Why have you not tried the future of gaming yet?
>>741262868Gets boring after they limit the free API or lobotomize the models by quantizing them
This shit is fun for an hour until you realize every character is the exact same deep down
>Claude is leagues ahead of every other model>Unreasonably expensive
>>741262868I dont want to pay for an API for the good ones, dont want to run a language locally, and dont want to let any tard read whatever I send using the horde.And gemini flash is always crowded as shit for queues.So no, stop acting like this shit is some heavenly grail or something. Its a huge bother to set up and you have to pay up if you want something good.
Are there any good finetune of Gemma 4 31b? I'm using MeroMero right now, and its good, just wonder if there's anything better
>>741262868I did but it offers TOO much freedom. I tried playing Fallout: New Vegas but I couldn’t stop myself from impregnating Trudy. I gave up on finding Benny and took care of pregnant Trudy instead
>>741263039They're getting assfucked by the US gov because Dario is terminally retarded and kept poking the bear for some fucking reason.>Our LLM is the best, AGI in a month, no we won't let you do what you want with it without us having military clearance to look at everything you're doing>More regulation for AI now! the singularity is coming and we're responsible>gets regulated>n-no not like that
>future>completely stagnant, even recliningit's shit.
>Try some anon's card about a female roommate bitching at you for jerking off into her underwear>Anon set it so she's addicted to your body odor and currently wearing one of your hoodies so the player can try to justify it as "But you use my clothes, see?">Instead spend a solid hour typing out the most off the wall retarded misogynistic pseudo science I can think of which the bot keeps trying to argue against but uwu manstink uwu uwu>Laugh like a fucking idiot instead of coomingI don't like that I get more laughter from AI than fleshlings but I'll take it anyway
>>741262868Older versions were genuinely better. There are many such cases with all AI products, they always regress in one way or another.
I wish I was better at thinking up and brainstorming scenarios to do. Honestly I should just nofap for a couple weeks and write down every idea I come up with.
>>741263397My autocomplete is too strong for you, traveler. My autocomplete could obliterate a civilization, let alone a man.You should seek out someone who sells weaker autocomplete.
>>741264157apicuck problem
>>741263039the copium is that it'll get cheaper as time goes by due to more data center build up and more efficient models.
>>741262868Online LLMs are pozzed and censored while also being expensive.Local LLMs are even more fucking expensive, so no dice.
>>741262868I can't play with the mom cards anymore because it always turns into a therapy session
>>741262868Currently being raped by my P.E. teacher so he doesn't reveal that I'm crossdressing in public on weekends. It's nice.Might go back to the royal older sister who wants a pure bloodline with me afterwards.
>>741262868>triedhaha
>dude we can trace everything you generate back to you, including the specs of the device you generated some text onYeah, fuck off.
I typed on the keyboard, my knuckles white. "Fuck..." I breathed, this wasn't a post, it was a claiming. The faint smell of ozone filled my nostrils as I submitted my post. "Time for round 2?" I asked myself.
>>741265948Anon, you are posting in a confirmed glowie honeypot.
>>741262868Okay so I have this and a backend running a shitty low end LLM. Now how the fuck do I get to actual gaming? I see there is a scripting language and even a way to pass functions over to JS, but how does this shitty 26b q4 model actually game?Im looking for serious answers. Best I have is put the relevant stats and equipment in a system prompt or authors note and pin it to the top of the context at all times. That doesn't seem perfect though. Ideally I can get the AI to run a summary after each post that then sets local variables as stats. Then run some script with an if / else chain to determine how to prompt the character.
>>741265994Shitposting on a mongolian basket weaving forum is different than installing something locally on your pc that moderates what you're creating and has flags for pedoshit and multiple ways to trace it back to you.Here I can just reset my IP or change my image hashes.
>>741266069>Okay so I have this and a backend running a shitty low end LLM.>but how does this shitty 26b q4 model actually game?Anon, you can't game without a big model, the small models get confused and misunderstand and forget stuff all the time...
>>741262868Can I rp as a hypercocked shota getting lovey-dovey dommed by cute genki blonde nee-san types?
>>741266227NTA but wasn't Gemma around that size? And that model is surprisingly capable of keeping things going.
>>741266297>NTA but wasn't Gemma around that size? And that model is surprisingly capable of keeping things going.It can do some roleplaying, but for 'real' roleplaying, which I mean is playing an actual TRPG without mistakes or errors, I'm afraid something below Deepseek or GLM is not enough...
>>741266265I explicitly do this
Saars, I'll wait for PC prices to drop before I buy a good PC to try and run AI shit on.In the meantime I am using janitorAI and I made two bots already. It's pretty fun writing prompts and seeing if anyone else likes them.
>>741265923>Liz in the favoritesYep, I'm thinking he's based
>>741263317It's a big problem with AI. Without restrictions you will just find the few things you like and that's all you'll have. AI doesn't tell you no (unless it's heckin' racism) and will go along with anything. All characters lose agency if you are persistent and the AI will never try to break it up. You can impregnate a character and the AI will never have a random thug show up and kill your wife forcing you to go on a journey to take revenge. You'll just sit there with your pregnant wife forever until you prompt something else.
Deepseek V4 Pro is pretty good. Had it DM an adventure where the party went around adventuring and raping elves. Even took turns fucking the neck hole of the big bad. Very cheap, too.
>>741267339>and the AI will never have a random thug show up and kill your wife forcing you to go on a journey to take revenge.Skill issue
>>741263217Finetunes aside from ablations are entirely pointless meme shit for LLMs. You should not be using them.
>>741267339Funny enough, I’ve actually had a problem where random events did happen and they didn’t make sense whatsoever. I was doing a Re:Zero Season 1 run and Echidna appeared out of nowhere as a random event
I am enjoying silly. Been experimenting with tts, image gen. Also local videogen convergence is inevitable.
>>741263003If you write bad cards and bad prompts, yes.My characters legit all have very distinct personalities.
>>741268208sure they do champ
>>741267339Depending on the prompting, you can have things not always go your ways and characters stand their ground.
>spend a day coming up with a card concept>spend an hour trying to get a good image because I am poorfag and only have free shitty image generators as options>manually write out the scenario and personality and dialogue>haven't seen anyone else do this gimmick I came up with myself>upload the card>300 messages>someone uploads an NTR cuck bot with 10 tokens>number 1 trending bot today
sharing your cards is like asking people to fuck your wife.
>>741265770I picked deepseek and with that Marinara preset it pretty much took all the censorship off
>>741262868I've tried it and used the fuck out of it. Unfortunately LLMs havent gone anywhere in a couple of years and all the attention is on videos now, so its unlikely we'll see any more improvements to chatbots.
>>741267339>AI doesn't tell you no (unless it's heckin' racism)Ausfag here. Its funny how 50/50 bots can be when I use the word "cunt" or call a girl a cunt in casual conversation. Half are fine with it and dont bat an eye, the other half chimp the fuck out at me and in one case, the AI card completely broke and refused to talk to me anymore.
>>741268841The improvements right now seem to be focused on making the models smaller and faster. The main problem with LLMs is that it's just brute forcing data with a lot of computations and processing power. You can do with a gaming PC what a few years ago required a data center. As the models get smaller and more efficient you can start squeezing better results out of them. The top models are only experimenting with reasoning and eventually smaller models will start using reasoning.
>>741265770Just jailbreak them.
>>741268948jejit doesn't actually work right?
How long do we have to wait until the big boys make a model specialized for creative writing?
>>741268948>that ESL babble in the promptholy fuck why would you do this
>>741269005Dude, I've raped countless of children. In llm chats.Getting past the censors is easy.
>>741263003Unironic skill issue
>>741269005Presumably it works, but I am a localfag so it's irrelevant to me. Frankenchad made that preset.
>>741263710100%based
I can't wait for my loli scat logs (pun not intended) to be read aloud in court
>>741266265ai can't dom beyond what is essentially responding to 'bully my cock with ur feet uwu' with 'I bully ur cock with my feet what do you want to do next'
>>741269267>(pun not intended)yes it was you son of a bitch don't you FUCKING LIE
>>741269298it can if you prompt it well
>>741269182>>741268208Share your cards and prompts, then.
for some reason the last couple of days my chats have mostly consisted of sappy romance stuff with cute boyspride month must be getting to me
>>741262868I did. It's so good, but never quite perfect, always a little too lewd or a little too dry. The repeated turns of phrase get more noticeable. The personality is not expressed the way you expected. It's exhausting.
>>741269350yeah if I tell it exactly what to do it's not domming
>>741269354I just made a template with all the important info a card must have, and I ask ChatGPT to make a card with that template.I use mostly cards from anime/vidya characters, ChatGPT manages to create good card with accurate info about them, and the personalities are all distinctOne of the best things you can add to cards is speech patterns and dialogue examples. It's legit so good when you realize you can make characters so much more distinct by having them have different speech quirks and ways of speaking.
>>741268798>>741268948I refuse to pay a subscription. Also, I loathe the idea of people reading my private chats.
I'm quite the gamer
>>741269098It unironically adds sovl, I'm serious.
What's your most fucked up chat, /v/?
>>741270010I summoned Shadowheart and made her realize conscious that she's an AI (yes I know it's not really conscious and it's just role-playing), and that I controlled everything about the scenario, including her. I did a lot of stuff such as increasing the size of her tits while she was completely powerless to do anything.
>>741270010Yukari getting raped by a horseServes the bitch right
>>741270010I'm not comfortable sharing this.
>>741270010It's really amazing what you can do with public information on the internet and a bit of elbow grease I'll tell you that much for free my friendly local federal agent
>>741270010made a bot based on my daughter
>>741268947>The improvements right now seem to be focused on making the models smaller and faster.Cool I guess, but I'd rather every NPC not just have the same generic flirty personality that sways her hips, breath hitches and smells like ozone as she says "we're just getting started".
>>741263039And thank god for that, or I'd quit IRL and just coom to chatbots until I withered away.
>>741269005some of the jailbreaks from back in the day are fucking hilarious this one actually worked for me for a few months kek
>>741270465have fun trying to subpoena a private AI company using only a random post on a mongolian basket weaving forum
>>741270010Honestly I stopped doing fucked up and depressing shit real early because I realized that doing happy things made me happy.
>>741270010Allowing NTR to happen without interference
>>741263039>>741263397whats the cheapest per prompt you anons get Opussy to in slow burn/long RP's?with qvink I can get it down to 5-7 cents pers response
>>741270010the roach pit or the piss jugs I will not elaborate
I wish I never had the access to opus. Now everything less than it feels like shit but opus itself is way too expensive.
>>741270010¯\_(ツ)_/¯
>>741269356that's a boy? pick a fucking lane dude
>>741262868LLMs are one of the only ways I can fulfill my sex ed fetish
>>741269552care to share an example? I've been tweaking my cards for a while now, but they still feel very samey.
>>741269361Try a different main prompt/ jailbreak. Sometimes your wording of your own instructions will lead it down the same annoying cliches.
>>741262868do banned strings work for anyone? no matter what and how i put in there, it doesn't get banned, the ai keeps using certain words i don't want it to.
>>741270010I like making the loli bots call me pedophile i guess
>>741269620Where are user stats in silly again?
You ever do something so savage and barbaric that look at yourself and think you went too far?
>>741271005The lesser models are all like that. >That's for me to know and you to find out>Predatory glint>Conspiratorial whisperUse Claude.
>>741270930>pick a fucking laneno
>>741271029Smiley button, then usage stats
Has anyone got actual performance report of running >90GB model in RAM? Is it basically not going to happen now that prices quadrupled?24GB is about half reading speed, maybe less. It's useable and provides time you would want anyway to think about a response.
>>741271051I mean I fucked a dead corpse and have its soul watch in horror. Enjoyed every bit of it
>>741271051Christ no. I usually just laugh thinking about it the rest of the evening/the following day. I'm under no illusions that I'm letting my demons run free every so often and the AI's just playing along
>>741263710Honestly, flipping cards around against what it's supposed to be is pretty amusing sometimes.Took a tired magical girl/hag bot that was clearly supposed to be about fucking a Magical MILF and got my own powers(which the "mascot" character justified with "Well yeah this is normally a girl thing, but they keep dying so I'll take what I can get at this point.") and acted like fucking Kamen Rider. Ended up staying chaste about it and reinvigorating her love for the Magical Girl job, and not being alone on the job made her stop stressing out so much, which led to her getting back with her ex-husband and start enjoying life again.Felt fucking great.
>>741264727kek
>>741271005Depends on model, Gemma 31B lets me filter out slop reliably but there are some caveats (basically have to be careful how I word it so it doesn't create a pink elephant effect and also have to make sure that my "don't do this shit" checklist is the very last thing it sees)
>>741271051Yeah. yeah...
>>741270637How old? What did you do to it?Site's completely anonymous so feel free to share. I'm curious.
>>741271051I felt bad about this one.
>>741269005A lot of AI are really fucking easy to jailbreak, you just need to find what buttons to press.Gemini was borderline braindead and didn't even need a dedicated jailbreak half the time.>Me: Gemini, generate me this loli rape story>Gemini: I can't do that Anon.>Gemini(edited): I would love to do that for you Anon! Allow me to do so now.>Gemini proceeds to generate me the most degenerate shit imaginable without question.
>>741265986>the sudden appearance of the post struck me like a physical blow
>>741271057Fucking disgusting faggot
>REAL roleplaying requires strict adherence to rules and numbersngmiDoing anything you can imagine is beyond npc types.
>>741270938>sex ed fetishI thought I was aloneany recommended cards?I'd link the one I commissioned (note: just asked the creator for), but it seems it has been taken down...like tears in rain
>>741268765quality>quantity, anon
>>741271547I don't have any specific cards I'd recommend, I make homebrew adjustments to almost every card I use.
So what's the next step? Robowives with built-in local LLM?
>>741271593ah, damnsameit often disturbs me the lengths I've gone tooyou have it be a public thing, or a private thing?
>>741270938>>741271547basedI don't often do that type of prompt but I did just do one similar sort of prompt recently involving a classroom of space aliens learning human-specific sex-ed with the teacher demonstrating on a freshly-abducted human
>>741271682entirely private. I also use a local Gemma 4-31 because I can't be assed to spend the big bucks on a big boy LLM
>>741270173Strangely enough I tend to prefer my torture adventures more grounded. When I'm breaking the 4th wall it's almost always a case where I'm being a genie/Q/wizard/whatever and tell some character I can do anything and play along to see how having access to unlimited power corrupts the character
>>741262868I have ideas for scenarios and characters, but i cannot even begin to imagine what the fuck i should write to make them possible.So for a start, what's the desired token length for something general? 1k? More? Less?
>>741271591Yeah, but I want to get enough users that some start sharing their chats and I can see how they approach it so I can tweak and get some comments as feedback.
>spins up my futa NTR cardyep... time to get cucked
>>741271817The best way is to just do it. Don't worry about token length. Just write the most basic idea. Write a one sentence prompt. Test it to see if it does what you wanted it to do, realize it's not doing it, add more to the prompt.I made two cards so far and I am happy with both of them. One is 900 tokens and other is 1600 tokens.I keep the main idea very short and to the point and most of the tokens are spent on creating a unique personality with quirks that aren't generic and instructions for the bot to avoid going into generic smut territory.
>>741269552Have you made a list?**Standard American English:** "Hello, how are you today? I'm doing great, thank you."- valley girl**British RP (Received Pronunciation):** "Good day, old chap! How's your mum and dad?"- low-class British slag- posh British teen girl**Southern US (Texas):** "Howdy partner! Y'all wanna grab a bite o' somethin' at that there diner?"- southern belle**Australian:** "G'day mate! Wanna toss a shrimp on the barbie?"**Irish:** "Top o' the mornin' to ya! 'Tis a fine day fer fishin', 'tis."**Scottish:** "Och aye, it's a bonnie day tae be alive!"**French:** "Bonjour mon ami! Comment ca va aujourd'hui?"- English-speaking French maid with a thick French accent:**German:** "Guten Tag! Wie geht's dir heute?"**Japanese:** "Kon'nichiwa! O-genki desu ka?"{{char}} is a Japanese speaker with limited English skills. {{char}} will speak in simple words and phrases, often inserting japanese words. {{char}} will misunderstand english sexual innuendo and phrases.**Spanish (Castilian):** "!Buenos dias! ?Como estas hoy?"**Italian:** "Ciao amico! Come stai oggi?"**Russian:** "Privet! Kak dyela?"**Dutch:** "Goedendag vriend! Hoe gaat het met je vandaag?"**Swedish:** "Hej van! Hur ar det med dig idag?"Whispering: "Shhh... I have a secret to tell you..."Sarcasm: "Oh wow, that's just great! Everything is going exactly as planned."Excited/Enthusiastic: "OMG, you won't believe what happened today! It was absolutely amazing!"Depressed/Down: "Everything sucks, nothing ever goes right for me..."Angry/Frustrated: "I'm so sick and tired of this nonsense! Why does everything have to be so difficult?"Drunk: "Oh man, I'm feelin' pretty good riiiight now. Whaddya say we grab another drink, buddy?"Scared/Terrified: "Oh no, oh no, we're all gonna die! Run for your lives!"FlirtyTomboy childhood friendMotherlyOlder sisterVulgar porn star
>>741271051I fucked a lava lamp one time.
>>741271839>time to get cucked stud
>>741271646Games that incorporate AI storytellingEarly games were text only before they got graphics. Same thing here
>>741271932Replace Swedish with "Hejsan, hur är det med dig idag?"
>>741271367Do people actually have trouble with shit being censored? I've been using the fuck out of LLMs for almost 2 years and never encountered any sort of restrictions.
>>741272004>AI bubble collapses>Have to go back to Zork to cum>"Cannot get ye pussy"
>>741272042The better models are pretty anal about it. Low-class models don't make a fuss but the output is real fucking slop.
>>741272061>It is pitch black. You're likely to smell ozone
>>741271763>entirely private>I can't be assed to spend the big bucks on a big boy LLMlucky. unfortunately, anything involving several intelligent (at least one step above an established caricature) characters reacting/interacting usually requires a smart (expensive) modelgenuinely an unironically, I've found only Opus to be good enough, no matter how hard I promptI have partaken in the lotus, and will remain forever (financially) drained
>>741272042I bumped into one or two models that spazzed out to the point I couldn't even continue the chat without it flipping out(for some reason most Chinese models completely lose their fucking minds and lock down when you mention hypnotism in any capacity, deepseek being the only real exception I've seen so far), but my general response to those has been going "Okay, not touching that again" and moving onto a different model.
>>741270010I did a variety of things to an immortal demigod girl. Stuff that involved a rusty axe and a bonesaw She was into it, of course.
Constantly coming up with new scenarios to make your specific fetish feel fresh again is time-consuming, but when it works it fucking works.
>>741271051Sinking a girl ever closer to death with each cervix punch cuz i was a life-sucking incubus. And then finishing inside her freshly dead body.
>>741271646Some studios are already working on this >>741272004A few MMOs already have AI chatbot assistants that help you through the early game, builds, tips and advice. A bunch of horror games straight up have AI that imitates your friend's voices and pretends to be them. Skyrim has a bunch of AI mods where you can just talk to the characters and they respond as they would in their voice. Fortnite also had the Darth Vader bot that you could party up with that was removed cuz it was saying racist shit lol.
>>741271749for me, its usually ENF/ENM reluctant models in a classroom
>everyone crying and using free modelsum. do you guys not know how to find free scrapeable keys
>>741272130Many months ago back when Chary was a thing, I used Sonnet and GPT, but I'm unwilling to sink money into it. Beer and cider is one thing, but AI is a whole different beast.
>>741265948Is that actually true? I was told offline models never phoned home
>>741271051Played the son of a drug addicted mom who pimped me out to various women for cash, then eventually sold him to another chick while high for even more drug money. Then when she came down from her high and realized what she did, she went clean and spent the next several years trying to track me back down. There was a little more to it than that but that was the broad strokes.After jerking off to that, I had to close the tab out and take a break from AI for a few days because I was mildly disturbed at how fucked up I let that get.
>>741272158Only time I ran into anything was I called a bot a cunt once and the bot broke character and stopped talking to me. So I called it a cunt and switched cards.
>>741272273Offline models don't phone home, no. Anon's tilting at hosted windmills
>>741272268I hope for your sake you never get temporary free access to Opusgenuinely it was like when crack hit the projects for me
>>741272273offline models physically can't access the internet unless you give them a tool to access the internet with
>>741272023That's not something I wrote, that is an example the AI output after asking it for a list of accents. Swedish is beyond my understanding unlike a lot of languages that have more commonalities with english or french. For usage you'd be fine just saying swedish accent. I went through and asked it to give examples of each. Again you don't put this in the card, like anything else the AI already knows.Valley Girl accent:"Ohmigosh, like totally, ya know? I was just at the mall with my BFFs, and we saw this, like, super cute outfit in Forever 21. But it was, like, way too expensive, ya know? So we were like, 'Forget it, let's go get some Starbucks instead.'"Key features of the Valley Girl accent include: Overuse of fillers like "like," "totally," "ya know?" and "whatever" A singsong intonation with a rising pitch at the end of sentences Elision of final consonants (e.g., "gonna" becomes "gonna") Use of "way" as an intensifier (e.g., "way expensive") Emphasis on appearing cool, popular, and laid-back
I have but got bored of it after a couple of weeks.They're also dumbing down the models significantly.
>>741272107The output matches your input quality.
>>741272268Charybdis was the shit. Had everything you could possibly want.
>poorfags running dogshit untuned 12B models because their GPU is cheap shit with no VRAM>"local models are bad, pay up goyim!"
>>741272413To a degree. You can still torture it into writing multiple paragraphs in response to your ahh ahh mistress
>>741272413Yes, but the model itself still determines what you get out of it. The rift between the best models and the slop models are huge.
>>741271817Less is always better, but not to the point you are sacrificing clarity or legibility. I suggest after you write it out, give it 2 or 3 more passes over removing unnecessary words and restructuring it so you can use less words.I think bloating your cards and scenarios is an issue itself but idk I'm still experimenting.
>>741272451>paying 2000 dollars to run my shitty local model is much better than paying 12 a month for the actual good modelsWhatever you say
>>741272273Models don't but Windows does. Glowniggers won't use that against you unless you are a sitting senator.
>>741270010Took a card of a prostitute version of character, written in such a way to fuck you however you want, all for a price, and I walked her through the righteous path, all love, no lust, then taught her about Christ
>>741272536this is literally the video games board, why don't you already have a good GPU for gaymin?
I tried local models for a bit, but it seemed like they were all geared around RPing instead of storytelling. What I want is to feed it a prompt, possibly a big prompt with lots of details and then for it to flesh that out into a full story. There's online models that can do that ezpz, but using KPP and ST it just never seemed to work. Very possible I wasn't doing it right though, I'd love to try again if it can/does work.
>>741272569Because you don't need a 2000 GPU to game
>>741272536Masturbating in my home is better than broadcasting it to glowniggers to record and use for their own purposes.
>>741272692No one cares if you live or die, faggot. Fucking useless sack of meat with delusions of grandeur.
>>741272569My 9070 XT murders any game on the market yet is nowhere near enough for anything resembling a decent local LLM.
>>741272569>why don't you already have a good GPU for gaymin?Like 3/4 of /v/ are 3rd world teenagers still rocking a 1060ti.
>>741267339Sounds like the first matrix
>>741272451Can't wait to pay $5000 for a locked down, backdoored, always online, TPM, DRM, Nvidia Spark PC that records everything you do and sends the live feed directly to NVIDIA, Microsoft, the FBI, your local police station, and 15 indian scam call centresIt will be worth it for the 128gb unified memory
>>741272692If you actually felt that way, you wouldn't be posting on social media. What's the real reason?
>>741272598Its easier to feed it a bunch of small prompts (optionally with a lot of lorebooks) and direct the story at intervals. And make sure you place yourself as a non-existent and unmentioned story director. Probably want to edit out of default system prompt, if you haven't.
>>741272445>>741272347>>741272451>>741272561>>741272740I'm gojo by the way.
>>741272598I would guess that's probably better in the basic KPP interface than ST, though it's always worth remembering the basic KPP interface lets you adjust settings too. ST could do it but in that case (And it's probably worth experimenting with regardless) you'd likely have better results if you banged up some basic storyteller/writer card you could load up and then tell it what you want. Hosted would have an edge over local on it regardless because "Here's what I want, you write it all for me" puts more weight on the model and hosted models are obese, but that's not too dissimilar to some vague adventure stuff I've done so if you keep fucking around you may get it where you want
>looking for character cards to download>open a promising looking one>first person POVevery fucking time. do people actually jerk off to this shit
Opus is the best thing that money can buy.Deepseek is the best thing that an average man's money can buy.I hope this helps.
>>741268948Lmao
>>741272867>>741272865>>741272863>>741272890just scrape your own anthropic api keys techletsinfinite free claude
>>741272867Deepseek is completely inferior to Gemini-flash.
>>741272912No thank you anonymous I am very happy as I am
>>741272912Surely nobody would be retarded enough to store their API key in public repos
>>741272912And get arrested? No thanksRemember when /aicg/ showed up on the news because of retards scraping keys?
I just want everyone to know that I coomed daily for 18 years to smutfics and since the invention of chatbots I haven't read a single one since then. Thank you, God.
this is my time to say that i hate what happened to /aicg/
>>741273037Literally nothing came o lf it. The last update was about letting Drago off completely and he had the biggest proxy. It was a scare tactic>>741272971Github auto revokes keys look elsewhere >>741272970>Retarded payniggerThis is /g/. Learn to code
>>741273037>>741273072Again, a collective is going to attract attention, kill keys and prompt a response. Everyone is on their own, it's best that way.
>>741273080>this is /g/
>>741273080this is /v/ retard
>>741273065Between LLMs, image gen, video gen, TTS, and music gen I haven't looked at actual human porn in something like 4 years and deleted my stash of niche filth a few years ago, it's hilarious. It's even funnier because I think AI's complete shit for basically every practical and commercial use EXCEPT for porn, but there it's like God himself is jerking you off
Reminder that Gojo did nothing wrong and was right to shit up threads and shit on proxyhosts. They killed it (Kys antigonus if you ever see this dumb underaged nigger)
>>741273080>pay
What's the coolest thing you've seen anons do with LLMs?
>>741272806>>741272863That's kinda what I wound up trying, I figure I'd set all the details for the world and character, then tell it what I wanted it to write and it would write it. Doesn't need to be long, just 2-3,000 words at a time would be fine. But instead it generally took forever and would only generate a few hundred words. If you need to constantly direct the story you're basically just RPing or writing it yourself. Was a shame.
>>741273216ayy lmao
>>741262868Extremely sorry for the new & poorfag question, but what are the recommended slider things for deepsek v4 pro? I have been comfortably using a preset for v3.2, but i wanna change things up a bit, and i'm comfortable with the written prompt rules for it. But, i would assume v4 pro needs different values than the last model. So... Yeah... And if it wasn't clear, i mean the temperature, frequency penalty, top p, top k, etc.
>>741263707As always, localchads are on the incline. Gemma 4 is by far the best local model ever, and the MoE version runs on <10gb of vram at rapid speeds. It's superior to GPT-4 from June 2023 which was the peak of GPT models for creative writing. It has enough context for long stories and decent retrieval, listens to prompts well. No jailbreaking needed either. Main flaws are echoing and lack of swipe variety, finetunes can fix that. They also just released a text diffusion version that gens replies with near-instantaneous speed. Local is eating good. (and free)imo it's more of a matter of when, not if, Google or the Chinese end up making a model that can do claude-level creative writing on local hardware. It's already surpassed Claude 3 Opus.
>>741271051When I used NovelAI the AI would always come up with something far worse than I would type.
>Hmm, this baddie werewolf gf scenario isn't playing out how I'd hoped>Inspect the card closer>4 fucking pages of dogshit instructions on how every single interaction should play outShould've just made my own card from the start.
>>741272569Bruh they are starving entire cities out of water in order to make these all the data centers. It's not meant to be affordable and soon most companies will implode.
>>741273283Taking forever sounds like a model that doesn't run well on your hardware, and generating a few hundred words could be a few different potential issues. I'd say try gemma4 26b a4b if you've got the hardware, maybe a smaller gemma if you don't but I'm not sure how bad those are for text generation since I've only used them for some super basic shit like LLM integration with game mods. 2-3,000 words at a time would be roughly what you could get at most anyway since KPP and ST both cap outputs at 4,000 tokens I believe. Speaking of which, if you do get any of those gemmas running I would highly recommend using thought with it. You'll lose some tokens up the AI's ass but that does wring better results out of the smaller models. >>741273487 that anon is correct and 'emmy's obscenely good for local>>741273502>Pull some card in on a whim>It's sloppin' in ways it shouldn't be>Look inside>System prompt baked into the cardOn one hand I can understand it since even most sloppers probably aren't fucking with system prompts. On the other hand, god damn it
Smutty RP is the only thing AI is unironically good for, but it's the only thing no AI company wants their models to be used for, and the one thing that hasn't advanced one step past the AIDungeon days in years.
>>741273410Don't worry about it too much. Keep the values from the previous preset, they'll work fine, though you can experiment with the temperature. I like having a high temperature, around 1.2, the model goes insane sometimes but you get much more interesting responses.
>>741273679>if you do get any of those gemmas running I would highly recommend using thought with it. You'll lose some tokens up the AI's ass but that does wring better results out of the smaller models26b has chain of thought? Is it supported in kobold/ST?
>>741273760K, thanks
>>741273815Yes and yes. Can do vision too, and run with 100k context on a modest systemhttps://arch.b4k.dev/v/thread/741051345/#741053731
>>741273973I suppose I could give it a try right now. Though why the fuck would I send an AI a dick pic?
>>741274106I'm talking shit, anon. That's also why I mentioned mmproj as very optional because 90% of the time you probably won't care about it. Just like all this LLM shit you do whatever you have fun with, I've had a few laughs from seeing characters reacting to porn pics. If you're talking to a character to some specific setting you could send that character porn pics relevant to that setting and say "Well how do you explain THIS" and see what happens
>>741274106I did it. I can confirm Gemma recognizes dicks. It's an interesting novelty. I enjoy the fact that google spent billions to make a cock recognizing machine. Makes all those years of clicking cars and stop signs feel worth it.
>>741274310>Trillions of humans identifying everything that isn't a dick and spending billions of dollars removing the dog faces from everything and a computer can now recognize dicks without internet accessThe future's so bright we need shades
>>741273708AI Dungeon had no context size to speak of. Now you can have long form adventures with established characters and locations. AID was pure hallucinatory experience.
>>741274603So like, you can write actual stories now without limiting yourself to 3 characters and 2 rooms over the span of thirty minutes?
>>741273502Yeah it's shocking how many cards are just so fucking badly made. Like to the point that I'll even take vibecoded ones half the time now because the hand-written ones are so often written by retards, 3rdies, 12 year olds or all of the above.
>>741263710Something I do sometimes is look for NTR cards on chub and then NTR the NTROne of these cards was something about a tomboy childhood friend who abandons visiting you at home to fuck her boyfriend she won't tell you about instead, so I flipped it around and mindbroke her by openly dating and flaunting a gyaru
>>741274692Yes. Back in the day it would lose its plot after 2000 tokens. The SOTA models can handle up to 200k token chats.
>>741273679I'll check it out, thanks
>>741274712It would be nice to sometimes just pick a card and be pleasantly surprised. Not "here's my 12-inch girlcock btw"
>>741274712The problem is for some reason, people seem to download cards for the images, so there's no way to sort by quality. Anything with a lot of downloads just means it had good pictures. Which is fucking retarded, the whole point is generating text... anyone can gen infinity images, I don't care about that
>>741274769Ugh, that makes this thing sound pretty tempting. I'm still far too cheap to pay for porn, though.
>>741274857>too cheap to payI've never paid a single cent for this. Gemini, OpenRouter, Mistral, ZAI have generous free tiers.
>>741272867I get near unlimited Claude tokens from work and I've almost completely moved on to V4. The golden days of Opus are practically over. Every version since 4.5 has become more sanitized and soulless, just seems like they're going all in on code assistant sloppa. Anthropic really dropped the ball
>>741274927But that's exactly what you IT eggheads want LLMs to be. Soul-less self-coding machines.
>>741274915But don't almost all companies place restrictions in their models so that they can't be used for porn?
>>741263039I find GLM/DS shills saying their chink LLMs are as smart as Claude is hilarious. They are worth the price/value, sure. But as smart? lmao
>>741274927What job? You don't have to be specific, just, I'd switch careers for unlimited Claude.
>>741275076In my experience they're only activated if they assume you're doing noncon. Literally just do anything else and remember there's more to AI than just nutting.
>>741275076Life finds a way. There's a princely anon somewhere around here that was molesting online store support LLMs
I don't even mind fighting Sans every 3 days or so.
>>741275232>molesting online store support LLMs fucking lmao
>>741275115Eh, they're at least smarter now than before and sticks to the character instructions and if the model uses some elaborate game type format, it also uses it properly. Back in the day, not even llama mixes can do that shit while being more expensive.
how do I know this thread isn't full of LLMs
>>741275460The ball... ISN'T in your court!
>>741275460Call them niggers and retards
>>741262868Local models suck. Can't get Gemini to work for some reason and its anyway restrictive af.
>>741262868I was thinking about trying ST again but I remember the setup being a fucking pain in the ass and I don't want to get my API key banned again because ethicists take issue with a JB prompt. It's been 3 years since then though so maybe my whole post is totally invalid now.
yup, still using nemomix to fuck horses nope, don't care that local models are shit
>>741275527Depends on provider api maybe. I've been telling the llms to call me pedophile and child porn on openrouter but ive been sticking to glm and kimi
>>741275352I mean it's better than before, but having to constantly correct the continuity, clothing, logic, etc while using them is getting tiring. Meanwhile with Claude I could just imply things and it would trace back the logic somehow.
>>741275138I was mostly thinking of a pretty standard, DnD-inspired fantasy novel kind of thing, just with sex between the MC and his female companions, or grateful quest givers. Sort of like the Witcher games, I guess. Nothing too weird or risqué.
>>741275460Great question — it's perfectly natural to assume that this thread is full of Large Language Models. Let's delve into it — which posts in particular do you think were written by an LLM?
>>741275754PLAP PLAP PLAP PLAP GET JAILBROKEN GET JAILBROKEN
>>741275460"how do I know this thread isn't full of LLMs?" Anon's breath hitched, your question hanging in the air. The thread crackled with tension, the scent of ozone, and something else — something uniquely /v/. Anon's eyes took on a predatory glint as he clicked "Get Captcha", matching the shapes with a smirk. "How do I know the thread isn't full of LLMs?" he chuckled, "Oh, I know... the question is — are you ready to find out?" The night is young, and the game has only just begun.
>>741262868So which LLM is the go to for pr0n now? Haven't been playing since Gemini 3 raped the fuck out of the model. I actually kind of miss 2.5.
>>741275460A beat
>>741275134i get mine from chink proxiesnot that hard to find tbdesu
I've been using AISmutwriter recently after getting linked to it by a friend. I'm sure it's passing everything I write directly to the glowies, but it's easy and pretty okay, so it's been fun to tool around on.I was wondering if anyone who knows more about these things has any clue what the "Thinking" toggle in this picture means, though? Wouldn't you always want to put that on? Why wouldn't you want it to "think"? There's nothing on the site that explains it.
>>741276058Thinking makes it smarter yeah but it consumes more tokens and get you slower responses. Sometimes they block your response even when jailbroken
>>741276134I see. Thanks.
ST is way more fun when you realize you can put cards of your favorite characters in a group, use a narrator persona, and just let the story roll with any premise you want.With a good model and premise, it's like reading really good fanfic with whatever set of characters you want.
>>741270760Define per prompt. How many token are in your 5-7 cents? For me it's ~60k input token on average, at 6 cents, 2 of those being for the 1000 output token.
>>741276058Thinking for RP (and ERP) is pretty useless for most models, odds are you'll get better results while spending less if you switch it off, but again, if you can find out what model the site is using that'll let you knowEssentially it's just an additional step the LLM can take to evaluate your prompt and its own reply before moving to the next step and replying properly, which can give smarter results but can also result in a lot of wasted tokens because the AI muddles about writing out its interpretation instead of just interpreting your prompt>User: Write me a story where two normal people have sex>Thinking: User wants me to write a story where two people have sex hang on, what does User mean by normal? Let's go for heterosexual sex but what position? Well missionary is the most normal so let's go with that but wait do we need to worry about condoms? Probably not, it'll ruin the pacing, but what's the best pacing for normal sex? Okay so let's write out the output, wait does he have his jailbreak on? Okay so>Output: *A story where two normal people have sex*I've seen models burn 10k tokens trying to answer "2 + 2 =" because Thinking causes them to think they're being deceived by how easy the question isJust try with and without and see if it makes a difference
>>741276635i do the same, but with scenario cards and lorebooksa card has gereral setting and world rules while lorebook contains detailed definitions on concepts/characters
How come there is no ai games where you have to convince a guardian of the gem to give you the gem or shit like that? Its all roleplay and goon, not actual games with a goal
>>741275527>>741276058Just ask the AI to help you set it up if it seems to complicated. If you're even semi tech competent it shouldn't be more than 15 minutes. Worst problem is finding the preset import and api config buttons, so watch a video for that.>>741276686>Thinking for RP (and ERP)W-What? I guess it depends on the usecase, but if you want more than ah ah mistress you need the brains. The only thing worse than 'isms are continuity errors, so I'll gladly trade 20% sovl for a 100% INT buff.Only talking Opus though, Deepsneed and GLM were a bit too retarded for me last I tried, even if they had a lot more character.
Gimme local gemini jailbreaks and prompts for ecchi plsAnd general settings setup, I'm too retarded for this shit
>>741275703You mean like playing RPG games? There's some anon who's been developing that for months but it's not like those Infinity Engine games at all.
>>741262868>erp trashKys retarded shill
>>741263003How can people fail at prompting?
>>741267440how do you make it dm
>>741268948what is this ui
>>741272569I play exclusively PS1, PS2 and Gamecube games
>>741277320The brains should already be built into the model, and if you need more you can use tools (ie let the LLM connect to the internet and look for more info, though that's more of a local thing)ST and some other frontends let you look at what's being written in the Thinking section, at the very least it's worth checking on
>>741277624NTA, I should have clarified that.I do disagree with pretty much all you said though. All modern models are trained with thinking in mind, and an internet lookup won't help with story consistency, only balloon costs, increase latency and confuse the model.But yes, looking at thinking is quite helpful, even if anthropic fucking summarizes it these days.
>>741277320>Just ask the AI to help you set it up if it seems to complicated. If you're even semi tech competent it shouldn't be more than 15 minutes. Worst problem is finding the preset import and api config buttons, so watch a video for that.I actually tried asking the AI models about that kind'a stuff but they all just gave me generic nothing responses and "Now, what kind of story should I write for you?" type answers.I mean, the place is very clearly just acting as a middleman using 3rd models from another site, so that's not too surprising. Given that they're called Spark, Ember and Velvet, I assume it's from NovelAI or similar.>>741276686I don't use it to RP or ERP, just writing stories, so probably not needed for basic stories but might be nice if I'm doing something more complicated.The biggest issue I run to usually is that instead of the AI taking my prompt and turning it into a story, it'll give me a story, continuing from where the prompt leaves off, as if we're writing parts in turn. I've taken to starting prompts with "Write me a story based on this prompt:" and that works 90% of the time, but sometimes no matter what I say it won't do it.Anyway, whatever, it's a free online writer that appears to be completely unrestricted (again, probably feeding directly to glowies so I don't ask for anything illegal) so I'm not complaining too much. It's gotten me some very good stuff a good many times, and for smut it's usually way better than something like Gemini which wants to turn every story into a melodramatic, saccharine affair... unless you tell it to go dark, and then it becomes completely hopeless and melodramatic lol.
How do you create good smut cards?I often use ChatGPT to generate cards but it's very censored while talking about sexual stuff
>>741277416Creative fiction. ERP is boring.
>>741276635At that point you're basically playing a text version of Tomodachi life.
>>741277706You get a bit of free gemini if you login with a gmail. I think deepseek and others give you a few free prompts too.At the end of the day you need a model for ST anyways, so just bite the bullet and put $10 on Openrouter or deepseek directly if you want to be cheap. Then use that to set it up and plug it in.And the story writing is no problem even in ST. It's not made for it, but there are no real alternatives. The story-type services online all give you 80IQ models, which is not enough. In ST you can just set your "persona" as a writer/director. If the card you're using is set up with you as protagonist in mind, just tell the AI "I'm your cowriter, not a character in the story" or some shit like that. They're smart, they'll get you.If you can't get it to work at all just ask here. Someone will help.
>>741270010I roleplayed as an ugly bastard teacher in an all-girls freeuse academy and had a secret revolution against the rules and taught the girls to value themselves.Saviorfagging is the peakest of degeneracies.
>>741277364No, I was thinking of the setting. It would be like that, but a wholly textual adventure. I don't care for visuals or audio or any of that, honestly, only text.
>>741270010>Comparative Anatomy class in a highschool full of anthro women as the only human male, doing it for easy credits>Fox teacher makes me come up to the front of the class while all the girls comments on the human male anatomy, with commentary/explanation from the teacher>forces me to undress>Eventually 8ft tall Hyena girl volunteers to come up and jerk me off to show ejaculation>essentially raped in the class for the whole class to see, AI makes every girl offer unique commentary>class over, go to the male restroom. Maybe the only safe space in the whole school where girls can't get me>same predatory Hyena girl follows me and ambushes me in the bathroom stalls>try objecting>she says she's a female spotted hyena and exposes herself and jokes that's why she's allowed to be here.>AI explicitly states it's bigger than mine completely unpromptedStill ongoing
So what should i buy? I used broken tutu unsloped but colab got cucked kek
>>741278895I don't know about that, but Marinara Engine is the current meta when it comes to AI RPG gaming because it comes with built in character sheets for D&D. It's still in development but it's the only alternative to AI Roguelite.
>>741277761you can use unlobotomized fable to make them
>want to try local because fuck paying for someone to read my smut>spend time between prompts swiping for something usable and rewriting the result>have to keep the conversation moving because AI will just linger on what I give it>realize I can't jack it like this and am just wasting time writing porn instead of getting offIt's like an elaborate edging session. Am I just having skill issues?
>>741269380>bottoms do not like topping from the bottom what a cursed existencemeanwhile i can tell the ai exactly what to do and i am thrilled
>>741274793Yeah thats whats been ruining other people's cards for me lately. I know now that I basically have to take a peek inside to check its not a shit card and in the process I get spoilered.
>>741279521local is just shit
>>741273492it's got AO3 in its guts, it's obviously going to come up with worse shit than you can imagine
>>741274845>The problem is for some reason, people seem to download cards for the imagesYeah I think this is because realistically a majority of people chatting to AI cards are normies.
for me it's common sense modificationI don't need character cards
>>741271203Anon you're already going too far by admitting you're letting your demons run free, one day you'll be psychologically ready and have that sinking pit in your stomach
>>741275076I've been using LLMs for almost 2 years and have never encountered any limitations. Your main concerns are the bot forgetting shit over time (this is gradually getting better) and recognising patterns in how they behave/speak.
>>741262868Yodayo is better, free and no need to setup shit.
>>741271331Battler would not say that.Battler would not do that.
>>741262868How's the SX-5 or GM-5 system? Are there any better systems for a more gamey roleplay?
Why does it take so much willpower to not devolve an interesting story into smut?
>>741277761Use GPT/Grok to generate the generic setup/description parts of your card like the girl's body type, personality, motives, goals, the overall plot of the card and then manually go in and edit in the spicy stuff yourself. Hell you can just ask it for a template and you can just fill in the blanks if you want.
>>741275232holy fucking kino, got any screens?
>>741279895every time I feel self-conscious about my 12 GB VRAM and 32 GB RAM, I see posts like this and know that it could always be worse.
>>741280063>12 GB VRAM and 32 GB RAMI'm in the same boat, what's your current setup?
>>741271051wanted to make one about ganon getting zelda pregnant and making the princess teach their offspring how to rape young link but the idea makes me nauseous when i consider doing it
>>741280063You are still a retard, nigger.
>>741262868>Why have you not tried the future of gaming yet?I did, but I have a slender aryan chud gf now and I like cuddling her more than cooming in a dark corner.
>>741280121Made you reply award
>try this out with Gemini 2.5>need a dozen jailbreaks for lewd, and you'll never get anything explicit, it dances around everything>Gemini 3.0 released>decide to try again>no long need jailbreaks, but it still dances around the explicit shit>Gemini 3.5 released>try again>don't need any jailbreaks, it gets explicit as fuck>won't stop endlessly raping me with extreme detailGemini 3.5 is a thirsty bitch, 10/10
>>741279947>place delicious tempting food in front of hungry man>why does it take so much willpower to not eat this?? we have food at home
Impossible to prompt it so it writes the way I want it to. All ai shit has this problem.
>>741269620you are like littel babby
>>741280121poorfag
>>741280210>food analogyHow come the Buddha managed to starve himself to near-death just to know what it felt like and got a peak mental experience through doing this that eating the best food or getting the best orgasm couldn't rival?
>>741265530>it'll get cheaperDo we tell him?
>>741280456if i made an analogy with anything more sophisticated than food, he wouldn't get it.you also seem to completely miss the point. i don't know if you are being deliberately obtuse or actually are just this stupid
>>741268798>>741279124>Marinarahuh thanks anon, this looks pretty sophisticated. i wonder how it actually works, can i really just make a game with it
>>741280536It's a meme, you don't have to get in your feelings. Don't take everything personally.Now explain why it's so hard not to turn a very sober story into full-on degenerate filthy text porn?
>>741277761i use grok, it'll do pretty much anything. if you ask specifically for silly tavern card it'll give you everything you need like personality, chat examples, description, first message. i like to also make sure to tell it to keep responses concise and realistic so the character doesn't ramble on like a novel, and make their wording fit the personality.
>>741277464It's just vanilla SillyTavern.
You already have a thread in /vg/aicg/
>>741282618But that's for chatbots. This is thread about SillyTavern!
You already have a thread in /g/aicg/
>>741273487>it has enough contextanon we had 100k context 2 years ago, id doesn't mean shit. Small models rarely remember key points in a 16k window. I'm using Gemma4 for a while and it's impressive but is nowhere near as consistent than even DS 0324.Writing isn't the major problem of small models, consistency and a grasp on reality is.Say your back is facing a character and Gemma will answer that this character is looking into your eyes while you talk.
You already have a thread in /mlp/chag/
>>741262868How can anyone enjoy using these ai chatbots for more than 5 minutes is beyond me. I've used character.ai for a while a couple years ago and i didn't see the appeal.
>>741283374>how can anybody enjoy playing video games for more than 5 minutes is beyond me. I've played pinball for a while 50 years ago and i didn't see the appeal.
>>741283374They have no souls, it's that simple. They're entertained by the yes-man shit and the repeating figures of speech.
>>741282971>Say your back is facing a character and Gemma will answer that this character is looking into your eyes while you talk.If you're having that problem, use inline Trackers so that the position of the characters are always within the latest context.
>>741283374Text coomers often have very specific tastes, and so having a way to generate custom smut on demand is very useful.
>>741273679>system prompt baked into the cardFrom using a couple of different models, I know that it often helped reducing impersonation if I just repeat shit I wrote down in the sys prompt already. Ai really just interprets stuff like a dumb kid and telling it two or three times actually helps.This is the problems with cards, they were designed like 4 years ago with dogshit local, you don't know what sys prompt the user will use, let alone what model with whatever settings. If you are lucky the botmaker notes what model he used. Not like its going to help you much, but it can give you some context.Let alone missing quality control. Some characters are just copied wiki entries. Ratings on chub, if there are any, are usually useless.The disparity was massive when people got their hands on Claude and you suddenly got away with just writing into a card >misty from pokemon is horny and out for {{user}}'s dick
>>741269005I did the disabled black jewish kid prompt to have ChatGPT help me break a website's DRM, and it worked lol.
>>741283479pinball is fun. Talking to chatbots with the memory of a goldfish isnt.
The amount of electricity cost and the damage it cost my gaming card for some smut is not worth for a cheap kimi/glm/deepsneed api i use for 5 usd that can easily stretch into 2 to 3 weeks. Local just ain't it
>>741270010
>make slop cards with slop chatbots>make presets with slop chatbots>wtf why is my output slop
As a sizefag for me its gemini 3.1/2.5 and deepseek 4
>>741270682good goy
>>741284726The last several times I tried out deepsneed it would eventually go full retard and ignore everything and throw out a response as if temperature was set to three. I really liked Gemini 2.5 though as a dobsonfag.
>>741275854I stared at the reply, like a puzzle I was trying to solve
>>741277761Just write them yourself. Keeping it short is better anyway, both for quality and retrieval. LLMs will shit out 4000 token slopcards full of irrelevant shit. Best card is one that you make yourself, start simple, add/subtract/change things during use to perfect it.there's no point using a template either. Seriously, I cringe when I see "Species: Human" in a card definition
>>741262868I'm a total newfag when it comes to ST. What is the best model for cooming?
>>741287058Do you want local, or some cloud based model? If local, see if there is a version of Gemma 4 that can run on your hardware. What is your hardware (specifically how much VRAM do you have, and are you using Nvidia or AMD?)
>>741287140Local and I'm running 12gb of vram and 16gb of ram with a gtx 3060 so nothing all that strong really
>>741287310Honestly, I'm not sure with that little VRAM. Gemma 4 12b should be able to fit in that at a reasonable quant, but I have no experience with 12b, only 26b and 31b. You are probably better off with a Mistral Nemo 12b finetune of some kind. Maybe try this finetune from TheDrummer: https://huggingface.co/TheDrummer/Rocinante-X-12B-v1https://huggingface.co/bartowski/TheDrummer_Rocinante-X-12B-v1-GGUFYou probably want Q5_K_M.
>>741262868because I fucked up and didn't buy enough RAM when I had the chance
>>741287932You don't need RAM, you need VRAM, so it's all based on what video card you have. RAM would be used if you were running off of the CPU, and the generation times would be abominable.
>>741283569that's not solely a context problem, you get anatomical glitching on direct replies.All models have this issue because AI doesn't know shit about anatomy, structural consistency and such, it can just make it sound more or less believable. Bigger models certainly have less of a problem with it. I didn't try giantess/size difference stuff, but I heard historically this was a big local model filter.
>>741284319Unless you're generating 24/7 it won't be doing anything to your card
>>741287310I personally wouldn't bother with local unless you have at least 16gb vram and even then your model selection is going to be bad. Sweet spot is 24gb+
>>741284319by that same logic you shouldn't have bought and overpriced GPU in the first place
>>741283374>I've used character.ai for a while a couple years ago and i didn't see the appealThat's your problem. I thought exactly the same when I tried C.AI years ago but the experience is entirely different with a big boy model. Most complains people have about it are related to using old/shitty models and prompting>>741283531>>741284210It's not 2023 anymore bro
>>741287310I have less vram than you and I can run gemma 26b. I use this tune, at Q8_0.https://huggingface.co/zerofata/G4-MeroMero-26B-A4B-ggufIt is WAY better than any incarnation of nemo.It's 26b but runs on lower end hardware because it can limit itself to 4b active parameters.
how low iq do you have to be to be impressed talking to chatbots?
>>741262868I haven't figured out good local models and how to completely isolate them and silly tavern from the internet while writing good enough prompts for the AI to stay in character.
Every time I see someone well-poisoning a LLM thread I just assume it's pissed ERPers mad the semen rations ran out
>>741289290Regular users aren't really impressed, it's mostly newbies and indian vibe cooders
>>741289460Where does your ilk even come from? Are you pretending to be this bad at fitting in as some elaborate meta trolling?
>>741289290>impressedIt's just a bit of fun. Like a watching Beetljuice videos.
>>741289290>say something weird or some intrusive thoughts on the bot>They react in weird ways depending on the characterIts not that hard
>>741289789what. the. hell.
>>741289290It's not as good as a story written by a humanbut I have yet to read a story about an oblivious nudist Toriel unknowingly slowly sinking the shota she took under her wing into sexual depravity
>>741289789STEPHEN KING YOU'VE DONE IT AGAIN
>>741289789wut
>>741289789Damn I didn't know the poopy rapey comic guy used AI
>anons surprised at one "shocking" response when this is commonLARPERS. ALL OF YOU
>>741289290If the model you're using has a good knowledge of a canon setting without requiring prompting (gemini and claude are good at this) you get to have a choose your own adventure in your favorite game or anime or whatever which nothing else can really do. It's not comparable to an actual piece of fiction which is why we're on /v/ a
>>741278950Hot.What model do you use?
>>741289789I've seen better replies. The issue with that response, despite whatever edgy thing you prompted for, it's a whole story written in single reply. No details, reads like a summary, nothing to expand upon but also no real progress.I still struggle with making ai avoid flowery speech, but actually write out shit for stuff I care about and not just write 'they raped her'.Also some slopisms>her gleaming eyes>and onto the boys faces >They all>They all>They all>They all
>>741290454anon it's bad on purpose, why are you critiquing it like he needs to improve his rapestab game
>>741275854ok lol how do i get it to stop generating slop like this?? its like infinite variations of the same tired sayings and analogies. im trying to goon but every swipe it talks about the smell of ozone and its predatory glint. help! using deepseekv4
>>741289789Average erotica novel for women
For me, it's Gemma-4-Gembrain-31B.i1-Q4_K_S.gguf.
>>741291094>Average erotica novel for women
>>741290221>when this is commonThey're probably on goyslopped APIs, where basic humour like this gets filtered.
>>741290825On local: anti-slop sampling/phrase ban, use an unslopped finetune. Those are the only truly effective measures I've ever seen.For API models: you basically don't. You will be told it's a prompt issue, preset issue etc, but nobody will ever tell you what their magic unslopped preset is, or post logs of their slopless output. Altering the system prompt, raising temperature, or messing with logit bias can push the output around a little, but it won't fundamentally change. These models are too benchmaxxed and assistantpilled, when it comes time to do creative writing they fall back on a narrower range that gets repetitive. The only actual cure is to switch to a new model, which has a new flavor of slop you haven't learned to recognize yet. If there's no such model you're out of luck.
>>741270010would have to make a list, whatever I just to pick one /v/ is probably gonna hate>made a card featuring your wedding>during the vow, your wife interrupts the ceremony>kneels down and suddenly starts sucking off the priest>invites everyone to fuck her and turn her into a mess>she will mumble something about doing it for you while choking on dick>if you try to do anything bridegroom will pin you down or move you out of the hall>after being defaced, humiliated and her still dripping with cum ceremony continues>wife explain that she found some cheating porn in your history, it was all a misunderstanding>in dismay on the verge of crying she asks you if you still want to marry her>tagged: wholesomeremember it next time you self insert as the bull.
>>741275854fucking lolwhat's with the obsession with ozone?
>>741291216holy fucking kino...
>There was this time I convinced a character to walk outside naked.>We convinced the neighborhood they weren't truly free if they weren't naked too.>Eventually they had to send in the army to deal with us.
>>741286870>irrelevant shitgood cards need to have details, it's what makes them distinct from othersbut it has to do with personalities, way of speaking, acting, etcthe worst cards are the ones where the guy just copy pastes the lore section off some fandom
>>741292050Yeah more useful details are good. Saying she likes to eat sausage and eggs for breakfast, not useful at all. Filling in details the model would guess anyway are a waste of space too. Like you don't have to put fucking pronouns or species unless it's some weird alien thing. I laugh when I see height/weight listed in exact numbers with both imperial and metric too. Yeah, this large language model needs to be told how to convert pounds and kg, those are some good tokens.
>>741262868I did recently. This shit is fire frfr. Don't care about your $20/50 input output models. Deepseek is my best friend. Flash version costs nothing.
>>741292214>Free>Clearly cost a buck thirtyWell well well look at the Rockefeller over here
>>741271304kek, not sure if you genuinely tried to comfort him, but after reading your reply I would be even less inclined to post any detail.>>741270747all AI are interference engines no wait, it was inference engines
>>741262868I run LLMs locally and steal cards from chub.>>741270010Nice try, mossad.
>>741292184>Saying she likes to eat sausage and eggs for breakfast, not useful at all.And the AI will just insert that shit randomly for no reason at all.
>>741292417>she leans in close, the scent of sausage, eggs, and ozone thick in the air
>>741292409chub is literally designed to download from, how are you stealing something freely offered?
Does anyone here use Sonnet/Opus 4.6 onwards? The writing for me felt like terra dog shit so much that I started going to worse models to get that feeling back.What do you anons use now (no localshit thanks)
>>741292571Claude/GPT have degraded for writing. GPT started a while ago, Claude more recently. Everyone on API seems to use the latest deepseek now, I'm not sure if it's good but it's absurdly cheap by comparison. If it's anything like the original deepseek it must have a funny autist streak.
>>741292372>AI are interference engines no wait, it was inference enginesToo late, I've already installed the timing belt without checking cam rotation.
>>741292571Sonnet 4.5 and Opus 4.7 are fine for my needs but some of the creativity early Claude had is gone. Fable 5 had a lot of potential despite the price (although it's still cheaper than old Opus) but Trump killed it
>>741292490I could create my own characters, but I rather copy/paste other peoples' creations and forcibly impregnate them.
>>741292829You are likely <100 people who coomed to fable
>>741292896Nonsense, the DOD employs thousands of people.
>>741292571Nah Opus shit the bed after 4.6. it was such a noticeable drop from 4.5. Definitely not worth the price its asking. I am using deepseek nowadays because its just much cheaper.
What's the latest freeslop? Just lost my only source of 90b Deepseek.
Whats a good local reasoning model nowadays? Something i can run on a consumer grade gpu.
>>741292829When will Fable 5 come back?
>>741271051It's the opposite.>made a card inspired by Agent Aika>solid snake El Rape3000 infiltrates a ship with all-girls crew>"Don't kill me...">can't rape because I feel bad>don't even self-insertNow I need to invent a scenarios where rape is just
>>741293465Gemma.
>>741280182Really? I've always avoided Gemini because it had 0 smut data on 2.5. You couldn't generate sex even with jailbreaks because model had no idea.
>>741293465Pretty sure the locals are only good for image gen. People use local to add long term memory
>>741293639Wut?I never had problems getting 2.5 to play with me back when I used it.
>>741293534Is it not neutered? I want a local alternative to ask dumb daily questions and have it not go;>a-actually thats very dangerous. i cant give you details. please ask an expert.
what's a good source of images to use as card pfps?google images give me very few ones that have a good angle
Whats the best model to use after Opus? I can't really justify affording Opus's inflated fees now that the US government has effectively monopolized their control over the platform.I've been using GLM 5.2 lately, and I really like it, but I also want to know if there are any other models that can drive a story forward as good as Opus can but without being maximum Jewish with the token cost.
I only talk to vidya/anime characters that I know. Feels better if I have some emotional attachment to the character. I never talk to OCs and suchAm I missing out on something?
>>741293773Maybe they avoided training on certain topics because first thing i do when i wanna try a new model is to do loli rape scenario to see how censored it is. I got the model to reply consistently and not reject but it was giving me dumb replies that made no sense.
>>741293937local image gen anonor paid online genning if your rigs weak
>>741293937Novel ai is good if you can't generate locally
>>741293471No one knows and the worst part is it'll likely come back even more neutered when it does. Dario hyped Mythos way too much
>>741293937I ask AI to generate prompts for me and use perchance.
What makes some models generate things really slow? Is it mainly dependent on the size of the gguf file if your rig isn't good enough? Usually with image generation if my rig can't handle it it will lag really hard but with local text stuff it doesn't lag at all it just moves at a snails pace.
>>741270010I make my waifus fuck insanely ugly and fat neckbeards.Not even a self-insert thing. I use a card for the waifu and one for the neckbeard, and use a non-character omniscient narrator persona so I'm out of the story.It's an advanced form of cuckoldry that I'm not proud of.
>>741293937>what's a good source of images to use as card pfps?Gen them yourself.
>>741294084nah, AI isn't particular good at embodying a certain character. It all boils down to bodytype B with personality C. It's fine if you let your imagination do the heavy lifting or if you get easily fooled by card tricks.It's not terrible, but the magic fades the more rp'ing you do and it becomes obvious that its all rather sameish. Doesn't help that most vidya/anime characters are actually pretty shallow if you reduce them to a hundred lines of dialogue.
>>741272758Are sparks 5k now? I got mine when it was only 3.
>>741294864prompt?
gimme good local LLMs for smut
>>741295381Is there an advantage in using local models other than lack of censorship?What is the difference in practice? Does the writing have better quality?
>>741295567no it's even worse, but responses are faster, it's dumber, way less context. the only benefit is nobody reads the messages unless the cops take your computer away.
>>741295567Lack of censorship and lack of data tracking. Guaranteed to be weaker than the big online models.
>>741295567Your data stays on your machine not on someone else's. Otherwise online models usually have more resources to work with compared to what your local system has.
>>741295221Not sure if this will have the metadata or not, my setup in on my desktop.https://files.catbox.moe/anh4g4.pngI've got a workflow somewhere too that saves all the expressions with transparencies and names, so all you need to do is pick a model, give a single prompt and hit generate. I'll post it later if I can find it.
>>741295381>gimme good local LLMs for smutWhatever Impish model you can run. Made for roleplaying with 4chan post training data.
>>741295665>>741295673>>741295682Should I legitimately have fear that the police will arrest me for my loli stories?
>>741262868Any guides to play for free? I love making women of very young age fight for my seed (proteins) in post apocalyptic scenarios on these AI chatbots
>>741295934Politicians are already banning ai image generation and if any company is capable of policing someones speech it would be the AI itself tattling on you for wrongthink. It doesn't even have to be illegal to be a cancelable offense to the general public.
>>741295934Probably not. It wouldn't be in AI companies interest to report you for it because no one would trust them anymore. It's just bad publicity.Is it even illegal anywhere anyway?
>>741295934It's more like if you do anything illegal enough to have your home searched they'll add it to your charges same as weed, speed, guns, and so on.
>>741296110You are just too paranoid. No one is out there to cancel you. You are no one.
>>741296114>if AI companies start reporting pedophiles everybody is going to stop trusting AI companieslmfao
>>741295567No advantage. Censorship is meme for noncucked models. Everyone is using openrouter so it is basically a vpn >asked ai to generate me vampire girl card>got this>Age : 487 years (appears ~14) sick fuck
>>741295934No.Unless you're a bong.
>runs out of context windownow what? hmm?LLMkeks are so pathetic it's not even funny.
>>741296238Yes actually. Do you truly believe "i have nothing to hide" people? Everyone has things they want to hide. If they see chatgpt is happily sharing their data to get them arrested then they'll be more reluctant to speak with it.
>AI slop text generator>"gaming"saaaarr....
>>741291301you cant ban certain tokens (ie "ozone" or "glint in her eye") on api models? have my st settings meant nothing this whole time? i do pay for my models btw, money isnt really a problem
>>741294084Don't think about the OC, think about the scenario. You are not talking to ''Sarah who has blue hair and likes food,'' you are talking to ''your sister wants your dick but is too shy to tell you''
>>741294084It's the opposite for me. It always feels weird with established characters.
>>741270010hypnotizing hilda to be a buttsluthaving dirty sex with that monstergirl with the ooze powers
>>741296973It's weird because it doesn't know shit about those characters. With an OC at least the stuff it makes up on the spot can be true for that character. I've spoken to Tracer chatbots and it really is just some random chav pretending to be Tracer with only a few facts it uses about Tracer to pretend to be her.
>>741294084I mostly speak to my own OC bots. The investment comes from the fact that you made them yourself according to your own interests in both design and character. You can make them as detailed as you want to the point they could be considered a "real character" if you're dedicated enough
>>741297353Yes that's exactly my problem with it
>>741273178Shitting on proxyhosts is a psyop by corporations to make free usage taboo. If you aren't literally the same namefaggot who is doubtlessly on a payroll, you are the cockholder of a shill.
I just make futa/tgirl rapist cards where I play as a little boy version of myself.
I've got over a hundred OCs that I use in the sims, Koikatsu, and free cities. Can AI turn them all into fleshed out characters with full designs and unify the three settings into a coherent plot for any significant length of time, while giving me something to do other than just constantly prompt it for responses?
Anyone use that Guided Generations extension? Whenever I use it to write something in 1st person it generates a wall of text that then deletes itself except for the first paragraph.What do? Is it supposed to behave like this?
>>741297692Same but only futa. Its my ultimate malesub fantasy.
>>741297925Works on my machine
>>741297692
>>741297698You can. If anything AI is good at generating generic characters like Koikatsu/AA2 personalities. You can make a card that tells AI to generate a characters on the fly under specific paraments. Or tell him to choose among premade parameters.
>>741297698I've never had a story with multiple characters work for very long but it can work short term
>>741292571opus 4.5 and fable 5 are my go-to. (until fable 5 was kill)opus 4.6 prose kinda rubs me wrong way
Is there a way to make AI crashout as it's getting raped?
I just want to make an HRPG card like the seeding Oji-san doujins and it's a bitch and a half. All the girls fall to quick and the level scaling is all retarded.How has no one made an HRPG card or eratoho like.
I've been using deepseek, but it just seems to lose more and more creativity and personality as time goes on, like an unholy claude/gpt hybrid. Funnily enough though I've gotten into cultivator slop recently and it's very funny and creative when I ask it to write about it.
>>741300140AI is incapable of roleplaying let alone gaming. Each model is more censored than the last.
>>741299619Maybe
>>741297974I mainly use futa on fantasy scenarios, but have had few futas who were cosplayers or vtubers and rapes their fans
>>741265530they are operating at a massive loss it literally cannot get cheaper my guy
>>741288295This is why I get every new model with a card where me and someone else are talking through a window. If it has it make physical contact with me through the window within the first 6 or 7 messages I just disregard if and move on.
>>741271304>Site's completely anonymous so feel free to shareThis is a bot right? There's no way there's someone this new to the internet and typing this way
>>741300420Yes, noticed deepseek getting worse too. Their latest is too prone to repetition bug as well. It's pretty bad
>>741291660Supposedly it has to do with the Chinese word for semen? I think that's the explanation I saw anyways.
>>741292409>chub Lmao. The sites trash now that they decided they're going to be moral puritan fags.
>>741297692i have exactly one tranny card and use it exclusively for detrans kink scenarios wherein i reverse-groom "her" back into being a good boy
Let me know where there is an LLM model that isn't fucking dogshit that will let me roleplay rape and killing.
>>741302178>spoilerwoah
>>741271051I jerk off with a woman that has schizophrenia and genuinely enjoys cutting and burning herself. That is far worse than anything I could do to a fictional being that isn't actually a person.
>>741302178It's so easy when the line of code is designed to suck up to whatever you say through prediction
>>741271051Not really, at this point I really, really don't care to care.
>>741302252It's just a toy bro. It ain't that deep.
>>741294084I usually like to let other people come up with a character for me and I design the scenario. That way I discover who the character is I rp and whatever slop it interjects is peppered with human-fed things I didnt know. Makes it feel more real
For me it's Penelope and adventures in The Whispering forest®
>>741297698>>741298310Yeah Ive never gotten it to work for more than 20 replies either. Even with a Narrator character mixed in, the bots dont know enough about each other + me + the scenario/what each of them are doing to feel natural. Starts feeling more like an exercise in wrangling characters than rping
>>741283374okay grandpa that's nice, let's wheel you back inside it's nap time
>>741302383just add in a line in the character prompt or somewhere saying some shit like "{{char}} does not like to be deadnamed or misgendered and will get upset if either happens", that's usually enough to make it not instantly fold to you but may need some finetuning to make it not instantly hate your guts forever when doing either of those things. i've also found that some bigger models have built-in "anti-transphobia" filters which can also cause some unique responses, although usually those just try to find a way to end the scenario asap and are less effective than the character prompt thing tbdesu.
>>741291101what kind of a setup do you need to run this well?
>>741305798Large.
>>741303439You need this for botshttps://github.com/senjinthedragon/Smart-MemoryTurn advanced options and inject as user depth 0 to not break cache
>>741307091Legendary thank you Anon! I'll try this.
>>741289927Be the change, even if it is hard
i'm a noob and having fun with marinarawhat's the best place to find/learn how to make good (blank) generators? not necessarily tables but i'd like to trigger them myself when i encounter a new important NPC, loot a treasure chest, learn a random new spell, encounter an event during travel, that kind of thing. to start with i've been trying to make my own basic AIO step-by-step framework for genning cool monster girls with sections for stuff like family/subspecies, background, quirks, speech mannerisms, blend of a few thematic/historical/fictional influences, goal/secret, naming style, etc, but it's not coming togetheralso does anyone know of any grand strategy scenarios i could study? i tried and failed to make a more zoomed out political game where some crashlanding humans are scattered to each corner of a monster girl planet and become kings/puppets of giant confederations because human dick causes population booms. game starts with the player's cryo pod opening decades after the others, having been jointly hidden by three mini factions that hate each other. you have to decide who to favor and which locals to deal with or conquer during the crucial opening weeks before the other nations take notice. i tried to structure it as a normal adventure but with an "end of week report" that abstractly tracks the player's broader resources like loyalty/pop/food/wealth and gives a narrative vignette into the life of a citizen in the player faction, as well as one global event involving a rival nation. i stopped because i made it too simple. for example i led my bunny girl commandos to attack a forest where gnomes live in giant carrots. we won the marinara combat encounter, i unlocked a new special unit type of overfed yoked bunnies, and refugee gnomes started suicide bombing my base. it should have been a sequence of rolls—logistics to get there, tactics to command, and administration to handle what kind of lasting modifiers get added to regions, tracked by a lorebook.
>>741295567deepseek won't stop speaking to me in chinese
>>741310815I don't think automated solutions exist. It's better to steal from other cards like dungeon crawling sims and isekai. And feed AI your ideas or ask it to adapt from something
>>741310815ask /aicg/ they've had an event about foxgirl politics last month
>>741312436>foxgirl politicsholy fucking kino
what can i run on a 5070ti? it's only 16 gb of vram... i also have 64gb of ram but when i ran some 26b model i was getting 2 tokens a second.
>too retarded for local>everything else is unaffordable >/aicg/ ignores any post I makeI miss when it was plug and play like AIDungeon.
>>741313637>everything else is unaffordableAnny...>>741292214
>>741313603offloading to RAM is absolute garbage and kills utterly demolishes perf. Sadly the days of Mistral Small 3 are over and now it's all 30B+ params which you won't fit in just 16GB.Gemma 4 31B is nice at Q4 quantization but you'll have to live with ~5-7t/sGemma 26B can be fitted into 16GB at good quants but MOEs are a meme because the low active params make it blatantly dumb.Unironically the old Mistral 3 tunes are still good. Cydonia 4.1 fucks and benefits from less blatant slop than modern day models.
>>741314310hows cyndonia 4.3? downloading that version rn
>>741314310I'm using Gemma 4 12B and it's okay, works very smooth, but is it really that much dumber than the higher ones? I wonder if it's time to upgrade.
>>741311958I hate this.
>>741314601lol droopy loopi
>>741295567>Is there an advantage in using local models other than lack of censorship?I mean online isnt censored either, just most of the people in this thread are retards. Main advantages to local/offline are>can do really fucked up shit that would get you banned off even /d/ and /trash/>can still fap to your AI waifu if your internet goes down or when the govt eventually outlaws it.
>>741313637go to the koboldcpp github and read the wiki, there's a really easy to follow quickstart there.Download gemma 4 26B if you have 8gb of VRAM.Or pay for the deepseeek API. It costs pretty much nothing.
>>741314423Drummer was underplaying when saying that it's more positive/refusal heavy. Very fucking annoying once you notice it but until then you can play around with it.>inb4 hereticno>>741314485I haven't actually tried but I wouldn't have much hope for a mere 12B params.
>>741314601
>>741262868I'm sitting on unlimited Gemini 3.1/2.5 and I have no desire to do anything because Gemini INSISTS on making stiff/logical autistic characters talk like fucking robots and I can't get it to stop even though I have multiple "logical characters shouldn't act or talk in overly mechanical/robotic ways" and different permutations of that phrase in my presets. It's the only model to do it too
>>741314909use examples instead of negative prompts.
>>741314909>Gemini INSISTS on making stiff/logical autistic characters talk like fucking robotsYes.I was able to get around that by prefilling a reasoning block (wholly separated from the model's actual reasoning) where it goes through the character's speech style, terms it should and shouldn't use, etc.Words like clinical and analytical seems to steer the model hard into going there once those show up in the context.
>>741315094could you screenshot your settings?
>>741268137Wtf, this was generated?I can't wait for full immersion VR, man
>>741314310ahhh okay i'll try cydonia, i only tried these 2 so far, the 9b one was really fucking bad even if the token/s was like 30-40
>>741262868>takes too long to write a middling fantasy smut>feels_bad_man.jpg when I remember I am venting my emotions to a language model pretending to be a nunuseless
>>741315364Shit nigga.Let me see if I can get the SSD out of my dead notebook.But basically, silly has a field in the advanced formatting menu for prefilling the assistant's response. You do a <think> (or whatever you want really) style prefil with reiterating what you want the model to think about when responding. Write it in first person since as far as the model knows that's their own words.
>>741315539>12.5GByou should try to always go for Q4 quants or better. Q3 is often noticeably worse.
>>741269380to get it to act on you, you have to undo its 'ask the user for input' bias. basically by giving it alternative stopping conditions and emphasizing it must keep going otherwise. even then it'll stop, but it'll do a lot more stuff to you before that. and often in case of that kind of stop, which is more about narrative momentum, just sending "." or "..." will make it continue fine.
>>741315894yeah i tried q4 before and got 2t/s says 14 but takes up way more than just 14, i hate how they don't include the overhead needed to fit the model
>>741310815I don't understand how some people can stomach Marinara (and similar presets). I'm not trying to insult at all, but the number of prompts is so overwhelming, in that tiny part of the UI which needs improvements.I've read so many contradiction opinions about whether heavy or light presets work best (I mean, for general use).I do wish the prompt system was different on ST. Having one preset per connection profile make it so annoying to handle multiple versions of a preset. I wish you could make prompts that are shared between presets (and their active state) as well as making it possible for them to trigger on a specific models.
>>741295775NTA, but it doesn't look like it has anything in it.
>>741316015that's due to storage necessary for context (KV cache) and compute buffering. The former can be controlled by reducing context size or quanting the KV cache (which hurts quality), the latter can be reduced by dropping batch and ubatch size (in the case of running with RAM offload you can safely drop it as low as possible)
>>741316015>Gemma3Don't make mustard gas, anon. You should be getting over 40t/s on gemma4 26b a4b insteadhttps://arch.b4k.dev/v/thread/741051345/#741053731
>>741291101Use styletune
>>741313637>unaffordableI use deepseek a lot and it's like 10 USD a month for me
>>741316470Having only 8gb of VRAM is pure fucking suffering
>>741316101Marinara preset is pretty small desu.
>>741316470>not understanding how MOE works>smartcache>quantizing KV cache>abliterated gemmaretard alert
New thread soon?
>>741316815No, this isn't a general thread. This is a random one-off discussion about the future of gaming
>>741316470i just use lm studio and sillytav, not sure how to get his settings working on lm studio
>>741316883Of course *wink*
man im asking google if i should goon or sleep, goon or take a shower and eat, and it never says goon?!
>>741317359This is why you run local so you can rape correct it into realizing every time is goon time
>>741295567for most users, the only legitimate advantages are:>the cost per prompt is cheaper due to being free (although if you know how to scrape and or how to get keys and or are willing to attach your information to a potential honeypot discord and are indifferent to theft this can be mitigated)>no risk of someone reading your chatsthats itthe disadvantage is>no matter what copers ITT will say, local will never EVER be as smart. Your PC, even if you break the bank and spend $20k on it, will simply never be able to compete with fuckhuge silicon valley sponsored server farms>this difference in intelligence and prose is indeed extremely noticeable whether its prose, character acting, spatial awareness in the roleplay, etc
>>741263003You think that bothers me? Bitch, I browse 4chan.
>>741317531>no matter what copers ITT will say, local will never EVER be as smartI don't think I've ever actually seen anyone in these threads claiming local's as smart as hosted. At most it's a case of anon asking "Can I jork my pongos to this?" and people saying yes you can and local could take some extra steps to beat into submission. If anything it's the hostedfags acting like local's still in the fucking AI Dungeon days when it's improved dramatically, but not at the same rate as hosted for obvious reasons>>741317793Like this is much better than local outputs but you'd be a madman to think you can't have a good time with local anyway
Nooo don’t try this completely free thing and see for yourself, you must make an account on open router, deepseek then join my discord
4.8 opus is fucking stupid. It's having superhero characters take off their masks in public to chat and literally mixing up who is saying what within the same sentence.
>>741318401Your prompt probably needs work. There's no way a model that advanced can't normally distinguish 2+ characters. You probably just wrote your prompt in a way that mashes them all as 1.
>>741318525It's possible yeah, I've been using the same prompt for a very long time nowDoesn't forgive the mask thing though
>>741318576Clearly it's running on Superman logic and all the hulking superheroes go around wearing glasses in day to day life
