Story 5: Build, build, build

Gemma 4 wasn’t working out, so Claude and I decided to try GPT-OSS 20B.

I know what you’re thinking. Is this guy crazy? Why isn’t he going for Qwen, the sweet spot for 16 GB? It’s good at coding, tool calling, and general knowledge. Best in class, they say. Call me what you want, but I couldn’t talk myself into running an open model made in China. It’s nothing personal. I just couldn’t shake the feeling that somewhere in code I couldn’t read, there might be something calling home.

Note: I did end up benchmarking Qwen locally in June 2026 (see Stats). It just never became Lucy’s primary model.

GPT-OSS turned out to be a pretty good choice anyway. Tool calling wasn’t bad, reasoning wasn’t bad, and I could fit it all in about 15.1 GB, even with a 131K-token context window.

So I kept rolling. But Lucy still wouldn’t give a straight answer. Either she made something up, or AgentGate blocked her and I didn’t understand why. Once again I was bored, wondering how everyone on YouTube had 8 GB cards doing handstands, building monster games, and basically running their lives on free tokens.

I kept building and documenting, but Claude started letting me down. I don’t know if anyone else felt this way, but around the time Mythos was a thing and Fable was coming out, Claude and I had a falling out. We didn’t always see eye to eye, and frankly, the building style changed. It became “Let’s spend seven hours gathering data, and then we can build the thing you want.” What should have been a Saturday-morning session turned into four weekends just to reach the next step of the Masterplan. I got bored, and the passion was fading.

Then, out of nowhere, GPT turned out to be a savior. Yes, I said it. Like a lot of people, I’d started with ChatGPT and never really liked it. After Gemini and then Claude, it was honestly a last-ditch effort.

Once I was there, though, I started enjoying myself again. GPT and I quickly understood each other, and its way of building and staying focused on the task really showed. I got so deep into my projects that I moved up to the $100 plan and built as fast as I could: 25 containers and all kinds of infrastructure, with plans to bring the NAS into the local AI and let it run my day-to-day.

But the inefficiency of my build was always there, haunting me. Eventually I realized the problem wasn’t the AI model. It was the build. Gemini, Claude, and GPT all had one thing in common: me.

The jankiest part of the stack was the guy at the keyboard.