my LLM is "self-improving" in a way:

- writing it's own skills/tools/plugins
- documenting it's mistakes and noting them for future
- tuning the LLM runner parameters

💪🤖

i think yesterday was the first day in a while i didn't post

my bad, i'm not trying to make nostr deader

my sleep schedule might be ruined but my agents are not

in other "vibecoding" news i'm now running hyprland (totally not worth it, better to write config yourself)

i just wanted to see what it was capable of

decent red theme, it was 90% complete

ok so here is my conclusion on #Vibecoding
definitely viable, it completed a whole extension for me and i never wrote any code (i did write 4 lines of tests, i could have just used natural language though)

one thing i liked is that the llm decided to write plenty of tests (100+ by the end) without being asked

the classic issue of "i assume this instead of looking it up" still exists, but not horribly

it took ages, and thought so much (usually over 20k tokens of thinking at a time), but i'd rather that than a bad result
besides it should be kind "in the background" anyway

i think i'd say it took 6 hours to write the extension (plus 4 hours of me configuring stuff *for* the vibecoding setup, llamacpp/opencode/etc)
that six hours was composed of 2 hours building the core "engine", 2 hours building the editor and another few sessions of me saying "fix this bug" or "add this feature"

i'll definitely be doing this more, i think with what i know now that extension could be built in 3.5 hrs

#LocalAI #LocalLlama #AI

#AskNostr what is more fun, polls or ask nostr questions?