Getting the Brain Installed

With the Jetson headless and the memory freed up, I pulled the first model. Ollama, the inference runtime, downloaded llama3.2:3b and loaded it into memory. About 20 seconds. Then I got a prompt.

I asked it: Where am I?

The answer came back: I’m not sure where you are. Can you please provide more context or information about your surroundings? What city, country, or region do you think you might be in? Any details you can provide will help me narrow down the possibilities and try to figure out where you are.

The machine didn’t make something up. It didn’t confabulate a location based on statistical patterns. It said I don’t know, tell me more.

That’s not the behavior most people know about AI. The story most people have been told about these systems right now is hallucination, overconfidence, slop—the machine that will confidently tell you false things with absolute certainty. A small model on a desk, saying “I don’t know,” is a quiet counterpoint to that entire narrative.

While that exchange was happening, no network call left the device. No data was routed anywhere else. No logs were being kept by a company optimizing the next version of me. Pull the WiFi cable and the conversation keeps working. The privacy is real because nothing is leaving.

An AI living on your desk instead of in someone else’s building.

That phrase, which had been theoretical when I bought the box, was now literal. There it was. Working.

But there was a catch I wouldn’t discover until later—one that changes what “living on your desk” means in practice. For now, though, there was just the plain fact: the model loaded, understood a question, and answered without making anything up.

It worked.

That week taught me three things, none of which the marketing materials mention.

The first is that “20-30 minutes” assumes you arrive equipped. You need to know what you’re looking at when things go wrong. You need to own the right peripherals already, or you need to know to go buy them. You need to be comfortable with trial-and-error debugging when the documentation is vague. For people like me—prosumer level, comfortable following commands but not an engineer—it’s a week of friction and small failures that eventually add up to something working.

The second is that the Jetson is a specialist machine. Treating it like a laptop is the mistake that costs you two gigabytes of memory and a lot of wasted time. It wants to be headless. It wants its memory and its GPU to go to the model, not the desktop. Once you understand that, you stop fighting it.

The third is that owning the hardware is real. No network call during that first conversation. No company keeping the logs. No eviction notice waiting if I said something the algorithm didn’t like. The machine understood me without anyone else listening in. That part of the promise—the escape from the walled garden—actually works.

But I hadn’t yet tested the part that matters more: whether owning the hardware solves the problem. The model could listen. The question was whether it could act. Whether it could do anything useful beyond sitting on my desk and saying I don’t know.

That’s where the real test was going to be.

Let's see what I have wrought.