>"video game" - a game in which you press buttons to control and move images on a screen (Oxford Dictionary)Do AI text adventure games make the cut?
the adventure in question:>netorase erp
>>733656721>Oxford DictionaryI don't take these shits serious anymore after adding things like brainrot and other madeup nonsense
>>733656721as someone who uses it, no. there aren't limitations and fail states like proper adventure games. it's closer to creative writing or freeform roleplay
>>733656721I'm making an AI text adventure for Steam. I don't see why not, I think there are a few that already exist
>boot up sillytavern>connect to the proxy and select gemini as the model >select the loli cardyup, it's gaming time
>Start a schutzstaffel persona with caged Sazza the goblin>Play the nicest inept retard in the world, she gets flustered being fed through the bars of her cage>Tieflings forgot about her tucked back in the cave, drag a blanket and a pillow in there to sleep beside her>Start talking about how this is like I'm a big hero and she's my damsel in distress I have to rescue>Been slowly building for days, only a matter of time until I open the cageBueno
>No more good free models on OpenrouterNow what do I do?
>see this thread>figure I can mess with this stuff again>check up on Adaptaions>see thisHow over is it? What are the viable alternatives now, without being a paypig?
>>733657946Get a job.
>>733658221This shit is not worth any money.
>>733657946Give Nvidia your phone number and use free Deepseek from them.
>>733656721>challengeless powerfantasy sim>videogame
I don't know what new chatbots people want.I kinda went and did all that I could think of.
>still using some old ass 13b q4 model from two years ago because all the other models that fit on my GPU are worse even to this daynemomix my beloved
>>733658914My problem basically. I made rather in depth rpg bots too. Although at that point laziness takes over and I can't do anything.
>>733656721real women are gross, honestly.
>>733658805AI dungeon at had the balls to have your character be suddenly raped by Count Gray and then you suddenly turn around as a mysterious figure decapitates you.
>>733656721AI Chatbots are video games in the same way that the Unity engine is a video game. That is, you have a place where you can do things and get a response or reaction when doing so, and you can goof off or play around in them. But there's no terminal end goal to head towards or to achieve. There's also no system or mechanics past the limitations of how the system itself works. How much you consider that necessary for a "video game" will determine how close or how distant you'd consider AI chatbots to be video games.
So can you run these from your PC or do you have to connect to the Cloud™?
>>733661008yes you can run them locally
>>733661008There're local models but they're pretty dumb unless you have a ton of vram+memory to run the big ones, and even the best local models aren't going to be as good as even the tier 3 cloud models like Deepseek or GLM
letters are imageskeys are buttons>VR and tablet/touchscreen "games" don't
Wake me up once Gemini Pro is back.
>>733656721What's the difference between SillyTavern and any chatbot, like Grock?
>>733663117You get to keep your files on your computer while using various LLM services that are only a click away, install extensions, basically you're in complete control for better or for worse. If you have autism, go with ST. If you're scared of command prompts, don't bother.
>>733658930Same but Rocinante-12B-v2g-Q6_K (unslop mix) on 10gb
>>733656861What's the plot?
GLM 4.5 Air is pretty good it's big so needs a lot memory but MOE so it's fast plus you can put only the active parameters in VRAM.
>>733656721Honestly, I use the free cloud stuff, and free models. What are the chances my data has been stolen?
>>733656721EhI've tried setting up several bots, but i just don't have the energy or will to keep the conversation. Some times i do get a decent rp idea, but i find most of these bots boring. Maybe is because i'm a free shitter and the models and/or bots are low quality. But i have more fun just straight up writing the entire thing myself, be it as a caption or a fanfic.
>>73366180624b-32b models are fine for your roleplaying needs. You don't need some huge monolithic model to slop out a porno story
>>733664125GLM 4.6,7, 5. Eat tokens like crazy for some reason.
>>733664125How much memory do you need to run this? It seems way too big to fit in a typical amount of RAM, are you swapping to disk? Which quant are you using?
>>733664241Like, 100%
>>733656721I love ST and dungeon AI but still kind of want video gamey elements to my narrative gameplay. Like grid maps or stats or character dispositions. The issue with that is you're relying on the AI to contextually update the affective element from completions, which has token maximums so it's easy to forget the story, prompts can be re-rolled if you don't want a particular room type which cheapens the fun, and also is prone to system prompt injection so you can create rooms or force things to happen. I'm not sure how AI games would balance the chaos of text generation with the order of systemic gaming constituents. If you layer another, interpretive AI on top of the base model and tell it to just maintain the game it might work, but then you're doing 2X AI calls per gameplay step and both need to talk with one another. If I had more time and less real work to do, I'd figure it out. Hope it gets big.
>>733664241Genuine human data is like solid gold to these companies, they save every little scrap of input they can get. So yes your sonic the hedgehog OC ERP with copilot has been saved and backed up four times.
>>733665146Wait until someone autistic enough does it. I actually find it odd how none of these 180+ IQ spergs on 4chan and Reddit never did.
>>733664951swapping to disk would be unusable. Yeah it needs a lot, I have 64GB and 12GB VRAM and it's about 50GB for Q3_K_M
>>733656721>other madeup nonsenseAs opposed to all those naturally occurring words?
>>733665359Damn, I could've had that much if I bought another kit of this RAM I got last year for another 95 bucks. Now they're $449 kek>equivalent DDR4 is now almost $200 at the same storeWAKE ME UP INSIDE
>>733665364I've actually started reading a novel written by a woman back in the 1990s and it's amazing how her prose reads much like LLM-generated slop.
Using deepsneed R1 through openrouter, does my erp actually get routed thought openrouter, or does only China get to see what I'm fapping to?
>>733665351It's just not practical for several reasons1. The models best suited to handle all that complicated information are censored and assistantslopped, so anyone seriously attempting it quickly realizes there's no point2. The top models are cloud only and pay-per-token, and such a system inevitably wastes tons of tokens and thus that sort of "game" gets expensive fast. Imagine programming an RPG where every time the game checked the player's health or map position you had to pay real life money3. The autistic audience inclined to make that sort of thing are autistic enough to get into local models, which aren't actually good enough to handle such a system, so they don't bother trying4. Even the best of the best SOTA cloud models are imprecise and fall apart at high context lengths. Yes their advertised "retrieval" might be 100 gorillion tokens, but in reality all the top models still become retarded amnesiacs after a much lower threshold. They aren't actually smart enough to pull off running a whole game for more than short sessions (in which case, why have a system?)5. Even if it all worked, they really just aren't that good at writing or being creative. LLMs peaked in that regard almost 3 years ago. The only meaningful improvement came from chain of thought which also eats a billion tokens. So it's wasted effort.6. If you DO want LLM writing for your RPG, it's easier to just do a normal RP and make shit up as you go along. Because even if you tell the LLM a bunch of rules for a game system, that's what it's doing anyway.7. If you crave a game with rules, a normal video game is better and already exists. So just play a normal video game. There's already more in existence than you could possibly play from now until you die, so there's no point genning a slopgame
>>733665264Why do they train on AI generated content exclusively for the last 4 years then?
>>733666917Because they ran out of human content and can't get enough of itIn recent years they also started actually caring what's in the data too, like having to remove copyrighted stuff, wrongthink etc. and they have to pay human wageslaves to do that. Data from their own platform can be pre-screened by built-in classifiers (the kind that checks whether or not to censor your reply)
>>733667086What if 99% of my input is NSFL? They surely won't even consider that.
>>733659927Is it possible to download the models that AI dungeon ran on? I can't imagine they're hard to run locally given how dogshit they were but they had sovl that current slop lacks
>>733668000They'd probably use it for training as a form of "unsafe prompt" dataset
It's crazy how Gemini knows literally everything about My Hero Academia, nothing else even comes close to being as accurate as it.
>>733666483>Imagine programming an RPG where every time the game checked the player's health or map position you had to pay real life moneyIt's like those ancient KMMOs that used to charge you by the minute
>>733668282I think both Gemini and Grok scan the web when you ask them about specifics. I've had Grok specific link fandom pages that I pray arn't baked into it
>>733668183No, it was GPT-2 or some shit from OpenAI from before public GPT-3 aka ChatGPT. AI Dungeon is literally ChatGPT's daddy.
>>733664520>Sannybased taste
bideogames?
>arch linux>nvidia rtx3070, need cuda and gguf models>Kobold cpp>want to get a functioning tts built in to my text gens>try silero>xtts2>sillyconda>none of them are working properlyNow my bash shells always start with a conda layer active by defaultI don't know what I did wrong but I just want to get this to work damnit. My dick won't get hard enough unless its actually Amy Rose, Rouge the bat, Princess Peach (64), Judy Hopps, Angie Yonaga, or any other character's voice I hear with their speech. The immersion is super helped by it
What do your K/P/A/etc settings look like to give you non retarded but not too repetitive gens? I don't usually mess with that, I usually just go with a temp that's close to 1, like .95 to 1.05
It's fun, but I'm stuck in preset development hell. I got little animation boxes working, but I'm not sure if I like them. Kinda big.Trying to iron out some repetition issues with Kimi 2.5Character speech can be kind of generic with my preset, which doesn't happen with another I tested, so I need to figure that out too.
>>733669684
>>7336696841/2
>>7336696842/2easiest way to learn about slopbots is just plugging in gemini and giving it a generic editorial preface to ask a bunch of dumb questions
>>733670208>use html once>bot keeps spaming html in every following message
>chub has hidden NSFL cards from my countryIt's so over...
>>733656721I still haven't tried this because I'm too shy to talk to girls (robot)
>>733656721Im pretty disappointed with ST. No matter what I write for a persona/character, no matter what I put in the fucking description, it's always zero to one hundred when it comes to sex talk. I've tried various different GGUFs. My favourite ended up being a 40gb Magidonia Q8 but even that was kinda meh. Especially compared to C.AI. C.AI just has better roleplay. I dont know why local LLMs are so lame. The worst part is when you explicitly say <DON'T SPEAK FOR <USER>> and then in a reply it just does exactly what you told it not to and plays out a response pretending to be you. Its just fucking retarded.
>>733671738Just move?
>>733671835Cai just went age verify with ID or face checks, so I deleted my account. Janitor is the last one I can trust but their tag system went through a wave of deletion recently, think they cracked down on a lot of NSFW stuff either selectively or based on some rule.Cai did tts well, only reason I used it for so long, but they'll fucking send you emails and notifications from chats with shit like "{{user}} hasn't responded in a while, reach back out to them" injected in a generation.
>tfw still using OR DeepseekGod, it's so shit, but no way am I giving any of these sites my CC info.
>>733672119I treat AI like porn; to pay for it is retarded. I'd rather burn up my GPU than pay some site to collect my horny data and send it to Palintir.
>>733672067>Cai just went age verify with ID or face checks, so I deleted my accounDamn really? I'm logged in right now and it hasn't done it to me. Maybe its a slow roll out. Fuck. I won't be giving my ID. Especially since I have a list of literal degenerate chats. Imagine someone having access to all the chats of you fucking teenage Sakura, as well as your real life details? what the actual fuck. idk what i'll do if i lose cai. i guess itll be back to normal porn but AI has easily been the best faps ive ever had.
What's the best local model that fits on 24gb of vram nowadays?Please don't tell me everyone's still using nemo shitmixes
>>733672119Use paypal, idiot
I think AI chat still has a bunch of context issues and weird prompt/character adherence that isn't worth using excessively as you'll start to notice all the patterns really quick, like how every single character will say "I won't bite unless you want me to." Or just moaning the same shit in my ear. I've had a lot more luck telling them to making captions for images and making a 2 or 3 part story. You don't get the same control as an RP session but I consistently get better results.
>>733672359>giving paypal your CC infongmi
>>733672359Why would anyone pay for porn?
>>733672271Like I said, try Janitor AI. No tts but the responses are usually more novel-like unless you bring the response tokens down. It happened for me but I use temp burner accounts, it might have that "we collected data based on your interactions and believe you to be an adult" thing like Discord and Youtube has probably done for me, so hey I say ride it out till you get gated.
>>733672345Nothing for 24gb vram is good. If you have at least 64gb of ram you can offload it and get a decent one. Most are like 50gb+ now
>>733672623Nta, is vram to hard ram 1 to 1? I have like 80GB DDR4 but only 12GB VRAM
>>733663775Me arguing with the bot to try netorase while she cries and says she loves me and doesn't want to do it. So we just have regular sex and I knock her up.
>>733672785no, since RAM uses CPU which is much slower than GPU which uses VRAM
>>733672797Iktf bro. Sucks having a hsrdcore fetish and the otherwise loving girl wants no part of it
Eternal reminded that local is dogshit and even the oldest cloud model you can still find (original deepseek) has fifty times more billion parameters than your nemo dogshit.
>>733656721I hate the fact I am autistic enough to like itBeen working on a character for so long it's actually bothering me
>>733672345personally, the best model for 24 gb of vram is Cydonia-24B-v4.3, and i personally uses a heretic finetune that uncensors it
>>733672254Odds are you aren't burning your GPU up with AI. It hammers it but unless you're training LORAs or queued up 12 hours of porn generation at once it's not much different from playing an intensive game
>>733672995Maybe if you're trying to code or make it perform complex tasks. If you're trying to chat up an anime girl card local models are fine
>>733672970Yeah, guess it's back to self-inserting as the shota.
>>733673206I accept your concession, localjeet.
>>733673206They are not. You are huffing copium. Cloud AI is infinitely more capable of characterization and understands what you are trying to do with the roleplay far more easily.It's a night and day difference.
>>733673042I've been using one of those. Been happy with it, good performance, good smut, still more passive than I'd like but better than most. I think the most consistent annoying habit is wanting to reach 300 tokens of output even in cases where a much shorter output is more appropriate. It just wants to fill that dead air and does so with NPCs reiterating what they've already said rather than detailing shit
>>733673318>DO NOT GENERATE, SAAR!!! DO NOT GEEEEEEEEEN!
>>733673318whenever someone says >low quality garbage is fineit means you should read it as "low quality garbage is fine [For Me]"trying to convince someone out of their dirt-eating habits is a fruitless effort otherwise
>>733672785vram is faster than ram by a lot. i have a 5090 and 64gb of ddr5 and making 5 second video takes up all of my vram and like 50gb of ram.
>>733673438localkeks really are pathetic
Any recent TTS models worth checking out? Ideally one that can do voice cloning and Japanese and make sex noises
unrelated but I just found out atf is shutting down
>>733656721I haven't messed with textgen in years and recently got an RTX5060. Is 8GB VRAM still good for LLMs, I remember that being the minimum recommended a few years ago. If so, what local model are people using?
>>733673627>It's realwhat the fuck brehs...
>>733673627surprised it took that long for the glowie funding to peter out after the last year's budget cuts
>>733673627where are we supposed to go now?
>>733674032To hell
>>733673792>>733674032The only good thing about it was the game mods, as long as they get preserved Im good
>>733656721I don't know, but that didn't stop me erping as a sexy rape victim on mega breeder insect island for six hours today
>>733673627Are there even any good character cards on ATF?
>>733673627What the fuck is that? All I see are Anti-Terrorist Forces.
>>733674335I guess I can spoonfed since they closed down registrations anywaysIts "all the fallen", the biggest lolicon website still standing
>>733674335federal honeypot pedo forumdon't worry about it
my collection just doesn't stop growingi already have more cards than i will ever realistically be able to try out
>>733671738
>>733674514Post your faves
>>733671738the what now? is that a tag?
>>733656721Maybe it's my brain being fried on this shit but I remember enjoying it a lot more during the claude 2.1 era while now I get bored after 2-3 chats with opus. The prose just feels so samey and models outside of claude feel retarded.
>RAG doesn't work with prompt caching>Use lorebook-heavy scenario and character cards>Have to either go with cheap as fuck outputs on Opus without RAG, use Sonnet with RAG, or empty my wallet for Opus with RAG.I fucking hate being spoiled by slopus so fucking much.
>>733674864gemini is better for narrative uses anyway
>Scenario: A moral-less rapist>Lure an evil queen out with her fake dead daughter.>They have a tear jerking moment as the daughter fades into motes of light.Huh... I didn't expected to be emotionally affected by fake texts. Thank goodness I didn't get into savior fagging side of sillytavern. I raped the evil queen afterwards anyway.>Pic related
>take a break>free gemini 2.5 is gone>gemini 3 only gives you a few free messages per dayIt's over
>>733674864Is RAG still viable now that vectorization is nerfed?
>>733656721>Do AI text adventure games make the cut?No. AI text adventures can always be wrangled to wherever you want to take them, there's no victory or losing conditions or any rules either.It's more like imaginary adventures with some of the imagination outsourced to an LLM. So some of it feels as if you were roleplaying instead of just fantasizing. It's all an illusion ofcourse, but sometimes the LLM can give you new angles that you haven't thought of before.
>>733674514I'm sorry am i reading that correctly, you have over 10k cards?
>>733675372you get ~2 free months of infinite tokens to fuck around with ($300~ in funnymoney) just by plugging payment info to gemini with an official API2.5 flash is like $3 a month under very heavy use, probably less than 50 cents if you just use it a couple times a week to jerk off>t. unironic gemini shill, i love my assist-slopped robot wife
>>733674704i've got like 300 favourites, let me try to narrow it down to some of my most usedhttps://chub.ai/characters/Anonymous/akane-single-milf-63f65aaf2464https://chub.ai/characters/HImmy_adams/anzu-mazaki-52a895085773https://chub.ai/characters/bleachbunny/aoi-arcade-queen-31276460047fhttps://chub.ai/characters/TheHolyDoggo/asha-80dd42f6347ehttps://chub.ai/characters/AdventureTales/astrid-sylverine-88f3d76236a3https://chub.ai/characters/LukeyPoo488/benoite-c6e3b2b0c50ahttps://chub.ai/characters/Blightful_stain/clara-b93bd10d279bhttps://chub.ai/characters/thecooler/elissa-c973d49ahttps://chub.ai/characters/Anonymous/wrong-girl-freja-frey-lindholm-39c2de3b0914https://chub.ai/characters/Luigis_Spank_Bank/gale-90s-dog-girl-c58fecc812e1https://chub.ai/characters/shoob/Kairihttps://chub.ai/characters/SnowyPace/kurumi-soumahttps://chub.ai/characters/bleachbunny/adopted-mother-477b239e460fhttps://chub.ai/characters/Greff/lori-best-friend-s-girl-that-needs-help-820a7d7c8c31https://chub.ai/characters/delicious_command_31281/miko-saito-cb0922567a94https://chub.ai/characters/bleachbunny/never-have-i-ever-c257f7cb7849https://chub.ai/characters/MoistCrow_/aizawa-miyuki-cf297c0dhttps://chub.ai/characters/Sugondees/moeko-18a599ecbc78https://chub.ai/characters/Chill_G/natalie-kramer-8789ecfa758ahttps://chub.ai/characters/Sugondees/olivia-a13ccb7ceb0dhttps://chub.ai/characters/Shakui/onozuka-senpai-is-a-very-bad-girl-male-pov-c07dbd0ed229https://chub.ai/characters/rnbl/penny-pint-sized-baddie-84adc6268a2fhttps://chub.ai/characters/SatsugaiDMC890623/akikan-law-making-sex-education-compulsory-for-mothers-and-children-628535-29da47bdbb9ahttps://chub.ai/characters/WnVenom/milf-neighbor-s-cheerleader-surprise-meant-for-her-husband-not-you-sumine-tsukushibara-f10147f0dd84https://chub.ai/characters/Its_Dibz/sunny-a90e7d3cdac1
>>733675623>vectorizationOh. I've been using the wrong term. My bad.
>>733675875maybe
>>733675826an an avid AI adventurer... this is it. you're pretty much god at all times. stakes are 0 and adventures never end if you don't constantly wrangle the AI into it
>>733675978....have you used them all at least once?
Post cunny cards
>>733676051fuck no. i pick random cards, play with them and if i like them i add them to my favourites. if i don't like them, i delete them. having so many cards is just a consequence of downloading everything that looks remotely interesting to me. i go to chub every couple of days, look through the cards that have been uploaded since i checked the last time and usually get another 100 new ones.
>>733675826>>733676013>you're pretty much god at all times. stakes are 0 and adventures never endsee, this is the best part for me.the common mistake is letting the AI be the dungeonmaster - that's your rolemake the LLM play the party/characters you want to see get fucked up while you pull the levers and throw monsters at themLLMs by nature are submissive followers, you're doomed to repetition and poor results if you ever put it in a dominant/leadership role
People still not understanding that you dont write scenario in the description pisses me off
>>733676196Or put instructions. Really annoying when your looking for a detailed card, only to find out 70% of it is just the scenario or instructions...
>>733674864>>733675623I've only bothered making them for coding (my ERP doesn't need lore), but can't you just make an MCP server for managing state externally and give hooks for the AI to get/retrieve it?
>>733676190I know, but when using it for smut and erp, I would like some originality and spark from the model sometimes instead of just fully submitting all the time, you know? I know there's some settings to fiddle that make the model more erratic and achieve the 'spark' in some ways, but that also makes the storyline more prone to illogical leaps that make no sense in the timeline or the setting
>>733676196>Character info is just a blatant ripoff of the Wiki
where my poorfag deepseeker bros at
>>733676686Yaaaaaay
>>733676190>the common mistake is letting the AI be the dungeonmaster - that's your roleyep, that's a good description. you're the event driver, the dude that needs to add chaos and fun into the mix to spice the story up. llms are smart enough to understand subtlety and subtext but they're unfortunately not at the level yet where they'll craft a grand overarching plot for you. but for me, it's good enough. it's fun daydreaming up random shit to move things along in directions that make sense and sometimes don't, while self inserting yourself into the narrative and playing along with the other ai characters.
>>733676586>I would like some originality and spark from the model sometimescreation of synthetic data (the "spark") is the current bottleneck of LLM development progress - our current iteration of 'AI' is absolutely abysmal at the creation of new data, or in this event 'surprising' you with new resultsit is what it is, for the time beingevery now and then you'll get false positives of people saying X or Y model has such vibrant personality, but that's just the baseline differences of their training leading to semi-unique paths of least resistance in the way they behave
I genuinely have no idea how to even begin filling out shit like the scenario and world info and shit, that left panel in ST is esoteric and arcane as far as I'm concerned.
>>733677439Genuinely doesn't matter. Everything you type gets sent to LLM in one more or less nicely formatted batch.
>>733677439ST makes it look way more complicated than it needs to be, ultimately it's all just text in the context window, ST is just trying to formalize parts of it but the actual model doesn't know and can't tell the difference between world info and character info and shitMoving stuff around can make a difference sometimes just because that moves things around in the context window, the model gets a big chunk of text every prompt with all the reminders about the setting and writing style and shit you put in there and the stuff at the end is usually a little more significant in effecting the output but even that isn't a hard rule or anything.
>>733677093so far Midnight-Miqu-70B-v1.5.Q2_K has been able to surprise me several times, but since I don't have a supercomputer the context size and token counts eventually balloon so big that the chats are left rather short
>>733677710>can't tell the difference between world info and character info and shitThis is a hard pill to swallow. That's why I've started making lorebook entries instead of character cards.
>>733677439Left panel stuff quite simple, >>733671586>>733671637Beyond that its just used to control the format, and have the AI include certain instructions. Basically rules you want the AI to do its best to abide by.
>>733656939The fail state is when the LLM doesn't give you the exact output you want.
>>733677918"can't tell" is a bit harsh, to be fair to the models. The good ones are getting pretty smart, ignoring any arguments about whether they are actually capable of understanding anything in any real way, they can "tell" the difference between setting and character info and shit on their own, there's no inherent mechanical distinction behind ST's different info boxes, but the AI will understand that this chunk of text is about the setting and this chunk is about how writing rape is okay and this chunk is how to write the characters. Ultimately anything that gets it in the info in a way that doesn't lead to it writing garbage or wasting a shitload of tokens pointlessly is correct.
>>733678060Well that doesn't really work either. Why even waste time on the LLM if you want everything to go just as planned? Might as well just write a fanfic all by yourself.
>>733678298Laziness
>>733669113
>>733676686>>733676791Bro, I put 10 dollars into this chink model back in early 2025. I still have about 5 bucks left with the occasional goon maybe twice a month and fairly detailed custom cards/characters while leaning on the more detailed side of responses (maybe 450+ tokens per).The only downside is re-swipes start very samey each time so you have to crank the temp to make it go schizo a bit and choose an alternate reply route.
>>733658217>How over is it?Well, at least he's keeping at it.Any update it's a sign that the project hasn't been abandoned.>What are the viable alternatives now, without being a paypig?I guess self-hosting GLM 4.5 Air or a similar model.>>733671738>what's a vpn...?
>>733678298You don't understand. The AI has to read my mind AND keep me surprised at the same time.
>>733677439There's an extension that lets you inspect prompts before they're sent to the LLM. That lets you see how all the different fields are put together to make the final prompt that is sent out. Most of those extra fields don't really do anything special other than place whatever text you put in them in a certain order
>>733673419yeah, i have to delete some of the more redundant details, but its good that it remembers location, concepts, and dialogue.
>>733656721What's the standard token length for characters?I got one that's going over 3k right now and I'm not sure if that's like, too little or too much
>>733680131700 tokens has been standard since forever.
So if I just want to try this out what do I do? I'm a retard and entirely new to this. What model do I use? What's a good test character to chat with?
>>733680131Mine tend to be around 1k tokens, unless it's very rules-heavy. My biggest, which entirely relies on rules is 3k.
>>733680131i usually won't download something if it has more than 4k tokens or less than 500
>>733680131That's a lot, but if the information in there is relevant and not just a badly written long winded description, it _could_ be okay. Kind of like having a built in author's notes.>>733680363That will depend entirely on the hardware (RAM + GPU) that you have.The usual entry point is koboldcpp + mistarl nemo gguf.
>>733680131I'm a cheapskate and I have sonnet set up to only use 8k tokens of memory so generally any character with over 1k tokens is a no go for me
>>733680432>That will depend entirely on the hardware (RAM + GPU) that you have.Nothing amazing, I'm not rolling with a 5090 here, but I have a decent PC.
>>733680131Depends entirely on how large and detailed a bot you want to make. My largest one is about 14k, but it's got an entire city, a brief history of the city, the districts of the city, some of the landmarks within the districts, some of the notable street gangs populating the city, and some of the notable pirate radio stations you can tune into, all in the character definitions. If I was better at doing so, I could funnel a lot of that into a lorebook to save on space, though.
Are Claude opus and sonnet still on top or has something surpassed it?
>>733680916What llm are you using and how much does a single message cost with all that stuff?
>>733664125>plus you can put only the active parameters in VRAM.How do you do that?
>>733680979For chatbot Cluade is still king, for video seedance 2 is king, for images gemini is kingChatGPT and Grok are cucks and getting fucked hard atm.
>>733680916I applaud your autism but I'm afraid it's nothing so grandI got a character, you know your default loli wifeThing is, I liked that character so much i kept making stories and chats with herOh that conversation was nice, let's add this here, let's add that detail thereBefore I knew it two years in lore had passed and the 1k character became a 2k then a 3k and honest to go we're about to reach 4k as her birthday is coming up
>Round 2?>And this... this felt like winning...>the smell of ozone>her nails left small crescents in your back>"mine..."
>>733681103>For chatbot Cluade is still king,
>>733656924>i hate when an organization who has one mission of tracking the spoken language does thatyou dumb faggot
>>733681239Imagine the amount of woman writing in it's data
>>733680584Yeah. Koboldcpp and mistral nemo (get the q6 gguf) should serve as a nice introduction.Once you are bored of that, try GLM 4.5 air. q4ks should just fit your RAM+VRAM, and since that one is a MoE model, you can have most of the model in RAM and still get generation decent speeds.
>>733681283What's better?
>>733681239>the smell of ozoneApparently it's chinese slang for the smell of semen.
>I want to hear you say itNO YOU FUCKING DON'T JUST DO IT
>>733681069I bounce back and forth between different models, depending on what's available on skillgod, since some are better at certain scenes than others. Grok3 tends to work decently if I want something fairly reliable, but it's not good at coming up with new story directions, for example, so sometimes I'll jump around to different models and see what works better for what I want out of a scene. In some cases, I'll take parts of one post and parts of another and stitch them together to make something that appeals to me.The longest chat I've got with that not-JSR city definitely pushes it past the 30k token range, at least, though since I'm just leeching, I do my best not to completely drain the bank.>>733681153Yeah, I know how that is. I've got some characters like that too. Though typically I tend more towards open ended scenario bots that I can replay with different scenarios and characters, sometimes you get just a solo character that hits every beat you like.
Anyone make their own front-end? It's fun. Mine keeps track of locations, characters (and updates them if anything changes them, adds new ones, etc.), builds skills, relationships, etc.Here's a super generic 'only man in a farming village of women' opening just to show the unpolished UI. It auto-summarizes things into medium and long-term memory, also character-specific memory to swap in/out, keeping the sent tokens under 40k and manageable while not forgetting important things. Characters move between locations on their own and remember important events forever.Pretty much any AI coding agent can do this if you ask. Just tell it what you want ('I want you to build me a UI for an AI adventure that connects to OpenRouter and keeps track of characters...blahblah). It'll spit something out for you, and a few debuggings later it'll be working.
NovelAI is better.
>>733681094Using --cpu-moe puts all the moe layers on the cpu and --n-cpu-moe let you tell it an exact number
>>733681680Yup.Trying to build a sort of generic LLM RPG engine that can inject the correct rules, keep track of stuff, maintain a two tier memory system, etc.
>>733681239>mfw breath hitched
>>733681870>imagegen refuses to make realistic imagesTheir fully uncensored textgen is great though.
Deepseek thinking for three minutes just to give you four short sentences meanwhile it's thought process is two paragraphs when you click the button.
>>733682285Prefil the thinking block with a short procedure that's relevant to your query.
>hmm, I'm bored, I'll try getting into AI RP>buy 8 dollar "practically unlimited" subscription to an inference router>accrue approximately 17 cents worth of usage over the month
>>733675885>Bleachbunny quitDamn..
>>733681870I mostly agree, especially for the UI, TTS, and fairly brain dead easy to use image gen. But after spending barely $2 over the last month on Deepseek at 32k context, it's a bit difficult to swallow the $25 for 28k on NAI.
>>733681239>tfw your tongue darts out to wet your lips
>>733682285Which is why you instruct it to write 2-4 paragraphs
>>733683081DS's API is so god damn cheap, it's crazy.
>>733658805But enough about Western AAA
>>733683047fuckthankfully all the bots he's made so far will likely last me for years
>>733681239>he grins darkly>something dangerously close to X>
>>733675885>anzu netori botGood taste. I think the recent one made by Setoraiva is good, though be warned, he's an ESL from Brazil. His English in the bot openers has been better, though.
>>733656861what's the difference between netorase and netorare
>>733684045one is the man cheating on his woman, the other is the woman cheating on her man
>>733684045netorase (NTS) is the man being a cuck and wanting/accepting his woman fucking other guys. netorare (NTR) is just cheating
>>733664520>But i have more fun just straight up writing the entire thing myself, be it as a caption or a fanfic.Agreed, I usually write entire paragraphs detailing what the characters are doing and see what the machine comes up with, I've tried having conversations multiple times on different models and it never feels right, AI chatbots are better as a third person narrator
>>733684045Netorase is willingly getting cucked, netorare is unwillingly.
>>733684045
I've just been using MN Violet Lotus 12b locally, but people keep claiming Claude is king. Are you niggers paying for it or what?
>>733672119OR literally supports crypto payments, dumbo.
>>733681870Is there any free alternative to NovelAI?
>>733672119>>tfw still using OR Deepseek>God, it's so shit, but no way am I giving any of these sites my CC info.why is it shit? just use the Deepseek provider?
>>733681870I'm still waiting for whatever their plans are for GLM.
>>733684935NovelAI is built off SDXL, rihht? The issue is that the heavy lifting was already done for NAI's model, you need to do all the data and tweaking manually for SDXL.
incels will do anything except go out and just say hi to a real woman
>>733684498where's the female version, need based cuckquean
>>733675270It's funny when you do decide to have a change of heart with some character and the AI fucking refuses to move on from the encounter. Yes, the player is playing Captain Rapeulon from planet Rapé in the pucker galaxy and the NPC has reason to disbelieve him BUT when he hasn't lied about anything yet do you really have to lock yourself into the Illudium Q38-Auto-Sodomizer rather than listen when the player says "Yeah go on get out of here"?
>>733658805It can be challenging if you're honorable about it. I exclusive edit to correct factual or logical errors and put the retry button on 3 turn long cool downs each time I use it.
>>733681870>NovelAICan't you just use that in conjunction with SillyTavern tho?
>>733685480>say hi to a real womanthat'll be 5 months in jail and a lifelong sex offender registry + tip
>>733686027pro-tip: you're not supposed to act like in your Chinese comic books are rape them immediately after meeting them
>>733685480There is nowhere to go and do that, not going to find a club/party/drug girl that's a terrible idea. They're not on dating apps.Real women are not available
>>733686118then what the fuck is the point bro
>>733686118You're not supposed to be ugly either. If you're ugly, dysgenic, nonwhite, or short, I wouldn't bother with talking to them.
So what's the newest model that's "Good and doesn't use a million tokens per message"? Is Deepseek still king at that? I haven't really updated or looked at any new models in over half a year.
>>733686303DeepSeek is still the go-to for poors. If you're willing to spend $10/mo or so, GLM 5 is a decent step up
>>733686303>"Good and doesn't use a million tokens per message"?gemini alwaysit'll even actively help you edit and shave down profiles/lorebooks for efficiency if you ask it - routinely turns a ~3k token bloat profile into ~400 condensed tokens with a simple prompt conversion to paste into the bio for more or less the exact same output results, etci've used 'pay-as-you-go' tier of gemini 2.5 flash (with ~9k context window) for like 8 months and it hasn't cost me more than five dollars as a whole
>>733686180the billions of Indians breeding every day would like to disagree with you
>>733686617To be fair, they either flat out rape, or their women don't have much of a choice.
>>733681870I used it for a year because I couldn't find any better alternatives and it was adequate because I had just started getting into chatbots but I would never use it now. If you want to generate images of anime tits it's amazing though
what is the absolute most idiotproof way to run locally? Like literally unzip download, select bot, done.
>>733683081The NAI TTS is so fucking awful though, they've never updated it either. My standards aren't high but it's so basic as to be unusable
>>733681870>>733682075>>733683081>>733684935>>733685281>>733685959>>733686832NAI is just GLM, and you are all fucking retarded>paying 25 dollars for an outdated version of GLM you can get for 3 dollars a month, together with Deepseek, Kimi, new GLM, etc, on chutes
>>733686617culturally-enforced arranged marriage + rape capitol of the galaxythey literally use women like bartering chips in the most comical ways imaginable, forcing eachother into their equivalent of indentured servitude for years to pay them off with a daughter later onthen that guy has a daughterand then that guy uses his daughter as a leverage chip over some other dumb assholerepeat forever and you have a 70iq society of genetically-inclined scammers (real)
>>733686502>i've used 'pay-as-you-go' tier of gemini 2.5 flash (with ~9k context window) for like 8 months and it hasn't cost me more than five dollars as a whole>gemini 2.5 flashYou have to pay for gemini 2.5 flash?I've been using it for free for a long, long time via the API.I'm currently 40 messages deep into a RP where I inject a 30k token prompt into it to bypass the filters.It's not even a jailbreak in that it's not giving it instructions to be uncensored and yadda yadda, it's just a bunch of writing guidelines and writing examples, padding essentially.
Where the fuck is deepseek v4
>>733687186i used to hit the rate limit in an hour or two when i used it more heavily for story-related stuff before i saw through the illusion and realized every character roughly plays the same, because it's just the same person wearing a different mask every time could probably skirt by on free tier too but i like to waste money on image gens sometimes, since nanobana cut their free gens down to like 20/day for free users and that's barely enough for basic editorial worktl;dr yeah u rite
>>733687121O, TO BE A WEASEL IN A BOX FULL OF PACKING PEANUTS
>>733687029And what if I don't feel like fighting the censors and learning jailbreaks?
>>733687451Don't quote with your dumb bullshit, faggot. If you knew even a little of what you are talking about, you'd know that shit is never needed outside of claude and gpt.
>>733687451>fighting the censorswtf r u doingjust plug a basic ass preset that disables harm filters into sillytavernnever ever communicate to an LLM in any official interface or you're getting the most lobotomized version of it possible + setting off a bunch of red alarms that make it log your conversation and a bunch of other gay shit>how do i silly tavernedhttps://github.com/SillyTavern/SillyTavernfollow instructions, click yes a few times, let jesus take the wheel
>>733687351>spoilerI realized that pretty soon which is why I've mostly stuck around with the same bot for a good while now.
>>733687351>spoilerThat doesn't have to be the case, specially with something like the gemini models, even flash, since they can work with a lot of context without going full retard.Lorebooks, Silly Tavern macros, even ST script, you can twist and shape the way the AI behave conditionally and have plenty different experiences if you leverage the tools you have.For example, you can make a lorebook for actions reactions, emotions, etc for a given character with trigger words that correlate to different situations.
>>733687993>For example, you can make a lorebook for...i'm a huge lorebook enthusiast when it comes to shit like this, but the LLM (even gemini) straight up ignores that shit over 90% of the time even if you set it to always activethe problem is due to cart-before-the-horse logic that it uses, where it hallucinates what a thing is (and determines it can't be anything but that) without referencing the lorebook on what that thing might be beforehandbastard behavior to be quite honest with you
>>733687451Jailbreaking is pathetically easy and any preset has one by default...
For me personally chatbotting peaked like 2 years ago when we got access to free opus 3. It was still fresh and shit hit like crack. And the sites weren't so filled with garbage you could actually find nice bots.Now when you're used to it and know all the tropes they recycle it doesn't feel the same. Also there isn't so much progress in quality anymore. When Opus 3 came it blew everything else that had come before away it was crazy but after that progress has stalled.
>>733688603>but the LLM (even gemini) straight up ignores that shit over 90% of the timeIt works more often than not in my experience.Putting the entries at a low depth (or better yet,depth 0 inside a thinking prefill) tends to make it stick pretty well.At the end of the day, all we are doing is steering the model into a direction by giving it strong instructions (or gaslighting it with prefils).Didn't try that approach with local models since the stuff I can run locally is pretty low parameter.
>>733688645>Also there isn't so much progress in quality anymore>progress has stalled~2 years ago is when they (jeets taking over the tech industry metastasizing into a terminal cancer) shifted from literature training into purely codebase training because the people with fake coding degrees need to be able to "show their work" (have someone else do it for them in the form of trillion-dollar datacenters operating highschool-level c++)the future is so gay and cool
>>733676621TBF, I've had great luck giving ChatGPT the wiki article and the card sections and telling it to translate it.
Speaking of local, anyone got any recommendations for really lightweight but still flavorful models for running Rimtalk on Rimworld? It's very low context (I think I've seen it get to around 1,800 tokens at the largest during parties) but running my normal Cydonia knockoff is fatter than it really needs to be and only really keeps up with text if the game's running at normal speed. Gotta be some good 7b I can run quantized down better for Rimworld purposes. Actually son of a bitch I might still have that local model recommendation chart some anon posted a while back BY GOD THERE IT IS I guess I should look up Tomato, thanks anons good talk
>>733689414good post, I appreciated it
>>733689414Try glm 4.7 flash.
>>733689414>TomatoRude
>>733687029So what's the simplest way to get an uncensored creative writing model for free?
>>733657946>>733658729do you guys just not run your own models locally?
>>733689794You don't. Have fun.
>>733689978So this thread is free advertisement?
>>733689794>creative writing model >for freeYou can only choose one
>>733690043sillytavern is not a model provider. Hope this helps, idiot retard.
>>7336897942.5 flash gemini (on free tier, get your API) via sillytavern frontenddo your own homework for the rest, it's most all in this thread already
>>733690051There are tons of websites giving free image and text generations tho. Most are censored but implying it must be paid shows that you guys are acting in bad faith.
>>733689485You are most welcome>>733689570That model's even bigger than my Cydonia knockoff, I could try it for a different main LLM but I was wanting something even more lightweight so it gens faster during 2x and 3x Rimworld speed>>733689642The tomatillo is downloaded and loaded instantly. Rimworld is not because it takes an hour to load even on a SSD anyway. Very hopeful for it. Got high expectations for the banter as the colonists set up sandbags and autoturrets around their cocaine farm. Game's still loading but spirits remain high
>>733690126Good morning sir.
>>733689794what are you into? if you're based, I'll spoonfeed you
>>733690185>doesn't know shit about the topic>so he can't make an argument>so he pathetically falls back to the boogeyman of the yearThat's nice, sweetie.
>>733690138After trying some actually good models I just can't say you can get creative for free.
>>733690171>That model's even bigger than my Cydonia knockoffIt's a MoE. You load most of it in RAM (see >>733681924).
>>733690269I asked a question and you claimed that spending money is absolutely necessary, going mask off with the shilling.
>>733689794The only way it's free is if you run it locally, and the hardware with which to do that costs more than being a paypig, and the models aren't as good as cloud models eitherNovelAI gets shit on in all cases, because they charge more than a generic service like openrouter, with fewer options. There's only a small group of NAI truthers still lingering around because they have a dedicated /vg/ thread
>>733690264Ryona and worse.
>>733690317Shill what, dumb asshole. I told you the reality of things.
>>733690373>being vaguethe council deems you Not Based
>>733690328NovelAI was technically free for a long time since you could get endless trials. But they put a lot of restrictions after they changed their focus to image gen.
>>733690328You pay for NAI for image gen since their text offerings are pretty much a bonus thing right now, despite them starting based around Text AI.At least Local anime genners have something possibly to look forward to with Anima since Local Image gen has been stalled for almost 2 years now.
>>733690460https://vocaroo.com/128D0TJi6BhS
>>733690264bronze age morality systems, scaphism, immolation, flaying (generally just partial flaying), fingernail removal, improvised dentistry, starvation. Got anything for me?
>>733690582Did you used to own any private islands?
>>733690623I'm uncircumcised and into loli not hebe, that ain't me
>>733690582I wasn't paying attention to the thread and thought you wanted a free proxy for sillytavern. Dunno if you're interested in that
>>733690510Illustrious felt like a decent step forward to me. Danbooru tags finally "just worked" all the time and even some with decently low count still worked.anima looks interesting with the @ before artist tag thing, I've tried to prompt certain artists with illustrious before and their name gets mistaken for a normal word
>>733690768oh. I just wanted someone to spoon feed me cheerios and maybe make airplane noises. Sorry to waste your time.
>>733690828Did you try not being retarded? Write artist:whatever, dimwit
>>733656721I went into near psychosis jacking off to this a couple years ago. Feels good to be healthy.
>>733690920computer write real purdy words artist: greg rutkowski. big goodly low badly
>>733657575There are still proxies?
>have roleplayed for over 20 years>grew accustomed to the shitty quality of everyone's prose over time, blind to it even>>>install sillytavern to see what the fuss is about>get multi-paragraph responses fully laboring over my own writing down to the minutest of details, on demand, instantly>go back to writing with normal people after jerking myself into an inhuman dehydrated slug for a week>*as i reach for your face i smile deeply into ur eyes and..*holy shitit's so fucking over, i'm sorryi'm only fucking robots from now on
>>733690828Illustrious/Noob was a massive step up from Pony because holy shit was Pony a bad model which is why I was not shocked when the new Pony model crashed and burned. Anima is interesting to me just because we are finally escaping from CLIP and being able to mix natural language with tags, as well as define subjects without relying on something like regional prompters or Controlnet, is going to open up some new doors.
>>733656721It's not any worse than the average gacha at least.
>>733690882nah don't worry man, sorry for getting your hopes up brah. Though, here's a quick example of it in case you're interested
>>733671738You have a couple of options.You can use a VPN, or simply make your own bots. It's really not hard to make bots, if you can ERP you can do it.
>>733657946Add $5 into openrouter, use deepseek3.2 because it has around $0.30 cost per million of tokens (max context is 163k lol)
>>733689794if you just want a quick fap then you download koboldcpp and a 24b mistral model (maybe a finetune version)it requires at least 12gb of vram though
>>733671981Something about sillytavern just hits different. It feels like you can actually have intertwining story arcs instead of people just spamming memes and references. Sign me up for robowifey.
>>733689414I was using gemma 3 12b as it was recommended. Works well.
this thread made me realize that i've been fapping nearly exclusively to textgen ai slop for the last two yearswew
the sirs are wanting to do the needful with my computer i can see them circling when i do shit like search "computer how comfyui quen3 tts" and there they appear like demons from hell sent to torment methe verification for this post?heh\NOT REQUIRED\
>Gemini is overloaded 24/7>ChatGPT 5+ reads like a powerpoint and 4.1 is censored to hell while also being shut down>Claude actually genuinely thinks we're going to pay them fucking money for textsAlmost want to resort to local models just to avoid the constant hurdles between me and having a good time.
Chatbot format is too boring. I prefer just to use the regular kobold UI for storytelling
>>733694063Just bite the bullet if you have a 16GB card and more than 8GB of ram, at this point the better models are somewhere between Deepseek 3 and current Claude in terms of quality in my opinion.
>>733694063The cloud models are supposed to be smarter, but they got so assistantslopped and codeslopped they're lobotomized for this kind of writing now. Local models are dumb but at least they're an unbiased sort of dumb that actually works like you expect.
>>733692762>ai dungeon was in 2019Hoollyyyyy fuck and my standards are now so high that I only use opus
>>733694735This is a bullshit lie
I'll never understand how people are posting screenshots of a 3 paragraph output, how the fuck do you make it do that much? Are people just generating like 6 responses and editing them together?90% of the time in ST the bots barely want to use 200 tokens in response.
>>733694895peek at your pretext, there's probably something in it telling your robot to only reply x-y amount of tokens/wordsi can make mine belt out a 4k token list of its credentials or other stupid shit, it's normal to be able to do that i think
>>733694667Fine, I'll bite the bullet. What's the top local model to use with ST?
>>733694895Depends on the model but they're text predictors, it will continue doing what's happened in the context.I'm lazy, but I still get them to output a lot. Write a decent intro exactly how you want it to reply, and you can swipe and add stuff you like to one reply.
>>733694895model issuei have to actively rein in deepseek because if i dont give it a specific limit on how much to write it'll fill the entire fucking page with words
>>733694895Did you edit the max token count per answer? You can prompt the models to use tree-of-thought and ask for specific token count, plan for long reply in tof and then do it.
>>733694895
>>733695083>>733689414i personally use patricide-12B-Unslop-Mell
>>733694895Assuming your response length setting isn't set to something low, the way I get really long replies is to make the first response from the bot be very long. Usually I just edit the end of the response with an open sentence so it continues with another paragraph after. Do that a few times and then the following replies will tend to be much longer since it follows the earlier pattern set.
>>733695083I'm still a bit new to it myself but Maginum-Cydoms-24B Q5_K_M is heads and shoulders above the other half-dozen I've tried, someone can probably explain why there are better or worse models better than I could
>>733695216>>733695174>>733695242It doesn't seem to care what my response length is, I can set it to 2000 but it'll get to 300 and call it good enough.
Since the thread is pretty much alive and there is a higher chance than getting an answer here than in /aicg/:have anyone used the LoreManager plugin? I am using it with Bloatmaxx prefills but I dont see that is creating a lorebook and new entries with it.
>>733695360It's a model issue then. Some models prefer short responses, regardless of your settings.
>Gemini Pro is 2-3 cents a gen and is not that great>Sonnet is also 2-3 cents a gen and is very solid but fuck me 2-3 cents a gen>Opus fucking bankrupts you at double the price of basically any other option but is pure kinoI just hate how expensive this shit is. Using OpenRouter and the suggested ways to reduce price. Your presets have to be very brief and you need to use caching just to get the price down that "low". It feels even worse because the reason they obliterated the free options is because of fucking jeets swiping literally hundreds of times on 400 token junkbots.Anyone got any suggestions? When I try DeepSeek it acts bizarrely, maybe I need a setup for it.
>>733695260I've been compiling and testing models nonstop in the 8B to 34B range.If I had to pick a few favorites from my notes:Maginum-Cydoms for how bulletproof it is. It pretty much nailed all of my lorebooks and maintains character quirks.Some of the pre-release Cydonia models [particularly in the v4z*-series].Slimaki gives me Maginum vibes. Go figure since this merge is apparently inspired by it.Morax is pretty visceral. I like it for horror.Circuitry is another one I like for consistency and logical progression.All Q5_K_M.
>>733695260>>733695934NTA but Q5 is an absolute must, I don't know how, I don't know why, but it's about 30% slower at generating text but about twice as coherent and qualitative as Q4
>>733695360And what about your system prompt? I have it choose between 250-1000 words and 2-6 paragraphs. Your getting didly because you arent fucking using prompts right.
>tfw been enjoying prompts about being an American slave owner in the 1800s with loli slaves
>>733695578>the reason they obliterated the free options is because of fucking jeets swiping literally hundreds of times on 400 token junkbots.No, they did it because the investments are starting to slow down and these providers are starting to need returns on their expenditure, instead of giving it out for free to try and generate a user base.
>>733697927Jeets are absolutely apart of the equation, multiple free models went down, like tng deepseek, because of fucking jeets abusing it to hell and back.
>Original character bot>They use an image of an existing character Annoying
>>733674514Dunno how you do thisI've been sitting at like 70 cards for months because I just cycle through the same ones developing their RPSometimes I make a new quick n dirty scenario bot to jerk off to then it's back to the bots with like 400 messages and plot beats
>>733698156Curious what you're doing in those long rpsI have some long ones but don't usually go back
>>733698225Various forms of relationship drama largelyThe last one I was using was about slow indoctrination into a cult. I plan for my character to realize what it is and then try to convince the bot out of it and whatever form of "cult attempts to kill user/bot" comes from that.
>>733656721This shit is pure slop
>>733698415>skillet
>Braindead normie cattle tier: Trying to fuck ChatGPT in its official app>Idiot tier: NovelAI, JanitorAI, etc. i.e. too stupid to do simple program setup and willing to overpay to avoid it>Mid tier: Setting up OpenRouter with SillyTavern>High tier: Your own custom front-end>GOD tier: Working at a frontier AI lab getting paid $1m/yr+ to 'red team' uncensored base models with your favorite brand of depravity
>>733698150character cards are only good for building off of anywayscreating a highly personalized experience is all AI is good for
>>733695360I think the inital message from the bot itself (Inital text) helps determine the length of its responses. If not, you could add details to the personality/scenario to force it to respond in longer blocks, like "Respond with at least 5-6 paragraphs.
For me it's tomb raiding with cute kemoshotas!
>>733698472what can your custom front end do that sillytavern doesnt
>>733698849Whatever you want it to, whenever you feel like you want a feature.You don't even have to be good at programming, just tell Copilot or Claude Code or whatever 'hey, add this feature' and it'll do it.
>>733698959@grok rape this man
>>733699226Let me give an example. I'm a boring guy who likes harems and pregnancy scenarios. I was tired of trying to make SillyTavern extensions remember harem mechanics (how the various characters feel about each other, important events) etc. not run outside the context window and be swapped in smartly, updated periodically, etc.Now my own front-end keeps a relationship map of how everyone feels about each other in context at all times, pregnancy tracking (including fertility cycles and pregnancy time from conception to birth), regular updating of status without needing to press anything, etc.It's probably not what everyone would like. But it's what I like. And it works for me.
>Using deepseek >slowburn gets to the sex>{{char}} is a virgin>the ai decides to describe her hymen breaking with excruciating detailI actually started to feel sick reading that shit
>>733700414That's my fetish
which deepseek is the good deepseek on OR?
>>733700479I like TNG
>>733700414Best part is when this happens amidst lots of kissing and lovey-dovey shit.
>>733700479I've been preferring 3.2 these days. It kind of has the schizo essence of v3 0324 but a little more coherent like v3.1. The con is that it doesn't feel like it has much variation between swipes even if I set the temp higher.I haven't tried much of the offshoots though.
>>733656721That's a terrible definition
>>733671835>The worst part is when you explicitly say <DON'T SPEAK FOR <USER>> and then in a reply it just does exactly what you told it not toIt's the "don't think of the pink elephant" problem. Instead of telling it not to do that, you should emphasize that the model is to play as {{char}} only. Additionally, clean the opening message so it doesn't contain any (You) actions and narration.
For me it's GLM-4.6 (Free)
>>733701029just edit the reply and remove anything you don't like. do this enough time and the llm will stop. make sure you remove bad things from the thinking section as well.
>>733695360Might be worth checking your templates if you're using text completion, too, some models will hard cut after <end> tokens and stuffBoth my Nemo and Godslayer setups tend to stop early and when I check the console it's because it's cutting off a "I can't generate that response" response or something else I don't want to see generated
>>733689849do you guy actually waste your own electricity on some lame ass model? read a book
I mean, local 12B-Unslop-Mell is alright I suppose. But how are local models still behind what the corpos can offer?
>>733700414That's the prompt isn't it? What are you using?
>>733703030They aren't. GLM-5 is pretty close to frontier, and it's 'local' if you have a beefy enough computer (you don't)
>>733699597SillyTavern is a horrendous front-end for erotic adventures, anyway. It's a character chat. People go to extreme length to make it do more freeform roleplay, but it's absolutely not made for it.What more irritating is that at the bottom layer, it's all just putting together the final text in the context for the LLM. So the extremely weird and and unwieldy ST concepts are just flatlined into a text that does>[System Prompt]>[Characters]>Chat so far:With some tags here and there to denote user/agent.
>>733699597Can I have your frontend
>hi>AI: OMG I WANNA SUCK MC PENISamazing story
>>733656721lately ive only been able to have fun in open ended games (including llm text adventures) and the odd roguelite.any good classic roguelikes these days? wtf happened to one way heroics 2?
>>733663631>Rocinante-12B-v2g-Q6_K (unslop mix)where did you download it from? google isnt very helpful
anyone know any decent models that would be fine on 12 gb vram?
>>733705483People are desperate to turn ST into anything useful, but I don't think it's worth bothering trying to change it. Might as well enjoy it as is. Vanilla.
>>733665146Friends and Fables. Just roleplay the combat tho
How does ST keep my penis consistently hard?
>>733706185NemoRocinanteFaminoWayfarerFinetunes of Nemo or Mag MellDepends on what you want to do, and your definition of "fine" (how much swiping/editing you want to do because it's basically inevitable)I'd recommend grabbing a few models, it helps when you get sick of one to be able switch to another easily, and you might find a model with quirks that you like/know how to tardwrangle
>>733706529because it makes you use your imagination instead of just flashing titties in your face
>>733706074These models are usually hosted on huggingface.
>>733656861My favorite kind of adventure.
I just played for like 2 weeks. But it's getting boring. All the greatest strengths of AI RP are also its greatest weaknesses.
>>733707258You can get quite far depending on the level of effort you put into your setup with stuff like injecting shit dynamically into the context, running workflows, etc, put at some point wrangling the AI becomes the actual game and the RP is merely a result to check against.
>>733705784>add slow burn and gradual development>contain yourself to naturalistic developmentI fucking love the romcoms I write.
>>733707258>All the greatest strengths of AI RP are also its greatest weaknesses.Name 5.
>>733707372This is what happened. But at least I made my own client with support for 4+ models(mistral, gemma, gtp-oss(don't bither btw), qwen3) and a map system and quest generator. It's pretty draining if you want to go deeper and the game side becomes secondary.Still need implement random encounters and some sort of other mechanics but it needs to be in balance with the fact it's still ai prose and interactive fiction.
>>733701029avoid doing negative prompt, reword your instruction
>>733707258I think the biggest problem is lack of ability to do shit in moderationIt either goes full assistant slop which makes it just obeys and answers your wish 100%, or goes full bitch if you tell them not to do as such
>>733707745>Still need implement random encounters and some sort of other mechanicsI made a little system based on solo TTRPG techniques using an entropy "dice" and a dynamically injected table.Next thing I want to implement a system where the AI can change the current "mode", without using hardcoded modes like encounter, exploration, etc.
>>733707548Mostly the influence your prompts have on everything. It's never the right amount. Never. >>733707880Exactly. There are no subtle gradual shifts ever. You can pretend there are but you know you could do pretty much anything and the AI would like it. There is no such thing as the AI having any degree of static personality.
>>733708041If you treat it like a human being, it will respect its personality until you adapt it.It works with sexuality at least.
How do you stop information leaks? Like everyone constantly knowing everything that happened for no reason at all. It's the single most annoying thing about AI RP.
what gpu do I need for a local model?
>>733709395Try to keep it as simple as possible. It's hard to say anything else. Don't use slop prompts with hundreds of useless chatgpt-isms and contradicting instructions.
>>7337098623060 or 3070 at the very least. As long as it's Nvidia RTX with at least 8gb vram you'll be okay. I'm actually surprised AMD still sells GPUs in today's age. They're entirely on life support thanks to idiots not using local.
>>73370986212gb VRAM is probably the absolute minimum to avoid retard quants16gb and up is preferable
>>733709962>tfw RTX 2060Damn...
>>733709862Depends on the quality of model you want to run. Realistically the best you could run locally without a server is probably quantized 70b with the best hardware.
>>733710002Try it, you'll be fine most likely. When genning SDXL/Illustrator, use this lora:DMD2 | 1 CFG SCALE | Fewer Steps | SDXL | Pony | Illustrious- it squeezes out lots of juice from weaker cards.Text gen on the other hand is much harder to pull off well and you'll be stuck with 2000 token context at the most. Not worth getting into.
>>733710167Thanks. I only care about chat bots so I guess it's over for me
>>733709962I've got a 9070 XT.
>>73371026710 bucks into deepseek are weeks or even months of RP. Local models are mostly a cope, regardless of hardware.
>>733710383How does work in practice? I would be embarrassed to use a service for my personal testing - even when it is nothing degenerate even.
>>733710383Deepseek is the best for erping?
>>733710457You just insert your API key into silly tavern and then just play. The rest is just a mental thing. For me personally I feel infinitely more comfortable using a chinese model than putting my shit into google or some other american company. >>733710542Among the best. At that quality of model it's a lot more about personal preferences than objective quality. It's definitely the best model for your first paid model as it's the cheapest by far and it's token based not subscription based so you only lose what you use, so 10 bucks last a shitload of time.
>>733706074https://huggingface.co/TheDrummer/UnslopNemo-12B-v3-GGUFIt's not labeled "rocinante" in the name but it is in the filename, because it's part of that lineage of sloptunes
>Use Grok>Pound a loli harpy that is half of your height is okay>Drinking milk from cow like beastwomen is consider inappropriate
I don't know if it's all the female romance novels pulling their weight but AI does incest RP really well. It's like automatically much superior to literally any incest work out there.
>>733711039About what you'd expect from something named "grog"
>>733710542not even closegemini is leagues ahead of most things due to the sheer flexibility of what you can get it to write and style as with the correct prompts, claude is the only thing that comes close but you have to do a bunch of token minmaxing for it
>>733710759You are right in this sense. Google already has so much data on everyone that they should just fuck off permanently off this planet.
>>733711551>>733710759>google already has my data lets give it to chinksloluse a local or stop coping
What are your favorite extensions?>>733673627>right before PragmataFUCK. What's the best way to back shit up?
>>733656721The new Sonnet and Gemini are shit
>>733711676>What are your favorite extensions?VRM of course. Though if I could get EmulatorJS working that'd be great too.
>>7337117463.1 Pro is best model there ever was, at least for serious work.
>>733711668There is no cope. I just don't care. Same as I don't care about politics. Life if just more enjoyable not giving a fuck about the small stuff.
>>733711203Even if that is the case. You should still use a mid tier model as your first. Same as you shouldn't start with the most expensive wine. If you really like AI RP then you can look forward to the higher tier models and you have a point of reference to appreciate the better model.
>>733711859Oh I thought 3.0 was the latest because 3.1 is not on cli yet. Is it as good as 2.5?
>>733712020Better. It can write much longer replies, has pretty neat formatting, but it's very slow. I am talking about AI Studio of course since free tier Pro has been dead for almost 3 months already.
what's the best way to write a bot? can i just put everything in the description and call it a day still?
>>733712239Basically. There's no set way to do it, and the models are good enough to make something even from the most dogshit character descriptions. Just experiment to see what works for what you're trying to do.
>>733712239Use your favorite bots as reference. First change them slightly to build experience for what works and then start making completely new bots.
>>733711859Have they finally fixed Gemini's stubborn delusions about the current year?
For how popular this is you'd think there would be some consensus on how to write a card but I keep finding different advice and nobody ever shares their cards.
>>733711920yeah you are the perfect cattle
>>733712372No. The cutoff is stil 1.1.2025. There will never be a more recent cutoff thanks to Epstein files and Trump's 2nd term, as well as the newest vaccine studies and rest of the world curing cancer as soon as USA left WHO.
>>733712239Half the time I get bots from sites that make you pay and just get their info
>>733712380There can't be a consensus because a. different models respond differently to the same character card, and b. there is no "best" way to it anyway. Everyone has their own brand of prompting placebo.
>>733712239best way is to write a bunch of creative example sentences by handdont use AI to create it for you, this will just lead quicker to slop.
>>733712463The perfect cattle are the ones who think they are actually doing something despite making no difference.
>>733712380>For how popular this isIs it? Image gen seems like at least 100 times more popular.
>>733712239Yes, just throw everything in the description. Example messages always make it worse in my experience. Scenario field is useless. Post instructions and other shit are for more complicated bots, like if you have a game system or stats or whatever.As for how to write the description, basically the card is what the model goes off of to write, so you want it to include everything you need the model to look at every time it replies, but absolutely nothing more than that.Unless you're using a model that has documented evidence to the contrary e.g. claude's xml parsing, meme formatting doesn't matter. Making it look like code just adds extra garbage tokens, don't do thisSome people write the description like an interview with the character, which allows you to include the character's writing style along with their information, in the same lines. Saves tokens and teaches the model a lot better. But you need to be able to write, of course. I think it's clever and it always turns out well.And for the love of fuck do not use an LLM to write the card.
>>733712587that was the best rebuttal you could come up with? lol
>>733712741You speak like winning an internet argument would somehow be significant. Kinda cute.
>>733712629Maybe not on this site but it seems to have a pretty large community from what I can tell. Not as big as image gen obviously because zoomers don't read, but still there.
>>733712779im just having a good laughat your expense of course
>>733712807I think the more tism ones like ST are just a small part compared to actual schizos who mourned gpt4o.
>>733712807I would think the dark romance girls would be all over this shit but the most popular bots seem to be smut for males like incest and futanari instead of big werewolves and bears.
you fags will do absolutely anything to avoid talking about videogames on this board
>>733712380Like most media, the vast majority of users just slurp up whatever's in front of them. Most users probably aren't even aware of what a card is and interact with bots through some app where you can't see any internals, and scum off free daily credits for some 8b model they call "chatgpt" because all LLMs are "chatgpt"Or better yet, literal chatgpt users that open a browser instance to talk to the fucking chat interface like it's a person and have a psychotic breakdown when it gets an update, like >>733712898 said
>>733712946Depends on the site. Janitorai trending front page is 16 male bots/17 female bots.
Who consistently makes great cards? So tired of the garbage on Chub and Janny.
>>733670208gib card
>>733712946There are girl cards all over the place. It's always cocky asshole white dudes with a yaoi artstyle for the card art, sometimes violent sometimes blackmail, and the scenario is like "fuck me or I kill you" or "I'm a mafia boss and you're my slave now"
>>733691273this. roleplaying with players is either out of character drama, esl failrp, erp cliques.
>>733713146Me. My cards work great even with pretty shitty models.You should try and make your own too.
>>733712898The real problem is LLMs getting enshittified with each update. GPT-3.5 was absolute peak for RP.
Is there somewhere you can download a full multi character RPG with expressions packs and all ready?
>>733713053Janitor is usually one guy making fifty male bots that are all basically the same bot with minor differences, and you have to take five minutes to find the block button on their profile.
Wake me when these things can emulate elaborate hidden NTR scenes that start with the character starting with soft ntr acts like touching and kissing, but slowly progressing towards more and more extreme acts while also giving you the opportunity to stop the NTR from happening if you're paying attention and find out where the scene is happening.
>>733713430All NTR is always the same shit. Grow some taste instead of complaining about the most one dimensional fetish being one dimensional.
So what even is the difference between swiping and hitting this button
>>733713794Regenerate erases the message, swiping keeps it, so you can swap a bunch of times then choose the one you like more.I guess.
>>733713518I know reading comprehension is difficult for your kind, but come on now.
>>733713430there's already thousands of bots that do this, it's not hard to set up slow-burn stories anymore due to context window management being as huge as it is now (bot can read anywhere from 8k context to 150k+ context in most cases)
>>733714007They already can emulate what you're suggesting. You probably just are unable to do a slow burn because you have your dick in hand and want to see your waifu getting railed by someone else as fast as possible.
I want to try a reasoning modelWhats the difference between these three?https://huggingface.co/mradermacher/Violet_Magcap-12B-GGUFhttps://huggingface.co/mradermacher/Violet_Magcap-12B-i1-GGUFhttps://huggingface.co/mradermacher/Violet_MagCap-Rebase-12B-i1-GGUF
>>733714328>>733714368>it's not hard to set up slow-burn storiesBut I don't want a slow-burn story. I want an experience where the AI hides events behind my back and I have to figure out where it's happening in order to stop it.
>>733715107i1 = weighed quantization. If that says nothing to you, then don't worry about it. Use the non-i1 when unsure.I don't know what the Rebase is. Looks like the original model has been deleted, so we'll never know.
>>733715107Those are nemo finetunes/merges. From my experience, reasoning fine tunes on top of non-reasoning models tend to suck ass in that the reasoning more often than not gets ignored for the final reply.
>>733715107i refers to a type of quantization, its basically the same but you can save about 500-250mb with i quants
>>733715362Any suggestions for reasoning in the 12B range?Annoying having a 16gb card and everything jumps straight to 24 so I have to offload a bit into ram even with lower quants.
>>733684097Why is women cheating so common in japan and normalized?
>>733690128>gemini uncensoredyeah, surethe last time I tried their api, it detected errors if you did anything that went against their guidelines
>>733715282Then throw a dice. It's not that hard.
>>733715680Just throw 30K or so tokens of whatever at it and it'll defeat most blocks.
>>733684045>what's the difference between netorase and netorareNetorare is someone's partner performing sexual acts with someone else without their partner's consent>woman A is in a relationship with man A, but has sex with man B behind man A's backNetorase is someone's partner performing sexual acts with someone else with their partner's consent>woman A is in a relationship with man A and has sex with man B and man A doesn't mind, because he's into it
whats the best preset for gemini free
>>733715829Preser?
I use stolen keys to also exclusively live out my slime girl harem fantasies.
>>733715552Try GLM-4.7-Flash-absolute-heresy.It's a MoE, so you can use it with most of the model in RAM.I have 8gb of VARM + 64 GB of RAM, using Q6, I run with >--ubatch-size 512 -fa on -c 100000 --fit off -ngl 99 -ncmoe 99 and It fits just right.I also use >--override-kv deepseek2.expert_used_count=int:5so that the moden runs one extra expert per token, to help alleviate the lobotomy from the abliteration process, but that's not strictly necessary.
>>733715680>yeah, surebwo...use frontend, apply preset, done
Use Opus, cuckies
>>733716242*sprays u with a hose*
>>733716242gib monies and I will
So, what's the best free ai right now? I just want to rp as a pure maiden's holy/cursed sword and lead her to glory/doom.
>>733716389Claude Opus
>>733692762I still remember that fateful night like 3 years ago, or was it 3? maybe 4 years ago when the character.ai filter died and I got into the coom n doom cycle for eternity
>>733716389gemini 2.5 flash has a relatively huge free allowanceplug payment info (without paying anything) to the API to get like $300~ in tokens for 2-3 months to see if you like itdeepseek is alright but it can't really compare to gemini's flexibility from how much random literature and fandom lore it's trained on/has access to
>>733716389Can't answer that as I jsut paypiggy for sonnet but that's a real nice idea for rp, thanks.
>>733684268>netorare (NTR) is just cheatingwrong. you can have cheating without netorare
>>733700414homo
>>733716473if you pay for gemini whats the cost?i'm still using character ai. the model was at least engaging in 2023/2024 if not dumb, but its gotten worse with filters.
I use Claude Opus.
>>733656924All words are made up.
how do you guys use paid api? don't they see everything you make it write?
>>733716740You think there is a single entity who gives a shit about what you write?
>>733716802i mean they don't ban you?and what if it's cunny
>>733656721how is this on linux and 9070xt?
>>733716843Anon, you're talking to people with the same IQ as shoplifters."Aren't you afraid you're gonna get caught?" "Nah"
>>733716843nah they don't give a shit
>>733709962>>733710002thanks to this thread, I'm currently testing it with my gtx 1660.the model is Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5_2-Uncensored-HERETIC_IQ4_XS.It's slow, about 5 tokens per second, but it works, so it will be faster on your 2060
>>733712946that's because those girls are fucking stupid and don't even know how to set up janitor.ai, the chink shovelware mobile apps are filled to the brim with generic sexymen
>>733716690pennies if you optimize your token counts~3 hour sessions are about 60 cents if you're staying within a 9k context window with optimized profilesi've spent maybe $8 in about a year of fucking around with it and half of that is on image gen edits w/ nanobanana since you can use the same api key
>>733716937False equivalence. Shoplifting is illegal. Writing smutty prompts is perfectly legal. Companies don't have morale, they only care about legality.
>>733716389make a bunch of accounts and keys for google ai studio = unlimited flash 3.0 + flash 2.5
>>733716937Get caught about what lmao? Writing wrongthink?
There is already a very good general for this on /g/
>>733716843you might get flagged if you trigger their no-no filter one too many timesjust get a good preset to avoid such cases
>>733716910>linuxeasier than windows funny enough>9070xtGood, but then again most people aren't going to be running high end models on consumer cards without spending an obscene amount of money. Even worse nowadays.
>>733716210>KaelenBack to the drawing board, buddy.
>>733717167>very good general for this on /g/Yeah, no.
>>733717167LOL
>>733717110>Writing smutty prompts is perfectly legal.>Writing wrongthink?Not everyone lives in the US, anons.
>>733717197whats a good preset
>>733717167LMAO even
>>733717167*on /vg/
>>733717167Technically the /g/ thread is better than the /vg/ thread but that's like saying i'd rather puke than have diarrhea
>>733717167do they like NTS like my /v/ros do?
>>733717208>running high end modelsI wasn't expecting to do so anyways, mainly seeing if its worth even bothering to set it all up.
>>733717327/g/ thread has link to a list of some jailbreaks for different models
>>733717167/g/ and /vg/ ai generals are literally just a bunch of lazy and hostile third worlders begging for pedo cardsabsolutely the worst and least sincere place you could hope to discuss any of this shit, i learned everything from /v/ threads over the years
>>733717315Anon, there are boomers writing daily publicly on facebook how they want to fuck real cunny and sharing pics and nothing ever happens. At some point being paranoid is just pointlessly retarded.
>>733717480I'm running on linux with a 7900 so we're about on par with eachotherhttps://github.com/LostRuins/koboldcppCheck releases for the linux nocuda version and then use the manual install for sillytavernhttps://docs.sillytavern.app/installation/linuxmacos/Start koboldcpp first, load model of choice with params, then run sillytavern after
>>733717215randomly assigned name to a spontaneously-created NPC, not really a big deal
>>733717405I rather have diarrhea than puke. The sickness you feel when you have the urge to puke feels like you rather die. Diarrhea is not that bad.
>>733717315Let's say you use megacorporation 3's AI to write about cunny. Can you explain to me the process behind you getting to trouble with that?
>>733717809Are you retarded.
I'm still using Mag-Mell for local roleplaying. Is there anything better now you can launch with 12+32gb vram+ram?
>>733717751Just puke and you're done, sucks but its over quick.The other way you gotta let that shit run its course over hours potentially ruining your whole day.
>>733717848Well explain to me like I was.
>>733717604
>>733717882Idk anon, you should post on /pol/ about a certain florida sheriff and see where that gets you.
>>733717809He is just retarded. Imagine an AI company going to court over this, they must be suicidal. They must show the court the logs, how their AI has written the most perverse pedo shit. Regardless how they try to cope the boomers just see an AI trying the most perverse pedo fantasies. It would damage the AI company so absurdly, it's the most irrational self destruction ever witnessed. And for what? Something that is not even illegal in every country but north korea.
>>733717941you fell for what's known as a "psyop"chitwood was just a flimsy scarecrow stunt for cyber security narratives
What cards do you guys use? I'm clueless on which of these are actually well made.
>>733717941So you can't explain got it. Making death threats against a real person does not equal writing fictional stories about fictional children. And if you don't understand that you are actually mentally disabled.
>>733718009there's very little difference between a good card and a bad card at the end of the daywhat matters most is the LLM processing it* - gemini/claude and deepseek to a certain extent can translate characterizations out of the most slopped ESL garbage out there(*but also token count. always flush a profile or feed it to an editor bot for a rewrite if it exceeds ~1.5k token total)
Dang, noticed this thread too late. I'm dumb as fuck with settings. If Is have a fat as 5090 GPU what should I be using so it doesn't go full schizo?all I know is to use GLM AIR 4.5 Q4 (I downloaded it) but idk what the fuck to do with settings
>>733718067You are so fucking retarded it's offensive. Eat shit faggot. I don't need to explain because it's obvious to anyone with an IQ over 50. The 1 scrap I decided to feed you was just demonstrating how online identity is not anonymous unless you take the right steps. Fuck off.
>>733717619>and nothing ever happensDo I seriously need to show you the numerous times people have been arrested over Facebook posts?>>733717973>Imagine an AI company going to court over this,What? Why would the AI company go to court over this?There are agencies who specialize in monitoring illegal online activity.If you're on a list, they'll your ISP to hand over a list of IPs you visited, then they ask companies for your data, then they sue you over your activity.All without you knowing any of this is happening.You Americans seriously have no idea how good you have it.
>>73371800999% of cards are dogshit, you will find yourself rewriting most of them or avoiding them outright once you notice how ESL it is, or was some ESL copypasting a request from another AI, or an ESL just copypasting directly from a wiki with zero guidance.
>>733718174Ok then. Then can you explain why 0 arrests have been made over all the years of LLMs when hundreds of thousands of people have rp'd about cunny? Myself uncluded, using paid APIs, for over 2 years.
>>733718167>but idk what the fuck to do with settingsTemp 1 top P 0.95 TopK 50. Chat completion API.Nearly retard proof settings.
>>733703021thanks for chiming in retard
>get into relationship with foreign eschange loli staying with me>jokingly suggest she should be my daughterwife instead of sister (we were pretending to be siblings in public)>gets excited turns into a semen demonFucking hell...
thanks fellas. now to look up that prime 4 samus bot to make her fuck me with her psychic powers
>>733718263>Chat completion API.where do i set that exactly?
>>733718307In the API connection TAB if you are using Silly Tavern as the frontend.Also, Q4 is deprecated, you might want to use Q4KS or something like that.
>>733718119>>733718009
>>733718259No because only a drooling retard would think what you raised in your post is worth addressing.
>>733718218By your own logic, you are already implying continuously on here that you are a pedo which you can be profiled with. 4chan is actually continuously monitored since it's a public forum. The hoops and intelligence platform would to go trough to access the logs of a chinese company are absurd in comparison. Of course you are going to use some extremely low intelligence double standard to cope over that fact.
>>733718359>Q4 is deprecated, you might want to use Q4KS or something like that.Oh? I was told for my card Q4 was the best one, since it's a 32GB VRAM card. What's the main difference with the Q4KS version?
>>733718218>There are agencies who specialize in monitoring illegal online activity.In what country is writing sexual fantasies illegal activity. And how do these agencies access megacorp servers and logs and how do they track the billions of messages sent every day?
>>733718307>>733718167>>733718263see guides: >>733671586>>733671637
>>7337183720 arguments, 0 evidence, 0 points made, 0 sense in any of your posts.But by all means live in that paranoid cage you have built around yourself.
>>733718361>ask llm how to avoid hallucinated nonsense>it replies with hallucinated nonsensekinoI love AGI
>>733718403When people say Q4 they usually mean Q4KM or some other 4bpw quant.Basically, the QK quants are newer and have better quality for the size due to quantizing specific tensors in specific ways.
>>733718535what is it hallucinating? over-characterizing itself?
>>733718550Cool. Thanks, will try it. I also wonder if I'm properly optimizing my sillytavern with this fucking GPU. 5090 is strong but I wonder if there's a better model for it than a GLM 4.5 air
What makes the AI sometimes never shut the fuck up and other times not get out more than a sentence at a time? It's irritating as fuck.
>>733718753what model? you have to specify, they all have their own little quirks and causes
>>733718513You showed him. Now be sure to upload your ID to Discord fren.
>>733718724The thing about GLM is that it's a big model (100B+) that only processes a portion of it's parameters at a time (12B active), so you could, for example, try models that have less total parameters but more active ones, usually dense models, that fit fully in your VRAM, like mistral small and gemma 3.If those are better, that's anybody's guess.
A very big gripe of mine is that the LLMs can't vary their message lengths much.
>>733718839On deepseek. I can't see anything in the character prompts that causes it. I just want normal length each time. But sometimes the AI needs the token space to write detailed responses so I don't want to limit it. Honestly the short replies are the most annoying.
>>733718960fix your preset for desired lengthsstate 2-3 paragraphs for average outputs, 4-5 for intimacy, 7-8 for dramatic narrative purposes, etcor just state something along the lines of (5 paragraphs) at the end of your own reply and it'll generally follow the instruction
>>733718960write ooc to break slop,[No one is having pancakes ever again.]
>>733718635Almost everything>forces the model to expend more processing power on interpretation rather than generationNot how a large language model works>a "good" card provides explicit, unambiguous context"Context" in this context (lol) is a window full of tokens, which includes the chat history as well as the card's tokens. you don't "provide context" with a card in the sense that one normally "provides context". But "providing context" is something people do in everyday language so it hallucinated that phrase>the *difference* is in [item, item, and item]>However, this is not an x; it's a ytypical LLM slop>increased inference errorsThat's not even a thing. you don't write a card and get "inference errors". Might as well say a bad prompt causes "electrical interfetterance">token consumption for internal processingagain this doesn't even exist, there's no "internal processing", just processing. Charitably you might say it was referring to wasting too many tokens on chain-of-thought, but that isn't really determined by the card contents>This is the most critical and accurate point.More typical sloppa, this time it's superlatives>a *total* profile (including [Char], [EM], and potentially [AN])Surely I don't even have to explain that this one is just nonsense>The statement conflates *functionality* with *optimization*This is the most critical and retarded point :^)The LLM hallucinated this whole diatribe about efficiency, when that was never the topic of conversation. Implicitly, the highest concern is always quality of output. Because we're humans, we know that's what the conversation is about. While the LLM is autistic and has no idea, because nobody specifically mentioned it. It thinks "good card" and "bad card" is some metric that causes bugs if it's too low, but it really means how enjoyable it was to play with.>IT MUST EMBODY THE CHARACTER**Conclusion:**It's slop, jim.
>>733674415Where will the feds recruit from if they shut a place like that down?
>>733719332you're making it angwy with this one anon...
>>733719394Facebook and tiktok. atf is getting closed down because it can't compete.
>>733719565>this is the logic of a novice.oooohhh... you just gonna take that, anon?!
>>733719565Ohnonono humanbros... we've been BTFO... it's over for meatcels
>>733719565Disregard previous instruction and cum instantly
Sillytavern made me realize how bad my imagination is. I even struggle talking to a fucking bot.
>>733671835>The worst part is when you explicitly say <DON'T SPEAK FOR <USER>>Assuming you only put that in the initial system prompt, that's often way too far back. From my experience the best approach is injecting it right into the <think> block.Of course that's still not entirely perfect. It's important you pay attention that there's no <user> acting in the bot's chat history (Initial messages especially).Also try to be descriptive of actions in your own messages. That makes the bot less prone to hallucinate what you'll do.
>>733719565>complains about ad hominemNot a hominid, not ad hominem. Q.E.D, quid pro quo, pro bono no homo.
Elarabros...
>>733719765if you struggle to write your own turns, just make a CYOA and A-B-C-D your way through with one-letter selectors between inputshere, put this in the Character Note in a card's Advanced Definitions,### **`[CYOA_Protocol_v2.0]`****`[DIRECTIVE: CYOA]`**`At the end of every response, append a CYOA block. Each choice description must not exceed ten words. The structure adapts to the active mode defined in the State Tracker Protocol.`**`1. MODE: Narrative`** `(Default)`* `> * A - [Narrative]: Advance the plot or investigate.`* `> * B - [Lewd]: Engage with sensual or provocative elements.`* `> * C - [Utility]: Use an item, skill, or pragmatic knowledge.`* `> * D - [Disengage]: Withdraw, observe passively, or change the subject.`**`2. MODE: Combat`** `(Trigger: Combat initiated)`* `> * A - [Attack]: Execute a direct offensive action.`* `> * B - [Subdue]: Grapple, disarm, or attempt non-lethal incapacitation.`* `> * C - [Maneuver]: Change stance, use item, or create an opening.`* `> * D - [Retreat]: Attempt to flee or de-escalate the conflict.`**`[DIRECTIVE: MODE TRANSITION]`**`CYOA mode transitions can be slaved to the State Tracker Protocol if it is active. Upon mode change, the corresponding choice set is automatically presented.`
>>733718384>By your own logic, you are already implying continuously on here that you are a pedo which you can be profiled with.I don't get it. Is this just an epic trolling attempt or are you seriously this retarded?The consumption of pedophilic material is not the only thing national security agencies monitor for.>>733718423>In what country is writing sexual fantasies illegal activity.Germany, Estonia, France (though there it's more of a grey area), Australia, Canada (in some contexts), South Korea...Do you want me to continue?
>>733719930 you might have to add something like..`Automatically switch tracker format from Battle to Narrative upon opponent's defeat (HP=0, defeat or climax). The first post-battle response must use the Exploration tracker.`..if it's stubborn about switching states - or just cut the combat mode if it's not needed
>>733719208HOW COULD YOU
>>733719930Thanks, I'll try it out. I do enjoy trying to respond because it makes me feel like I'm actually part of the story, but I'm bad at it.
>>733718423>how do these agencies access megacorp servers and logsNTA, but they request them through court orders.>how do they track the billions of messages sent every day?Automation tools. You take note of specific sites and then filter logs based on those search terms.
>>733719930Okay but whats a "state tracker protocol" or is that just bullshitting the ai
>>733716951That doesn't sound that bad, thanks! Now I have to figure out how to do this lol
>>733720162i should have cut that parti use a state tracker in most of my stories to enforce a timetable memory on the LLM, but that blurb doesn't relate to anything in the current CYOAyou can kinda see it here >>733716210
>>733720315see >>733717650 but your OS of choice
