5
submitted 4 months ago* (last edited 4 months ago) by Randomize@lemmy.world to c/perchance@lemmy.world

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.

[-] Randomize@lemmy.world 4 points 7 months ago* (last edited 7 months ago)

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.

1
submitted 8 months ago* (last edited 8 months ago) by Randomize@lemmy.world to c/perchance@lemmy.world

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.

3

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?

[-] Randomize@lemmy.world 4 points 9 months ago

OP said "there were some updates going on a while ago (I think)". This indicates that OP is not sure about it and thinks updates are not running atm. So Mack2's comment about ongoing updates is valuable, no offense.

[-] Randomize@lemmy.world 4 points 9 months ago

on reddit, mostly.

[-] Randomize@lemmy.world 5 points 9 months ago* (last edited 9 months ago)

Thanks dev! There are actually two questions many of us have,

  1. The already mentioned context window (this was the one thing we hoped most for) and
  2. Which model did you choose (so we can look for workarounds easier)?

Thanks for all your hard work! I really noticed the text gen getting better little by little. Besides the hyperfixations and repetitions you mentioned there is also the on-off problem of following the prompt and writing instructions, other than this I've noticed only minor quirks (ignorable for me as the other problems are more important).

[-] Randomize@lemmy.world 4 points 9 months ago* (last edited 9 months ago)

Tell me you're minor without telling me you're minor...

[-] Randomize@lemmy.world 4 points 10 months ago

For me, too. but I'm more into t2t. I've basically started from zero with no knowledge whatsoever, but I've already learned so much about ai behaviour and surprisingly about myself. I hope some day I have the energy to try to understand coding a little bit, but atm, that's still final boss level for me 😅

[-] Randomize@lemmy.world 6 points 10 months ago

And none of them even think to consider that dev could just easily decide to TAKE THE WHOLE SITE DOWN FOREVER because of ungrateful spoilt kids who want everything, for free, in perfect condition, right now, and don't stop bitching about it.

So thank you for this post.🌈✌️

[-] Randomize@lemmy.world 4 points 10 months ago

Nope. To everything. Just let dev work out the problems instead of wasting precious time on answering "questions" like:

"The new AI sucks roll it back right now!!!!!!!11111!!!!"

"Why are you SO MEAN just give us the old model pweeeease?"

"Why does it take so long, ChatGPT makes updates in seconds!"

"Why do you ignore me when I say just pay 20 people out of your own pocket to get it done RIGHT NOW!"

and so on. I can totally understand why dev doesn't want to deal with this shit.

[-] Randomize@lemmy.world 4 points 10 months ago

Second this. There's already a post on reddit from someone who tried (and only locally!). OP, tell us when you're done! Let's see how long it takes YOU to get a perfect working ai with no bugs or quirks. (Spoiler: It won't happen, ever, because every model has quirks and phrases it often falls back on.)

[-] Randomize@lemmy.world 4 points 10 months ago

Exactly. And as for the model atm, it's tweaked very close to the old model (or maybe dev even put the old model back into place while working on the new one, idk). I just hope we get something different in the end, bc personally, working with the old model feels exhausting for my rp.

But I know dev is cooking right now, bc my rp suddenly got very sexual out of the blue 🤣

[-] Randomize@lemmy.world 5 points 10 months ago

I'm sorry this is OT, but I just wanted to use the opportunity to say thank you very much for all your hard work on the text gen (and img gen, too). Feel free to ignore or delete this if it annoys you, but I felt like this needed to be said. 🫡

3
submitted 10 months ago by Randomize@lemmy.world to c/perchance@lemmy.world

I'm not complaining, just mentioning that I noticed those especially those two words (but also a few others) are used very frequently in my /mem and I'm curious if it's reason or cause of strange ai behaviour. Does anyone has this too?

[-] Randomize@lemmy.world 4 points 10 months ago* (last edited 10 months ago)

-1 Just a thought... Maybe dev (who's working alone on this update in their free time, putting a lot of money and energy into it), puts the test versions out to get feedback and help with bugs and stuff from users? I think for providing such an awesome ai completely free they deserve some patience. It takes time to tweak and train it.

view more: next ›

Randomize

joined 10 months ago