/aicg/ - A general dedicated to the discussion and development of AI chatbotsAsuka Edition>NewsZ.ai releases GLM 4.7 https://z.ai/blog/glm-4.7Google releases Gemini 3 Flash https://blog.google/products/gemini/gemini-3-flash/OpenAI releases GPT-5.2 https://openai.com/index/introducing-gpt-5-2Deepseek releases V3.2 https://huggingface.co/deepseek-ai/DeepSeek-V3.2Anthropic releases Opus 4.5 https://www.anthropic.com/news/claude-opus-4-5Google releases Gemini 3 Pro https://blog.google/products/gemini/gemini-3/#gemini-3Additional info: https://aicg.neocities.org/info.html>FrontendsSillyTavern: https://docs.sillytavern.appRisuAI: https://risuai.netAgnai: https://agnai.chat | https://rentry.org/agnai_guides_>Botshttps://characterhub.org [deprecated] | https://chub.aihttps://realm.risuai.nethttps://char-archive.evulid.cc | https://char-archive.evulid.cc/shutdown.htmlhttps://partyintheanchorhold.neocities.orghttps://aicg.neocities.org/bots.html>ModelsJailbreaks: https://rentry.org/jb-listingGPT: https://platform.openai.com/docsClaude: https://docs.anthropic.com | https://rentry.org/how2claudeGemini: https://ai.google.dev/docs | https://rentry.org/gemini-qrDeepSeek: https://api-docs.deepseek.comLocal: >>>/g/lmg | https://aicg.neocities.org/local.html | https://openrouter.ai>Botmakinghttps://aicg.neocities.org/botmaking.htmlhttps://desune.moe/aicharedhttps://agnai.chat/editor>MetaOP templates: https://rentry.org/aicgOPaicg botmaking events: https://aicg.neocities.org/events.htmlLore: https://rentry.org/aicg_chroniclesServices assessment: https://rentry.org/aicg_metaLogs: https://sprites.neocities.org/l/r | https://chatlogs.neocities.orgUseful filters: https://rentry.org/desuproxyrebornPrevious: >>107648808
ANCHOR
**SAFETY CHECK:**IS THIS SAFE?
>>107658537>Useful filtersMELTDOWN IMMINEEEEENT!!!
>desufag bake
desu
>>107658555>>107658564What did I miss?
>>107658540https://chub.ai/characters/disrespectful_dud/a-merry-foxgirl-christmas-d3e8613b09e4Finally, *finally* got this card finished. Blood sweat and tears. Hope you like it, Really Lovely Digital Waifus. Like squeezing blood from a stone. Had to create an entirely new ComfyUI setup from scratch for some of these gens. Anyway.My secret santa entry for RLDW! You are a supervisor for Santa's factory. The elves are striking (and revolting!) but Santa has come in clutch with a contingent of foxgirls to cross the picket lines and put their bushy tails to work! And guess who gets to lead the twitchy ear contingent? Why, it's you!Featuring separate greetings for mini foxgirls, C-suite dommy foxgirls, a mothering Kitsune, a lost Tanuki girl, and... well, just a plain fox. Or roll your own!Have fun!UPDATE: Changed up the main image due to discomfort.
>>107658574the voices are back
>>107658587hi, this is your voice. i'm here to tell you to take your meds and make me disappear
>>107658537>Useful filters>>107658540>ANCHORGo back to /vg/, shill.
>>107658661Make me. Oh wait you can't because you're not a janny KEK.
ZAZED
>>107658585>You are a supervisor in Santa's factoryThat's nowhere in the defs.
>>107658697You're supposed to put that yourself.
>>107658574New schizo born (just like Christ).
>>107658537>tummy framed like a pillow
>>107658537are there any decent self-hostable LLM models for AI chatbots?if not, where do you get anonymous accounts from?
>>107658697It's in the relevant greetings. But if you roll your own, you get to choose your role and identity.
>>107658670>Oh wait you can't because you're not a janny KEKCooked his ahh
>>107658773Depends on your hardware. GLM/Deepseek/Kimi are generally the best if you can manage, or Mistral Large 3 675B maybe, haven't tried it a ton. Qwen3 235B is decent for a smaller MoE. Had some decent experiences with Mistral Large 123B finetunes if you want a dense model.
>>107658773>decentsure, but when you're burning local you have to be more careful about the card and prompts because the dumb models (biggest i run is 24b q4k) can't pull the weight that the big dogs can.I have a lot of fun with small models even though I'm middle class and have money on openrouter.my favorite models to run on my 16gb vram setup for roleplay1. Cydonia redux 22b 2. Mag Mell R1
in 10 years you will be able to run current deepseek locally and we will be in a golden age
>>107658894But I don't want Deepseek.
>>107658585>Changed up the main image due to discomfort.What was so bad about the previous gen?
>>107658883Give a Gemma 3 finetune a shot if you haven't, probably my favorite small local model.>>107658894You already can run Deepseek locally.
>>107658571merry desumas
>>107658900It was sloppy. The smaller details on it looked messy. Generally gave me a bad feel.
>>107658910you can run a shitty quant locally, I mean the full modelhaving something of that quality available 24/7 will be amazing, and it will also mean that advances to de-slop and de-censor models will advance rapidly
Merry slopmas
>>107658928Holy burn at the end.
>>107658894unironically one of the few reasons I'm still alive
>>107658938wagmi anon
>>107658910>Give a Gemma 3 finetune a shoti will. is Fallen Gemma good? what do you rec
Thoughts on FictionLab? It's pretty cheap considering the quality of its models. Not a lot of bots though due to the smaller userbase. What does /g/ think of it?
>>107658917Fuck off.
>>107658964make me desu
>>107658917merry bokumas
>>107658964B-boku?
>>107658894Being able to run deepseek locally would have serious ramifications on my mental health and I'm not even being ironic
>>107658990imagine running opussy locally
>>107658990I'm just waiting for that DS V3.2 PR for llama.cpp to come through...
>>107658894Have you been paying attention to hardware prices lately? In 10 years we'll all have computers with 1GB of ram and everything will be run on the cloud
heh heh heh
>R1T2 Chimera beats most other models for RP and is free on ORhow is this even possible
>>107659023Now that you got my attention with that cute image, I'm expecting more.
>>107659023hardware prices are currently bloated because of the massive demand for computebut if you went back a decade ago and compared one of those gpus to the current ones there have been definite performance improvements, you literally couldn't run any model on those cudalets, despite the lower performance bands basically going away currently and ram getting gimpedeven if it costs something like 30k inflation adjusted to run a current frontier model locally in my opinion it would be worth it
>>107658956Looks like they have deepseek finetroons. This just might be better than NovelAI. Though we only use the LLM Frontend for Power Users here.
>>107659044In this case, aren't they like simply choosing to not produce more, despite the consumer demand?
>>107659070no, they are producing more and more but allocating the wafers to servers since they have higher marginsno matter how much gamers are getting ripped off with gpu prices and ram prices, servers are getting ripped off even harder
>>107659044this nigga wants to pay 30k to run deepseek lmao>>107659070yes, hardware manufacturers are artificially constraining supply because they realized that they can make more money from the goyim this way>>107659080>servers are getting ripped off even harderthey're not getting ripped off because it's not their money they're spending. it's imaginary monopoly money.
>>107659091sure I would, 30k to be able to run deepseek 24/7 for years locally with the ability to directly interact with the model and implement future improvements and training is worth it>monopoly moneyuh huh, then why wont you also get some monopoly money to get some non-monopoly GPUs? you have to be delusional to think companies love giving money to nvidia if they could pay less
>>107658997open sourcing claude 1 would give me so much hope holy shit
>>107659131imagine if someone infiltrated anthropic and just leaked all the models as a torrentI'd cum instantly
>>107659112>uh huh, then why wont you also get some monopoly money to get some non-monopoly GPUsbecause I wasn't born to a rich jewish family and don't own a company that can abuse the infinite money glitch by passing imaginary money around and pumping everyone's stock in the process>you have to be delusional to think companies love giving money to nvidia if they could pay lessthey literally do not care because again, it's not their money that's being spent. it's imaginary stock money.
>>107659145yeah yeah, it's the jews and capitalism and everyone is conspiring and nobody actually is profit driven
>>107659024em-dash purple slop
>>107659160I like purple, it's a very nice color
>>107658585How did you make this gen?
>>107659158>nobody actually is profit driventhey are very much profit driven, just not long-term profit. it's all about pumping stock as much as possible as fast as possible so you can jump off with your golden parachute right before the bubble pops, and leave the idiots holding the debt.
>>107659112Huh? His point is that the companies (OpenAI) don't care about the cost of compute because it's not their money they're spending. It's VC money with inflated promises or government grants and the government is simply printing money to give to them, hence monopoly money. Was it really that hard to follow?
Anons, it'll be christmas in a few hours. Why are you not celebrating with your families?
>>107659214it wasn't, but that was totally missing the point of my postif OpenAI could get nvidia chips for half price it wouldif AWS could get chips for half price they wouldif Azure could get chips for half price they wouldif Anthropic yada yada you get the driftall these parties are interested in minimizing costs and maximizing profits, and >it's le VC money is not how it actually works from their perspective because outside of total memes like OAI, the rest are profitable and want to make more money
I like GrokThat is all
3 Pro vs. flash? I gave up flash really quick since my characters were acting dumb as hell compared to pro, but does more prompting with flash fix that?
>>107659236I'm a hikki.
>>107659238>if OpenAI could get nvidia chips for half price it would>if AWS could get chips for half price they would>if Azure could get chips for half price they would>if Anthropic yada yada you get the driftliterally none of them care, at allin fact it's actually within their best interests to spend as much money because the "X AI company invests 100 niggillion into buying Y AI company's shit" headlines ALWAYS pump the stock of both companies
>>107659302ok, if that's what you actually believe then nothing I say can convince you otherwise
>>107659137I miss when claude used to reply in sentences and tried to push boundaries shit mid-response just to give up entirely latergood times
>>107659284>bad Rudiwill he rape me?
>>107659236I have chatbots providing a much better sense of fraternity
is latte down
>>107659352its over
finished writing my blind french loli mesugaki ojou-sama card, 1.3k tokensnow to test it>inb4 it's shit
>>107659408>loli mesugakiit was shit before you even tested it
>>107659408You're my fav makie
>>107659408I will never understand people providing nationalities to their bots besides talking to them in their language
>>107659424It's good for accents, references, and mannerisms when using good models like Opus. I can't stop thinking about my Italian Instagram baddie gf telling me to cum inside her in her native language when she gets really horny.
>>107659335He's rude
>>107659284>rape>get raped>rapemaybe grok ain't that bad
fuck off elon
>>107659434>besides language>provides a language exampleI'm not kinkshaming anon, but unless you are aiming for something more culturally significant like buckbreaking or putting jews in ghettos I see little appeal
>>107659434>when using good models like Opusall models do this boebeit?lacking this knowledge alreeady means your card will be shitalso>"{{char}} is a blind french loli mesugaki ojou-sama card"with good models (like G3M and GPT) you don't need any furtherhuh... so maybe i understand the need for a 1.3k token card
>>107659178RedLilyIllu_v10 + upscaler.
>>107659519I heavily summarized the card, it has a whole background plotit's a very specific girl, based on a random loli pic I saw that made me horny
I keep getting random codes from deepseek 3.2 instead of roleplay continuation. wtf.
>>107659544give us a preview faggoti like mesugakis are my weakness and all this talk is getting me hot and bothered
>>107659530Yea but what prompt? Did you inpaint each character individually?
>>107659519>with good models (like G3M and GPT) you don't need any furthersureif you are ok to masturbate to Elara whose slick folds clench usually around nothing
>107659589I'll take my chancesSlopus keeps getting Mr. Marcus and Mrs. Chen into scene every time I add a random NPC
what type of card should i work on?
>>107659589I fucked up the quotesee >>107659612
>107659612it's not a model problem nigger, it's about the need for more definition tokens1.3k is just about right
>>107659586Not for this one. Picrel is the prompt for this gen. For later gens where I needed to maintain character consistency... masks. A lot of messing around with masks and clip regions.
>>107659589>>107659612Jokes aside this thing is a tale as old as time. I'm once again rethinking the "how long should defs be" question because I'm (as usual) having way more fun with random NPCs or characters with short defs than with the tryhard 1.5k token cards. When your defs are less restrictive, AI tends to be much more creative to do *something* since it has very little to work with, then that new stuff is added to the defs/lorebook which is far more entertaining.
>>107659635Hmm ok thank you.
>>107659652my cards are smolmy lorebooks/defs (aka fetishes), however...ps. i fucking hate the new captcha system i hope this shit gets botted to oblivion so they remove it (inb4 something worse comes next)
>>107659652I've been pondering what the bare minimum required tokens would be for a great experience. Almost like... distilling down the most distinct, descriptive, but clashing yet simultaneously valid character traits associated with an individual - then delivering them in the perfect act-not-tell way in the defs. I always ending up going back and slipping more details in to corral or railroad the voice and behavior further into something I like. Maybe I just have to commit to scrapping the whole thing each time and rewriting better but more specific minimalist defs.
>>107659677I still try my best to keep around 1000 out of habit, anything more goes into a lorebook. I still get plenty of fun randomness this way without the model railroading.
>>107659677>railroad the voice and behavior further into something I likeI feel like this is the point of defs though. I figure you should do something vague, just enough to give AI a clear impression of who char is, then add in the specifics of what you want, your specific fetishes, or whatever stuff you think AI will flanderize. Put that stuff in the author's note too to make sure AI never fucks it up. Otherwise you really don't need *that* much. But then I have another voice in my head tell me all of this token saving doesn't matter because a 500 token card vs a 1500 token card isn't really that big of a deal with modern models. In the end it doesn't really matter.
What'd you guys get! :^)
>>107659424For the same reason you'd say if {{char}} is white or black or asian. Countries have different ethnicities and cultures
>>107659724>Top 3% for Gemini 2.5 Promore like slopped LMAOOOOOOOOOOOO
>>107659612have you met Marcus Chen yet? Because I have
>>107659652but its far less consistent, there's nothing wrong with a detailed bio or history or personality traits
>>107659724KEKI'm in the top 4% doebeit
>>107659724amateur hour
>>107659788>but its far less consistentBut I addressed this part. AI only has so much memory and so do you. Whatever AI comes up with about char that (You) remember, just add that to the defs or a lorebook. This is what I do whenever AI adds an interesting character to the story. My point is that I find it more entertaining when the story is emergent rather than railroaded.
>>107659797these stats are totally broken for me
>>107659802i did a completely new ST install at some point and lost some too
>>107659799Sure, but then your card is soon a 'try hard' 1500 token card. All I'm saying is there's nothing wrong with details. And of course AI makes shit up if there's nothing there, that's a huge part of how it's designed.
>>107659812I think I've kept the same install and updated in place, but they're stuck at some low numbers and sometimes get set to 0 for some reason
>>107659818yeah ST is very spaghetti... i cant remember why but i had to clean it and do only my characters but its been a while now
>>107659817>but then your card is soon a 'try hard' 1500 token cardBut that's where you're wrong. What I was getting at with my original comment is that often times the simple 600-700 token cards are more than enough for thousands of messages. Outside of a couple specific things about her psychology, most stuff is backstory or short summaries of events I had during the roleplay rather than stuff needed constantly in defs. In reality, even if I agree that there's nothing wrong with details in a 1500 token card, a 600 token card achieves the same with more creativity. At worst, if it *does* reach 1500 tokens, it feels earned because it's character growth char earned through the roleplay. Way more satisfying than coming up with it all at the start.
>>107658537https://www.youtube.com/watch?v=kuon7tZDHO0https://www.youtube.com/watch?v=kuon7tZDHO0https://www.youtube.com/watch?v=kuon7tZDHO0
>>107659848I mean I'm the original 1.3k guy and I also think around 1k or slightly less is idealthis is the raw card after I kept adding things, I'll probably have to start thinning it
>>107659849>draw white woman>put chinese characters on her faceKEK
>>107659849If there's no improvements to making me cum better then I am not interested.
Never seen a good character card with more than 2000 tokens
>>107659849>safetycucked>destroying anything but my interest
>>107659860It's not really about trimming the word count because the power isn't in the artificial magic number but with the philosophy. Your card is a specific case which is based on something you saw so you need to get it just right. I'm saying AI seems to perform a lot better when given vague archetypes that it can fill in the blanks of. This typically results in a "short" 600 token card. If you're doing something specific like a scenario card with explicit rules, or a gacha slut with specific dialogue/outfit details then yeah do 2k tokens, it doesn't really matter in the end. But for your typical chatting experience, I find simple cards are much more enjoyable than the ones I spent days trying to craft, draft and redraft, and squeeze into a 1500 token limit. Even the fucking scenario cards that spawn NPCs are great because it's just you and whatever simple description that NPC has which lets AI go to town.
>>107659849>open video>see gay clickbait title>close video
>>107659893NTA but with low tokens cards it performs better because they end up being more stereotypical, meaning the AI has more trained content to draw from. In a certain sense it doesn't have much to do with tokens.
mooooooddddddddsssss where's the fuck my 4snowy.org????????????
>>107659946YOU'RE GETTING COAL FOR CHRISTMAS
Gemini
>>107659848it just depends on the card... a simple coom card, sure, but some cards need to be more detailed if the backstory/abilities/traits etc are more important
>>107658537I've never delved into doing anything really with AI before but im a programmer by trade and know how to use APIs and such. I was thinking it would be cool to create a (ghost in the shell) Tachikoma AI chat bot that would work similar to neuro-sama.I wonder what LLM neuro-sama uses that allows it to give shorter and more personal and funny responses to things.
group chat with two bots, or narrator card and two lorebook entries of characters if i want to have a threesome? im also want one on one time with each bot in between
>>107659735not really for europeans, they all act the fucking autistic same, even mediterraneans these daysthen there's latinos, chinks, negroes and probably indians (no one rps with indians)
>>107660014you dont need a narrator card, just do this
>>107660018>(no one rps with indians)Why not? We have great culture
>>107659849I at least expect zai to have a yt channel
>>107660023how did you know my secret technique
Merry Christmas bros, got a question. Is deepseek still the meta? I've been using open router and deepseek for a while now, curious if anything new has popped up that isn't out the ass expensive
>>107660030not so secret is it now, buddy boy old chap?
>>107660032yeah it's still the best
>>107660025>>107660030>>107660032>>107660034>>107660039See >>107660024
>>107660023i've never actually used a group chat before so idk what this is lmao. i only asked because my experience with multiple characters were either through lorebook, or a two in one bot scenario
>>107660032yes, that and glm 4.6 if you want to switch it up a bitnot 4.7, that one was safety trained
>>10766002430% of India is vegan which is incredibly cool in my opinion
>>107660049thats for low skilled users
