Date: 2026-03-18 11:52:28
Holy shit.
So I got on Nate B. Jones' trail again, this time it's the 'build a second brain' project, which morphed into the 'build an open brain' project, which of course in my hands morphed into 'build a second brain, but everything must be local'.
It's actually kind of working. The biggest problem I'm having with it is that it makes shit up when it doesn't have the facts to solve a retrieval prompt correctly. Instead of telling me 'No, you do not have a cardiology appointment this week', it will have me scheduled for a heart transplant on Saturday morning.
While it hasn't done precisely this, you should see the batshit narratives it cooks up around my wife, about who it actually has some datapoints. I'm kinda taking a break from active development at the moment.
That doesn't mean I'm not thinking about the project, though. I want to give it a review, in my typical style, which tends toward a bit of comedic cynicism, tempered by some insight and wonder, all swimming around in a sort of engineer soup that seems sometimes to define me. I bet you thought my AI wrote that shit, but it didn't. That was totally me, being tot'lly 'umble, of course.
The project, OpenBrain.local, has four moving parts at the moment. brain_server.py, injector.py, brain_watch.py, and quick_inject.py. all of it is flask, all interoperates using APIs. The brain server exposes external API. it also parses down requests.
The heavy lifting is performed by qwen3.5:latest, a 9 billion parameter model, and lfm2:latest. Qwen acts as a front-end, and lfm2 acts as a sort of general purpose tool. Qwen is the head, eyes/ears/mouth/interpreter to lfm2, the hands. It works surprisingly well, is not terribly difficult to adjust and modify, but it needs work to be more than a really elaborate and effective demo.
The problem is, it's been vibe coded with google gemini, which has brougvht me this far with it -- but as complexity in the project grows, it gets exponentially more difficult to keep the model from gutting the system of some feature or features to 'simplify the project for the sake of troubleshooting". It does not mention that it has gotten rid of logging; or nearest-neghbor syntactic results from a 781 dimension semantic search; that capability just quietly goes away, and you only realize its happened when the logs go silent or the quality of results suddenly plummets to a level since hour three of the project's life.
This is costing time, energy, motivation, and most importantly, it is leaving me without access to a tool that is providing others of my peer group a great deal of advantage.
It's also costing money, as I am paying 20$/mo subscription to gemini pro.
Its a barely useable proof of concept at this point. I could rest an go back at it with gemini in a few days; I don't like the sound of that, though, that's how projects go to dark corners to die. I could also take it to Claude, get a code review, and pick up dev there. I don't like this, because Claude has a lot of exposure in the physical world right now, and I don't want to become engaged and have Iran blast a hole in the anthropic datacenter in Dubai and lose all my merbles.
It's especially agravating because this personal assistant/open brain thing is a complement to another unfinished vibe coder project, the bluesnake-ai agentic project. It's in a state not unlike this one; maybe not quite as far along as open brain local.
Maybe put it on github and let some other vibe coders go after it and make pull requests. Sounds like a lot of work though, and that would depend on any of them actually giving a shit.
When we get down to that point of 'bespoke, disposable software', the concept of collaborative development of really anything takes a hard hit; the AI systems make it possible to collaborate with yourself, for all practical purposes.
The jury is still out on how successful it will be at the scale at which I am working; but all it will take is for the next generation of open weights models to hit ollama. Improvements in the backend make every thing better up front.
Cheers and shit.
home/profile/status/shortform/soapbox/documents
produced via mdfoundry (c) James Stallings 2026 - all rights released under the BSD license.