alphakamp on Nostr: I am diving back into ai hard again. (ADHD hyper focus, apologies to other hobbies ...

I am diving back into ai hard again. (ADHD hyper focus, apologies to other hobbies and previous hyperfocus and all the people involved). While I'm glad I took a few months off from AI, I'm really enjoying guiding the local Hermes along, giving it tools grabbing from a bucket of openai business accounts for a good usage window to get some coding done without paying high api costs. Glm-3.5-flash in openrouter is so cheap it feels like it is free compared to the big guys. Honestly glm is just as good as 5.6-sol.

The task I worked on today, I had attempted many times in my AI adventures. Past attempts I was never satisfied with. I have a wiki on dated version of doku wiki on a container template that could not be updated. In had it login to wiki via browser and saved credentials in hermes. Once logged in it saved every page. I then had it deploy a new doku wiki an import the whole wiki. Fecking brilliant, been avoiding bothering to do that job for a min. But in wasn't satisfied with that I want my agents to have direct knowledge of that wiki, then used llm-wiki( Karpathy's LLM Wik) to pull and convert the doku wiki to something the agents can parse quicker. Sol had a plan to build a whole synchronized version of the wiki. Thinking of switching to just poling the doku with api. It decided to put the wiki in sqlite and the wrote itself a client to parse it. Seems pretty damn fast. The filesystem tool is faster than api tools? Possibly I'm going to explore more. Next is getting more in depth with our system monitoring/alerting and ticket system. In have in house MCP for the ticketing system that I had codex write.

Ultimately it is great to build in house tooling running on our own premises. Next steps get some locally run model for main Hermes orchestration. Many choices in that space, I think if we had 32gb vram and qwen-3.8 variants, we could really be cooking. Anyone got spare vram?