Nobody taught a spam filter what spam is by writing a dictionary of bad words. It read millions of emails that people had already marked as spam or not spam, and it noticed what the spam ones had in common. That is AI learning in one picture.
Training is the process of showing a system examples until the patterns stick. The result is a model: a giant set of learned numbers that turns new input into a good guess.
Patterns, stored as numbers
During training, the system makes a prediction, checks it against the right answer, and nudges millions of internal numbers when it is wrong. Do that billions of times and the numbers start to encode real structure: what a face looks like, how a sentence usually continues, which emails smell like spam. Those numbers are called weights, and they are what the system actually learned.
One strange thing about this: nobody can point to where a fact lives inside those numbers. The knowledge is spread across billions of weights instead of sitting in a file somewhere. That is why you cannot just delete one fact from a model, and why changing what a model knows means training it again.
Learning stops before you show up
The models behind modern chat tools finished training before you ever typed a word. When you chat, the model is not learning you. It is applying what it already learned to what you just said. Anything it seems to remember about you comes from the product around it, which later posts in this series open up.
This surprises almost everyone. When you correct a chatbot, it feels like you are teaching it, and inside that conversation it does adjust. But close the tab and the correction is gone. The model itself only changes when the lab trains a new version, which happens months apart. What products can change day to day is everything around the model: the instructions, the memory, and the tools.
Why this beats written rules
Rules break on cases nobody anticipated. Examples cover the messy middle of real life, where spam does not use the obvious words and faces show up at odd angles. That flexibility is why the same recipe now writes text, recognizes speech, and generates images.
The tradeoff is that learned rules are hard to inspect. When a spam filter gets something wrong, an engineer cannot open a file and read the bad line of code, because there is no line. The rule lives spread across millions of numbers. This is what people mean when they call AI a black box, and it is why AI gets tested by watching how it behaves instead of by reading its code.