Nobody wrote the rules — the AI wrote them itself

Here's a mind-bending fact about modern AI: no engineer sits down and codes the rules for recognizing a cat or finishing a sentence. Instead, the model is trained. This is the difference between programming (a human writes explicit step-by-step instructions) and machine learning (the model discovers the patterns on its own from examples).

The engine behind this is gradient descent. The model starts with millions of random internal numbers called parameters (or 'weights'). It makes a guess, compares that guess to the correct answer, and measures how wrong it was — a number called the loss. Then it nudges every parameter a tiny bit in the direction that would have made the loss smaller. Repeat this billions of times over mountains of data, and the random numbers slowly organize themselves into something that 'knows' language or images.

Think of it like tuning a giant radio with a million knobs while blindfolded: after each turn you're told 'warmer' or 'colder,' and you keep adjusting until the signal comes in clear. The AI never receives hand-written rules — only feedback about how wrong it currently is.

This is why AI can do things its creators can't fully explain: the knowledge lives in the learned pattern of weights, not in any line of code a human typed.