Setting Up Your Own Model

Welcome to the desk. Part One asked only that you read. From here on you build, and we start with the piece everyone dreads and almost everyone finds easiest: standing up your own AI, on your own machine, that belongs only to you.

This is the chapter where the argument stops being an argument and becomes a thing that hums on a table in your office. By the end of it you will have a working AI you can talk to, and you will have proven it private with your own two hands, the internet switched off and the answers still coming. People expect this part to be the hardest. It is the one that surprises them.

The one idea you need about hardware

There is exactly one technical idea in this chapter, and once you have it the rest is downloads.

A model has a "brain," and that brain has to fit into your computer's fast memory to run. A bigger brain is a smarter model, and a bigger brain needs more memory to hold it. That is the whole theory. Memory is the one number that matters. You will see it written as RAM, or on a Mac as "unified memory." When you read that a machine has 16 gigabytes or 64 gigabytes, that number is the size of the desk the brain has to sit on. A small brain fits a small desk and is quick and capable enough for plenty of work. A bigger, smarter brain needs a bigger desk, and that desk is most of what you pay for.

Hold "memory holds the brain" in your head and every choice below gets easy.

One more thing before you spend a dollar. A Mac is the simplest path, for a real reason. On a Mac the memory is shared across the whole machine, so a Mac with 64 gigabytes can run a large brain that a normal Windows PC cannot match at the same price, because on a PC the limit is the smaller memory built into the graphics card. A good Windows machine with a strong graphics card works too. But for the least fuss and the most brain per dollar, buy a Mac. That is what these examples use.

Three machines you can actually buy

Start with what you already own. A fairly recent computer with around 16 gigabytes of memory can run a small model today, for free, just to find out whether you like working this way. When you are ready to buy, here are three honest setups, one per firm size, with what to get, what it costs, and which brain to load.

A note on the prices. As I write this, a shortage of computer memory has pushed the higher-memory machines up in price and made them harder to find. Treat these numbers as a snapshot, not gospel, and check the current price at apple.com/mac-mini and apple.com/mac-studio before you buy.

The solo lawyer. A Mac Mini with 24 gigabytes of memory, around $999. Load a small model in the 8-to-14-billion range, a Llama, Qwen, or Gemma of that size. It handles summaries, first drafts, and questions about a document with ease. For one person it is a real workhorse, and it fits on the corner of a desk.

The small firm. A Mac Mini M4 Pro with 48 to 64 gigabytes, around $1,900 to $2,200. Load a mid-size model in the 27-to-32-billion range, and if you went to 64 gigabytes you can run a 70-billion model, a real step up in smarts. This is the daily machine for a few people, strong enough to lean on for real work.

The medium firm. A Mac Studio with 64 to 96 gigabytes, around $2,000 to $3,500 depending on memory. Load a 70-billion model with room to spare, and serve it to the whole office over your network so everyone talks to the same private brain from their own desk. One strong model, shared by all, with no per-seat fee, ever.

Not sure where you fall? Start at the solo setup or the machine you already own, get convinced, then move up. You do not have to buy the biggest thing first.

Setting it up, in three steps

Here is the part that sounds like it needs a server room and does not. No command line. Three steps, and they look like installing and using any ordinary program.

We use one app for everything in this book. It is called GPT4All. It runs your private model, and it can also read your own documents, which is the second tool you will build. One app the whole way through, nothing to juggle.

Step one: install the app. Download GPT4All from nomic.ai/gpt4all and install it like anything else. It is a normal app with windows and buttons, built so a person who has never touched this can use it. When it asks on first launch whether to share anonymous analytics or your chats, say no. You are here for privacy. Start as you mean to go on.

Step two: pick a brain. Inside GPT4All is a list of models to download. Pick one in the size matched to your machine above, a Llama, Qwen, or Gemma in that range, and click download. A one-time download, like getting an app onto your phone. If the model feels slow once it runs, there is a setting to let GPT4All use your machine's graphics chip for more speed.

Step three: start talking. When the download finishes, you type into a chat box and the model answers, right there on your machine. That is the whole thing. You now have a private AI.

The moment the fear ends

Do this the first time you have a model running, because it is worth more than any reassurance I can write.

Turn off your internet. Switch off the wifi, or pull the network cable, so the machine is truly cut off from the world. Now ask your model a question. It answers anyway.

Sit with that for a second. It did its work with no connection to anyone, no company on the other end, no warehouse of strangers' computers. There was no one in the room but you and a machine you own. That is not privacy you take on trust from a policy. It is privacy you can see, switch off, and watch hold. This is the moment I promised you in the first chapter, and for most people it is the moment the last of the fear lets go.

This is your vault

That quiet machine is now the vault from Chapter 4. Levels two and three, every task that touches a client, live here and go nowhere else. From here on, every tool you build runs on it, which means everything you build is private from its first day, without your ever having to think about it again.

There is one small surprise waiting in the next chapter. Your very first build will not use this model at all. That is on purpose, and it teaches the most important habit in the whole book: knowing when the smartest move is not to reach for the AI. Turn the page and let us make the first thing.