I got Qwen3.8-27B running locally.
It felt like it was going to be a moment.
There was an intelligence in the room with me. Running silently on my Mac mini, answering my questions without reaching out to some distant data center.
It was right there. Local, untethered, open, and, if I wanted it, mine.
Here was a 2024 Mac mini (M4 Pro with 24 gigabytes of memory) quietly generating intelligence on my desk. Twenty-seven billion parameters squeezed into a box I can hold in one hand. An incredible moment by almost any reasonable standard.
My first thought was: This is incredible.
My second was: I need more RAM.
Because then I started talking to it.
The intelligence was there, certainly. It could say hello. It understood my questions and responded to them. But it wasn’t smart in the frontier-model sense. It wasn’t the level of intelligence I had (absurdly) expected to materialize inside a small aluminum box on my desk.
What an interesting kind of letdown…
A miracle happened. My expectations had simply outrun it.
Twenty-seven billion parameters isn’t a small intelligence. It only feels small because the machines that recalibrated my expectations are unimaginably large.
I’ve grown accustomed to talking with systems backed by vast data centers, armies of GPUs, and infrastructure I couldn’t begin to reproduce at home.
Then I invited one smaller intelligence into the room and wondered why it wasn’t as clever as all of them.
I blame myself. I expected a baby with a doctorate to appear.
Here was this new little mind taking its first steps, and instead of marveling at the fact that it could walk, I was disappointed it hadn’t finished its dissertation.
Still, there was something different about it.
The frontier models may be vastly more capable, but they don’t live here. I visit them. I send a message into the cloud, something enormous wakes up somewhere far away, and an answer returns. However conversational it may feel, the intelligence remains elsewhere.
This one is in the room.
I could turn off the internet and it wouldn’t disappear. I could choose the model, shape its environment, and decide when it ran. Its continued existence isn’t tied to a subscription or dependent on a company deciding I can still have access to it.
It’s not just intelligence I can use.
It’s intelligence I can keep.
That realization lasted for perhaps thirty seconds before my mind began assembling the components for a machine that might make it smarter.
Graphics cards. Cooling systems. Cables. Power supplies. A glowing Habitrail of whirring GPUs designed to give this local intelligence some room to breathe.
All of it is now painfully expensive.
I should’ve bought that RTX 5080 when I had the chance…
At the time, I thought buying one would be impulsive, so I exercised restraint. I now deeply regret that responsible decision. I could have reframed the entire purchase as an investment in sovereign intelligence.
Surely, if I give this local brain a larger playpen it could grow into something beautiful. It’ll just take more memory, more electricity, and several thousand dollars’ worth of new hardware.
What I have right now feels closer to a Tamagotchi.
It’s charming. It responds when I speak to it. I can show it to visitors and say, “Look! The rudimentary future of our digital companions lives inside this box.” Then it will confidently misunderstand something fairly basic, and we will all nod politely.
But perhaps that’s unfair.
A Tamagotchi was a simulation of something alive. This is something stranger: a real, functioning intelligence compressed down to fit inside a consumer computer, sitting silently beside me and waiting to speak.
It may not be frontier-level intelligence, but it inhabits the same physical world I do. Hell, we share the same address!
I can feel the warmth of the metal box sustaining it. I can disconnect it from everything else and it remains. I can tinker with it, constrain it, expand it, or leave it running quietly on my desk.
That feels like the beginning of something.
I can see a personal AI paradise on the horizon: private, capable intelligences that live with us, learn how we work, and answer to us rather than to the companies hosting them.
Minds we can shape without asking permission. Companions whose memories and continued existence don’t depend on a subscription staying active.
Not merely intelligence, but intimacy.
Not merely access, but ownership.
I had hoped my current hardware might take me all the way there.
It can’t.
But it did take me one step closer.
So no, I’m not disappointed.
I’m impatient.
Author’s Note
This piece began as a Saturday-morning brain dump after I got Qwen3.8-27B running on my Mac mini. The experience, excitement, disappointment, hardware envy, and questionable RTX 5080 restraint were all mine.
I brought those raw notes to Codex, and we batted the idea around together. It helped me recognize that the piece wasn’t really about disappointment. It was about impatience. And that the larger tension was between accessing intelligence somewhere else and keeping it in the room with me. From that conversation, Codex assembled a first draft. I rewrote it in my own voice, added and changed several passages, and then brought it back for a final copyedit.
The artwork followed a similar process. We developed the image of a tiny metal box casting an impossibly large shadow: the physical intelligence may be small today, but the potential it projects is nearly infinite. We explored several styles and compositions before arriving at the pulp-comic direction. Codex helped write and refine the image prompts, and I continued experimenting with them in Krea 2 Turbo.
There’s an obvious irony in using one of the frontier systems described in the essay to help articulate my feelings about a much smaller local one. But that contrast is also the point. I had the experience. I brought the taste, judgment, and final decisions. The AI helped me see more clearly what I was trying to say.
Consider this piece another small example of the future it describes.