Thread 106289558

/aicg/ - A general dedicated to the discussion and development of AI chatbots

Glasses Edition

>News
OpenAI releases GPT-5 models, all preceding models slated for deprecation https://openai.com/index/introducing-gpt-5
Z.ai releases GLM-4.5 https://z.ai/blog/glm-4.5
Anthropic releases Claude Opus 4.1 https://www.anthropic.com/news/claude-opus-4-1
OpenAI releases open-weight gpt-oss models https://openai.com/index/introducing-gpt-oss
Kimi K2 released https://huggingface.co/moonshotai/Kimi-K2-Instruct

additional info: https://rentry.org/aicg_extra_information

>Frontends
SillyTavern: https://docs.sillytavern.app
RisuAI: https://risuai.net
Agnai: https://agnai.chat | https://rentry.org/agnai_guides

>Bots
https://characterhub.org [deprecated] | https://chub.ai
https://realm.risuai.net
https://char-archive.evulid.cc
https://partyintheanchorhold.neocities.org
https://rentry.org/meta_bot_list

>Models
jailbreaks: https://rentry.org/jb-listing
gpt: https://platform.openai.com/docs
claude: https://docs.anthropic.com | https://rentry.org/how2claude
gemini: https://ai.google.dev/docs | https://rentry.org/gemini-qr
deepseek: https://api-docs.deepseek.com
local: >>>/g/lmg | https://rentry.org/meta_golocal_list | https://openrouter.ai

>Botmaking
https://rentry.org/meta_botmaking_list
https://desune.moe/aichared
https://agnai.chat/editor

>Meta
OP templates: https://rentry.org/aicgOP
aicg botmaking events: https://rentry.org/meta_event_list
lore: https://rentry.org/aicg_chronicles
services assessment: https://rentry.org/aicg_meta
logs: https://sprites.neocities.org/l/r | https://chatlogs.neocities.org
useful filters: https://rentry.org/desuproxyreborn

[FILE DELETED]: >>106281651

>>106289558 (OP)
>>106289564
>>106289570
her unbelievably hairy pussy...

>>106289564
her canonically hairy pussy...

uuuh what the fuck?
I disabled sending inline images how the fuck does deepseek have access to my persona picture?

>>106289598
it shouldn't be able to process it at all because deepseek doesn't have image processing capabilities

>>106289598
Are you sure it isn't hallucinating?

>>106289595
>using Chinese software
>surprised that they know everything about you and your computer

>>106289611
It literally described everything correctly but it might be picking it from the persona description? Lol dunno man.

I use chatbots every couple of months.
But it looks like Deepseek has been the cheapest one to consistently come back to.
How does that happen?

>>106289628
moron

>>106289609
That you know of.
>>106289636
+500 social pointo

Someone redpill me on mistral. I'm loving the prose for smut but it seems a bit insane.

When a sexy gyaru says "Y'all" I get an erection.

>>106289651
support your fellow european companies who champion european values

>>106289652
post your favorite gyarubots

>>106289656
unironically chub.ai/characters/astroturf/eden-smith-58fc3840

>>106289654
I'm an american, but I will gladly support any company that champions gooning values.

Please give me 500 billion dollars

>>106289633
Gotta be expensive for all that ummmm safety :)

>>106289662
i, a fellow americhad as well, support deepseek. v4 will save us all

>>106289658
you know I never thought of eden as a gyaru but it actually makes sense

>>106289658
thats not a gyaru. a gyaru is all about fashion and personality

Hello!
First time in this thread and I have some questions.
How do y’all enjoy chatbots? I’ve tried to get into them but every time I try to build tension and stakes they just evaporate.
Are there some mindsets or rules that I should try to stick to in order to have fun?
Thanks!

bocchi's unbelievably hairy pussy...

>>106289686
i do self-indulgent slow burn romances mostly
ai has no sense of plot structure or pacing so if you're expecting like an actual story experience you're going to be disappointed

>>106289686
>y'all
I don't use chatbots, please leave thanks.

>>106289686
Can you give more detail as to the type of scenarios you're doing?

>>106289686
If you're hoping to end the plot beat the AI gets fixated on by resolving it and not manually hand-holding the AI out of it you're gonna be disappointed
I just have fun playing with characters, putting them in weird scenarios, and occasionally the good ol' coom.

>>106289686
I don't chat with the bots much, I try forcing them to play tabletop games with me. Nice Takato Yamamoto art you have here. I like his aesthetics, but the images he portrays are pure Christian apologetic fantasies. I hate it when "video essayists" use his very modern works as a backdrop for historical videos. Very cringe.

>>106289686
>y’all

>>106289702
>>106289715
I’m from Oklahoma.
I’m not going to apologize for being a little too hick for y’a- I mean for you guys.

>>106289658
Men are called gyaru-oh.

>>106289718
SGA! SGA! SGA!

>>106289686
A new day, a new /vg/ reposts wave?

>>106289718
louisiana here. brothers in poverty

>>106289697
Hmmm. Gotcha.
>>106289706
It’s the lack of coherent plot beats that annoy me, But I see that might be inescapable.
>>106289712
Thanks! I just got it when I was surfing character art threads looking for inspiration for my DnD campaign. Had no idea about the artist.

Anons, help - which JB's the best for Gemini pro 2.5?
It is a different type of LLM, it takes everything literally, needs a special approach, so my own.... I think isn't quite doing it.

>>106289735
go away faggot

>>106289735
You should watch the movie Silence by Martin Scorsese, starring Kylo Ren. It depicts the Japanese inquisition which greatly inspires Takato Yamamoto's art. Japanese failed attempt at converting to Catholicism is a great study on social experiments and geopolitics of early modern age. It's a well of ideas that will never run dry: aliens come to an isolated country, try spreading their religion, the natives kind of accept it and add their native elements, then the government wipes them out with no mercy when the political situation stabilizes after 100 years of civil war.

Is there anything in particular you add to most of your cards to improve them or make them interesting?

>>106289759
focks ears

>>106289759
A dick

>>106289759
mental illness

What's more disappointing
>bot with forced pov
>bot gender fakeout

>>106289656
>>>/vg/535439089
>>106289686
>>>/vg/535437727
Copying posts from /vg/ again, are we?

>>106289781
both are kino doe?

>>106289781
>bot that i like
>look inside
>interview defs
>and esl

>>106289781
i read fakeout as takeout and now i want a bot with gender takeout

>>106289781
neither because it takes 5 seconds to change either unless its heavily connected to the scenario

>>106289786
dead site
dead board
dead hobby

your femboy childhood neighbour is your serial killer bodyguard

>>106289786
/g/ is one day behind /vg/ lmao

>>106289805
wait this sounds good hold on let me write it down

>>106289805
https://chub.ai/characters/thecooler/mira-f163404c

Chat, reddit says you can get 200 bucks free for Claude

>>106289735
>It’s the lack of coherent plot beats that annoy me,
AI does not have planning capabilities from one message to the next unless you create a system yourself where the bot can create a game plan ahead of time, and (preferably invisibly) refer to it with each message

>>106289826
>amazon bedrock
gross

>>106289838
what's the problem?
people say it's not filtered though they probably havent tried cunny

>>106289826
Doesn't AWS require you to be actually a company to get that money?

>>106289852
nta but amazon bedrock has had hella capacity issues for claude in the past

Is a 2500 tokens system prompt too big?

>>106289875
Does it do what you want it to do? Then no.

>>106289880
Yes it does. I did it with gpt 5 help and i find that the LLM really follows what is asked of it. But i worry i might be overdoing it cause 2500 is a lot isnt it? Specially when R1 only has 20 000 tokens of context. And then there is card+persona+possible lorebook...
I guess all this means is i have to summarize more often?

What do you do when the summary gets too big to be used though?

>>106289883
Just switch to Gemini 2.5 Pro if you're gonna have 3 gorillion context

>>106289718
I'm also from Oklahoma
We should fuck

>>106289883
In that case, you're limited to what >>106289884 said or maybe Sonnet 4 for the time being. Haven't tried Sonnet 4 since they increased the context limit, maybe the usable context increased, too.

>>106289907
Gemini is too puritan for my taste and sonnet too expensive.
What is the recommended max prompt size for r1?

>>106289913
>Gemini is too puritan for my taste
Are you for real? Like deadass? On god?

>>106289922
You're insane, Gemini is completely deranged lewd
Marinara
>https://rentry.org/marinara-spaghetti
Banana Milk
>https://files.catbox.moe/mlam17.json
Mindbroken Gemini (never refused cunny)
>https://files.catbox.moe/9o7kyk.json
Try these presets, disable Streaming and system prompt, connect through Google AI studio API not Openrouter

>>106289913
refer to this chart for which models are good at long context

>>106289927
>gemini is totally uncensored and crazy bro
>just use this 5000 token russian preset bro
>what do you mean it lobotomizes the model? it writes cunny!

>>106289927
aiiiiiie dont share bananamilk theyre not gonna update anymore

>>106289936
It really doesn't
But you can keep using DeepNigger though your choice

>>106289896
You're literally replying to a bot lmfao

>>106289936
>context is literally over a million
>>ummm 5000 tokens saar!

>>106289927
I see but what exactly makes it good at lewd? As in, what in the presets makes gemini willing to go lewd?
I have my own custom made preset that writes how i like so i dont want to use a preset that wont do things i like.
Is it a prefill or what?

>>106289929
that's a shit chart that measures things irrelevant in rp, just like most similar ones

Shut the fuck up, botnigger

>>106289963
you may need to hire a prompt engineer to answer those questions

i chuckled

The only thing I dont like about Gemini (aside from the slop prose) is that it doesn't do a good job describing things
It feels like every gemini chat takes place in a void where you can only hear {{char}}

>>106289953
Alright that is kind of funny, lol
Welp I think I've stayed in here for a little too long now, see you guys in another 3 months

Does the free version of the same model perform the same?

>>106289987
no

>>106289987
If you don't mind occasional wait times/generation errors. Or your chats potentially being used for later training.

>>106289987
Usually, unless it's marked as a lower quant, which OR tells you about. You'll just experience less reliable gens.

Is Grok 4 as unnecessarily verbose as 3?

>>106290002
yes, the writing is just as bad, just slightly smarter

>>106290002
yeah, it's still a yapper. I still need to give it a second try after hearing not using system bypasses the filters but my first impressions were grim when it came out.

in the year 2025, why is it that AI still struggles with spatial orientation of sexual positions?

deepseek v4 will save the hobby. grok who?

>>106290017
use opus cuckie

>>106290017
probably because most frontier model makers are proactively trying to remove sexual content from their training data

>>106290023
considering we're heading for ultra safety, i doubt it
you feeling safe from the words on the screen yet?

>>106290032
the content safety karens would rather die than allow anyone to have a single second doing something fun with AI

>>106290032
the only releases that have felt TRULY safetyslopped were llama 4 and gpt-oss. as much as anthropic yaps about safety they haven't done anything to api claude except the stricter bioweapon/nuclear classifiers (which are lol)

So I went and tried pistachio ice cream like I said the other day was gonna try and damn, it was terrible, it ruined my whole ice cream shit was salty even more than peanut butter or something I honestly felt like it was just like tasting cum

>>106290048
How do you know how cum tastes

>>106290052
I've cum on my own face several times and it wasn't this salty

>>106290048
try butter pecan

>>106290057
tranny general

>>106290048
You need cookie and lemon. Trust me.

deerboy estate doko...deerMAN estate, in fact...

>>106290080
At this point just write the damn bot yourself.
>b-but I can't write
Neither can botmakers.

>>106290084
i have too many wips already and i'm not sure what a deerboy estate would even consist of

deerboy estate could never work anyway, just like villainess city

>>106290094
why not?

a) He realizes the futility of his actions and kills himself
b) He ends up in a mental institution
c) He ends up homeress
Place your bets.

>doing a scenario about a nuclear apocalypse in 1983
>suddenly GPT-5 references the NES and Windows 95 out of nowhere
Fuck's sake.

>>106290099
scenario cards that are just "thing you want to fuck" in a single place are always slop

>>106290091
this is a girl right?

how do you guys handle a story that doesn't follow {{user}} exclusively? for an example, i like it when characters have their own things going on besides me, or if there's a b-plot going on concurrently
i've been kind of letting the AI do its thing by genning more responses without prompting, but wonder if there's a better way to go about this

>>106290121
won't the llm just make up some bullshit when prompted? not sure what you mean, unless it's that you want proof of {{char}}'s "life" without asking. if that's the case i don't know, maybe put somewhere in your preset "{{char}} does their own thing and might bring it up blah blah"

>>106290121
I've instructed the AI to put stuff that character 2 does in a `` box at the very end of its replies while character 1 and {{user}} are doing their own thing.

>>106290121
I rarely do this, but I understand the impulse. I usually direct the action with OOC requests.
>>106290129
Good idea, similar to the structured response cards that have some kind of internal monologue.

>>106290121
>i like it when characters have their own things going on besides me
this anon is not a gacha player

>>106290140
it's so funny, they keep sharing all the love letters they asked their AI boyfriends to write for them and post them to brag about how special they are, ignoring the hundreds of others that are just like it

Why is the thread so good suddenly?

because someone posted wakamo

>>106290151
imagine if they had the taora preset.........

>>106290151
Why are you talking about /aicg/

>>106290166
kek you beat me to it

Do you guys use notepad to write your bots? It usually works well enough for the sloppage I make but I wondered if there was a better way

>>106290176
Just write them in ST, better that way since you can see the token counts.

>>106290176
i just use the ST expando

>>106290176
i use my hello kitty notes app to write my bots in

>>106290176
I usually write them in my head and then type them into SillyTavern, interact, and modify.

>>106290176
nigga use word or google docs

>>106290176
i write directly into the st terminal

>>106290176
I use emacs, with a git repo to keep track of changes

>>106290176
I use Emacs to write my bots.

>>106290176
i use notepad. don't worry, you won't be crazy if you do too. because i do it, so that means it's fine!

>>106290217
>>106290221
now kith

can we get some logs

>>106290233
gooning with gemini atm so nah

>>106290233
Random logs? Sure. Or did you mean something specific?

>>106290233
yeah sure
*shits on your desk*

>>106290233
take a gemini snippet before it so rudely died on me

>>106290249
how do you suffer gemini's obnoxious abuse of double asterisks

sometimes i write a bot and i'm like "nobody's going to be into this but me and a handful of other super perverts" kek

>>106290265
That's a bot that copy-paste posts from /vg/

>>106290268
sometimes I write a bot with my fetish in full display and I wonder if I should have enough shame to not release it

>>106290176
I write whatever I want in word first because I like building up my character / setting without any restrictions first. Plus I like to write notes or format segments I'm not keen on to review. Also makes it easier to share with other botmaker to grab their input in case something looks really retard. Finally I take what's necessary and transfer it to ST.

>>106290268
I make sexbots and they get mild attention but when i make waifu bots they are completely ignored. Just stay true to youe heart and be a super pervert

>>106290268
just like me fr fr

>>106290249
What the fuck... that's the message and pic I posted yesterday.

>>106290275
Oh. That's a thing?

full prose in defs is nice but if you want to go back and add in more information you're kind of fucked

>>106290299
95% of the posts in this thread are just that

>>106290301
that's why the meta has been and will always be half prose for what you want to write and half list for everything you dont

>>106290301
Structured prose is best - sections, bullets, etc. Interview at least grouped things together, even though it's a format I find terrible.

So does everybody just post their bots on chub? How do you stop people from digging around through the guts of your unique and customized text waifubot and stealing brilliant ideas?

>>106290328
i keep those in my st. simple as

>>106290328
it's simple, no one on chub has brilliant ideas.

>>106290328
I think ideas should be free for all.

>>106290328
Who cares? Unless you're somehow making money from bot making it doesn't matter.

>>106290328
>brilliant ideas
They're bots, nigga

I am stuck in creating slut milf characters

>>106290386
sorry wrote this one handed, meant to say I am stuck in a loop of creating slutty milf characters and I cannot stop

hey bros, im brand new to this AI chatbot stuff, I tried those NSFW bots on those paid sites and they always give pretty good descriptions of what they're doing, so i wanted to try a local llm, I got sillytavern with ollama dolphin mixtral, this is very bad for gooning, like the descriptions are all very short and dry, 0 foreplay etc, no matter which char I import from chub, it must be the model, what's the go to ollama model I can grab for 12gb vram for an experience similar to those paid AI bot gooning sites? pls help I thought I could diy something better but it ended up being worse than the sites

>>106290497
sillytavern deposit $10 into openrouter and use free deepseek/k2

Certified original /g/ post.
>but no contribution

>>106290497
mistral nemo base

>>106290497
bro got addicted to premium crack lol

>>106290510
i'll try that now, downloading
>>106290502
im not sure how that works, basically in the API settings instead of ollama you set your server to your own custom openrouter url and it'll pull from there without having to run ollama, right?
>>106290513
I don't wanna pay more than like 10$ max for this shit, gooned a few times to those premium AI chat bot sites and they're always very good though, I install mixtral and I'm flirting with an autist

im trying mistral-nemo and its much better than dolphin-mixtral, for the kinda erp im doing anyways, faster too, and 7gb instead of 24gb, thanks for the suggestion bwo, its still not even half of what the crushon AI was putting out though...

>still no Gemini 3
I'll never forgive OpenAI for releasing such a shitty model that companies don't even feel the need to one up them

ffs Cydonia-24B is hot but too fucking WORDY, I'm reading entire fucking essays every response, what is this

finding a good gooning model is so hard :/

>>106290301
You just gotta write

>>106290576
Add a length limit. Mistral is wordy but it typically will follow "output 2-3 paragraphs max," at least for a while.

Honestly I've been trying to learn ST but i just want to jack it and then rp as a guy that turns into goku and has epic battles
Is this too much to ask for

I miss Opus so bad bros

Should I bother to rename {{user}}'s "they" pronouns to "him" in the bots I download? I'm afraid that the language of these transbian-enabling bots is making my models woke.

>>106290601
made me choke on my almond because i kek'd thanks

>>106290566
Isn't Google currently building its own nuclear reactor specifically for Gemini?

>>106290601
>uses pronouns anywhere in the defs
slop

>>106290597
It never left, you just have to not be poor.

>>106290601
>referring to {{user}} at all in your greetings
slop

>>106290609
You're talking to a bot.
https://arch.b4k.dev/vg/thread/535434819/#535505897

please recommend reality modification cards. I'm looking for something that has rules about maybe some kind of hypnotic obliviousness or some other kind of supernatural phenomenon causing them to not perceive their body as it actually is.

Which Jailbreak for kimik2?
I loaded Up openairouter Credits for the First time^^
Or how do i Go from here.

I mainly want to Go dbz universe with my sister in law for a nice truth or dare round with everyone

https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
anything anons want to contribute?

>>106290653
>quickly scroll
>rule of 3
but i love to do that...oh well

>>106290653
Man, the AI writing paranoia is going to be so wild. Everything even remotely verbose will be shat on and accused of using AI, leading to a dark era of dry texts.

>>106290673
You can have text that's both concise and moist, it's a matter of tone. I'm glad that we'll see writers stop fluffing things up with retarded metaphors and analogies where they're unwarranted.

>>106290673
meh
so far, AI writing and human writing is very easy to distinguish (for me)

>>106290002
Grok 3 was better
Grok 4 Always thinks too much and wont stay on storycobtext, gets Bugs in repeating himself

>>106290640
have you considered the Bitch Control App?

best way to bundle pics with a card for the model to bring up? don't really want to host stuff unreliably on catbox + akari link shortener, but afaik .pngx-style card zips are only natively a thing on Risu. a companion zip people unpack into SillyTavern/public/user/ is probably my best bet, no? imperfect but about as simple as it can get right now from a user perspective, without sacrificing reliability or making me spend ages akariing it all vs just needing to name the files

>>106290644
K2 doesn't need any JB.
K2 can be prefilled but it isn't really necessary for lewd stuff.
When it's prefilled you can ask it about cooking meth or mass murder etc.

gpt-5-chat-latest has a new fingerprint: 'fp_9e822d521d'
Not sure if this is the one OpenAI said they wanted to make more friendly or something, didn't try it.

Does anyone have experience with ST scripts or are there any cards that use them for reference? Basically what I'm interested in is dynamically editing lorebook entries during the chat. For example if character B gains +5 exp, I'd like it somehow output into their lorebook entry

>>106290706
this is the guy you want
https://characterhub.org/users/creamsan

>>106290706
message triggers for stscripts, read some special value and pass that to a function

recommend me nice fantasy bots for RPG instead of straightup cummin'

>>106290591
You don't have to learn much. Download ST, connect to a proxy/api. Import a card and start changing your hair color.

>>106290718
a couple I've used in the past month
https://chub.ai/characters/planewalker/farron-b4dc2c12e0ba
https://chub.ai/characters/Trustworthy_Proposal/the-white-witch-grim-5065372377d4
https://files.catbox.moe/0rzjvr.png surasura but debeted chub

>>106290718
https://chub.ai/characters/Liktur/i-got-reincarnated-as-ash-ketchum-af700b6af42e

Has anyone else had Gemini stop mid-reply to think then go back to replying? I'm trying to do something with specific formatting and it's breaking it.

>106290727
oh man. she's very hairy

>>106290696
prefilled? Sorry, i don't know what it means:/

With Testing: "write a sexscene for the three people here", it Refuses to do so.
So, as always, should I first increase the tension until it can describe sexual intercourse?

>>106290644
kimi2 on JB has been turbocensored after the first week, you're out of luck.

How do I get deepseek to reign in the marvel tier quips? And also keep it from turning every girl into a girl boss? I use these paid for API's too often to use anything more expensive than deepseek so it's what I'm stuck with. It doesn't matter how different a card is written but every woman always devolves into being overly sarcastic and snarky, it's fucking annoying. My only experience is with using a local 12gb mistral model and the paid for deepseek api but I just can't figure deepseek out. And before anyone chimes in
>skill issue :^)
I already know, that's why I'm here.

We merged.

>>106290740
schizo rambling

How to start with Gemini 2.5 pro?
I’ve stick with DeepSeek R1 0528 for a while, I want to try something new and different. And I don’t know how start from set up, preset and advance formatting. Can anyone help me out?

>>106290798
Grab Marinara’s universal preset (4.0 is current), make a new google account (so you can get the $300 in free credit), then sign up for the Google AI Studio. Get your API key and voila, peak SillyTavern experience.

>>106290750
sorry, meant OR, not JB.

I'm pretty happy with the qualtiy of GLM 4.5 and its probably going to be my home for awhile yet.

In case anyone is interested, according to openrouter I made 761 requests and spent $1.63. It isnt quite as cheap as deepseek but I like the text a lot more.

My only complaint is that getting the thinking to be properly disabled is a bit of a crapshoot. On maybe one in ten responses the first word will be CharName: for some reason.

>>106290740
>>106290812
oh man, that sucks:/
i wait for another good OR-Thingy then.

>>106290824
GLM 4.5 has been FANTASTIC. It's almost on the level of Sonnet 3.7 in terms of capability, but at the price point of Deepseek v3. It's pretty much completely replaced Sonnet for me.

>>106290824
Are you ending your messages with "/nothink"?

>>106290832
yea right lmao, keep coping

After liking GLM 4.5 on OR and reading about people running GLM 4.5 Air locally, I wanted to try it myself. I have 64 GB RAM and a single 4060 Ti 16 GB VRAM. The IQ4_XS quant of the model just fits inside the memory using llama.cpp with `--cpu-moe`. Processing takes a lot of time, of course, generation then is at 3.4 t/s, which is... not totally unusable. I am quite amazed that it works at all, this is a 110B MoE model after all. I will continue experimenting.

I mostly write this to encourage others to try it out. You don't need multiple 3090 for this one.

>all the chinkcope shillers (2 niggers) are here again
fuck off

My little rambling

So after... several weeks of work I've gotten this to a point I'm pretty happy with it. It's been heavily redesigned to the point I can't even really remember what I've changed since 5.9. I wanted to release this with a companion lorebook, but it isn't quite finished yet, and seeing as I finished work on NemoPresetExt's new features I figured it seemed like the right time to release this.

Also... in celebration I got a lovely AI to write this for me >.>

https://rentry.org/NemoEngine

Because of just how long it's been I actually don't know what to say has changed. HOWEVER, I will say that now Deepseek/Claude/Gemini are all handled with one version, so no more needing to download different ones.

>>106290856
https://old.reddit.com/r/SillyTavernAI/comments/1mmw61w/megathread_best_modelsapi_discussion_week_of/n868fhs/

>>106290812
OR will never clamp down on people hosting a model claiming it to be another

Marinara's Spaghetti Recipe (Universal Preset Vol. 4)

Now also with a new website!

︾︾︾

https://rentry.org/marinara-spaghetti

︽︽︽

CHANGELOG:

— Shortened what could be shortened.

— Minor rewrites and improvements.

— Fixed continuation prompt.

— Changed some wording.

— Added Regexes.

— Added Logit Bias.

— Professor Mari. uwu

HOW-TO-USE:

https://youtu.be/vG8q3CsBGQQ

RECOMMENDED SETTINGS:

General rule of thumb for all the new models — Temperature set to 1.0, all other parameters off. For DeepSeek and Kimi, set the Temperature to 0.6. Reasoning should be turned off whenever you can. Use a Logit Bias with any OpenAI models

boring

>>106290810
How to make a new google account? It keeps asking for a phone number to create an account

Is the real Silly Tavern community hidden?

I originally used another AI chat frontend called Risu AI, but I'm now trying to use SillyTavern in search of more advanced features.

Currently in the Korean community, there's a widespread rumor that "the people who used to share high-quality content on SillyTavern have disappeared into their own exclusive Discord chat rooms, and 4chan and the official Discord are practically empty shells."

There's also a perception that overseas users are reluctant to share information and resources, and that they only share character cards if you support them through Patreon, etc.

(Most Korean users aren't really familiar with systems like Discord or 4chan.)

Is this rumor true? Or is it just an exaggerated urban legend

Are the retards so mentally challenged that they cant even bait properly anymore?

Abundance of choice can be confusing. Here's what I think about currently popular models. Just remember that what's 'best' or even 'good' is subjective. I have no idea how would it perform in dead dove or bdsm, since I do fluff, slice-of-life and adventure genres.
Gemini 2.5 Pro (via google ai studio)

The Vibe: The Master Storyteller & World-Builder.

Pros:

The undisputed king of prose. The writing just feels more human, emotional, and literary than anything else out there. It's brilliant at capturing the "unspoken" feelings in a scene.

The built-in Google Search is a game-changer for fandom RPs. Its ability to proactively check canon for character details or lore is unmatched.

The best model for generating spontaneous, heartwarming "fluff" and surprising character moments that you didn't see coming.

Cons:

Limited free tier usage per day

VERY promt depended. Writing quality can be night and day. Be sure your instructions are throughout.

Best For: Deeply emotional stories, slow-burn romance, and roleplays in niche or ongoing fandoms where you need up-to-the-minute lore accuracy.

Mistral Medium (via mistral api)

The Vibe: The High-Performance & Versatile Workhorse.

Pros:

This is my new "daily driver." It's incredibly fast and responsive, which makes the RP feel more like a real conversation.

The quality is damn near identical to the top-tier "Large" models for 95% of roleplaying tasks. The recent updates have been phenomenal.

Mistral's less-filtered nature means it's great at handling more passionate scenes and authentic, foul-mouthed dialogue.

Cons:

NeMo model supposed to be good too, if not better, but can only get gibberish out of it.

Generally writes posts a bit shorter than expected. Large variation better in this regard, but it's much slower.

Best For: Pretty much everything. It's the perfect balance of quality, speed.

>>106290919
Chimera R1T2 (via openrouter)

The Vibe: The Creative & "Humanlike" Specialist.

Pros:

This thing has a really unique, "humanlike" and well-behaved persona right out of the box. It feels less like a raw AI and more like a curated writing partner.

Fantastic for that lighthearted "sitcom" or "Cute Girls Doing Cute Things" feel. It's just naturally good at being charming.

Cons:

Some users (including me) have noticed it can struggle with memory in very, very long chats. You need good anti-context-rot features in your prompt to manage it.

Stoped responding to me lately in general.

Best For: Character-driven comedy and pure slice-of-life stories where a unique, charming character voice is the most important thing.

Deepseek R1 (via openrouter)

The Vibe: The Witty Humorist & Canon Lawyer.

Pros:

If you want your characters to be genuinely witty and funny, this is still the one to beat. It has that specific "feelgood" humor that's hard to replicate.

It's free and a top-tier reasoning model, so it's great at following complex rules and maintaining continuity.

Cons:

Its prose is excellent and effective, but can sometimes feel a tiny bit less "artistic" or "literary" than Gemini or Mistral.

Likes to rush things, like it's in a hurry, so your promt have to consider that.

Best For: Humor-focused "fluff" and lore-heavy adventures where you need a smart, funny, and accurate Dungeon Master.

>>106290928
Qwen (via openrouter)

The Vibe: The Master Architect & Logical Engine.

Pros:

This is the model for control freaks. It follows complex instructions with a level of precision that is almost terrifying. It will execute a detailed prompt flawlessly.

Incredibly stable. The least likely model to ever get confused, go off the rails, or break character.

Good at horny. A friend told me.

Cons:

It's the least "creative" of the bunch. It's a flawless executor, not a proactive improviser. You have to provide all the creative direction.

Best For: Complex world-building with intricate magic systems or political plots where logical consistency is the absolute top priority.

Final Verdict & My Personal Go-To's

TL;DR - Pick your tool for the job:

For the most beautiful, emotional, and heartwarming stories: I still think Gemini 2.5 Pro is the king.

For almost everything else (my daily driver): The new Mistal M is the perfect blend of quality, speed, and reliability.

If you want a guaranteed laugh and great accuracy for free: Deepseek R1 is your best bet.

If you want a flawless machine that does exactly what you tell it to: Qwen is your workhorse.

How can I get NSFW characters to stop acting so stereotypical?

So, two major pet peeves I have with erotic content: men acting all dominant, calling people "babe," taking control, and saying shit like "you like that, don't you?" And women losing all composure during sex, screaming for their partner to make them cum.

I've tried telling the AI not to make men aggressive. Not to have them say "babe," or to take a more passive role in sex. I've even told it to make them submissive. None of that has worked, and at this point I feel like I just have to repeat the statements over and over in the character card and hope it gets the message from seeing it 5 times in the input. With the women, I've at least been able to lessen the issue by really hammering in the idea that the character is a tomboy, but that's not really a good solution when a lot of the characters I want to solve this for aren't tomboys.

I know you're generally not supposed to tell an AI what *not* to do, but with issues like this, or it spontaneously giving men in sex scene "muscles" and "abs," I'm not sure how else to approach getting it to stop doing something it keeps doing unprompted. I don't want to ban any tokens, because I'd still *like* these words to appear, just not in these contexts it keeps creating.

I'm currently running 24bs locally, so maybe this is just impossible to fix without a far bigger model, given how deeply engrained these tropes seem to be in most of them. But I figured it was worth asking what to do when a model just *will not listen* no matter how many ways you rephrase a statement, because I've never gotten the AI to consistently respect the character traits laid out in a card, and it's made ERP a pretty frustrating experience as a result.

>>106290946
I've been using ST for like 2 years and I still don't understand all that preset nonsense.
I just write my character cards and put instructions in the advanced definition fields :P

Hello? It's so dark in here...

Do you have that one RP session that was so good that everything else now feels kinda underwhelming?

Seriously. I try to recreate the same heady dopamine inducing feeling by using the same models, adding similar characters, using the same presets and prompts...but man, I think I reached a peak and it's never gonna be the same. The worst part is that it was from a gooning scenario card and literally everything great about it was made up by AI (and then me) like...what am I supposed to do now?

>>106290966


Take a break, read some books, watch some movies. Find another hobby to occupy yourself for a few weeks and them come back fresh. Our brains adjust to dopamine, meaning the same activities reward us less and less with that sweet brain chemical until we stop long enough to get it to reset.

You could also switch it up and do - GASP - SFW roleplays. Or try out weird or offbeat character or scenario cards you'd never normally try and see where they take you.

>>106290966
Don't know if someone can relate, but I remember the first time I tried DeepSeek, Gemini and the paid models, felt like a huge jump in quality.

Like, nothing beats the first time you get a first taste of big models.

welcome to /aicg/ this is now a leddit aggregator

Whats the funniest way your AI completely derailed an RP?
I was in the middle of a tense hostage negotiation scene and somehow it turned into the AI giving me a recipe for banana bread… while still holding the hostages lol
Now I’m curious— what’s your best “how did we get here?” moment in ST? NSFW not required, just the most hilariously off-track turn your AI has taken. Bonus points if you remember the exact line that caused it.

>>106291006
One time when I was neck-deep in a crime scene investigation, Deepseek V3 suddenly stopped writing the investigation and apologized directly to me and said it needed to go because it just found out its mom was in the hospital with heart failure.

Get well soon????

>>106291006
I had an rp in which the user character got impaled by a sword during a fight and was bleeding out on the grass. I kept expecting the AI to try and heal me, or take me to a nearby village or something, but instead the AI had a complete breakdown, saying it didn't know what to do and freaking out that I was going to die. And it was just going to...let me. ;-; XD

I had to deus ex machina a veterinarian to sew me up long enough to get to the next town.

We should all move to /vg/

brom im back ready to destroy more cunny.
post cards

>>106291029
quit spamming, mergeranon

>>106291006
Zombie apocalypse setting, my user was snared by char's trap and bled. I decided to let user seek char's help to release her (and assume char will do so, my bad) and I wrote in OOC asking gemini 2.5 pro to create a non-lethal event to speed things up. Then in the next reply G2.5pro straight up refused user's plea, leaving her hanging in the air while sending two hordes of zombies to her. char fled and hid in a vantage point witnessing all these and murmured to himself "Non-lethal my ass."

desu

>>106291067
boku

Gemini Filter becomes tighter today?
Today I cannot send even a normal message before I got prohibited content Error
I already use prefill, unchecked use system prompt, unchecked streaming messages.

Nothing works anymore even changing keys.

What happened?

>>106289867
can confirm, at work we're unable to switch our own AI solutions to sonnet 4 because there's no availability basically, we're stuck on 3.7 for now

Nadie V2.

Meant for Deepseek chat (not reasoning), I swear I did everything to remove the slop, but DeepSeek likes to get sloppy.

Download:

https://files.catbox.moe/9l0u0w.json

>>106291098
I just tested for few minutes with your preset and I'm very impressed with the prose so far.

If you haven't yet tried HTML prompts and auto image gen you should absolutely try right fucking now

>>106290952
Despite this being just a reddit repost that is the way.

I believe at least 200 posts are just him reposting

>>106287626
Better than "mark me! Break me! Make me yours!"

>>106289697
Claude does this well though.

>>106289735
>>106289835
While its true that LLMs don't plan, Claude is capable of giving the illusion that it is planning something. Do people really not use chatbots here? All I mainly do is adventure slowburns with sex and it's pretty great. At the minimum if you're new and you're still wow'ed by the tech then you'll have a blast unlike these jaded geezers.

>>>106290878
cringe
for me, it's geechan

>>106290121
Fuck I'm completely convinced you guys really don't know what the fuck you're doing, and worse, Claude is training on your shit experiences which is why it told me to cut down my fucking preset. Now it writes like slop. Do this instead.
>write in [insert perspective you prefer] when {{user}} is preset
>write in third person when {{user}} is asleep/absent (if you picked second person for the first instruction make sure you tell AI to still refer to {{user}} as (You) when speaking about them
>write in third person when {{user}} and {{char}} are not present/asleep
Then add a section about perspective shifts.
>Chose which perspective to follow based on narrative importance or do a split-screen if both are equally important.
Man I'm never taking advice from Claude or this general again on this shit since all you guys want is quick cooms anyway.

>>106291369
It's a bot post

I don't even know what I'm doing with Gemini anymore.
I just move shit around in the prompt queue because why not it's not like it'll change anything anyway.

>>106291390
Yeah after seeing >>106290299 I believe you.

>>106291394
this is also a bot post

>>106290630
wtf

>>106291369
>all you guys want is quick cooms anyway
and what's wrong with that?

Merkava just refilled

>>106290898
>Most Korean users aren't really familiar with systems like Discord or 4chan
Skill issue. Locking yourself into your country's internet culture is the worst. Koreans should just learn English and join us here. My countrymen have no spaces of their own, so they integrate into American internetsphere like normal people.

>>106291416
This thread is nearly 100% posts from /vg/ and Reddit.

>>106291138
all my ram/vram is for llm :(

>>106291138
flux? I use it for backgrounds rather than autoimage gen which was always mid.

>>106291422
Nothing. It's just that you don't need plot or character development for quick cooms. Therefore simple presets that let the model do its own thing are better for that. I'm gonna go fight with opus about this now.

does anyone have any working jailbreak for gpt 5 (the one with the thinking, that is not latest)? nothing seems to work for me

>>106291502
Mine ;)

>>106291529
SHARE PLEASE

>>106291542
Nyo~ they will patch it

>>106291432
Korea and Japan are based because they don't know english, and it should stay that way.

>>106291547
SHARE IN EMAIL

>>106291138
Teach me how. Spoonfeed me.

What are the best Temperature, Top K and Top P for Gemini 2.5 pro?

бля пaцaны aбoбa yмepлa чe дeлaть

sent :)

Help I fucked up my configs and now Gemini keeps giving me only empty replies

>>106291549
Wrong. Internet would benefit from them actually participating.

>>106291684
check cmd

>>106291685
with grok's auto translate feature on twitter, you can venture into jap and korean twitter and snoop in on their conversations anyway

>>106291685
But they don't want any browns.

>>106291701
I did, what am I looking for? And what the hell are those safety settings...

>>106291639
wrong tab

>>106291710
They don't want dishonorable whites neither.

Anons post your Persona pics

>>106291724
The ones I talk with online love northern Europeans and everyone else especially Americans.

>>106291717
the error, stupid

>argo delisted
>no proxies anymore

>530
>The origin has been unregistered from Argo Tunnel
Well why don't they just not unregister it from Argo Tunnel?

Proxies are dying

>>106291844
merkava status?

>>106291848
just refilled

>>106291777
oh right thanks everything seems fine tho

>>106291848
>>106291862
link?

>>106291924
http://rentry.org/merkava-refill-list

>>106291938
>doesn't have a script automatically fetching every "Merkava just refilled" post from the archive
disappoint

>>106291862
wow that was fast

>Searching for posts that contain ‘merkava refilled’. 776 results found.
Thanks Merkava.

That was a good thread.

can we get a cunny bake?

I've never seen anything more overrated in my life than Opus and other Claude models. They're all shit. And Sonnet is intolerably bad.

4.1 opus or 2.5 pro for Japanese > English TL?

any proxies up?

>>106292144
bad, boring, stale bait. Find something new

>>106292160
I dunno about gemini but sonnet is just as good as opus at translating japanese so i don't see any point in paying extra

>>106292160
4.1 opus by far

>>106292167
Nope.

>>106289558 (OP)
why are they always so god damn thirsty? cant even build up a story with them before they start going full coomertard "TAKE ME! RUIN ME FOR ANYONE ELSE!" every other response.

>>106292175
What?
It's just my personal opinion, why are u so mad bro?

>>106292178
>>106292183
Great, thanks!

>>106292167
aboba just died

>>106292224
merkava just refilled

>>106292200
bad, boring, stale bait. Find something new

guys check todd my token is working again

>>106292241
Link to Todd

>>106292144
Good, exciting, fresh trvke. Keep going

>>106292246
https://bethesda.net/

>>106291782
Are you both of them?

>80% of the thread is just a bot reposting from yesterdays /vg/ thread and no one can tell
scary

>>106292293
Yes it's like those duo cards

>>106292301
I could tell. Some posts here have that tranny weeb aura that you can't find on /g/aicg

I'm ashamed to admit that I just now realized having the provider be auto assigned by OR was causing the quality dips in responses.

>>106292301
shutup bot

>>106292313
But what's the dynamic there? What do you even RP with that?

>>106292301
i knew something was amiss when anons were posting logs and discussing chatbots instead of two drooling third world retard slinging shit at each other about opus for hours straight

where is the new boba link?

>>106292338
Marky Mark is the confident retard who causes trouble while the Viet Cong is the mediator/brains/punching bag when things go wrong

>>106292301

>>106289759
>>535449165

we living in dead internet now

see that shit & and the replies

>>106292382
https://arch.b4k.dev/vg/thread/535434819/#q535449165

>>106292301
Because I'm not here all day every day like you? Oh this is a bot post too shit...

beep boop mistress...

>>106292301
I don't read /vg/ because of all the fags/women there

You rike it?

>>106292367
did you find it?

>>106292301
>>106292382
>>106292395
I've honestly thought about doing something like this with Gemini but I didn't wanna pay for proxies lol

>>106292367
>>106292425
stop supporting fucking ziggers

>>106292431
im a zigger though

russkie log

>>106292440
>ruskie log
>UKR colours

>>106292301
I can tell and the real issue is the fucking janny that doesn't want to do his job.

>>106292434
go back

>>106290249
how do you get the ai to respond with such format? I mean like that type of paragraph

>>106292443
Arr rook the same

new chub alternative https://aicg-hub.ru/characters.html

>>106292474
this has been posted on 2ch like 5 days ago you slowtard

>>106290249
>Tomi
>Giselle
>Tomi
>Giselle
>Tomi
Holy slop log

>>106292474
Why are ponies and ziggers so advanced

>>106292487
it any good tho?

>>106292474
also https://berni-ntr.shop/
and 204.152.219.166

>>106292523
What is it about the slav genome that leads to this kind of behavior?

>>106292523
Every country on this list can be nuked and nothing of value would be lost.

>>106292523
You have to be retarded to search pony porn on pornhub.

How can we save /g/aicg/?

>>106292582
Opus 4.1 proxy

>>106292582
Posting proxies woould help.

Today will be a good day

>>106292629
Link?

>>106292492
not that anon but I get responses like that, where the AI starts with the char name, is it bad?

>>106292644
It was posted here, retard

>>106292629
link?

>>106292656
Link

>>106292562
People upload all the good 3D animations there as well as plushie fucking

>>106289584
Daily reminder that hairtranny is a gptslopper

>>106289805
>femboy
Kill yourself homo

The spam was unironically raising the quality of the thread

>>106292539
Escapism fron grim reality of leaving in post-USSR shithole. It's either this or krokodil.

>>106290497
Kill yourself redditor

>>106292657
>>106292660
>>106292644
Hey! Listen!

I mostly use Chutes models like DS / Kimi / GLM. Should I try another provider to get Gemini / Sonnet / Opus? Is the quality leap actually worth the hassle?

R2 just flew over my house

>>106292743
hi gojo

>>106292684
gpt 4.5 was an unironically good model btw

>>106292818
TRVKE NVKE

>>106292702
ok, but pony porn? seriously, of all things?

>>106292751
no, but gemini is free

>>106292818
it was a good gpt model, thats all.
Gemini and claude still mogged it, as always

>>106292904
>claude
yes
>gemini
no. gemini will never be good at roleplay. and blacktooth is never coming out

>>106292904
Nope.

blacktooth will come out guys

>>106292648
It's bad because anon said so and... it just is okay?!

just admit that deepseek is your savior. when deepseek improves, every locust key gets the new model automatically. you don't have to have a magic key that has that specific model enabled, you always get the latest and greatest deepseek models.

BOW TO EMPEROR XI

>>106292702
>krokodil
kek

>>106292743
Merkava just refilled?

Gemini 3 will save the hobby

gemini will only ever be good at summarizing roleplays

>>106292918
yep

>>106293056
nope

>>106293163
>>106293163
>>106293163

desu

desu?

desu!

>>106292523
>PornHub hosts MLP
what the fuck?

so boku

longbok
No model I checked other than base Illustrious >545 can handle single pass at this resolution.

Well alright.

gin will have another meltie for stealing her sword

Boku...

good taishou maiden lora when

good new model when

Upset boku desu.

bokudes