What my kid asks Wikipedia

There's a Mac Mini on a shelf in my office with the entirety of English Wikipedia sitting on it. Not a link to it. The thing itself, as a ZIM file, served by Kiwix on the LAN and indexed so a model can pull from it. Next to that is Qwen 3.6 35B A3B, a mixture-of-experts model that keeps about three billion parameters active per token, which is why it's quick on a machine with no real GPU to speak of. In front of both is OpenWebUI, with a system prompt I wrote for exactly one user.

That user is eight.

Why offline, specifically

I'm on frontier models every day for work and I'm not a purist about this. But when it came to my daughter having her first real conversations with a language model, the question I kept landing on wasn't "which model is smartest." It was: where does what she says go?

With the setup on the shelf, the answer is nowhere. It goes into RAM, comes back out as an answer, and that's the end of it. Nothing she's curious about at eight lands in a training set or a log on somebody's server. Not because those companies are villains - because she shouldn't have to think about that at eight, and I shouldn't have to trust a policy page on her behalf.

The guardrails are a document

The system prompt is about a page. It's not clever. It says: you're talking to a kid, you're reading from an encyclopedia, cite the article you're using, say "I don't know" when the article doesn't say, and don't make things up to fill the gap. There's a section on the questions that aren't a machine's to answer - the answer there is "that's a great one for your dad," every time, no exceptions, and she has learned to roll her eyes at it.

I'll be honest that a page of prose is not a safety system. It's a set of manners. But the model's only source is an encyclopedia and its only reader is in the next room, and for that threat model, manners have held up fine.

What she actually asks

This is the part I didn't see coming. I figured dinosaurs. There were dinosaurs. But mostly it's been why questions with no bottom. Why is the sky that color. Why do we have to sleep. Why are there different languages. Why does the moon follow the car.

The encyclopedia is good at the first layer and pretty honest about not having the second, and the model, told to be honest, passes that along. "The article doesn't say why" is a sentence she's heard a lot now, and I think it's the most valuable thing the whole setup has taught her. There's an edge to what's written down. Past it you have to go look, or ask a person, or sit with not knowing.

She's also figured out she can ask it to explain something again but shorter, or again but with an example, and that it never gets tired of her. That patience is the feature. Not the knowledge. I'm oversimplifying, but for a kid I think that might be the whole feature.

The part I'm unsure about

I don't know if I'm doing this right. It's entirely possible the correct amount of LLM for an eight-year-old is zero. The honest answer is she'd be asking these questions to a phone app in four or so years anyway, and I'd rather she meet the thing on a shelf in our house first, with me in the next room, before she meets it on a billboard.

Time will tell. I skew optimistic with this stuff and I'm aware of it.

Anyone else building something small and offline for their kids? I want to hear what they ask it 🙏