How does the internet fit inside an LLM?

Before we can chat with an LLM, it has to be trained. Imagine giving it an enormous library of writing: articles, books, websites, code and other examples of how people use language.

The model plays one game again and again. Hide the next piece of text, make a guess, compare the guess with the real answer, then adjust slightly. Training repeats this across a vast number of examples.

The finished model does not carry a tidy copy of every webpage. It is closer to a heavily compressed version of what it saw. Think of squeezing a library onto a small pen drive. You cannot open the original books, but the model has kept many relationships between words and ideas.

It learns that Paris often appears near France, that recipes tend to contain ingredients and steps, and that questions are often followed by answers. Those relationships are held in the model as numbers.

If the training library was packed last year, the model does not automatically know what happened this morning. It needs fresh information from a search tool, document or database.

Compression can also lose detail. A model may reproduce the shape of a fact without preserving the exact fact. That is how a fluent answer can still be wrong.

There is no tiny library hidden inside the model. The suitcase picture only helps us understand how a lot of training material can shape a much smaller set of learned numbers.

Training gives the model broad patterns. It does not yet teach the model how to behave like a helpful assistant. That comes next.