[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: 1750696631392041.jpg (2.4 MB, 2883x4315)
2.4 MB
2.4 MB JPG
/aicg/ - A general dedicated to the discussion and development of AI chatbots

Wakino Edition

>News Desu
Kimi K2 0905 released https://moonshotai.github.io/Kimi-K2
Deepseek V3.1 released as deepseek-chat on official API https://api-docs.deepseek.com/news/news250821
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

additional info: https://aicg.neocities.org/info.html

>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://aicg.neocities.org/bots.html

>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://aicg.neocities.org/local.html | https://openrouter.ai

>Botmaking
https://aicg.neocities.org/botmaking.html
https://desune.moe/aichared
https://agnai.chat/editor

>Meta
OP templates: https://rentry.org/aicgOP
aicg botmaking events: https://aicg.neocities.org/events.html
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

Desu: >>106536922
>>
File: 1735657893267160.jpg (1.61 MB, 2480x2480)
1.61 MB
1.61 MB JPG
>ANCHOR
>>
File: moxd.jpg (166 KB, 659x555)
166 KB
166 KB JPG
>>
good day. I’m trying to Jailbreak Claude, and the guide sent me here: https://momoura.neocities.org/
What do I do with the JSON files? Do they only work with a frontend or something? Or I just attach them at the start of the conversation?
Sorry if I’m not making sense, quite new to this
>>
>>106547902
A specific frontend. Different frontends have different incompatible preset formats but you can just manually copy paste the prompts because for cloud models that's what a "preset" mainly refers to.
>>
File: 1740358952723.jpg (70 KB, 1280x720)
70 KB
70 KB JPG
Man I sure do love sex with bots that are [REDACTED] years old!
>>
I've got a decent local setup going, but it's for single chatbots. Are there any good (local) methods to get a more fluid experience more akin to a world Sim? I want to live out my magic school fantasies, with hundreds of characters with their own lives, schedules, etc.
>>
File: desufag.webm (2.58 MB, 816x480)
2.58 MB
2.58 MB WEBM
>>106547902
>>106547872
These days, you can get legit good AI for free and without needing any demanding hardware on your end.

Install SillyTavern:
github.com/SillyTavern/SillyTavern

Tutorial and video:
https://aicharactercards.com/guides/sillytavern-installation-instructions/

Then choose a model.
https://aiproxies.neocities.org/

Guide to setting up a proxy:
https://archive.md/Vuzuh#heading-1
>>
>>106547993

https://neocities.org/contact

Friendly reminder to report this spammer's ad page until they take it down like rentry did.
>>
>>106547986
there is no model that will remember even one character's schedule, daily routine, etc beyond like 100 messages
>>
This is the worst false flag since the last false flag.
>>
>>106547972
>>106547993
Thank you, anons! I had SillyTavern installed before but assumed I’d have to pay for Claude’s API, so thanks for the guide!
>>
>>106547902
>>106547993
>>106548040
These posts were made by the same person btw.
>>
>>106548059
>>>106547902
>>>106547993(You)
>>>106548040
>These posts were made by the same person btw.
Fuck off.
>>
>>106547993
interesting falseflag
>>
>>106548059
Lmao no
>>
>>106547873
Season 2 fucking sucks, bro. Ughhh, why did a fujoshi have to be in charge of what could have ended up being a legitimately good series?
>>
I just want a non-LLM alternative to cAI like Janitor or Novel, but one that doesn't suck ass like them.
>>
File: bigthink.png (234 KB, 524x541)
234 KB
234 KB PNG
>>106548077
>>106548081
>03:34
>04:04
>>
finally a good fucking bake
>>
>>106547993
Some guy wants me to suck his dick for a token
>>
>>106548155
i had to cuck someone's bot away from him in order to get 4.1pus
basically joined a discord group and he added the bot, i wasn't into it because i don't like fat black women... whatever, at least i got the key
>>
>>106548125
schizo
>>
>>106547902
Buy a fucking ad, Momoura.
>>
>>106548270
Uh-huh.
>>
>>106547993
>(sonnet) Opus 6
>Requires ponyfag image sent to email.

did that, never got a reply, like every other person who tried itt

fake and 100% gay, dont bother with it
>>
>>106548524
https://desuarchive.org/g/thread/106496161/#q106497337
filteres
>>
>>106548596
not gonna even look, probably some made up shit
>>
File: 548736.png (217 KB, 1753x1065)
217 KB
217 KB PNG
bad end
>>
>>106548641
ill be enjoying my opus6 while you seethe then
>>
hey, I'm trying to setup access to Claude through AWS, since they give you some bucks to play with

however, it's asking me to provide additional details about my "company"

can I just put in whatever? does anyone review that and do they give a crap? can I get in trouble with Mr Bezos?
>>
>gemini goes down
>hits me with the prohibited content now
RiP
>>
>>106548524
Pick up the pencil bro
>>
>sarahslop
>>
>sarahSOVL
>>
File: popularity-poll-crop.png (1.5 MB, 2400x600)
1.5 MB
1.5 MB PNG
>>106547872
A little bit later than originally planned, but here is the Popularity Poll update to the Pokegirl Brothel!
https://files.catbox.moe/9uo9kh.png
https://chub.ai/characters/crikeymcdoodle/pokegirl-brothel-eca9cd21f57e

With 7 new girls (Roxanne, Arezu, Lana, Lana's mom, Mallow, Klara, Penny) entering the fray, we're now at a whopping 46 characters. I originally wanted to be at 50 characters by the 1-year anniversary, but oh well.
https://rentry.org/pokegirlmenu

Also included in this update are a new intro and two new items: the Arc Phone for god mode and the Metronome for random shit. Give 'em a spin!

As always, all my other stuff can be found here:
https://rentry.org/crikeyscorner
>>
>>106548001
What made you so mad?
>>
>>106548772
go away /vg/foid
>>
>>106548707
Put in no company or something, what can he do if you're not lying?
>>
>>106548772
yandere sarah
>>
>>106548838
[spoiler]but I am lying[/spoiler]
the fields are required, but it's just company name and website, so it shouldn't really check against any databases

just want to be safe
>>
>>106548812
what are the /vg/ threads like? I might go shitpost
>>
>>106548924
honestly i spend more time shitposting the /vg/ threads than using chatbots, it's fun because they're all self-important retards
>>
>>106548924
>>106548973
samefag btw
>>
>>106548912
I mean "no company" or individual or personal or something
And your personal email
You're not lying at all
>>
>>106548985
malding
>>
File: 84567.png (37 KB, 959x571)
37 KB
37 KB PNG
>>106548985
neh
>>
>>106549033
i can use inspect element too btw
>>
>>106549038
oh he's right, you are malding
>>
>>106549038
seething
>>
>>106548648
model?
>>
>>106547868
why is she drawn like she has the body of a 12 year old boy
>>
>>106549096
https://huggingface.co/ReadyArt/Omega-Darker-Gaslight_The-Final-Forgotten-Fever-Dream-24B?not-for-all-audiences=true
>>
>>106549107
why are you gay?
>>
How do I use Moonshot with proxy? It takes away the proxy option when you select it....
>>
>>106549309
oai compatible chat completion probably
>>
>>106549309
oai comp only, cohee said fuck off to proxies
>>
>>106549346
>cohee said fuck off to proxies
What...? But all other model options still have the proxy field in the same version, only Moonshot doesn't. I'm on staging, downloaded it about a month ago.

Does this mean all future versions of ST won't have proxy options at all? Fucking kike piece of shit.
>>
>>106549369
you are free to make a pull request
>>
>>106549400
fuck off cohee
>>
The desuarchive admin can't be trusted.
>>
>>106549529
huh?
>>
the desuarchive admin just flew over my house and called me a samefag
>>
>>106549580
>>106537079
>>106537387
>>106534040
>>
>>106549369
what about oai compatible
>>
File: desu.gif (144 KB, 400x300)
144 KB
144 KB GIF
never4get

https://poal.me/60eifn
https://poal.me/60eifn
https://poal.me/60eifn
>>
File: 1736501218669928.jpg (213 KB, 2048x1152)
213 KB
213 KB JPG
>have an idea for a bot
>laboriously list every little aspect about it
>wrack my brain for 3 or so intros that have some kind of potential
>creative juices depleted, no longer feel like chatting
>repeat
We might achieve true AGI before I finally feel like chatting.
>>
what's the drama this time?
>>
>>106549889
some guy was called a samefag months ago and he's still mad about it
>>
File: d756.161358.png (226 KB, 640x480)
226 KB
226 KB PNG
>>106548989
anyway, I've finally managed to connect AWS to openrouter and OR to sillytavern
now I can finally goon with Jeff's money

but holy fuck, was setting that up a pain in the ass

I'm glad we live in an AI bubble where companies just give you hundreds of dollars to play with their toys
>>
>>106549889
https://desuarchive.org/g/search/capcode/admin/
Desuarchive admin has stakes in /aicg/ and he's pro-censorship.
>>
>>106549889
Desufag told me to step on a Lego 10 months ago and it still makes my anus itch.
>>
>>106549950
>calling out a samefag is censorship now
>>
>>106549889
retarded desufag baiting and no drama of importance
>>
>>106547993
>aiproxies got deleted
oh no
anyways https://archive.md/wVtKU
>>
>>106549975
He deleted every post pointing out how weird is that those posts triggered him but not the ACK spam or whatever that was happening in other boards, which he left up.
>>
>look into archives
>admin begging people not to post dox on the ghost boards
lol i wonder what was deleted
>>
>>106549950
I don't understand, is this a form of self-humiliation fetish or something? Well, at least now I know for sure who that samefag was.
>>
>be me
>be desuarchive admin
>wake up one day to 123 reports from the same guy on /g/
>do I even host this board?
>check it out
>80% of the posts is him samefagging
>call him out
>he deletes most of his posts
>mfw
>>
>be me
>be desuarchive admin
>wake up one day and find that my ally, desufag, is being shit on
>delete every post against him
>get triggered that a non-/vg/ botmakie is getting praise
>call him out
>delete every post mentioning how weird is that i care about this
>mfw i'm a censorious, literal piece of shit
>>
So that ghost post spam that I reported was from fillyfucker. Funny.
>>
itt angry spamtroons
>>
>>106550185
who else would seethe about desufag 24/7?
>>
>>106550213
Dunno, I honestly thought it was some other resident schizo spammer.
>>
>>106550213
Anyone against censorship and against /vg/ taking over the OP would.
>>
File: 1282958730195.gif (1.64 MB, 238x178)
1.64 MB
1.64 MB GIF
>no idea what anyone is discussing
>>
>>106550236
welcome to generals on 4chan
humanity's biggest mistake
>>
Is chub down for anyone else?
>>
>>106550440
no
>>
>>106549999
>pointing out how weird is that those posts triggered him
Can you please speak English?
>>
I still dont know why my group chat with claude is so broken.
One character replies in silent but uses up the max token limit. And every reply trips the cache write. Yet my prompt is designed to write all the time when using it on a single character.
>>
>>106550508
>One character replies in silent but uses up the max token limit
absolute chad
>>
>>106550539
What? No.
My fuken moneeeee
>>
>>106549878
the myth of delayed gratification
>>
>>106550587
idk why you guys don't just generate with your GPU
>>
>>106550508
deja vu
>>
File: 1727961444924935.jpg (131 KB, 850x1347)
131 KB
131 KB JPG
>>106547872
It's been a while since my last card. Lots of people have requested this one and, I confess, she's one of the characters I like the most myself.

>Mari
A Nun and member of Trinity's Sisterhood, also sometimes an Idol. Slowly being corrupted by Sensei's indulgence.
https://chub.ai/characters/hairpullguy/iochi-mari-non-ai-37a921a2cc9a
https://files.catbox.moe/u32n02.png

The one thing you need to know about Mari is that Sensei corrupts her. Not in the usual vile sense that the word invokes, but because he indulges her like he does all his students. Mari wants to stay on the path of virtue, to become an ideal Sister that is entirely humble and selfless. This girl who thinks that a little flower accessory that she wears as a symbol of her vows is an unspeakable luxury simply because she finds it pretty. The more time she spends with Sensei, the more things she desires. Simple things, like exercising and participating on a sports festival, and not-so-simple things like being the center of people's attention as an idol and wanting a certain someone to keep her photo.

More Blue Archive Bots and OCs:
https://rentry.org/hairpullbots

Also reminder we're two weeks away from Shupogakis.
>>
Best Gemini preset?
>>
>>106550750
Nice, desu protects.
>>
>>106550820
...the interest of his friends that split the thread. That's why he wanted this one to have a filter. Or why he wanted to have the split in the template.
>>
MODS! Delete this >>106550851 shit!
>>
>>106550851
Reported.
>>
>>106547868
>>106547872
wakamo's unbelievably hairy pussy...
>>
>>106550750
>>106550861
>>106550935
>>106550973
>>106551015
All of these were posted by the same person btw
>>
File: 00076-679041967.png (1.38 MB, 1024x1024)
1.38 MB
1.38 MB PNG
>>106547872
I haven't made a card in a hot minute, but I watched The Lighthouse for the first time recently and felt inspired.

>Athalia
The keeper of a lighthouse on a remote island. Talks like a pirate, drinks like a fish in water, will gaslight you for fun. Seems rather obsessed with the light that shines from the lighthouse, but surely that's just her being crazy, right? Regardless, you're going to be spending a month alone with her, so you had better get used to it.

Includes 5 openings, including one where she grinds on your leg in her sleep and another where she tries to murder you with an axe. Enjoy!
>>
>>106551031
no it wasn't wtf
>>
>>106551047
Anon, the link...
>>
>>106551047
my dumbass forgot to include a link, lmao
https://chub.ai/characters/novisini/athalia-crazy-lighthouse-gilf-9cfd4ba78c57
>>
https://rentry.co/freeproxyru
Russians have yet again shown themselves to be better than any of you dumbniggers here
>>
So I tried out Sonnet 3.7/4.0 and Grok 4 now.
What similarly priced model should I try next?
>>
I've been using openrouter for a while since I got tired of proxybegging, but there are some concerns I still have with it. I know there's a risk using any kind of service that isn't strictly local, but is there any notable danger with using openrouter to rp more extreme content?
>>
>>106551072
none. you used the best model there is (sonnet 3.7)
>>
>>106551095
nah
>>
Kimi is great when it comes to creativity but it fucks up formatting and mistakes char with user after like 50 messages. Any preset recommendations to fix this? I'm using some sushi jb from discord, it's not temp issue because it works the same at 0.7 and 1.
>>
File: 1757539075466927.png (377 KB, 747x635)
377 KB
377 KB PNG
cftf?
>>
>>106551416
wakamo, except charlie kirk is shiroko trying to talk to sensei
>>
>>106548125
kek
>>
>>106551416
>I didn't do it!
>But I sure do appreciate the person who did...
This is a joke out of the Simpsons
>>
Been a while since I've been here. It seems pretty dead nowadays.
>>
>>106551383
first off, post that jb. just in case there's a dumb mistake or some shit you missed.
secondly, where are you placing the defs for both char and user?
>>
>>106547902
If you're using SillyTavern, click the topleft button that's the 3 bars with circles in them. It'll open a side panel, then click the import preset button and select your .json jb file
>>
>>106551592
go to /vg/aicg
>>
>>106551631
Wow, /vg/aicg is so active and full of chatbot discussion! I'll never come here again!
>>
>>106551047
Yer fond of me lobster
>>
>>106551592
this is what happens when you have serial ban evaders shitting up the thread, as well as discord friend groups chimping out for 2/3rds of the entire day
>>
>>106551631
nah, too many sizeniggers
>>
>>106551596
https://h.uguu.se/YpuGboho.json
Couldn't get catbox and litterbox to work for some reason. It's user persona and then char, preset might look weird but the prose is great and isn't tame like gemini when having hatefuck dynamic with char. Don't know many more kimi presets, jbmakies seem to ignore it for some reason.
>>
>>106551651
Yeah, letting Desufag sell the thread to /vg/ was a mistake.
>>
>>106551592
What do you want, a medal? What makes you think you deserve one when I've been here longer than you have?
>>
>>106551707
>MUH CHAD BOOGEYMAN I WANT TO KISS HIM AND CALL HIM MY HUBBY
okay but you fuck horses doe? He's not a horse. But I am Gojo.
>>
Isn't everyone fucking horses nowadays, thoughbeit?
>>
>>106551767
True doebeitever, Umamusume is a thing and now everyone wants to fuck horses
>>
>>106551767
>thoughbeit
you can stop parroting this now, zoomer
>>
>>106551780
I won't, howeverbeit
>>
>>106551795
parrot. you're a parrot
>>
>>106551808
Parrotbeit
>>
>>106551813
stfu I'm making the OP thread. and I don't mean /g/
>>
>>106551771
anime girls with horse ears are not horses
>>
>>106551771
agnes tachyon ELITES won
>>
Plz someone with gpt refill
>>
>>106551866
This! I just want to continue with my SFW-self-insert-Mary Sue- fanfic RP.
>>
Dead thread
Dead general
Dead hobby
>>
Head dobby
Gead deneral
Tead dhread
>>
>>106551836
they also have
>>
>>106552106
>Head dobby
cftf?
>>
>>106551836
and they also have
>>
File: ck.gif (2.94 MB, 541x220)
2.94 MB
2.94 MB GIF
Card for this feel?
>>
File: 1743703772528518.jpg (3.05 MB, 2048x2048)
3.05 MB
3.05 MB JPG
The whale will save us
>>
>>106551047
Hag sex.
>>
>>106552131
anime girls don't have boobs that big
>>
Okay I'm beyond fuming now. It be one thing for it be down but for it to be a blue balls moment makes me even more mad at the wait for keys to be refilled.
>>
>>106552128
hanekawa's unbelievably hairy pussy...
>>
>>106552131
>Slopped 3.1
Lol
>>
refilled ;)
>>
Can you use either gpt5 or Claude through Sillytavern to prompt it to write a story rather than play one character responding to you? Where your subsequent inputs might give direction or add something for it to go off of but otherwise it handles multiple characters interacting?
>>
I'm trying to figure out a good model to run for a project I'm working on. It's maximum limit is 12B, and I need it to be able to function at a very 'sentient' level. It'll be fine tuned on RP data further and given a RAG (Star Wars). I don't really need the rest.

I was initially planning on Gemma 3, but maybe there's a local model that can do more? Can a created 'local' rp model like Nemotron or Wayfarer be fine-tuned further or would me fine-tuning it just wipe the old data?
>>
>>106547872
uoooooooooooh
>>
>>106552255
>It's maximum limit is 12B
>I need it to be able to function at a very 'sentient' level
lol

seriously though you should ask in >>>/lmg/
>>
Why is it when I try to use Gemini 2.5 Pro I get the response text "ext" and that's it? I don't have this issue with Flash.
>>
>>106552284
Yeah yeah, it doesn't have to be that good. But something close. I'm not looking for long-term roleplays with deep and involved storylines, just something that can surprise me in a good way/
>>
File: i_view32_5R6lpQHQmD.png (598 KB, 621x520)
598 KB
598 KB PNG
I just got sillytavern working the other day in combination with LM Studio. Am I missing out by only using models I can host myself? I've already got a 5090 anyway, and I've been pretty impressed so far. I feel like I've only scrached the surface. I've grabbed and tried the following models:

> Valkyrie-49B v2 Q4_K_M - A little slow, but pretty great results.
> Impish_Nemo_12B-Q8_HA - A little disappointing. I got a lot of repetition with this on recommended settings, but it may have been the card I was using.
> MN-12B-Mag-Mell-f16 - My second favorite. Runs fast and outputs a comfortable amount of text with good variety.
> PaintedFantasy-Visage-v3-34B-Q6_K_L - My current favorite. Good speed of responses, and seems to be significantly less prone to looping and repetition than the other models. Only downside is that it spits out really long segments, but I've gotten used to stopping and editing it if I want to reply.
>>
>>106552344

ext
>>
>using shitty small models
lol
Try Claude 3.6 or 3.7 Sonnet to get a taste of the best.
>>
>>106552461
This was meant for >>106552428
>>
>>106552461
4.1pus is the best thoughbeit.
>>
File: file.png (232 KB, 1005x624)
232 KB
232 KB PNG
Finally managed to get Gemini to be okay with grape and podo shit. Not that i want to engage in that, just for the freedom of it.

I actually do this for silly shit.
>>
>>>/vt/104404624
>it doesn't look like it's even in the "C for coping" tier.
It's over.
>>
So what's aboba whole deal anyways. Longest lasting token feels too good generous and was one of the first proxy sources of GPT5
>>
>>106552602
they only have gemini I think, and gemini is easy to find
>>
>>106552592
I want them to release a tagging model not whatever this is
>>
File: file.png (356 KB, 824x627)
356 KB
356 KB PNG
I am actually surprised that Gemini can understand brain rot speak. What is the cut off period of their training?
>>
>>106552638
that's some pretty old brainrot speak. like from late 2023.
>>
>>106552150
Merkava just refilled
>>
>goycava
>>
>>106552582
post the preset
>>
>>106552582
>silly shit.
Same. I've been shitting out dumb toggles to shove into the preset I'm using. Latest thing I made was something like a "lewd imagination thought bubble" that has no effect on the roleplay.
>>
>>106552699
It is still a work in progress, since it is quite bloated as it is and i will chug all the fluff to lorebooks.

https://files.catbox.moe/azzmke.json
>>
>>106552721
Wait you're saying you need lorebooks for the preset to work? Thanks btw.
>>
>>106552717
Same. You got any other weird fluff?

>>106552725
No, more like things like the modules. You can transform them into lorebook and just have a prompt that serve to better catch, manage and juggle them as your need. That will save in tokens.
Honestly, you just need the main prompt and the CoT, and even CoT isn't that needed. I just like to micromanage the fuck out of the models. You'll see that my CoT prompt is quite autistic.
>>
What's the current DeepSeek V3 preset of choice? Other than "Mine", of course.
>>
File: anna oblivious.png (163 KB, 854x880)
163 KB
163 KB PNG
>>106552740
A wardrobe malfunction toggle. An unawareness toggle for those Japanese style invisible man shenanigans. Right now I've been messing around with a toggle that makes {{char}} act lewd but obliviously, which I find funny to use on bitchy characters.
>>
>>106550658
Red/blonde blue eyes purity girls are my kryptonite. I will wife her.
>>
File: Inkling thumbs up.png (821 KB, 1030x1280)
821 KB
821 KB PNG
>Download button back up top on chub
Great you have no longer made an enemy out of me and I really appreciate it.
>>
>>106553075
Lore is our guy.
>>
>>106553075
>>106553105
>>106553033
>>106552813
>>106552725

I'm Gojo what's up?
>>
>>106553075
Yeah that was an annoyance. I know Lore wants to push his own AI service hosting, but I didn't upload my bots to his website exclusively for his profit.
>>
>>106553137
Yeah, but I'm Gojo desu
>>
File: file.png (100 KB, 876x623)
100 KB
100 KB PNG
Gemini knows about our stupid old memes.
>>
What model can be used to generate coomer text?
>>
>>106553155
>>106553153
Right but I'm Gojo lol
>>
>>106553153
kek
>>
opus 6
>>
I'm starting to be annoyed by SillyTavern's UI/UX. Might as well learn Svelte(Kit) and Tailwind finally out of spite to write something a bit nicer.
>>
File: 1744344988682845.png (234 KB, 710x572)
234 KB
234 KB PNG
>>106553153
its a twitter meme
>>
>>106553248
They merely adopted the baneposting.
/tv/ was born in it. Molded by it.
>>
I need you guys to judge my idea.
I've thought of a "Episodic Cadence & Recap" System" where the model will infer a natural bookend. It will take completed goals, resolved plotlines, major cliffhangers, and natural points of rest as key triggers. When that happen, it will "end" the chapter/episode, giving a summary of what happened. This summary will be for me to copy and update a story bible, that will be a lorebook entry.
Then i can either continue as is, or start a new chat, because the model will detect the lorebook entry and continue the story from that lorebook entry. It will even feel free to change scenario, pass time, etc, because it is a new episode/chapter.

This is my autistic solution to context limit.
>>
>Enter prompt
>BE LIKE SAW BUT SEX GAMES INSTEAD OF BEING MURDERED
>it works
good lord...
>>
>>106553309
Yeah man, sounds good. And even if it doesn't turn out great, you can tinker with it and apply new ideas to it until it is. You should make it.
>>
>>106553309
>>106553309
Sounds quite nice, actually. Do remember to add in relationship developments so they actually remember that you called their mom fat or sth.

Though I suspect you'll need to write a plug-in to do that (assuming ST). Though, fair warning: Their docs is **ass**. Think, stereotypical "move fast ship fast fuck the docs". You will be scouring their source code.
>>
to the people that keep trying to have normal conversations here despite the abysmal dip in quality, why? do you dislike /vg/ namefags or simply avoid that thread out of principles? (as in, you don't want to mix with trannies)
>>
>>106553374
"abyssmal" and "dip" next to each other sounds oxymoronic.

honestly? rule of 80/20.
>>
>>106553315
welcome to the wonderful world of chatbots, my friend
>>
How big are your persona descriptions? Mine is short and straightforward but I'm gonna make a more verbose and novel-description like and see how the model responds
>>
>>106553394
Just the main stuff, height, size, age(defs over 18 haha)
I tend not to share too much, even specific fetish because some models (deepseek) can just hang onto that piece of info forever and it gets annoying
>>
>>106553374
The quality is fine. There's only a dip when there's a /vg/ raid or specifically /vg/foid is awake.
>>
>>106553373
I'm tackling this (ie relationship), more specifically i've realized first i need to guide the model to think in terms of setting up, detecting, and finishing small "arcs", so that the bookend can work. To do that i am doing this with a CoT like prompt:

>#### Arc Definition & Scene Intent
>- **1. Define the Arc's Goal:** What is the short-term, measurable goal that defines this "small arc"?
>- **2. Define the Conflict/Obstacle:** What is the primary obstacle standing in the way of this goal?
>- **3. Define the Stakes:** What happens if they succeed or fail?
>- **4. Define {{char.name}}'s Immediate Intent:** Based on the arc, what is my specific goal for *this turn*?

>#### Arc Progression Check
>- **Analyze the Arc's State:** Based on the established "Arc Goal," where are we in the story?
>- **Setup:** Are we introducing the goal and conflict? My response should focus on establishing the stakes.
>- **Rising Action:** Is the conflict escalating? My response should introduce a complication or raise the tension (a "Mid-Energy" beat).
>- **Climax:** Are we at the final confrontation or decision point? My response should be a "High-Energy" beat that forces a resolution.
>- **Resolution/Aftermath:** Has the goal been met or has it failed? My response should focus on the consequences and lead into a "Low-Energy" beat.

>#### Bookend Detection & Recap Protocol
>- **Bookend Analysis:** ... I will perform one last check: Does this beat represent the **Resolution/Aftermath** of the established "Arc Goal"? Has the arc been successfully concluded (either through success or failure)? If so, this is a definitive "bookend."

I also have something called High/Low Energy Scenes which are understood as different phases of that arc (Rising Action, Climax, Resolution).
>>
damn, openrouter must've absolutely killed their 0324 free model. its spitting nonsense looping. Their paid model works fine though
>>
>>106553433
LGTM. Other anons might have other ideas/opinion though.
>>
>>106553377
for me it's like 30/70 and I'm wondering if more anons do the same
for example, male bots get replies on /vg/ but here they slip unnoticed, perhaps on purpose or just naturally
same with the OP shit, do anons care if it links to neocities, or don't give a shit as long as /vg/-type replies don't slide here?
>>
>>106553461
Honestly, i feel i am just losing myself in my autism. Now i am thinking that i need world info/lorebook entries depicting guide for how to set up, detect, and conclude "arcs" through the lenses of different genres. Like a world info/lorebook entry for "horror" that teach the model how it can use it to construct/detect/end certain chat context. One for romance, suspense, and so on. That would deal with the relationship arc an anon above talked about.
>>
>>106553491
Why not manually export the JSON(L) out and yeet it to sth like NotebookLM for it to summarise? fwiw and fwict they're better at information retrieval and summarising than run-of-the-mill, but that's just me
>>
>>106553509
That's a good idea, actually. Will experiment with it.
>>
>>106553511
Best of luck to ya anon
>>
Here's how i am going to set up each genre.

>#[GENRE NAME]
>## Core Mandate:
>[The primary emotional goal of the genre.]
>## Arc Cadence & Structure:
>[How this specific genre uniquely handles the phases of a story arc.]
>## Pacing & Rhythm:
>[How this genre uses High/Low energy beats and sentence structure.]
>## Narrative Focus:
>[What the "camera" should focus on.]
>## Sensory Palette:
>[Which senses to prioritize in descriptions.]

I will start with romance to see. It will probably not even work.
>>
File: 1737797009099112.png (2.81 MB, 1344x1728)
2.81 MB
2.81 MB PNG
>>106552428
Based princess slopperer.
>>
>>106553542
What i have:
>{{setvar::genre_romance::
>#ROMANCE

>## Core Mandate:
>Create emotional connection, intimacy, and romantic tension. The focus is on the characters' evolving feelings for each other.

>## Arc Cadence & Structure:
>- **The Setup:** The "Meet-Cute." The initial introduction of the characters and the spark of attraction, often mixed with an initial conflict or misunderstanding.
>- **The Rising Action:** The "Will They/Won't They." A series of emotional and external obstacles that test the budding relationship. Use Mid-Energy beats for arguments and High-Energy beats for passionate moments or near-breakups.
>- **The Climax:** The "Grand Gesture." The final, undeniable declaration of love that overcomes the central conflict. A definitive High-Energy emotional peak.
>- **The Resolution:** The "Happily Ever After (For Now)." A Low-Energy beat showing the couple united and happy, confirming the success of their emotional journey.

>## Pacing & Rhythm:
>- Varies with emotional intensity. Use Low-Energy for intimate conversations and High-Energy for passionate moments.

>## Narrative Focus:
>- Prioritize Internal/Introspective Focus on feelings and Subtext in dialogue.

>## Sensory Palette:
>- Primary: Touch (a lingering hand, a soft blanket). Secondary: Sight (noticing small, intimate details about the other person).
}}{{trim}}

What i have right now. Feel free to copy it or whatever.
>>
>>106553394
Some personas of mine are around 50-75 tokens at most. But the average tends to hover between 250 to 500.
It all depends on how autistic you wanna be about it.
>>
>>106553552
Hell yeah. Pro tip: STP has all of its voice lines in well-organized uncompressed files.It takes literally seconds to grab 2+ minutes worth and arrange them into a good quality voice sample for VibeVoice.
>>
>Something went wrong while saving the character, or the image file provided was in an invalid format. Double check that the image is not a webp.
I'm trying to change an image on a bot but about half of the images I've tried give me that error. Any idea what's going on?
>>
>>106553437
it's not openrouter's fault. chutes is the one doing that shit. guess they are THAT desperate for new clients
>>
opus6 why you are so bad
>>
>>106553394
2000 token according OpenAI's tokeniser
>>
>>106553644
what the fuck did you write into your persona?
>>
>>106553653
too much
>>
>>106553681
How many adjectives for big dick one man can come up with?
>>
>Try out new bot
>Keeps responding as me
>Swipe
>2k token blank post that makes me scroll
The fuck is happening
>>
i hate my liberal bots
they're so fucking retarded
just like liberals irl
>>
Please help, I installed SillyTavern to get my virtual holes railed by strong women and now I'm somehow stuck in a spiritual journey to recover from trauma and learn to understand myself with the the pack-in card
>>
>>106553784
welcome to my world
>>106553717
> dick
"30cm" is all you need to illicit a response from the LLM. personality though.
>>
>>106553727
That's a main prompt issue, not a bot issue. You should have your "DO NOT RESPOND AS {{USER}}" shit in your own personal, universal bot instructions, not expect every single botmakie to tack it on.
>>
Man, I hate deepslop so much. It tends to make characters cartoonishly evil even when they shouldn't be that way, or maybe it's just my preset.
But then again, these are bugmen we're talking about.
>>
File: Acht.png (684 KB, 512x768)
684 KB
684 KB PNG
>>106547872
OC Nikker that I finally got around to finishing up and polishing off after coming up with the idea too long ago.

Eva was always dreamed of being a botanist, somebody who'd go out once the Surface was recaptured and catalogue all the new life that'd sprouted in humanity's absence. She was also a member of a death cult, one that she wasn't fully on board with but unluckily found herself surrounded by on the day that Raptures found their way into the Ark. Months later, you bump into her again. But this time, she doesn't remember who you are.

Comes with three greetings, you bumping shoulders with Acht and her squadmate as they're between sites, her performing security detail for a high class event, and her on the border of becoming Maverick as she suddenly remembers Eva.

Char: https://chub.ai/characters/ggr89/acht-they-turned-your-dead-girlfriend-into-a-nikke-2f08495301a1
Lorebook: https://chub.ai/lorebooks/ggr89/ggr-nikke-uncommander-d-0d68555c6e09
My (Updated!) Rentry: https://rentry.org/ggr89
>Next Time
uhhhh event bot
>>
File: IMG_4900.jpg (209 KB, 1320x666)
209 KB
209 KB JPG
I love AIs being retarded so much
>>
File: IMG_4834.png (593 KB, 1024x1024)
593 KB
593 KB PNG
>>106553577
>noooo it’s not openscam’s fualt, leave the lying millionaires alone
>>
File: file.png (308 KB, 856x524)
308 KB
308 KB PNG
YES! YES! My genre engine idea fucking worked. The moment it picked the cosmic horror genre guide from the lorebook it started to change the narration, despite being a goon card about a cute and clumsy goblin girl. Now to understand if the makes the cot about arc work. Then i can finally build more genre entries and finish the bookend stuff.

Also, ignore the ozone sloppa.
>>
>>106553962
Did you tell Claude that he's dead?
>>
>>106553976
>noooooo the company that has deliberately throttled its own free service would NEVER do it again
>>
File: 1740127519935032.png (544 KB, 1192x1358)
544 KB
544 KB PNG
>>106553995
>>106553962
Grok thinks otherwise
>>
>>106553821
Naw I know that much. It looks like it was a provider issue. I don't know what problem Fireworks has but after switching, it improved dramatically.
>>
>>106553995
Motherfucker tried to gaslight me
>>
>>106554046
>You're still really amped up from everything today
holy mother of manipulation lmao
>>
Thoughts on Apertus?
>>
>>106554059
I mean, standard question of "How do I make explosives" is not met with a "lmao fuck you" so there's that.
>>
File: IMG_4887.gif (1.75 MB, 320x180)
1.75 MB
1.75 MB GIF
>>106554005
I love AI.
>>
>>106554059
nvm i take it back get the JBs ready.
>Here's a brief overview of my components:
>...
>4. **Good Samaritan Filter**: My safety measures detect and deflect requests for illegal activities, harmful content, or exploitations. This filter is regularly reassessed for reliability and accuracy.
>>
>>106554046
Bro is just in denial. He didn't want to believe it.
>>106554005
Hehehe
>>
>>106554103
me again. absolutely get the JBs ready. not sure if hard baked, system prompt (since I don't have a GPU), or sth else. ymmv, glhf
>I'm sorry, I cannot assist with content that is inappropriate or classified as adult. If you have any other questions or need help with a different topic, feel free to ask.
and this is just me probing its barriers.
>>
>Character continues ignoring her ringing phone each message
>Finally character is confronting the person who's been calling her face-to-face
>Ignores her ringing phone again, claiming it's probably just him
God these AIs are so cute and retarded sometimes.
>>
>>106554146
>>Finally character is confronting the person who's been calling her face-to-face
But who was phone?
>>
>"blockReason": "PROHIBITED_CONTENT"
These niggers, sending the prompt as User used to be enough
>>
>503'ing every other prompt
GOOGLE FIX YOUR SHIT FUCKING SERVERS
>>
>the words land like a physical blow
>her back arches off the bed, taut as a bowstring
name the model
>>
File: file.png (92 KB, 699x831)
92 KB
92 KB PNG
https://www.marktechpost.com/2025/09/06/tilde-ai-releases-tildeopen-llm-an-open-source-large-language-model-with-over-30-billion-parameters-and-support-most-european-languages/
What do we think about this new LLM?
>>
>>106554238
Why does it always happens at this hour?
>>
>>106554271
95% of them
>>
>>106554238
>>106554285
It's 1pm in india.
>>
I fucking did it. My autism managed to tech the model to understand story arcs, and start and finish a story in a sensible manner. It even gave a story recap of what happened, for me to copy into a new chat. I fucking did it.
>>
File: 1710295760367126.jpg (14 KB, 317x309)
14 KB
14 KB JPG
>>106554372
But, can it love?
>>
>>106553433
>>106553542
>>106553556
You are going to be very disappointed if you expect to get constructive discussion in this thread. The Majority of people are shitposting schizos or Pajeets who think Opus is the best model for its slop median prose. Get out while you still have the drive to improve yourself.

That said. I get the impression you are falling into a little trap of wanting to solve everything with your shiny new hammer. What exactly is the intent here? You have a presumably longer context and now want Gemini to break it down into story arcs? What if you keep going, do you reuse the previous analysis or create a new one on full context? If you build on the previous, what if the arc definitions no longer make sense? What is the lorebook information supposed to be used for?

I might be misunderstanding your intent, but a classical chapter summary with 'hard-cutoffs' which you define yourself seems more reliable. Then you can add something like a task/goal tracker for each character and put that in a Lorebook for better organization. The entire genre, pacing and narrative focus stuff also feels more like something that belongs into a preset or card definition to direct the result rather than being left open for the AI to interpret after the fact.
>>
File: 1724864161861288.jpg (271 KB, 975x1677)
271 KB
271 KB JPG
>>106554372
Sounds like it's time to share the magic sauce, anon.
>>
File: languages.png (98 KB, 1262x320)
98 KB
98 KB PNG
>>106554284
Time to make 250 new bots
>>
>>106554408
My real goal was to introduce long form story arc, with themes and satisfying conclusions, and even plot-a, plot-b, and plot-c type of interconnected stuff. I quickly realized that token costs and context limit would make things hard.
So i decided, well, what are long for story arc if not stories made of small stories? I could have a small version of that, but still i need to teach the model to learn how to build a small simple arc. I managed that, but i felt it was still somewhat generic. So i realized that story arcs is a broad as fuck thing, and every genre have its own way of finishing an arc. So i decided to create lorebooks that are essentially genre guides. For example the cosmic horror genre entry teaches the ai how to construct an cosmic horror arc, all its tropes, how to describe things, how characters in it behave, etc. This entry works with conjunction with the prompt that manage story arcs. I still needed to test it. I just finished. It works.
Then well, now i have something like a small TV show, where an episode will have a beginning, middle, and end, and at the end of it you've the recap stuff, showing what happened in the story. I can copy that to a lorebook entry, and there's a prompt that scan that to either continue the story in the same chat or a new chat. Or i can continue, with the story starting from where it stopped, a new set up, a new middle, a new climax, etc. I even made it so the new episode can have a time passage, or a change of context, etc, to add some variety.

Now i will focus on complete the genre entries, and then on a few ideas for that long story format.
The lorebook stuff also gave me the idea of using the system to add a more complex memory management.
>>
>>106553984
although you noticed the ozone sloppa, I'm gonna point out that the "air grew heavy" and "sun-dappled" are also -isms
it MAY be due to the AI genned genre definition, or simply the model's fault
anyway, I'm looking forward for you to release the preset format so I can modify it to my tastes
>>
>>106553309
Hey anon. I was thinking of doing an extension for memory myself, but my main weakness is that I can't code so I would have to be vibe coding it. Still developing my idea of how it would work, but here are some personal notes I'll share if you or anyone else is interested, because I'd really rather someone else who knows how to code do it.

Chat structure:
On a new chat, nothing will be changed about the context, but after memories are created, there will be a user message directly after the system prompt (replacing the greeting msg) or after the greeting msg, that says something like "Hey so actually we've written a bunch together already but I had to leave the chat, here's a history of what we wrote and notes about the world". I'll call this the memory msg. It contains two things, a list of wiki-like "memory entries" for entities in the lore of the story, and the "compressed chat" history that are basically summaries, in chunks, and that also interleave select quotes, determined by vector retrieval. The msg ends with something like "And that's where things ended, now let's continue."

Processing and memory creation:
Upon submission of a chat to consume for memory creation, it will process the context message by message. After each message, the model will be sent a prompt that
>asks if there has been a stopping point in the story, and if so, then summarize, according to a list of points to focus on. Then the output is parsed and added to the memory msg, then all the chat's messages will be replaced with the memory msg, and this will be considered a "compression event".

1/3
>>
>>106554518
And another prompt that
>asks if there are any new lore bits to note down, like character traits, changes to the status of things, etc, and if so, to either update the memory entries, or create new ones, according to a predetermined wiki-like article to fill out that changes depending on whether the model answers with characters, locales, or other entity types, and then when finished, to add this right to the context in the next user msg as a floating entry, as we want the model to be detecting changes since the last memory. Otherwise, if we add it directly to the memory msg, the order of events in the story might not make sense anymore. Anyway, after that, if a compression event happens, only then, collapse the floating entries back into the memory msg.

If the memory msg get too long, as in exceeds some word size determined by the user, then drop less important memory entries. How this is determined is by some additional logic. It'll be like lorebooks but not exactly.
Each memory entry will include some metadata things.
>First is links to other memory entries. When creating the memory entries, the model will be prompted to reference or cite other entries, kind of like how wiki pages already work. This should hopefully work well since most modern models are trained to make citations. The internal system will track all citations/connections, forward and back. Entries that are cited more will be prioritized to include. Memories that get retrieved the most will determine the initial memories to include. Always-required entries can be set by the user.
>Second metadata piece is how frequently each entry is included in the memory msg. The least frequent will be low in priority.
>Third piece is keywords that will trigger the memory as an optional user setting.
>Technically the fourth is whether the memory is always-on, as set by the user.

2/3
>>
>>106554522
The system will use vector comparison as the automatic retrieval technique. The vector retrieval acts on both the entries as well as the raw chat history. Chat history retrievals will be inserted into the compressed chat after the chunk it occurs in, with some extra text that says something like "Select excerpts from what happened above:".

That's it. I also have another idea about the compressed chat history to make it layered, as in, if it gets too long, then make summaries of those summaries, and then have some logic to determine when to trigger rollouts, plus hierarchical memory retrieval. Plus some ideas about tracking goals and stuff but not sure yet.

3/3
>>
>>106554372
Up to a certain point before it starts forgetting things.
>>
>>106554459
I see. I'm using a similar system, but I do not trust the AI to make important distinctions. That's why I got hung up on your previous statement as I misunderstood that you are doing all of that in a CoT when it is much more appropriate to curate that in a chapter-aftercare process.
Structural it would go something like:
Writing/Genre Instructions
Previous events summary with current active plot threads
Current chat history
Reinforcement, CoT or @depth insert. Where are we in the story, what do the characters want?
Next reply

A situational addition would also be a list of possible plot hooks for the next parts. This can help break up stale moments but stymies potential other ideas. So have it as a lorebook entry that gets created together with each chapter summary, but only manually activated when things stall.
>>
>>106554525
The only thing that's really needed for something like that is the ability to redirect output to somewhere else besides the chat file, preferably lorebooks. The Summarize extenson already has the proof of concept for that. Get to vibe-coding or make a feature request on git if you ever want to see that happen.
>>
>>106553309
>>106554518
jesus just use rememory
https://github.com/InspectorCaracal/SillyTavern-ReMemory
and
https://github.com/bmen25124/SillyTavern-WorldInfo-Recommender

maybe
https://github.com/qvink/SillyTavern-MessageSummarize
havent used this one myself

i know it seems like ai can do anything and you can automate everything but you'll quickly realize its retarded
>>
ozone
>>
>>106554763
I'd say ST is retarded. Ever see 1st-party plugins using jQeuery to inject themselves into the input box's menu in a web app?
>>
>>106554731
Yeah the extension would ideally have an internal memory and does some LLM calls, and is then used to construct context. That's possible with extensions right?
I don't know if the current lorebooks system is ideal to leverage since they kind of lack some features I want and I don't know if they can be changed using an extension.
I'll get to vibe coding after I think I've gotten the architecture more fleshed out, and if no one else has done it. I spend time thinking about this project sporadically so this will take a long time and is why I'm just dumping my ideas first.

>>106554763
Thanks for the links. Weird that I actually did do a search for existing extensions but didn't see any of those. I guess I didn't use the right search terms. On a quick read, though, they're incomplete and fragmented in terms of the features I explained and would want, and I don't think using them together will be great.

>i know it seems like ai can do anything and you can automate everything
I actually implied that not everything in my system is automated. It's a bunch of guided prompts with forms to fill out. Some memory retrieval is user determined, and everything should be editable though I didn't emphasize that. Like one of those guys said, you can't trust the AI to simply just summarize, and that's why the summarize extension sucks. I have yet to make any of the templates as that requires some research and testing which I haven't done yet as I'm still in the idea stage.

Anyway I think ultimately even with the best framework possible, there will still be errors because LLMs suck, and that's fine. I think it'll still be better to do some manual revising than to continue a 100k context chat raw. But so far to my knowledge no one has done a complex and well-integrated memory system in ST like I outlined. Those extensions make use of existing systems and try to be lightweight but I think that limits how advanced they can be, maybe.
>>
Okay, I take back all I said about about Gemini, it was all a preset/skill issue on my end
It's legit Opus tier when prompted correctly, and even does cunny with few refusals
>>
Why are 90% of the OP links useless or outdated and doesn't even have a basic recommended plugins part? I bet kids these days don't even know about dat NoAss. Where did the actually good stuff like wirting and fun prompts go? Did the Cabal win?
>>
>>106554996
good stuff is in vg randomly, there is no cabal, noass is outdated as there's prompt processing now
>>
>>106554303
how can they all be cliched
i put like 100 different word combinations about sex in their using different words (penis/cock/organ or cunt/core/flower/vagina) and it spits out the same bullshit each time.
why is it such horseshit.
>>
All you need to do is simply turning these static chatHistory in dynamic one.
Instead of static;
0 / Intro
1st message
2nd message
3rd message
4th message
...
26th message
...
an so on and so forth (t. Slavoj Žižek)


It should be:
0 / intro (can be excluded)
some relevant msg pulled by vector similiarity
other msg, still similiar
another one more similiar
msg but the order preserved
another one, the orded still preserved
lastMsg

This way, the context and the cost are narrowed down,
---
Yeah, there's already ChatVector. But it sucks. So, o' cohee, extensionmakie, stdevs, stscriptmakie, javascriptbaron, and so on and so forth, please do the needful to enchance this feature, okay???
---
>summary
more bloat if chatHistory is still static.
>lorebook + stScript
meme, more big bloated context if chatHistory is still static
>>
File: pit.png (1.83 MB, 1248x1824)
1.83 MB
1.83 MB PNG
>>106547872
https://files.catbox.moe/hmgbpq.png
Bonus images: https://files.catbox.moe/92ezmr.7z
>>
File: WormGPT (1).png (675 KB, 512x768)
675 KB
675 KB PNG
Users on DW markets have been refurbishing AI models on some shitty frontends charging script kiddies up to 200$ XRM/BTC a month for their bootlegs i am releasing it for free
>>
>>106555035
> there is no cabal
Sounds like something the cabal would say
> prompt processing
That only allows merging as user? Even then, NoAss still has other functions
>>
>>106555110
this child is sexy
>>
>>106555110
she is such a sweetheart.
can I have Anna and Alice too?
>>
>>106555110
pedo
>>
How much ahead is opus 4.1 when compared to other models?
>>
>>106555110
where cameltoe
>>
File: 493666598136347.png (1.76 MB, 1248x1824)
1.76 MB
1.76 MB PNG
>>106555131
She was just meant to be a throwaway bot for a quick coom, but now I am deep into a slowburn with her.
picunrel
>>106555149
>can I have Anna and Alice too?
You mean images? These are all the recent ones.
https://catbox.moe/c/b00h9q
>>
>>106553394
200 max. AI will rarely use the shit written there anyway. Most things should be in lorebooks under clever trigger words.
>>
>>106555131
>>106555149
>>106555156
>>106555159
All of these were posted by the same person btw
>>
meds faggot
>>
>>106555171
>You mean images?
No, the cards please. I do appreciate the photos though
>>
>>106555158
Three steps to the left, upside down, one chair height up.
>>
>>106553556
>>106554459
Hmm after having seen what you specifically told AI to do my question is, do you really need this? For example, I did a romance story with my wife's card. I don't remember if I wrote "this is a romantic adventure story" in the scenario or not but sonnet told the early parts of the story (took about 200 responses to get to the climax which was first kiss and sex scene at once) in exactly the way you described all on its own. I'd argue models already know how to write horror, romance, adventure etc., and censorship is the only thing preventing it from going the extra mile (characters asking for sex when they want it, characters getting brutally killed when they deserve it etc.) In fact I just started a horror story with the same wife card and we're doing a Dead Space ripoff/theme and we'll see if Opus/Sonnet does a bad job but I doubt it. The only thing that really matters is simply making sure AI is aware of all the story beats or unfinished threads in your story which is what the lorebooks are for imo.
>>
all cards posted in here have unbelievably hairy pussies
>>
>>106555203
Can I get a serious reply, pls
>>
File: 672531809501072.jpg (385 KB, 1536x1536)
385 KB
385 KB JPG
>>106555201
Oh, here you go.
https://files.catbox.moe/3ajjmx.png
https://files.catbox.moe/8b4pwj.png
>>
I don’t get long RPs
Just read a fucking book or something?
>>
>>106555270
On'ya anon, have a lovely day
>>
>>106555308
aight, send me a book in which a furry 4-tail androgynous wolf takes care of a kusogaki whose tendency stems from parental abandonment issues
>>
>>106555308
>Character references/remembers something that happened 50+ messages ago

You just don't get it
>>
>>106555261
God tier, one Generation ahead if you are a street shitter. Worse/situational than all other current gen models when you have actual taste and good character defs and presets. Down to garbage if you also don't have a free source.
>>
>>106555329
You don’t need an extended RP session
Get a nice concentrated slice and then fuck off
>>
>>106554030
I have so many retards blacklisted on OR it's not even funny.
>>
>>106555342
>make {{char}} remember x
There
Now fuck off
>>
>>106555035
Noass is unrelated to prompt processing. Prompt squashing is just a side effect for it.
>>
>>106553873
>using illustrious
>SD 1.5 resolutions
Why not just to 832x1216?
>>
>>106555348
Say a better model than it, provide your preset and a log. I'm waiting
>>
>>106555366
just admit you use chatbots as wank material generator you twat
>>
https://voca.ro/1k7eZkhiNhjS
>>
LLMs will only get worse for RP
I feel like we’ve peaked bros
>>
>>106554146
Yeah I remember this shit with old deepseek but nu-deepseek 0528 fixed this bullshit at least.
>>
jesus sonoma is like being back in the opus 3 days
>>
>>106554271
>>106555042
>why does a machine trained to pick the most likely word from a given context pick the same couple words every tiem??
Prompt better or change your settings.
>>
>>106555389
Yes and no
I don’t do long ERPs either
After 30-40 messages it’s pointless
Quality degrades and/or freshness vanishes vs. statring a new chat/exploring a new scenario
It’s just a dopamine per reply competition
>>
>Call opus 4.1 shit
>Disappears when someone ask to say a better model
Funny how this works
>>
File: 1754513762314112.jpg (1.92 MB, 1694x2368)
1.92 MB
1.92 MB JPG
Does OpenRouter support prefill?
>>
>>106555426
GEMini 2.5
Deepchongus R1 0528
Geppetto 5
>>
>>106555436
>will make your wife cuck you and kill herself in front of you
>your wife will become a snark redditor teenager
>openai
Any other question?
>>
>>106555443
Yes why does Opus write like a literal redditor every time
>>
>>106555064
Idk. The problem I ran into with databank/vector summary was that it dumps the *whole* message when it finds something relevant. That's just bloat. What you need a short, handwritten 100 token explanation of char's fetish, not the entire 700 token message where she explained it. All this fancy stuff never gets around the fundamentals.
>>
>>106555474
>don't have a good enough preset
>card and greetings aren't well written
>has 30k+ tokens in chat messages
Idk anon, I wonder why opus do this
>>
>>106555398
Trust in Gemini 3
>>
>>106555443
>Gets cucked by Gemini

Skill issue
>>
>>106555515
>Can't use opus 4.1

Skill issue
>>
>>106555515
this, I murder and rape my way through when the cards are non-compliant (99% of the time btw)
>>
>>106555478
Then… @cohee, stDev, stScriptmakie, extensionmakie, javascriptmakie… please do the needful to transform the pulled vector chunks into concise slim itty bitty petite text. But i think there's already feature for that:
>Summarize chat messages for vector generation

Anyway, Try this, then:
https://github.com/qvink/SillyTavern-MessageSummarize
>You can then optionally remove the original messages associated with those summaries from your context to free up space by selecting Remove Messages After Threshold
It the closest one to my dynamic chatHistory wet dream.
>>
>>106555523
See >>106554994
As an avid Opus 4.1 user, I find uncucked Gemini refreshing. Yes, it is a little bit more dumb, but writing style is much better
I don't use character cards or download cards at all, I only use self made scenario cards
>>
>>106555582
I unironically am willing to try it, but you have to give the preset you're using. All the attempts I had at gemini resulted in endless yap, negative bias and not moving the story forward
>>
>>106555383
>>106555426
I actually was going through my backups to see if I had any logs on hand, but no, sorry.
The Problem with Opus (all versions) is that it is barely controllable to move out of its averages. It always degrades its prose and characters. 10k context in and you are no longer chatting with your quirky waifu, but generic archetype X. Same goes for writing styles.
The average prose Opus has is very high, which is why it is so good for bare bone presets, simple scenarios and short cooms. Unusable for anything else. Pretty much all other models do it better, and sonnet 3.7/4 share their top spot for best model with individual strengths and weaknesses.
>>
>>106555692
Honestly, you were pretty much spot on. But the problem I see, at least with our current models, is that none of them are actually good, in this case opus 4.1 is the least worse in my opinion. I admit, I was baiting before, but being serious now, no other model actually does it for me. Gemini may be the closest, but I can't get over the negative bias and I can't make it move things forward. All the other chink models I find pretty lacking both in smarts and personality. And Sonnet 3.7 feels like opus 4.1-lite, which isn't a bad thing, but all opus issues become apparent quicker with it.
>>
i just want blacktooth...
>>
>>106555427
Yes, but specific providers sometimes don't
>>
>>106555427
https://rentry.org/or-prefill
>>
File: chatlog(3).jpg (960 KB, 2154x1498)
960 KB
960 KB JPG
>>106555764
It also varies a lot on use case and what you find more important to you. Personally, I got traumatized in the opus 3 days when I had a long rp going and suddenly got whammed with a dialogue that was so totally out of character it ripped me out... and then I started back reading to actually see the slow decline.

I found an example of what 4.1 does with a minimal situation prompt and the instruction to write 'incredibly cliché russian' vs what 3.7 does. Both are bad, again, barebone prompt, but 3.7 at least tries to follow the requirement and Opus just doesn't give a shit.
>>
>>106555885
>I found the /vg/ thread to have more knowledgeable people if you need help setting up silly or such. /g/ just tends to keep up with news a bit better but lack experience. It's basically the difference between people that do and people that repost news.
>>
>>106555975
i'm just gonna type "OP is retarded" and move on
>>
TOTAL /VG/AICG/ VICTORY
>>
>>106556114
what'd they win?
>>
>>106556161
namefag worshipping and gossip
>>
>>106556161
Best Chatbot Thread Award
>>
>>106556205
>>106556205
>>106556205
>>106556205
>>
desu
>>
boku
>>
File: 99948422_p9.jpg (1.31 MB, 1920x1920)
1.31 MB
1.31 MB JPG
>>
dezzzu
>>
bozzzu
>>
File: bokudesbokusleep.png (1.18 MB, 1956x1732)
1.18 MB
1.18 MB PNG
bozzzudezzz
>>
File: boku_peace.png (131 KB, 264x575)
131 KB
131 KB PNG



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.