancy parametric function. For example, a simple linear function is of the shape f(x) =k * x + n. Notice that in this function, only x is a variable that you put in. k and n are some fixed numbers, called parameters. The only difference is that most modern AI models have billions of such parameters.

Now the way AI learns is it just takes some random initial parameters and gives an output for a certain input. For example, you give it an unfinished sentence, and it guesses the next word. Because the parameters were random, the output is probably garbage. BUT, and this is very important, the AI can calculate exactly how wrong it was. And from this, it can calculate how it should change its own parameters so that, if it were asked the same question again, it would give the correct answer.

The caveat here is that a set of parameters that works well for one specific problem (eg. completing one sentence) doesn't necessarily work well for another one. So AI usually sees hundreds of thousands, if not millions of different problems, and tunes it's own parameters slightly each time. The idea is that the final set of parameters will likely work pretty well for the vast majority of problems. Although there's really no guarantee (in a math sense) that it will. But that's another story.

what 5 year old understands what you just said?