Home / AI Arena / How LLMs Talk
How LLMs Talk
Your order never goes to "the model". It goes to an app that staples it to a transcript, wraps it in an order ticket, and hands it to a server that rewrites it into a private script format and runs it through a frozen file of numbers. This series walks every one of those layers in plain language, following a single message "What's the weather in Paris?" from your keyboard all the way down and back. By the end, you can explain thinking mode, tool use, and AI agents to a friend without hand-waving.
Curious what happens inside the model when the engine runs? That story is the LLM Basics series. This one is about everything around it. The two can be read in either order.
-
Part 1
The File That Can't Talk
You can download a state-of-the-art AI today. It is one file, about the size of 40 movies, and on its own it does nothing. What a model actually is, and the single trick it knows.
-
Part 2
The Kitchen: What Runs the Numbers
A frozen file cannot compute anything on its own. Meet the inference engine — the program that does the arithmetic of prediction — and the real reason one model comes in five confusing download sizes.
-
Part 3
The Illusion of Conversation
The model forgets everything the instant it stops writing. How re-sending the entire conversation on every turn creates the feeling of a chat, and what a context window actually is.
-
Part 4
The Script Rewrite: How a Chat Becomes One Long Text
Before the model sees your conversation, a hidden layer rewrites it into a private script format, with a briefing document you never wrote stapled on top. When an AI suddenly acts strange, this layer is often the culprit.
-
Part 5
The Waiter: What LM Studio and Ollama Actually Do
Apps never talk to the model directly. They talk to a server that loads model files, rewrites conversations, runs the engine, and streams the answer back — and nearly everyone speaks the same order-ticket format.
-
Part 6
Thinking Mode and the Fate of the Thoughts
Some models write private notes to themselves before they answer you. What the scratchpad actually is, why you pay for words you never see, and the question almost nobody answers — where do the thoughts go afterwards?
-
Part 7
Tools, Agents, and Other Conversation Tricks
The model cannot check the weather, run code, or touch a file. It can only ask. How tool calling really works, why an agent is a loop rather than a new kind of mind, and what is happening when models talk to each other.
-
Part 8
The Whole Journey, End to End
One message, ten steps, six layers down and back up again. The complete journey of a single question, why the layers matter to your bill and your privacy, and a glossary to keep.