I had a revelation that I haven't truly written every line of code for a project in years. I wonder what the last line of code I write will be?

I think the last project i genuinely wrote every line of code for was in 2023. Made a small app where you and someone else could swipe left/right on local restaurants until you 'matched' on where you wanted to go to eat.

Since then, i feel like AI has slowly taken over most of the coding, while i've designed the systems. It's pretty weird to think something...

Finally got OpenCode to run on two GPUs at one, and added a queue to help manage them more effectively. Decided to have one gpu dedicated to opencode work, while the other is for experimenting with llm based projects (currently a pokemon emerald harness).

Qwen 3.8 27b is genuinely blowing me away. It's like having opus from may 2026 running locally, for free, 24/7

Shoutout to the heroes in DevOps out there.

My day job is an AI Engineer, and i've grown to appreciate my DevOps/Platform/Security team so much more through this process.

Currently building out the orchestration framework and security around the locally run homelab. By far the hardest part (minus that thermal issue lol) has been new model deployment rollouts, dealing with drivers for ROCm, docker/harbor, full failure recovery, and system monitoring and alerts.

This is stuff I never deal with in my day...

I'm on a quest to let my locally-run llm system run for 18-24 on genuinely difficult projects with no oversight. Currently at 4 hours, and i'm hitting a wall as far as architecture goes.

To make this more fun, I started thinking of ideas to build towards.

One thought was to create a system that was good enough to beat the 3rd gym leader in Pokemon Emerald (beating Wattson the electric type leader seemed fitting).

At that point, the system would have to have both ephemeral and persistent memory,...

I've had a blast toying around with local llm, and I'm realizing how much my current orchestration needs to improve.

I'm able to hone it in within context (with some summary checkpoints) on a ~4 hour runs, but anything past that is having poor results.

Now, I'm working on the long-run horizon orchestration. It starts with working on the core unit which i'll label an 'event'. One event will log a point in time when a significant action took place, how it influenced the system, and events with it....

The local llm homelab is way surpassing what I had expected. Qwen 3.8-27b is much more capable than I imagined. It's roughly on par with sonnet imo, and i'm pretty trusting of it's code output for small and targeted tasks.

Also being able to fit 128k context into one 32GB card has allowed me to run multiple agents in parallel - something i was not expecting based on some initial benchmarks I read.

All in all, the open weight model is capable, small, and gets the job done without much tinkering....

The homelab LLM is finally up and running. V620s running all out, cooling solved, and qwen 3.8 27b is providing slightly worse responses than sonnet. All is right in the world.

This thing is loud as hell though, need to figure out a way to dampen the sound of centrifugal fans without restricting the airflow. Like a hairdryer in my office.

The local llm build continues. Cooling is by far the most unexpectedly annoying part for the graphics cards.

I decided to swap from axial to centrifugal fans to get better performance, but hit a bit of resistance.

The centrifugal fan is 4W higher than what my mobo allows for without potentially damaging something. I attempted to just underpower it, but on start up the fans hit 100% power (but quickly die down). Still, i don't want to hit that risk every time i start the rig.

So ordering a...

Never thought the most time consuming thing about a local llm setup would be cooling.

I bought some V620 cards for the home lab and these suckers run HOT. I designed and printed a hilariously big fan shroud/adapter for a case fan (140mm) to push air into it, but they still ran at about 90C even after putting more thermal paste on.

Everyone on reddit was right -- i should have used a centrifugal fan, not an axial fan. Snagged a centrifugal from a local shop and currently printing out an adapter for...