3 ms·

Since we are posting short gists about the absurdities of prompting, I must share my own gist about how garbage prompting is today for LLMs: https://gist.githu

• wruza 2y ago
  But do LLMs work like this? I don’t understand your code to answer myself, but when I change the last word in the “notebook” tab of tgwebui (free chatless generation mode almost everyone is not aware of), then generation picks it up much quicker than if I change a word somewhere in the middle. This fact hints that alternation will cost much more in inference times, unless I’m missing something in your code. For some reason it doesn’t happen in SD, I believe because each “step” is a separate process that is affected by a restart anyway. That said I agree that LLM “prompt engineering” is much less actual engineering than in SD and is more “prompt writing”. Advanced SD feels like a fighter jet cockpit compared to LLM’s city car dashboard.
  • Der_Einzige 2y ago
    As it turns out, I wrote a whole extension for txgui that implemented some of what I wanted... https://github.com/Hellisotherpeople/llm_steer-oobabooga https://github.com/Hellisotherpeople/llm_steer-oobabooga