Step back and the picture is simple.

A model is two things. The program is the shape, written by humans. The weights are what training produces: billions of numbers in a specific configuration that make the network useful.

To produce those weights, someone downloaded a huge chunk of the internet, cleaned it, and converted every word into , numbers the network can actually process. The model trained by doing one thing, over and over: predict the next token. Through that single task, it learned grammar, facts, logic, and language.

That produced a base model, good at continuing text but not at being helpful. on thousands of human-written conversations shaped it into an assistant. That process also taught it to sound confident, which is why it sometimes produces fluent, wrong answers without any sense that it's wrong.

When training is done, the weights are frozen. Using the model is just fast arithmetic on fixed numbers. The expensive work happened once.

The knowledge inside is real but strange. It's distributed across billions of weights, none of which means anything alone. It can generalize to things it's never seen. It can also be confidently wrong, with no internal flag that knows the difference. And it cannot explain itself, because the knowledge doesn't live in rules. It lives in numbers.

That is what a trained model actually is.

But networks this deep were not always trainable. For years, one obstacle stalled the whole field, and getting past it is what made models like this one possible. That story is where we go next.