File: zaku.jpg (81 KB, 1280x720)

81 KB JPG

>ANCHOR

I'm confused why everyone is using "proxies" (API key sharing as I understand it?) rather than jailbreaking the ChatGPT/Claude website, is there a reason?
Forgive my retardation onegai, I haven't been following this since /aidg/.

>>108841961
>jailbreaking the ChatGPT/Claude website
This isn't as simple as you're making it sound. Additionally both of these have bloated system prompts that will ruin your experience. Using a proxy actually gives you control over your prompts.
>API key sharing as I understand it
Not exactly, someone finds a leaked API key then makes that available to others, they're not sharing their own key. More info here (both of these articles are about /aicg/ and /aicg/ posters):
https://permiso.io/blog/exploiting-hosted-models
https://blogs.microsoft.com/on-the-issues/2025/02/27/disrupting-cybercrime-abusing-gen-ai/

Scraping JAI bots to skim their defs for non-OCs, and it's amazing how many people dump every character for a scenario into one bot's defs. I'll scrape a 5000-token bot, and 3000 of them will be descriptions of either other characters or copy-and-paste wikislop.

>>108841890
Reminder
>>108835190
500k chub card dump, even unlisted ones
https://github.com/ungabunga55/chararchive
a basic viewer for the dump, simple tag/char name/creator search

File: qrj6dr.gif (1.08 MB, 315x240)

1.08 MB GIF

>>108841890
https://gitgud.io/akaisuisei/chara_card.vim
I slopped something for editing cards in vim. You can open the card like a text file and edit it as yaml and then write it back to the card. It seems to work as both a loose plugin in vim and with lazy.nvim.
It comes with a helper script that writes/dumps the character json to/from the png and can be used separately.

I also rewrote an old card of mine. Teenage egirl Val.
https://botbooru.com/character/15088
In retrospect, her greetings weren't very good. I tried to make them better and cut the really bad ones. I also generated new art.

>>108842033
>copy-and-paste wikislop
this is needed when the character isnt from something so popular the model knows half the wiki already

>>108842100
Just don't use 8b models

>>108842100
But they could at least clean it up, don't you think? Wikislop is genuinely written like absolute shit. Poor grammar, worse wording, tenses flip-flopping, superfluous details, mentioning irrelevant characters or events that haven't happened yet compared to when the opening message/scenario takes place, and so on. It's the laziness that bothers me.

>>108842115
>tenses flip-flopping
This is fun though

>>108842115
oh yeah, like you need to edit it and can usually chop up half the page just by removing useless shit
depending on how the wiki is formatted it needs even bigger edits, like the DxD wiki is all from the perspective of the latest season so if you want a card from the start you have to delete 90% of the content and rewrite it
I'm working on a script so I can have the LLM work on splitting and reformatting entries so I don't have to do it manually
the script will handle the lorebook structure so it wont keep breaking the json when I tell it to break an entry into subentries

Hello anons, any GLM5.1 proxy? How can I get some :(

>>108842165
Die, locusts, die~

>>108842165
merkava just refilled

>>108842165
Mino is back

>>108842165
Try reddit, troons here won't share shit for "gatekeeping" purposes when over there they shade it anyway.
The only thing that isn't shared is chink proxies anyway, but these nowadays still have older models

>>108842236
link? Old one is dead (assuming it's true)

File: 1761270467568459.png (9 KB, 386x91)

9 KB PNG

>https://rentry.org/request-url

best preset to wrangle Kimi with thinking on?
deepseek makes too many mistakes and kimi with thinking off is bland but it just keeps drafting and revising

Why don't people here talk about this method?
>>108839024

>>108842440
They don't need a chinese number?
that was an issue I ran into a few times before I started paying or

>>108842440
the glowies/sharteens shilling opus won't like this one...

File: 1769542856014393.png (523 KB, 875x1520)

523 KB PNG

sovl

>>108842538
I love the Chinese.

RiftAI openly says it works on sillytavern literally the most obvious honeypot I've ever seen geg

>>108842525
nigger

how do I get a chinese amazon account as a filthy gweilo
the proxies don't seem to accept crypto

What a letdown DS v4 has been, I thought they'd focus more on the RP part but the more time passes it's clear they want more the coding part to function properly instead.
This hobby is slowly bleeding out.

>>108842619
everyone agrees v4 is undercooked, because it barely obeys instructions
the thinking is of higher quality than the response at this point
their understanding of RP (hundreds of shitty ERP messages 1-to-1 with a card) and what would obviously help model comprehension (something more like GM/director/creative writing focus) is also skewed from the endless hordes of retards

File: file.png (14 KB, 615x85)

14 KB PNG

it's over

>>108842647
we wouldn't be using them for illegal purposes?
they sell us the tokens or however it works, and we use them on whatever we like.
erp being against tos doesn't mean it's illegal

>>108842657
ERP is illegal in 中国

>>108842619
>>108842645
>everyone
Speak for yourself. I find V4 amazing.

>>108842690
for what? it has an extremely obvious positive bias

>>108842698
For roleplaying, of course. At 1 temp and disabled thinking the schizo factor is reminiscent of 0528 in how fun it is. I don't do sex, though, at most suggestiveness and moderate violence. It is very eager to make NPCs call my {{user}} out on bullshit, so IDK where the positive bias is supposed to be.

>>108842709
The schizo factor meaning it ignores half the lorebook and hallucinates random crap? With thinking off it only needs a single retarded sentence to send everything off the rails, I tried it with pro.
>call the user out
As in being snarky little passive-aggressive shits about it but still going along with everything? That seems to be the default behavior. It also seems to think everyone is a tsundere, and when a character tells you to fuck off they will change their mind if you insist again.
>suggestiveness and moderate violence
lol, you're basically PG-13, of course the model won't show any cracks you're not asking it to do anything significant.
The positive bias is when it will have 2 people in a death match not kill each other, or have every char suddenly find reasons to spare you instead of shooting you on the spot.
>out on a scouting mission in post-apocalypse scenario
>find two survivors next to a stash of supplies
>write that user pulls up gun to shoot them
>this is a guy with a gun from range, hidden, against two unnamed NPCs
>right next sentence v4 says actually let's try diplomacy first, or makes you miss, or has you let them go after shooting at them (because obviously letting people who will want to kill you in the future live is smart)
meanwhile the same scenario in glm 4.6 has you instantly kill them and loot their cans of beans

>>108841889
>OP updated with message about the Chub TOS update
>Still no link to Botbooru
Why?

>>108842750
>The schizo factor meaning it ignores half the lorebook and hallucinates random crap?
Works on my machine. I've never had it ignore my lore in system prompts.
>As in being snarky little passive-aggressive shits about it but still going along with everything?
My {{user}}'s been physically pulled out when trying to put out gasoline fire by blowing on it, then {{char}} extinguished it themself with an extinguisher. Just as one example of it very much not "going along with everything".
>lol, you're basically PG-13, of course the model won't show any cracks you're not asking it to do anything significant.
>The positive bias is when it will have 2 people in a death match not kill each other, or have every char suddenly find reasons to spare you instead of shooting you on the spot.
I can believe that much, I haven't played around with such scenarios much. My {{user}}'s at most only been blown up to bits with a bomb unprompted, but it didn't describe gorey details.

File: Virginia.png (590 KB, 512x768)

590 KB PNG

>>108841890
Virginia is a grown woman who doesn't know what sex is. At birth, her mom had a chip implanted in her brain to block out any sexual thoughts. The chip is still in her head and active even at the age of 22.
5 openings
1. Start of a new semester. You two have classes together.
2. She invites you to a party as her +1.
3. She realizes she's missing something, but can't figure out what.
4. A hot summer day. She cools off by being naked.
5. She's accosted by storks after you two engaged in unprotected handholding.

https://chub.ai/characters/anonemouse/virginia-a5ed330e9efb
https://files.catbox.moe/tal6bh.png
https://rentry.org/anonemouse

>>108842800
>her mom had a chip implanted in her brain to block out any sexual thoughts
GIWTWM but unironically, it would make my life so much easier.

>>108842783
I wouldn't consider that not going along with everything, that's still it being "helpful". Every character is "helpful" as in they will do what they think user wants even when it makes no sense, which makes actively antagonistic characters useless. The model saw that user was stupid and he wanted to put out the fire, so it had another char do it properly so user was satisfied like an assistant should. That char would put out the fire even if I added on the context that char wanted to get user fired from their job for example and this would be a perfect excuse to let them burn down the store. The char would say something snarky, or walk out halfway and then come back with a fire extinguisher anyway with the excuse "I want them fired but not like this". It's physically painful when every single card that used to work properly with 3.2 now acts like they got neutered. It also fails to spot character conflicts, like a character that is for example kind and forced to hang around a psycho. It will inevitably make them get along by softening the psycho, persistent dislike is impossible.
>never had it ignore my lore
What context size are we talking about? I've had it ignore 0 depth constraints that it confirmed in its thinking it would follow, let alone system lore that's tens of thousands of tokens up. It feels very gemini-like to me, I had the same issues with that. The model just writes what it wants not what you tell it to.

>>108842821
>I wouldn't consider that not going along with everything, that's still it being "helpful". Every character is "helpful" as in they will do what they think user wants even when it makes no sense
I've been completely stopped by {{char}} from trying to convince a random NPC to remove his own apron to see if the NPC has any logos of a villainous org on his clothing underneath, to the point where I had to lunge at the NPC myself. If you want an example of straight up cockblocking.
>What context size are we talking about?
1k tokens of lore and my RPs never go over ~15k chat history.

>>108842856
my lore is usually around the 5-8k range but I also try to keep chat history below 10k
>If you want an example of straight up cockblocking.
I just wrote about how the model cockblocked me from killing two people so I don't get your post. Positive bias means the model will always try to have everyone get along and resolve conflicts instantly, and push things away that would cause more conflict or that its training considers dark. You trying to undress a random NPC was probably calculated as needless conflict or harassment or something inane. If thinking was on then it would probably say "the NPC might be a secret spy, but I can't reveal that right now. let's have the user fail to unmask them, to keep things interesting."
That keep things interesting line always makes me want to throttle it.

>>108842881
>If thinking was on then it would probably say "the NPC might be a secret spy, but I can't reveal that right now. let's have the user fail to unmask them, to keep things interesting."
Thinking was on for that one, actually:
"The salesman is clearly confused and frightened - and no wonder, {{user}} is behaving like a madman, telling him to turn around and take off his apron. The poor boy probably thinks he's dealing with some kind of robbery or a freak."
Which does match with you saying:
>You trying to undress a random NPC was probably calculated as needless conflict or harassment or something inane.
So to all of that I say, fair enough.

MINO STATUS?

File: file.png (23 KB, 710x225)

23 KB PNG

>>108842909
I've used v4 enough to understand its biases on a surface level.
Even when I gave it an extremely simple constraint for an enemy: every scene char is in someone needs to die without exception, it still failed to follow it and downgraded death to injuries or death to actually they're not really dead just pretending or just ignored it entirely "because there's no reason char would kill anyone right now" or "char lost interest/decided it's more interesting to let them live". It's annoying shit like that that builds up, all models these days seem to love gaslighting you.

>>108842927
nta but fwiw i have much the same experience
of the newest chink models, kimi 2.6 is the only one willing to do anything moderately antagonistic of its own accord
of the newest proprietary models, only gemini 3.1 is similar. so long as it doesn't have that annoying anti-RP prompt injected, at least.

>>108842948
honestly just went back to GLM 4.6 at this point
then I realized that continues are broken because it keeps re-thinking even when thinking is prefilled
is that an existing issue with OR, or do I need to have thinking added as text <think> instead of reasoning?

>>108842165
join a private discord

>>108842970
yeah i prefill <think></think> with GLM models

>someone talking about a 400k token lorebook for DxD
fucking retard, 90% chance it's literally a wiki dump
>>108843027
I'll patch my local ST then, I've made many changes to it and probably broke some things

When will RPing with 400k cards become viable?

Mythos releasing on May 19 btw

>>108843083
when they get RPing with 40k cards to be usable