How Language Models Work

Lesson 08

Tools

It never gets smarter. It just gets handed better text.

Every time you hear “the AI searched the web,” this is what actually happened. The model writes out a request as ordinary text. Normal software spots that text, does the real work, and pastes the result back in front of the model. That is all a tool is.

An octopus folded from deep aubergine paper, eight arms curling outward.
On its own it cannot add up and cannot look things up. The arms are other software: it asks in plain text, and the answer is handed back.

1Pick a job it can’t do on its own

Do thisPress Next step four times. The same four beats every time: you ask → the model writes a request → other software runs it → the model reads the result and answers.

2Step through the loop

3Everything the model can see right now

Part two

Two ways to get a fact into an AI: bake it in, or hand it over

Training (lesson 12) melts facts into the model’s settings. Powerful, slow, expensive, and frozen solid the moment it finishes. Retrieval is the document lookup you just stepped through: the app finds the right paragraph and pastes it in front of the model. It changes nothing about the model and everything about what the model can see. People call that second one RAG. The difference shows up the second a fact changes, so change one.

4Change the rule and watch which one keeps up

5Baked in by training

you: What's our refund window?

6Handed over by retrieval (RAG)

pasted in front of it just now

you: What's our refund window?

TrainingRetrieval (RAG)
To change itRetrain the model: big computers, weeks, a lot of moneySave the new document. Done in seconds
How fresh it isFrozen on the day training stopped (lesson 07)As fresh as the documents you keep
Can it show you the source?No. The fact is smeared across a trillion settingsYes. The paragraph is sitting right there in front of it
Best used forSkills: language, reasoning, general knowledgeYour facts: rules, prices, documents, anything that changes
How it goes wrongConfidently out of date, or invented (lesson 05)It grabs the wrong document, and the model reads it anyway