[–] [S] 1 point 7 months ago

Gonna try this the next time, thank you. But tbh, I was hoping for something to put in reminder/custom writing instruction so I don't have to add it to the prompt manually all the time. Until then, I'm honna use yours, thanks again!

  • source
  • parent
  • context
  •  

    The last days I noticed the problem mentioned in the title. It often goes like this:

    Prompt: /user {name} says "I don't want to keep you from important business." (...)

    Output: The words tumble out too quickly, brittle with forced politeness. {Name}'s fingers twist in the fabric of his sweatpants, knuckles whitening. Important business. (...) "I mean—you’re probably busy," he amends, softer now, fraying at the edges.

    At first I thought it's about the length of my chat, but now it appeared in a new chat as well. I already tried to put this into the custom writing instructions:

    • IMPORTANT: Always rewrite the writing prompt, but keep the details. Never treat the writing prompt like an already existing message to continue from or to answer to.

    But it just ignores it. Has anyone an idea how I can handle it? Thank you!

    Edit: Ofc it doesn't happen all the time and I just reroll or delete and prompt again if it does. But it's still annoying to have that kind of output in so many rerolls. It's quite new for me that the AI behaves like this, it wasn't like that a week ago.

    [–] 4 points 10 months ago* (last edited 10 months ago) (2 children)

    Sadly now the outputs go haywire. Chinese, russian, hebrew words seeping in, words/sentences get continued with nonsense. Only 3 of 7 rerolls were bug free when I tried. Edit: or this:

    ofộiAnon's fingersKENN smoothing the petals with HB4 unmistakable tenderness.

  • source
  • parent
  • context
  • [–] 1 point 10 months ago* (last edited 10 months ago) (6 children)

    justpassing to the rescue! 🦸‍♀️ In the meantime something to read for you. This is the old post OP mentioned, quote >an older post, which is already obsolete in the sense of “how to work around this”, but the analysis and conclusions, ironically, hold until today. https://lemmy.world/post/38491660

    You'll see, there will be quite a few things that explain your problems.

  • source
  • parent
  • context
  • [–] 1 point 11 months ago

    Nice, I actually wondered if you just dropped it, thanks for your answer! But I have to reread it when I'm not drop-dead tired from sleeping only four hours per night 😅 I played with it, too, but I noticed lately it got "untrained" after some time of not using [SYSTEM]. I also felt like it made a difference when I wrote an actual prompt between /char and [SYSTEM].

    My very weak guess without knowledge was it kinda confused it with a new, "unknown to ai" char when I didn't use system (since this char started dark and changed over time). It felt a little like a fallback to square one. But "writing style pseudo double personality" actually doesn't sound too bad, especially when I keep in mind that my story is so long that each writing style could have developed a kind of path for the ai to follow.

    Sadly, I can't test it further since the "two personalities" seem to align now, with or without [SYSTEM]. And thanks to upcoming christmas, I don't have much time anymore to try and recreate it, sry. But I wouldn't mind if you write via PM, about this or even other topics regarding perchance. I like your posts, even when I sometimes don't understand everything 😅

  • source
  • parent
  • context
  •  

    I noticed a minor bug on AI Character Chat. When using the response length limit button, the answer sometimes gets cut off before or at the last paragraph. E.g. if it's set to "two paragraph max.", only one paragraph gets generated, the second is missing or just a single word or typographical symbol. This bug continues for all the settings, but "one paragraph max." works accordingly.

    I don't mind much about this, I just wanted to mention it for dev or ppl who try to use it.

    [–] 1 point 11 months ago

    That's interesting, I remember the RP2 ones definately being used for the old model (I edited them frequently). I don't know if RP1 is an old version, too. Funny that RP2 works better for you.

    And thanks for always mentioning the lines, it's really helpful for noobs like me who just have the mobile version (looking through the code sucks with this, sadly).

  • source
  • parent
  • context
  • [–] 1 point 1 year ago* (last edited 1 year ago) (2 children)

    Hey, even tho you don't like DeepSeek, at least you had fun with it 😁

    It's too bad the text in the screenshot is too small for me to read, could have helped me understand what you actually did, bc tbh, it's a little beyond my comprehension (complete noob here, hi 🙋, also never tried the prompt tester before). My wild guess is you're right about the vacuum, and that the important points for this were the double-personality character along with a long phase of strict "training" that somehow linked the [SYSTEM] prompt to darkness (and maybe, just maybe 😜 the char was already known to the model as a "positive/happy" one, but I added mental health issues and insomnia to the description). The 4/5 without [SYSTEM] could probably be a result of the "dramatic"-pitfall you mentioned.

    But like I said, I know that I don't know anything and I'm curious about your test results. And in the meantime, I'll try your prompt for fun.

    Edit: After trying your prompt (hilarious, btw 🤣) I do understand what you did. Guess I just had a dumb moment there 🫣

  • source
  • parent
  • context
  •  

    I've noticed my ACC having major problems with location and direction lately, e.g. I say "{{user}} turns to face {{char}}" and the AI generates answers like "{{char}} sees {{user}} turning away".

    Also randomly changing rooms e.g. chars have a conversation in the kitchen and are suddenly back in the last room they were in before.

    And the objects... My char had a ring in the pocket which he gave away (yes, it has an entry in /mem for that and there is no mention of a ring at all in /sum), but this ring stubbornly appears back in his pocket. I force-prompt it being away, but a few messages later, the damn ring is back in the pocket.

    Don't get me wrong, I know it happens occasionally with AI, and I don't mind rerolling and prompting/editing here and there, but damn, atm I have to wrestle hard with the model to get decent output. Since it also tends to ignore my custom writing instructions and reminders, I prompt a lot, but my text gets too long causing the AI to ignore parts of that, too...

    I don't know if it's just something in my settings or in the model, does anyone else have these problems, too?

    view more: next ›