agentic
Yes this meme has finally reached me, and turns out it's not a meme after all! It began by Werdy recommending opencode to me, and that is such a fucking horseshit disgusting software that I instantly sharted myself in my mind (Node crapware with 500 fucking megabytes memory usage excluding any LLM, and a disgustingly retarded 10k token system prompt written by a certified retard). So I took an existing concept of mine (generating LLM actions in YAML), and threw together shitcode.py (unreleased) which performs all the agentic stuff in 500 lines of Python. After ironing out a nice (short) system prompt, I can report that not only is it working, but it's actually fucking great!
What I do is that I completely ignore the chat-completion endpoint together with shitty Jinja and the model-specific toolcalling syntax, and simply bang the prompt together, expecting syntactically correct YAML, simply resubmitting if the reply was invalid until a good result is received (at this point, earlier retries are removed ofc). A good LLM must be able to generalize at least this much! Let's see some results:
With this workflow, the incessant "Wait", "Actually" outputs start to make sense. Wouldn't use these models for normal conversation though.

Úgy hozta a sors,
hogy tegnap volt szerencsétlenségem belekapcsolódnom a Budapesten zajló vulgáris entartete kunst-jellegű utcai performansz televízióközvetítésébe, amit - mivel a rendezvény támogatója a gyerekek nemiszerveinek szétnyiszálására szakosodott zsidó maffia - én csak trooncert néven emlegetek, már amikor feltétlen muszáj. Nos, ettől ugyan nagy változás nem állt be a politikai beállítottságomban (Dirlewanger-től valamelyest jobbra van a skálán), de ha Moloch követői ennyire el vannak pofátlanodva, akkor a legkevesebb, hogy még 20 évet kivánok a best squad-nak.
