Starting From a Clean Machine
Building My Personal AI, Step 1
My iMac was wiped clean.
No apps. No files. No old settings. Just a login screen and a lot of empty space.
At first it felt like a setback. Then it felt like an invitation.
There's a phrase in Zen, shoshin. Beginner's mind. Approaching something as if for the first time, without the clutter of what you think you already know.
A clean machine is beginner's mind in hardware form.
Why build on my own machine at all?
I could do all of this in the cloud. Rent a server, sign up for a few services, and be running in an afternoon.
But I keep coming back to the lesson from the personal computer: what made it matter wasn't speed. It was ownership.
If I want an AI that knows my life, my memory, my values, my people, I want that living somewhere I can see. Somewhere I can switch off. Somewhere that's mine.
So the first step isn't the AI.
It's the room the AI will live in.
The driving range
In golf, you don't learn on the course. You learn on the driving range.
Same clubs, same setup, every time. You hit hundreds of balls. Nothing counts. You're just building something you can trust.
I named my setup the same way. The Driving Range is a short standard every one of my Macs follows:
One place for all work
One version of Python, managed the same way
No tools that fight with each other
A quick check that tells me, line by line, whether a machine is ready
It sounds dull. It's the opposite. When every machine is set up the same way, I stop thinking about setup and start thinking about the work.
Three machines, one mind
I have three Macs, and each has a role.
| Machine | Role | Think of it as |
|---|---|---|
| Mac Studio | Runs the AI models | The core |
| iMac | Desk work and learning | The study |
| MacBook | Coding and travel | The bag I carry |
The Mac Studio does the heavy thinking. The others ask it questions over a private network between my own devices.
It's a small version of the architecture I wrote about in One Self, Many Agents. Many places to work from. One core.
What actually happened
I ran the setup on the clean iMac. It installed everything. Then I ran the checklist.
Four failures.
My first thought was that something had broken. It hadn't. Three of the four had one cause: I ran the check in the same terminal window I'd opened before the setup. The window was still living in the old world. It hadn't heard about the changes.
One command, exec zsh, which simply restarts the shell, and three failures disappeared.
The fourth was a small tool for connecting my Macs over a private network. The iMac doesn't need it yet. So I changed the standard to say so.
Then I made a mistake of my own. I edited one line by hand, and put it under the wrong machine. The checklist caught it on the next run. Not me. The checklist.
Final result on the iMac:
29 checks passed. Zero failures.
What surprised me
How much of "AI work" is really housekeeping. Before a single model runs, you need the right Python, the right folders, the right paths. Get one wrong and nothing works, and the error messages rarely tell you why.
And how calming a checklist is. It doesn't care how confident I feel. It just reports what's true.
There's something in that. The mind says "it's done." The checklist says "let's see."
The how
Everything I ran, in order, is in the build notes. The scripts, the checks, and the mistakes:
If you have one Mac, you can follow along with just that one.
What's next
The room is ready.
Next, I invite someone in: a model, running privately on my own desk.
I'm curious what it will feel like to ask a question and know the answer never left the house.
Next: A Local Model.