Every question you ask a cloud AI is stored, logged, and often used to train the next model. Your drafts, your doubts, your business ideas, the medical worry you'd never say to a person — all of it flows to someone else's servers, tied to your account, kept as long as they like. You're not chatting with a tool. You're feeding a corporation the raw material of your mind, and paying for the privilege.
Here's the part almost nobody tells you: you don't have to. The same kind of model that powers a cloud assistant can run entirely on your own laptop, offline, seeing nothing, sending nothing. No account, no logging, no training on you. Just a mind in a box that answers to you alone.
It sounds like a weekend for engineers. It isn't. In one evening — a couple of hours, most of it downloading — you can have a private assistant running on the machine you already own. Here's how, in plain words.
Why local at all
Two reasons, and they compound.
Privacy that's actually real. With a cloud model, "we respect your privacy" is a promise you can't verify. With a local model, privacy isn't a promise — it's physics. The data never leaves your machine because there's nowhere for it to go. Pull the network cable and it still works. That's not a policy you have to trust; it's a wall you can see.
Ownership. A local model can't be taken away, price-hiked, censored, rate-limited, or discontinued out from under you. You downloaded it; it's yours; it runs the same next year as tonight. In systems terms, you've removed a single point of failure — the company between you and your own tool. Not your keys, not your coins; here, not your machine, not your mind.
The trade is honest: a model on a laptop is smaller than the largest cloud giants, so it's a bit less sharp on the hardest tasks. But for writing, summarizing, brainstorming, coding help, translation, answering questions — the daily work — a good local model is more than enough, and it's yours.
What you need
Less than you think. A reasonably modern laptop or desktop from the last few years. Eight gigabytes of RAM runs a small model; sixteen or more runs a comfortably capable one. A dedicated graphics card makes it faster but is not required — modern machines, including recent Apple ones, run these well on the main chip. No coding skill. No command-line wizardry, unless you want it.
The models themselves are open-weight — free to download and run. The two families to know:
- Llama, from Meta, open-weight and widely supported — a strong general all-rounder.
- Mistral, from a European lab, known for being small, fast, and punchy for its size.
Both come in sizes. Smaller ones run on modest machines and answer fast; larger ones are smarter but need more memory. Start small. You can always size up once it's working.
The evening, step by step
Step 1 — Install a runner. You don't wrangle the model directly; a "runner" application does the heavy lifting and gives you a clean interface. Several free, well-regarded tools exist that install like any normal app — one of them lets you download and chat with local models through a simple window, no terminal at all. Pick one, install it, done. Fifteen minutes.
Step 2 — Download a model. Inside the runner, browse the model library and pick a small-to-mid Llama or Mistral to start. Click download. This is the slow part — the file is several gigabytes — so start it, then go make tea. When it finishes, the model lives on your disk. Permanently. Yours.
Step 3 — Talk to it. Open a chat window in the runner and type, exactly like you would in the cloud. But now: no login, no history uploaded, no company reading over your shoulder. Ask it to draft an email, summarize a document you paste in, explain a concept, help with code. Watch it answer — from your own machine, seeing nothing beyond your screen.
Step 4 — Prove it's private. Do the test that makes it real. Turn off your Wi-Fi. Now ask it something. It still answers. That moment — a capable AI working with the network off — is when the abstraction becomes concrete. Nothing is being sent anywhere, because nothing can be. You're holding the proof in your hands.
> Our record. A cloud model is Heka — the creative word — rented from a temple that keeps a copy of every word you speak into it. Powerful, yes, but never yours, and always watched. A local model brings the creative word home: the same power, no watcher, no copy taken. In the weighing of Maat this is a rare clean trade — you gain a tool and give up nothing of yourself. The lens closes. The word stays in your house.
Where this fits your bigger move
Don't oversell it to yourself and don't undersell it either. A local model won't replace every use of cloud AI overnight, and for the very heaviest tasks you may still reach for a large hosted model now and then. That's fine. The point isn't purity — it's shifting your default. Make the private, local assistant the thing you open first, and the cloud the rare exception instead of the everyday habit. Most of your daily AI work — the drafting, the thinking-out-loud, the private questions — moves off the surveillance grid and into your own walls.
That's another brick in the house of Maat: the tools you think with, owned by you, watching no one.
Do this today
You can't download and install everything in the next five minutes, but you can take the first real step now: pick your runner. Look up one of the well-known local-AI runner apps for your operating system, and start the download of the application itself — just the app, the small part.
By the time you sit down tonight, it's installed and waiting. Then you grab a model, brew that tea while it downloads, and by bedtime you have a mind in a box that answers to you and no one else.
One evening. Your assistant. Your machine. No lens. Build it.