Pretty damn proud of this one. Omarchy Beam is officially on the Omarchy Plugin Store. The idea was ridiculously simple: Copy any link. Press Super + Ctrl + Q. Instant QR code. No opening a browser. No QR generator website. No extra steps. Just copy → shortcut → scan. It started as a tiny idea to solve an annoyance I kept running into, and now it’s something anyone in the Omarchy community can install. Small tools. Real problems. Open source. This is why I love building. Scan the code for quick access to the plugin! Thank you @ncfrontiersman for the assistance last night! #Omarchy #Linux #OpenSource #BuildInPublic

Guys! Make sure to chime in and listen! Great things are talked about in here.

I. Am. Back! The Hurricane really set things back but man it feels good to be back at 80%

Finally have enough service. Hurricane was nuts. No power or cell service. Hopefully it’s back on soon.

Open a webserver on your phone to transfer files from your Omarchy Computer and any Mobile Device on your home network! plugins.omarchy.org/plugin.h…

Used my own plugin today on another machine to move files fast... Im so proud of it. Its called Omarchy Beam. Its live in the plugin store and here on git, plugins.omarchy.org/plugin.h… github.com/TouchWorkStation/…

3 fun facts about Mercedes-Benz!

What did you guys do this week?

Finally! Omarchy Vault is live and in the store! Turn a spare drive into a private cloud on your home Wi-Fi. Press Super+Alt+U to upload from your phone by QR code, or copy files and press Super+Alt+D to send them to your phone. Local only, and it only runs while you use it. plugins.omarchy.org/plugin.h…

Tonight’s goal. Take the same learning process and apply it to a completely different model. Open source customizable models might be the future.

Tonight I tested an idea I’ve been chasing for a while: What if a small local AI model didn’t need to become bigger to become more capable? Instead of forcing more intelligence into the weights, I started experimenting with giving a frozen local model access to deterministic local capabilities that can execute, test, search, verify, and repair its work. The results were interesting. On a synthetic MBPP+ repair benchmark: 58% → 90% Same model. No weight changes. No additional neural tokens for the repair stage. That result had an obvious caveat: the repair engine understood the same kinds of mutations used to create the benchmark bugs. So I moved to QuixBugs — human-authored programming defects that we did not create. Raw local model: 15/40 solved — 37.5% Same exact model + frozen deterministic repair capability: 20/40 solved — 50% Again: No larger model. No fine-tuning. No additional LLM calls. No additional neural tokens. That does NOT mean the model suddenly became smarter. It means the system became more capable. And I think that distinction matters. The direction I’m exploring now is simple: Small local model → personal memory/context → capability selection → local tools/execution → verification → corrected result Instead of downloading a bigger AI, maybe we eventually download the capabilities it needs. That feels a lot closer to the kind of truly local personal AI I’ve been trying to build. Still early. Still testing. But tonight was the first time the idea felt measurable instead of theoretical. #LocalAI #OpenSource #AIResearch #LLM #MachineLearning

.... Tonight... The mid way test results so far. On 100 MBPP+ tasks with synthetically injected bugs, deterministic local repair raised the same frozen 3B model from 58% to 90% while adding zero neural tokens.

Day one of live use of a small piece of software at the dealership was a success! Now we tweak and do it again tomorrow!

More testing tonight. For reference, the machine I am using is a cheaper gaming laptop. 8gb vram, 16 gb ram (12 technically as using WSL) AMD processor. Tonight’s goal, Take the same model and cut tokens and give it a push in the right direction. Every test, every step I am logging to GitHub. If this makes an improvement on a very small model, imagine what it could do on a larger one. Stay tuned!

Last night’s experiments changed the way I’m thinking about small local AI models. What if making a model more capable didn’t just mean giving it another skill.md file or a bigger prompt? What if it could actually learn from what worked, what failed, and what was verified — and carry that improvement forward? Same small model. Same limited compute. More capable over time. Still very early research, but the results were interesting enough that I’m going to keep pulling on this thread. The goal isn’t a model that has more “skills.” It’s a model that gets better.

Tonight’s local AI experiment got interesting. Using the same frozen 3B model, I tested whether external capability logic could improve task performance without changing the model weights. On one small Python repair benchmark: • Base model: 50% • Compact capability layer: 60% • Verification-loop capability: 80% The tradeoff: the 80% result used more tokens, so this is not an efficiency win yet. But it does suggest something worth chasing: Small local models may not need to contain every capability inside the model itself. The model can stay small, while external tools, validators, memory, and task-specific capability packs do more of the heavy lifting. Still early. Still controlled testing. No “AGI on a laptop” claims. But I’m definitely continuing this tomorrow. #LocalAI #OpenSource #LLM #AI