My AI Journey
A running log · last updated Oct. 10, 2026
He who waits in the dark stays in the dark. But he who journeys forward eventually finds the light.
I'm not a programmer. I build PCs, I work with my hands, and not long ago AI was just something I scrolled past. Then a few videos caught my eye and I got curious. This is the honest log of where that curiosity took me: what worked, what flopped, and what I learned along the way. I'll keep adding to it as I go.
The spark
It started with a few short videos on Facebook about people using AI to trade stocks. I didn't believe half of it, but I wanted to know which half was real. So I started poking around.
What I learned: Curiosity is free. Acting on hype isn't.
My first AI: Gemini
Gemini was the first AI tool I really used, mostly as a smarter search engine. Tracking down car parts, figuring out maintenance problems, double-checking things. It shaped its answers around what I was actually asking, which a regular search never did.
What I learned: AI is only as good as what it can find. Double-check anything that matters.
AI on my own machine
Being a PC builder, I had to see if I could run AI on my own hardware. AMD has free tools for running AI models right on your PC, so I tried them out. Running a model on a machine you built yourself is a different kind of fun.
What I learned: AI doesn't only live in the cloud. Your own PC can run it too.
Came for stock tips, stayed for everything else
Investing was the whole reason I got into this, so I went looking for a sharper research partner and found Claude. It turned out to be a lot more than a stock screener with a chat box. It explains things step by step, argues with me when I'm wrong, and owns it when it's wrong.
What I learned: Treat it like a coworker, not a vending machine.
Letting AI near the trade button
Then Robinhood opened Agentic Trading, which lets an AI like Claude research and place trades in a separate account. I set it up with bumpers: a walled-off account, a small balance, a stop-loss on every short-term trade, and nothing happens without my yes. My first two trades both hit their stop-losses within days. Small losses, exactly as planned. That's when I knew the bumpers worked.
What I learned: The rules matter more than the picks.
A horror channel that runs itself
I also wanted to see what AI could create. I started arkofdarkness, a YouTube channel of short horror stories made with Faceless.so, and it pretty much runs itself now. Along the way I also tried Amuse, a free app that makes AI video right on your own PC. It was fun to play with but too unreliable for real work, and it's what convinced me to build a machine just for AI. Faceless.so is the one that stuck.
What I learned: AI can handle the busywork. You still decide what's worth making.
Building this site
I build PCs, not websites. But in October 2026, with Claude walking me through it, I launched Just Spilling the Bits in about a week: news, reviews, a newsletter, and a weekly market board. Things that would have taken me months of studying got done in evenings after work.
What I learned: You don't have to know how to do it. You have to be willing to start.
The home lab AI server Up and running
Being a PC builder, the next step was obvious: a machine built just for AI. It's done, and it's up and running. The heart of it is a used Dell PowerEdge R730xd, an enterprise server with two Xeon E5-2680 v4 chips (28 cores total) and 128GB of RAM. The muscle is an Nvidia Tesla V100 with 32GB of memory, a data-center graphics card built for exactly this kind of work. It runs Ubuntu Server 26.04.1.
The full parts list:
- Dell PowerEdge R730xd, 2× Xeon E5-2680 v4 (28 cores), H730 RAID controller, 128GB RAM
- Nvidia Tesla V100 PCIe 32GB
- Custom riser-to-EPS 12V power cable (the V100 won't run without it)
- Dell 1100W 80 Plus Platinum power supply
- 6× 1.2TB 10K SAS drives, plus one 900GB 10K SAS drive
- SanDisk X400 256GB SATA SSD
- 256GB NVMe M.2 SSD on a PCIe adapter card with a heatsink
- 2.5" drive caddies to hold it all
Almost every part came used, which is how a setup like this ends up within reach of a guy working a full-time job. The waiting game was the power cable, since that card doesn't get a drop of power without the right one. Now it'll be home base for running AI on my own hardware and for making the how-to videos below. Benchmarks and real-world results are coming in a future post.
The Working Bits
The Working Bits is live: a how-to section here on the site, plus a YouTube channel to go with it. Plain-English, step-by-step guides for real life, from spotting scams with AI to building my own AI server. Every how-to is something I actually did myself.
What's next: lab02
Update, Nov. 20, 2026: lab02 is up and working.
It's an old desktop board with an Intel i5-2400 and five graphics cards on PCIe risers, about 43 GB of video memory all together: a GTX 1080 Ti, a GTX 1070 Ti, an RTX 2060 Super, a GTX 1060 and a CMP 50HX, the old mining card. Two more cards didn't make it: the RTX 3080 crashes the machine, and the CMP 90HX isn't even detected.
Each card runs its own small local model: a 14B model on the 1080 Ti, 7B models on the 1070 Ti and the CMP 50HX, a Mistral 7B on the 2060 Super and a 3B model on the 1060. The one on the CMP 50HX stays loaded all the time and answers questions from my project notes.
What it does: lab02 is the text-thinking machine. It summarizes logs and run notes, sorts and tags render results, drafts descriptions and lists, and checks drafts before I use them. lab01 keeps all the rendering: stills, video and voice. When lab01 hands a text job to lab02, its GPU stays free for the render, and my Claude budget isn't spent on routine reading and sorting.
What it doesn't do: images and video. The cards sit on x1 risers with small memory, so that would be slow. Those jobs stay on lab01.
My rules: everything lab02 writes is treated as data and gets checked before it's used. Its model ports are closed to the rest of the network. Its files are backed up every night to the USB drive on lab01. And nothing gets installed on it without my OK.
New today: the 256 GB NVMe drive is set up, so models can load from something much faster than the old boot SSD.
What I learned: mining cards are picky: two of the seven wouldn't work. And the real bottleneck is RAM. The board has about 3.3 GB usable, which is tight for five models. A single 8 GB laptop-style stick was $19 on Amazon, so that's the next fix.
What's next: the RAM stick, then picking which data sets the always-on model learns from, and maybe a newer card down the road. I'll keep logging it here as it happens.
Got your own AI story? Email me. I read everything.