So… summer went quickly.
Kids have been back in school for two weeks already. And now I have two kids on Utah Valley University campus, which makes me feel old.
And wistful.
But whatever, life goes on.
— § —
And for better or for worse, right now a lot of “life” is AI and AI systems. At work, that’s my responsibility set in large part, which means that at home, that’s also what I do.
Current stuff:
-
Just did some hackery to make vLLM work on RDNA/NAVI21 (i.e. Radon Pro v620s, as I’ve become increasingly weary of llama.cpp’s broken cache management. That repo is at: https://github.com/leapdragon/vllm-rdna2-recipe.
-
Speaking of Radon Pro v620s, I’ve now got four of these. The last one still has to be mangled and bangled into the case, but it’s a mechanical problem at this point, we have a 1600w supply and risers to enable display and SAS out of other orifices (M2 slot, almost forgotten x1 slot).
-
Now have both Openclaw and Hermes running, as well as a dedicated tool host that is its own metal with decent specs but oddly what I mostly seem to do is sit around generating code with Kilocode (and then getting teased for using Kilocode).
-
Working on a reimplementation of a key favorite app from Newton OS. Yes, that Newton OS. There are parts of that OS that my brain is still shaped for and at this age I doubt I’ll ever recover.
And in general, it’s amazing to see Linux suddenly setting the world on fire. Apparently 2026 will be the year of Linux on the desktop, which tells us that I am exactly 33 years ahead of the curve.
— § —
I should post the tech stuff above on LinkedIn. “Hey I just ported vLLM to a new GPU” or “hey here are the things you need to know about agentive AI on local inference in 2026” or whatever.
But the thing is that a whole bunch of my sanity is maintained by not having to be a part of the corporate high school, keeping some semblance of dignity and integrity about me.
So, like a lot of other things I’ve done, it’ll just fade into the ether. I just can’t do it. I mean, I really hate LinkedIn. Like… really . hate.
But anyway, there’s a ton of code being written and hardware being hacked on. I haven’t been this personally embroiled in tech since the late ’80s and early ’90s. We’ll see where it goes.
— § —
In a way, the “where it goes” question is kind of important, because I’m soon going to be an “empty nester” and when I am, I’ll really be one. As in, empty.
Because I don’t live with anyone else. I don’t have a significant other. All of my best friends live out of state. It’s not clear to me what I’ll do, or why I would be motivated to do it.
I’ve passed that point in my life where work and ambition can be the things that get me up in the morning. So it’s got to be other things. But I’m not sure what.
— § —
Also amusing is that most of my interest in and attention to AI is really in a ’90s way, i.e. I’m still a Gen X computer guy at heart. That means:
-
I really like banging on the hardware, things like trying to fit more and more into a case and trying to get exotic hardware to work and to integrate with other exotic hardware
-
Though I ought to be sitting here building agents and harnesses, I have so far found that kind of boring
-
What I am finding interesting is using AIs to generate code, not autonomously, but working together on app projects that have no real purpose any longer thanks to AI
I burn millions of tokens at times, but it’s not hundreds of millions and it’s not fully autonomous, because I want to be the architect; I want to manage what gets created, moment by moment.
In short, I’m stuck in the ’90s, but that’s sort of okay.
I even went and got a Cherry keyboard so that I have old fashioned somewhat-long-travel keys, and I’m sitting here with a trackball instead of a mouse.
— § —
Some model testing, subjective impressions:
Qwen 3.5 122B A10B: I really love this model. I wish it was newer, and had received the benefit of everything that’s been learned about training over the last half year. It’s code is perfectly fine, but it lacks taste and vision when it comes to tech work. With that said, it is a great reasoner, planner, and chatter, one of my favorites at any size, that knows a great deal. I’m not running it right now, but I have a big soft spot for it.
Qwen 3.6 35B A3B: If you run this at Q8, you’re getting a blazing fast model that writes great code and has great taste. It’ll one shot a lot of things. It’s not quite as good at getting into the real multi-turn nitty-gritty of hard-problem dev drudgery; at those times it can feel disengaged.
Qwen 3.8 27B: For coding and project tasks, this is the best local model I’ve ever tried. I thinks a lot. But it’s the first local model that gives you that “frontier feeling” when working on a codebase. If you try it, you’ll know what I mean.
Laguna S 2.1: Don’t bother. Qwen 3.8 reasons a lot, but then generates perfect output that surprises you with its quality. Laguna S 2.1 just generates output. A metric ton of thinking output that… goes nowhere. It will nervously and neurotically think itself in circles for many hours and if you force it to complete and generate, it will generate something that well, well behind frontier quality and also well behind Qwen 3.6 35B A3B.
Nemotron 3 Super: Nah. It’s not special and more importantly, it “feels” just like a bot. Ok for code but lacks the “strokes of genius” that you’ll see when coding with Qwen, and just doesn’t have much that’s interesting to say.
— § —
Somehow Saturday is almost over. The weekend just barely began and I’m already thinking about the work week.
— § —
What am I going to do with myself?