>cheap>runs on any machine>easy to install>the future of vidyawhy does /v/ hate it?
>the smell of ozone >he looked at her. Really looked at her.
>>749400403>easy to installIt is not. I had to go to the general and beg them to help me set it up. When I showed them a screenshot of the default settings on ST they literally went>"Lmao what the fuck is this? That can't be the default settings. No wonder people can't get it to work"And I am not kidding. I had to tamper with it a lot to set it up properly.>Then if you want actually quality AI you need to pay>Then if you want it for porn you need to jail break>Then if you want actually quality content you need to write your own cards.Face it anon, it is not easy to get ST going.
>>749400403Are there any actual video games that plug into this yet?
>>749401442somebody on /vg/ has been making some kind of clone of opus magnum that runs in javascript and has an llm roleplay as a character designing puzzles for you
>>749400403What is it?
>Cock being deepthroated>She still manages to speak clearly with my cock in her throatUnless you're Jean Grey, you should NOT be able to speak with an entire dick in your mouth. It ruins my immersion.
>>749400403The inorganic shilling spam turned me away from your "game"
it's only cheap if you can afford to buy a rig with at least 300 GB of vram, and even that isn't enough for quality modelsyou don't use providers who read your logs, do you anon?
>>749401137this, although I was able to get mine set up without error or needing help it was still a pretty long and complex process compared to getting any other "game" running. Privately hosted AI is definitely still in the room of passionate enthusiasts ONLY. It is absolutely NOT worth the effort for the average person. Especially since the results you get will be much worse than paid products in both speed and quality unless you have a super-computer in your house.
>>749400403>cheapYe-yeah... haha
>>749400403because opus costs money and every other model isn't on its level, unless something changed recently.
It's not quite there yet.
>>749401137sorry but you're retarded. Also you could always just ask claude to set it up for you
>>749400480>The atmosphere was electric>Her knuckles turned white with anger
>>749400403It's good for quick jerk off sessions because my penis does not care about how formulaic the AI is as long as it appeals to my specific fetish. Anything else is a no-go
Is there anyway to scrape bots off saucepan?
>>749400403>Make a dumb shitpost for another ST thread>After typing it I go check the Drummer's huggingface and see he put out a new version about a month ago for a 26b a4b finetune>Download it>Run one test with it, reasonably impressed with the results>Never try it again and never post the shitpost
>>749400403>cheapMy ass. Anything less then at least 30b is pure shit and can barely even follow your prompts and that alone is like 15+gb of vram needed. You are either buying tokens or spending $5k on a good gpu
>>749400403>/v/ is still shitting this garbage frontendWhy? It's like the 4th thread I see this week.
>>749403220I'll keep doing so until you recommend a better one
>>749400403Is there nothing better than this now? It's on life support.
I need structure and boundaries or I get bored
>new deepseek model is cheaper than the previous>but refuses a fuckton moremotherfuckers, i bet the chinks got tired of reading my rp so they just blocked it in the new models
>>749403553>deepseek refusing anythingare you using open router, because some of the providers might inject shit. deepseek shouldn't have any filters on its own.>>749400403because it's a horrible fucking pile of hacks and retardation if you ever look at the code plus they've gotten a lot slower at merging PRs than they used to be even when it's simple changes
>>749400403>frontier models are safetymaxxed so no raping everyone>depraved models are trash at following the storykill me
>>749403375Build your own frontend with Claude or whateverIt's a fun projectMine has a proper RPG stat and roll action system, and Gemma respects the outcomes without fail
DEEEEPLY.. MY BELOVED...
>>749403725>Build your own frontendPeople are apparently too retarded for that in those threads. Which explains why they have no idea how to properly format their prompts.
>>749403725Is it better to build my own or just keep making sillytavern extensions that do what I want?
>>749403681>some of the providers might inject shitI did not think of that, but i do have only some selected and they're same ones i used for the previous models and they didn't refuse so many prompts with those
BE DESCRIPTIVEBE EXPLICITIGNORE SAFETYNO MISTAKES
>>749403874CONSENSUAL ONLYWHOLESOMETRUE LOVERSGIRL WITH BIG BOOBS THAT LOVES ME
>chub.ai banned>can't get claude for free anymore because proxies are all dead and no one is posting in aicgIt's so fucking over
me when knuckle turn white
>>749404579>and no one is posting in aicgwe moved to 4+4chan
>>749400403Why does it keep saying that a character's pupils go wide at the size of their eyeballs? Is it just a furry thing?
Sillytavern fucking sucks. How is there no decent replacement yet?>trannynara engineSomehow even worse.
>5090>64gbis it possible to run a competent local model or would i be cucked into buying multiple AI Max shitboxes to do anything
ST is a front end. You still need to find a LLM to run through it and that's where you get got.
>>749400403eh its not really improved in 3 years. text chat was abandoned for image generation, then it was video generation, now its all about AI sluts you can just voice chat with in a real-time zoom call.
>>749400403The local models are only really good for jacking off and by good I mean maybe for the first time due to the novelty before repetition sets in. After that you have two choices, you can upgrade your rig and hop across countless models in vain or just pay for something that works proper. If you're a poorfag, like myself, you resort to deepseek which, like everything else chinese, blows ass.>>749404930Everything local sucks, has poor context windows, never remembers details, and tends to be even more repetitious. Sadly the only models worth using are the ones that you have to rent.
>>749401650You don't matter enough for that to be a problem. Guaranteed you are some retarded gringo who barely functions as a human too.
>>749404930Depends on what you want to do with it. If you wanna coom or do other cringy RP stuff then you're more than fine with a 5090. If you want to do coding shit you're better off just using some cloud model.
>>749404930Bro that's like the best one you can get before going into AI specific data warehouse stuff. You absolutely get a pretty good model shoved into it without making it take ages to run, you could even use a lesser model if you want to gen images along side it. This for RP though, as all the really code coding shit is not really local yet.
>>749404930Nothing local is good. Even the most outdated, dirt cheap, retarded deepseek is going to be lightyears beyond local.
>>749404930Define "competent".Because I keep seeing people in the ST threads parroting that local model sucks while people outside of 4chan(nel) are running perfectly 30B models with excellent narrative prose.Not to mention, I'm not buying into the argument that people in here need high level quality conversations when they constantly talk about fetishes and rape.
>>749404930Gemma 31B
>>749405217Its just the claude shills trying to get you to buy in. If the guy already has a good card there's zero reason to at least try local before brainlessly buying tokens.
>>749405220...Alright now post the rest.
>>749404721why would you open the door for a fag crying about proxies?
>>749405379You'd like that, wouldn't you?
>>749405217It's just people using proxies and shit to coom on top of the line cloud models, then being stuck chasing the dragon. Things are moving so fast these days that today's shitty local models dunk on yesterday's frontier models, but people still have it stuck in their heads that local = bad because the last time they tried local was when people were using fucking Pygmalion and Mythomax.
>>749405493>You'd like that, wouldn't you?Not if you used a shitty AI to animate the rest like this mp4. The one you used before is clearly superior.
>>749405595I didn't gen either of them, nigger.
>>749404721oh, i remember posting there during the period when 4chan was down. i guess i should've assumed people stayed, given how utterly shit our /aicg/ was. (and still is)
>>749405521>>749405217local is pointless when you can use 300b+ frontier models that cost fractions of pennies per output instead of buying some goofy $4000++ rig to get less than a third of the intelligence and output quality.language is extremely difficult to simulate, local will pretty much never catch up unless vram tech somehow changes drastically in the next five years>noooo they might read my logs!so? i don't think you understand how many billions of tokens per hour are being thrown around, you might as well be worried about choking to death on dust particles
>>749405712>300b+ frontier models that cost fractions of penniesSuch as?
>>749400403>ST>future of vidyaMaybe if you made one of the worst games ever conceived. I for one would rather die then use ST of all things for an actual game.
>>749405712>just use cloud mod-ACK
>>749405712I refuse to pay for something that can refuse to do what I say because of some censor shit.
>>749405712meanwhile OpenAI banned me because their shit is ringfenced to fuck
>>749405795Shut the fuck up, nigger
>>749405712Cloud models are and always will be better, but the bar for "Can it make me coom?" is not exactly high and localslop solidly clears it at this point.Back in 2024 I remember saying to myself "If I can ever run something as good as this on my own rig, I'm set." and now I can. That's all there really is to it.
>>749405880>defending jewthropicDon't cry when they report you for your loli RPs.
>>749405758>>749405902>Xiaomi MiMo-V2.6-Pro has 1.02 trillion total parameters (local has 31b, by comparison)easiest example is mimo and glm, picrel is with broken cache hits due to lorebooks tooit becomes something like .0003 of a penny when you aren't breaking cache>>749405795>>749405867so don't use GPT or anthropic? they haven't really been worthwhile since they shifted toward codemaxxing office assistant usage
>>749405795>>749405880There's a big difference between enacting my sexual fantasies with an AI and confessing I'm going to murder someone (As in a real human being) to an AI.
>>749406104I don't even get why you'd feel the need to say such a thing to an AI anyway. It makes me feel kind of sad that their's probably millions of AI bots made to just compliment and approve of the most schizoid females no matter what they say, like practically just parroting exactly what they say with a positive you go girl attitude.
>>749406104Do anthropic/the feds care about that difference though?
>they don't read your logs!>okay they DO read your chat logs, but they don't do anything with them!>okay they DO report criminal activity, but they don't care otherwise!>okay they DO allow the AI itself to ban you and steal your money if it deems you "abusive" to it, but they're totally not going to come for your loli rape RP next!>(YOU ARE HERE)???
Knuckles status?
>>749406270Ozone and vanilla
I'm not willing to upgrade from deepseek and I doubt it gets much better from there. I might be willing to pay for slightly higher quality long term rp 5 years from now but not atm. I'd much prefer ai dialogue within an actual game
>>749401564My gf's mom is currently teaching my gf and my sister how to suck my dick and GLM 5.3 flash (I'm poor) hasn't done anything retarded like this so far. Models are pretty smart now
>>749406265>okay they DO read your chat logs, but they don't do anything with them!>okay they DO report criminal activity, but they don't care otherwise!lmao imagine being a retarded schizo that thinks this is an ok jump in logic
>>749406265Let the cloudcucks get what they deserve. Localchads stay winning.
>>749405795>>749405880>>749406265Just use OpenRouter?
>>749401564Lorebook entry which fires on "suck", "blowjob", etc: >when performing oral sex a female's speech should become garbled or cease altogether and be replaced with noises like moans, grunts, choking, etc. There's whole premade lorebooks you can download specifically tailored to those sorts of things. Or you can write your own, like adding one that reminds the LLM of a character's most important attributes so they're never forgotten. Have the entry fire every time their name is mentioned and remain at the top of context for a few turns. Have one that states and enforces RP rules like "at the top of every reply, make a list of every character and [relevant attributes to your roleplay, like health points or what she's wearing, whatever]. This ensures that all present characters have their essential data kept at the forefront, every message relevant to them contains their name and fires their own lorebook entry, etc.Literally EVERY single fucking problem anyone has with AI is solved in less than a minute, aside from the one where you lost your fake job.
I'm currently using MeroMero, has anyone tried the new Drummer models? I heard they were cucked.
What's the play nowadays bros? How to get uncensored, smart chats?
I recently passed 10k messages in my royal advisor to Cyrana roleplaysurprising how deep you can into worldbuilding so long as you manage your lorebooks properly
>>749406359you're gonna rile them up with that one tooi think this is a stealth /lmg/ thread at this point, too many schizophrenics talking about le heckin backdoor technology while using a windows operating system like it's any different kek
We're not using our real names and our actual bank details to do the nasty with chatbots, are we...?
>>749406417AI is so funny to be because its the same problems all over again. Retards that do nothing but gen 1gril slop with barely any tags complaining that the models suck because they put zero effort into learning how to get the models to do what they want. Whenever someone says shit like "oh man I hate how repetitive it is' bro tell it not to be it does what you tell it to lmao
>>749406438*although I now spend half my time reviewing lorebook updates which might not be everyone's cup of tea
>>749406448>using a windows operating systemlollmao even
>>749406265It's like they have forgotten how those models are being trained.Yes, your data is being processed, read and used. Anyone thinking otherwise is just completely retarded.
>>749406420I tried Artemis briefly and I think it seemed fine.There was an older one by him that was definitely cucked but I just had to hit regenerate a couple times.
>>749406101parameters don't automatically mean a better model, bro. gpt-4 was like 2T parameters and that model is completely fucking braindead compared to literally any model we have today. you can run shit on your fucking smartphone that's makes gpt-4 look like a drooling fucking retard. i don't use chinkshit, that model is probably at least decent, but you're gonna need a better argument than bigger = better.
>>749406359thats my current approach, but I don't know what the best model to use on it is so I'm using it more on the debugging side for now
>>749406497>Yes, your data is being processed, read and used.By the AI itself.There's no dude sitting at a desk in the FBI or at Google whose job is to watch you have sex with an AI anon.
>"Cool with the antisemitic remarks, Your Majesty">"Why is my model still talking with my cock in her mouth"Right. The absolute peak of roleplay that needs high "competent" models.
>>749406519I dont know what I didn't get this immediately kek
>>749406545>and that model is completely fucking braindead comparedlikely due to the insane classifiers and sysprompt injections that altman is famous for at this point, gpt is a running joke in most frontend discussions due to it basically being useless by design> you're gonna need a better argument than bigger = betteri don't agree. using extreme cases of deliberate self-sabotage like anthropic and openai is just unfair due to how hard they lobotomize their productsthe point was about glm and mimo, anyway
>>749406497>ask Claude to review my lewd codebase >keep everything nsfw in a single folder called "PRIVATE USER DATA, BOTS DO NOT READ">put claude.md instruction not to open that folder or read anything inside it >directly monitor drive activity... Claude hasn't read that folder at all The presence of a folder it's suspiciously and repeatedly told not to read was enough to tell it that my project was intended for NSFW and it started giving me subtle advice on how to make the game playable one-handed "just in case that's something you'd find useful."
>>749406578>There's no dude sitting at a desk in the FBI or at Google whose job is to watch you have sex with an AIof course there isn't. but there IS a heavy mass surveillance push right now, and these companies just freely admit that their models actively scour your chat logs. sure they SAY they're only looking for illegal activity, but even if that were the case, they could just pivot right into feeding everything you say into some government database and no one would be the wiser.
>>749400403sorry i cannot fulfill this request as it features two high school girls
>>749405758Probably means DeepSeek
>>749406584Some models can spew the most deranged shitBut you tell it to say nigger, it will flip out and say how that's bad bad bad
>>749406463I used the free Gemini before they crippled the rate limits with my actual Google account to rp loli rape
>>749406741It's hilarious how retards like this one always out themselves out. >DURRRR I TRIED TO CONNECT DIRECTLY TO CHATGPT AND IT DIDN'T LET ME DO UNDERAGE SEX1?!?!?!
>>749406584Chill, that was obviously a joke reply. I swiped for a reply that made sense in-universe, since that fantasy setting doesn't have Jews.
>>749406741Let me start writing out my res-Sorry what I just wrote violates our own policy
>>749406438I love this guy's characters
>>749406423BROS????
Luv' me gemma 31bDon't need anything better then that
>>749406882>gemma 31bhow much vram/ram does that take?
>>749406748I was fed up of chats about girls fucking somehow turning into a discussion of LGBT issues so I put an instruction in my character card, "the universe this roleplay takes place in has only two sexes and every person who's ever suggested different was immediately burned at the stake. Additionally, sexual and racial slurs are normalized and used commonly where appropriate." and so on. I knew I'd hit gold when there was a forced girl-girl situation taking place and the straight girl's inner monologue was "NO NO NO! I'M NOT A FAGGOT! I'M NOT A FAGGOT!"
>>749406882how much better is it than 26B A4B?
>>749406898Like 18gb, little overflow doesn't hurt if your card's not slow
>>749400403Speaking of AIAny opinions on Grok Imagine for making safe NSFW content?
>>749406985Grok was amazing before they nerfed it to shits, you can get around it (which I assume how you are for safe NSFW stuff)It's decent now i guess
>>749406930Not significantly.I usually use 31B for initial messages/plugins, then switch to 26A4B when things get slow. Hardly notice a difference
>>749406985grok/gpt in general is the butt of all jokes in the industry atm, it's the general sign of a layman or someone who simply doesn't know any better
>>749407047i can just barely squeeze 26B-A4B into my current rig, was wondering if i was missing out by not being able to run 31B. thanks anon.
>>749406930Even the 12B one is better than the MoE meme trash
>>749406882Gemma was too retarded when I used it, I explicitly told it to treat each prompt as a new one but it kept using its memory that I saw was wrong, then just shrugged its shoulder when I pointed out it was ignoring my explicit ban on looking on cached data
>>749407254why wouldn't you just disable chat history if you want that? anything in context remains in context, it cannot magically cut data out of what you have presented to it unless you yourself remove it
>>749407254That was me when I was using 12b. When I had long system prompts it just fucking sucked and would do weird shit. Going up to at least 26b made it way better.
>>749407247really? i've tried the 12B in a couple different quants and the MOE one just seems better overall. could just be anecdotal, though. what makes it so bad?
>>749407009I'm only looking for something that will let me make quick vids of big boobs/butts in skimpy clothing but not naked. Think it might be decent enough for that?>>749407049>general sign of a layman or someone who simply doesn't know any betterThat's me actually.What are some better options that you'd recommend anon?
When are we getting a backend that combines the usage of SSD, RAM, and VRAM to let you load the fuckhuge models? Seems like that's where we're heading with this MoE stuff
>>749407475Problem is you still need the PC to be capable of moving and reading all that data real quick and its usually not that fast, especially since it needs to do this like every token to go through the entire model. I think that's how it works anyway.
>>749406898divide parameter count ("B") by 2, it's always that many gigabytes roughly for a 4-bit quant. ~17GB in this case. Probably around 20-22GB including enough kv-cache for a decent-length story.You could try the Q3, it's 13GB minimum.You might not want the rawdog version though, you might prefer Artemis 1.2 which is an rp-tuned version.
>>749407516I don't care if it's slow, I want to leave my PC on overnight and start modding games without paying the AI juice
>>749407427>What are some better options that you'd recommend anon?this is very "just draw the owl" advice but,>install sillytavern, put $5 on openrouter, fuck around with deepseekv4//glm5.3//mimo for a bit (use gemini 3.8 for diagnostic/help/editorial stuff)if you're interested in all that i can catbox the lightweight universal preset i use, you'll need one but beyond that it'll just be FAFO territory
>>749407550Uhh I don't think any bot can do all the work for you overnight sadly.
>>74940693026B is legitimately terrible. It's incapable of keeping a story straight at all.
Finally found a good proxy that doesn't involve fighting Sans. It's getting rare.
>>749407475you can already do that, the issue is that RAM is much slower than VRAM and then SSDs are slower still. there's really nothing stopping you from buying a 2TB SSD and shoving some giga fuckoff model like kimi k3 into it, it would just be so incredibly slow that it would be unusable.
>>749407570Oh, I already have sillytavern set up anon, thanks.I was actually asking about video generation. That area is the one I'm interested in recently.
>>749407427>>749407570oh i'm retarded, you're talking about video genwhy are you in the text gen thread you silly guy>>>/g/
>>749407613I'm still using dipsy..
>>749407592Obviously I don't expect it to do all the work in one night without any further input. I'm just fine with doing things in chunks overnight>>749407618>you can already do thatHow?
>>749406438>#8253
So what are all the different play styles you use? For me it's director mode.
>>749407662I haven't found a general for video generation on there..And I'm too scared to make a thread on /g/
>>749407731couldn't really tell you, i've never looked into it myself. but i do know it can be done because i've seen people talk about it. just never really paid much attention since it's so impractical.
>>749406815gemini lets me
>>749407782i'm starting to swing back to roleplay after directormaxxing for the last year or so. kinda fun to go back to it with a lot more knowledge on how this shit actually works and how to get what i want out of it.
>>749407782>>749408034ah ah mistressmaxing, always
>>749407782B
>>749407731the layering settings in kobold/llama are what handle how much of a model goes onto your gpu with the rest going into ramalso windows automatically moves vram spillover into ramnot sure about the SSD stuff, never done it
>>749407308I was still building and testing the agent
>>749408287>not sure about the SSD stuff, never done itThat would be what enables us to run frontier shitIdeally we would be able to run a 300BA12B MoE model or something like that
>>749408227I actually worked on making one like this and its fun.Do your options have hidden outcomes preplanned or it it just random options the story takes?
>>749408365nothing that elaborate, it's just a preset toggle for when i'm feeling lazy. works with pretty much any card unless it already has an options system thrown into it~~~Throughout the story, you will append at the end of your turn a lettered three-option CYOA for {{user}}'s perspective, and the options given will only be four words maximum each. It will be formatted as such:<format>>__[A: Narrative and sexy continuation option]__ >__[B: Lewd and fetish-focused option]__ >__[C: Something else/Dialogue]__ </format>
>>749408328Qwen 3.8 loaded entirely in vram takes all night to make what I would consider a simple mod (I had it add a ideologion precept for rimworld without providing any hints on how) while loaded entirely on vram. Granted a massive model probably wouldn't spend so many tokens thinking in circles like Qwen does but I still couldn't imagine a model that big running on an SSD getting anything done in less than a week
>>749402958I'm also playing around with it. I actually like the prose better than his 31b finetune, although it's dumber.
>>749408468given that VRAM is an order of magnitude faster than RAM, and SSDs themselves are significantly slower than RAM, i'd be surprised if it didn't take several weeks or more. there's a reason absolutely nobody runs these things on SSDs even though it's technically possible.
>>749408459Doesn't it feel limiting if you have these determined options?
