Before the Model, the Use Case
Building My Personal AI, between Step 1 and Step 2
The workbench is ready. Three Macs, one standard, one private network.
The next step was supposed to be simple: install a model.
Then I asked myself a question I couldn't answer.
Which model?
And right behind it, a better question:
For what?
---
## How did the PC take off?
I went back to the personal computer.
It didn't take off because it was powerful. Early PCs were toys for hobbyists.
It took off because of one program. VisiCalc, in 1979. The first spreadsheet.
Before it, a business planner who changed one number had to recalculate the whole ledger by hand. Hours of work. After it, you changed one cell and everything updated.
Suddenly you could ask "what if?"
People bought an Apple II just to run it.
The pattern, as I see it:
- A task that is frequent
- A task that is painful
- Done by the person who owns the data
- Made instant, with a "what if" they couldn't ask before
So I wondered: what is the VisiCalc of personal AI?
---
## What people actually struggle with
I looked at what the research says about everyday problems, and how people use AI today.
A few things stood out:
- Money is the number one source of stress. About 40% of Americans put it in their top two.
- Life admin eats real time. Bills, insurance, documents, appointments, subscriptions. Many people lose hours every week.
- About 64% of US adults use AI, and a quarter use it daily. Parents use it far more than non-parents.
- But for paying bills and navigating health care, AI use is low. Only about a quarter of people who do those tasks use AI for them.
And the reason surprised me. It isn't that AI can't do it.
It's trust. People don't want their bank statements and medical bills sitting on someone else's server.
That's when it clicked.
The most painful problems are the most private ones. And private problems need a private AI.
That's the whole reason this series exists.
---
## My own list
I made a list of what I'd actually use, then scored each idea: how often, how painful, how private, and how much a local AI helps.
| Use case | How often | How painful | How private | Local AI helps |
|---|---|---|---|---|
| Life admin | Daily | High | High | Strongly |
| Money | Weekly | Highest | Highest | Strongly |
| Trading research | Daily | Medium | Medium | Yes |
| Learning | Daily | Medium | Low to medium | Somewhat |
| Health and care | Weekly | High | Highest | Strongly |
| Family logistics | Daily | Medium | Medium | Somewhat |
| Travel | Few times a year | High, briefly | Low | Not much |
| News and music | Daily | Low | Low | Not much |
For me, three things rose to the top:
Trading research. For learning and research only. The decisions stay mine. Life admin. Bills, email, documents, subscriptions. The things that slip. A sales and business-development assistant that a small business can run itself. An established sales method, with my own touch: a bit of the Challenger approach, a bit of empathy, a bit of AboutU.
And one bigger realization.
Right now, the user is just me. But a personal AI soon becomes a household AI. And a small business is just a household with employees.
So the road ahead has three phases:
Phase 1: Personal. It works for me. Phase 2: Household. It works for several people, safely. Who can log in, who can see what, who did what, and the rules around it. Twenty-five years in networking taught me those questions matter more than the features. Phase 3: Product. A small business can run it on its own. Now, the model
With the use cases clear, choosing the model got easier. Each one needs the same few things:
Read documents (statements, emails, filings) Use tools (calendars, files, data) Reason (what changed, what matters, what if) Fit in memory on the Mac Studio, with room to spare Be fast, because a slow assistant doesn't become a habit Be free to build on, including for a product someday
The Mac Studio has 48 GB of memory. macOS takes some of that, so in practice it's closer to 38–40 GB.
Here's what I compared:
| Model | Download | Strength | Trade-off |
|---|---|---|---|
| Qwen 3.6 35B-A3B | ~24 GB | Fast (only ~3B active per word), reads images, uses tools, open license | My main pick |
| Qwen 3.6 27B | ~17 GB | Every parameter works on every word, slightly better quality | Slower |
| Gemma 4 31B | ~20 GB | Solid all-rounder from Google | Slower (every parameter works on every word) |
| Gemma 4 E4B | ~7 GB | Small and very fast, fits the iMac | My second model |
| Llama 3.1 70B | ~40 GB | Very popular | Too big for 48 GB, and older |
My pick: Qwen 3.6, the 35B-A3B version.
The name hides the interesting part. It has 35 billion parameters in total, but only about 3 billion work on each word. It's a team of specialists where only the right few speak up. That's why it's fast.
It reads text and images, uses tools, and can think step by step. And it's released under an open license that allows commercial use.
My second model: Gemma 4, the small E4B version. From Google. Small, very fast, and it fits on the iMac if the Studio is ever off.
Two models from two different companies. One for depth, one for speed. And I'm not tied to a single vendor.
The part I keep coming back to
The model is the most replaceable part of the whole system.
New ones arrive every few weeks. This choice will be outdated within a year.
What stays is everything around it. The workbench. The memory. The values. The way I decide what it may do on its own.
I wrote earlier that the Self doesn't change, even when the thoughts do.
It seems the same is true here.
The model is a thought. It will pass.
The how
The full reasoning, the comparison, and how to switch models later are in the build notes:
Why this model →
Next: Step 2. Bringing the model home.