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.
contact the webpage
No, never ask the llm what it is, it doesn’t know, but since there is the imbedded order “always answer” combined with a kind of “don’t say you don’t know”, it just outright makes up stuff. Ask it multiple times (in new chats) and it will give different answers. Dev never revealed which model it really is, but based on some indicators at the beginning of the update, the most common guess in the forums are it’s a deepseek version.
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 😅
Do you see the post by “perchance” pinned on top of this forum? click on the name to access the profile. There’s the button “send message”.
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).
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 🫣
Patience, or stubborn af, or both.😂 It’s actually not that bad (for me), surprisingly. I was able to let it run freely again last time.
Just a few thoughts to that [SYSTEM] thing since you want to test it: I put it in every damn message prompt (it’s pinned) and let my char go through a long, dark phase as we met, so this might have built the pattern. Also, this chat was a experiment about multiple chars in acc, the one with the darkness trigger is NOT the main char. Warning, attempting to trigger this may require patience! 😝
Absolutely! I have to confess, I started with perchance like… uhm… somewhere around a few months before the update, being a complete noob with absolutely no knowledge about AI or coding.
atm, I wrestle with a story on acc that is 6.03MB (!🤣) by now. I know I stepped into every pitfall but one (luckily got out of a few, too), and was close to dropping it, but now I push it for fun (yes, I’m a persistant nuisance 😅). I have to prompt almost every answer and it’s really, really bad, telling me the char which talked to me in the last message is deep asleep and it’s just all over the place all the time.
But I think I noticed a funny pattern by now. I wrote the char as complex (two-personality pitfall). Since it tended to ignore writing instructions and reminder completely most times, I continuously put “[SYSTEM] write two paragraphs or more.” in the prompt. By now, I think each time it triggers “darkness”, but when I just hit the char button without prompt, it’s “normal”. Not 100% sure about this pattern, just thought it’s kinda funny after seeing it a few times.
First of all, thank you so much for this post! I didn’t even realize how many pitfalls I fell into (clearly up to the point of being unrepairable). It’s well written and mostly easy for me to understand (“mostly” bc I’m not native english, but I guess I just have to reread it a few times).
On your final thoughts I agree with strawberryraven, for me the old model was almost unbearable due to the “happy sunshine”-mentality and how hard it was to get it even remotely away from pretending that life’s always perfect (e.g. Anon admitted trauma to explain their behaviour in certain situations Char: “Oh, that’s bad, I’m sorry for you. Let’s play a game, shall we?”)
And OT, I love how the combination of OP and strawberryraven showed a picture-perfect example of a mature, constructive discussion. It’s rare here, thanks!
I’ve noticed it, too. Atm I prompt every message. Check your writing instructions (helps a little bit) and maybe emphasize chars being (very) friendly and never aggressive (in char description and prompts, but tbh, I didn’t try that myself yet bc I’m at a point in my story where I have an important argument with my char.)
Edit: If necessary, go completely over-the-top with the description, we had to do that to create negative traits with the old model, maybe it’s now the other way.
It’s fine for me, maybe there’s something in your description/prompt/latest messages causing that? When dev tweaks temperature to dark, I usually notice on the spot since my char descriptions are not all happy sunshine, but I only had that yesterday. I just waited a few hours until it was good again. 🤷♀️
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.
Ah, yes, I wrote my comment before the post got edited ^^ Too bad, actually, it would have been nice for slow pace stories to tweak the temperature gradually away from the initial one. But thanks, also for your fork, it really helps me learn about AI at my pace and on my own conditions (I learn differently than many other ppl) :)
Agree, +1 I just wish I knew how to add this to the petrafied ACC for my existing chats, that would be awesome. OC, if you’re really bored, can you explain a little noob like me how to do this?
I’m not sure what exactly you mean. You can create a new character and just delete the chat in the menu on the left, the character won’t be deleted. If you want to add it to an existing chat, you need to click “options” (bottom right) and then “add character”.
on reddit, mostly.
Thanks dev! There are actually two questions many of us have,
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).
Jup. Man kann ja auch nach “Hot”, “Controversial” oder “Top (…)” sortieren.
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.