The Unallocated Drive

I found a hard drive last week that had never been used.

Not lost. Not corrupted. Just sitting there — unallocated space, a library of movies from years ago that never made it off the external drive and onto anything. A relic of a past version of me who clearly had big plans for a media setup and then got distracted. It happens.

The obvious move was to copy it across to the JARVIS machine and get Jellyfin running. Simple enough. Fire up rsync, let it go, come back when it's done.

Except when you're running an hours-long rsync to a headless machine over SSH through a Tailscale VPN, you can't exactly just "check on it." I was switching between the Pop!_OS GUI on the tower and my MacBook terminal — not because it was necessary, but because I didn't want to walk away. I didn't want to interrupt it. I was babysitting a copy operation.

That's when it hit me. That right there is exactly what JARVIS is supposed to fix.

The Babysitting Problem

What actually makes an assistant useful isn't that it does things for you. It's that it lets you stop watching. You palm off the task, walk away, and come back when it's done. Go live your life. Do something else. The assistant handles the waiting.

A simple Telegram message when the volume mounted would have freed up the rest of my evening. Instead I was hovering. And the irony is I have all the pieces to do exactly this — the Telegram bot is connected, I can write a Python script in twenty minutes. I actually did write the script. I got the message confirming the volume had mounted and felt slightly ridiculous that I'd been watching the terminal for two hours before doing so.

This is a small thing. But it pointed at something bigger about what I'm actually building. JARVIS isn't useful when it's a chat window I open and close. It's useful when it's quietly running in the background, nudging me when I need to know something, and letting me forget about everything else.

That's a different design goal than "better AI responses." That's ambient intelligence. And I don't have the VRAM for it yet.

Silicon Valley Was On In The Background

Got Jellyfin running. Media connected to the TVs in the house. And I celebrated by putting on Silicon Valley.

I needed the reminder to actually enjoy what I'm doing. Because if you haven't watched that show while trying to build a self-hosted AI system on consumer hardware with a GPU that's being politely described as "primitive" — the satire hits differently. Everything they laugh at in there actually happens. The gap between the ambitious architecture document and the reality of what you can actually run is a very real place that many of us live in.

I'd been grinding at this with a sense of urgency. The Jellyfin moment was a useful reset. This is supposed to be fun. The constraints are part of it.

The Gemini Experiment (Brief)

With the brain side of the system still struggling — I knew llama3.1 wasn't going to work on the current GPU, confirmed openclaw needed at least 12GB VRAM — I started asking myself the question I've been avoiding: can I get started with a cloud model while I wait for better hardware?

I tried Gemini. The least expensive tier I had available. Five minutes of genuinely good conversation. Then the paywall. Fallback to a cheaper model — another two minutes, then the same wall.

I cancelled the subscription.

Part frustration. Part forcing function. If there's a comfortable paid fallback available, I'll keep leaning on it. The cancelled subscription is a commitment device. I need the GPU — and now I have one fewer reason to delay getting it.

What's Actually Blocking Me

The honest state of play: JARVIS exists as infrastructure, not intelligence. The plumbing is solid — VPN access, Telegram connected, Jellyfin running, Python scripts firing messages to my phone. The Fortress is taking shape. But the brain — the model — is junior because the hardware can't support anything better.

This isn't a software problem. It's not a configuration problem. It's a VRAM problem. And it has a known solution that costs money.

The question I'm sitting with now is what I can do well at this tier, rather than poorly at the tier I want to be at. There's a version of JARVIS that runs entirely on small models and scripted logic that is still genuinely useful — not as a thinking partner, but as an ambient system that monitors, notifies, and logs. Maybe that's where to focus while I wait.

I did look closely at a Mac mini today. The silicon is capable, the form factor is right, the setup would be cleaner. I said no. It's spending money for convenience, and that's not the same thing as solving the problem. The constraint isn't the machine — it's the VRAM. And a Mac mini doesn't fix that.

So. GPU it is. The question is when.

Next up: the hardware decision — what I actually need, what it costs, and why I keep almost pulling the trigger on the wrong thing.