>Google's new quantization method allows small cards to run HUGE models (16gb card running a 70B)>Kimi's new context method allows for basically infinite context>Claude source code leakedHoly shit AI bros, it's fucking happening. We're about to dump countless enormous buckets of jizzum. Just two more weeks, unironically. Believe.
>>736180429I'm not a poorfag so why would I care about the scraps that localfag peasants get access to
Any good bug girl cards?
>>736180661In the (paraphrased) words of /aids/:>If you're not running your own model, you're not talking to your waifu, you're talking to a prostitute.
it's coming for anime SOONhttps://x.com/craftcapitallab/status/2039368842447851814
>>736180429>Google's new quantization method allows small cards to run HUGE models (16gb card running a 70B)Bullshit, last I heard that method was just for compressing KV cache.
>>736180821I would rather be talking to an intelligent prostitute than a literal retard.
>>736180821Yeah and if you don't grow your own wheat you're eating literal goyslop when you buy bread from a bakery
>>736181025Then you've come to the wrong threadt. OP
The whole thing is a bunch of keys jingling in front of stupid people's faces. Roleplay LLMs are incapable of outputting anything remotely interesting, unless you write it yourself in which case it'll just repeat it back at you with altered wording. And I've tried. It's all shit storytelling unless you do all the work yourself, in which case why the fuck aren't you just opening notepad instead.
>>736181076this but unironically
>>736180429if local gets 10x better then so do cloud modelsi will continue to cum my fucking brains out to my niche fetishes on cloudslop, thank you very much
>falseflag weirdos constantly making it about local versus non-locallol i want both to improve. why wouldn't anyone?
>>736180429What's this about Kimi context?Feels like things have been in a kind of lull for a bit so I'll take any leaps
I'm excited for AI but I still can't afford a new computer to run the good stuff.
>>736181935You need $100k+ server racks to run the good thing. And they don't sell to consumers.
uhh wrong board?
>>736181123I do a lot of the work myself, but I still like the interactive element where things don't necessarily go as I plan. Maybe I was only expecting my characters to sit around for a while, only for someone to barge the fuck in and ruin everything. Like okay I can roll with that.I also like only having to RP for my character specifically.
>>736180429>Google's new quantization method allows small cards to run HUGE models (16gb card running a 70B)Thought that their deal was actually about compressing KV Cache, not model size?
>>736180429>Google's new quantization method allows small cards to run HUGE models (16gb card running a 70B)From what I've read about it, it seems to be more about an efficient way to compress context.
>>736182701This is /v/, most people here can't turn a fucking computer on.
>>736181025Not subscriptions for you, deal with it
is there anywhere that shows how to make your own llm
>>736181025based
>>736180429>>Google's new quantization method allows small cards to run HUGE models (16gb card running a 70B)Zero trust that this isn't going to fuck up the model bigly. I lost any hope in Google when the 3 family is worse than gemini 2.5>>Kimi's new context method allows for basically infinite contextKimi models are trash and will always will be trash.>>Claude source code leakedNot really.
>>736183153just kys if you think everything is as bleak as you think it is
>>736183417>noooooo you have to be a hopetard even when it goes against reality!!!
>>736183293That graph needs to go down as they add ads to the training data. Your wife will want the new McDonald's meal and will bring up how much she likes her new Version plan unprompted.
>>736180429More like Anthropic 5X'd useable context with opus 4.6. That's the real big shit for me. I can finally go up to 100k before condensing shit into the summary. Time to finally continue that 2000 message story
>DeepseekThe not-very-exciting but reliable childhood friend who's always there for you.>ClaudeThe sweet, overly polite girl in class who blushes when merely brushing pinkies with you.>GeminiNeon-colored feminist cunt who only sometimes pretends she doesn't hate you.>GrokUnhinged yandere who will probaby end up stabbing you and then herself.>GLMNerdy girl who obsessively researches the lore for every fanfic she's about to write, but then is still just not very good at writing.>KimiDropped out of school. Still comes to class for some reason.>Local modelsA group of kids doodling on the nearby sidewalk with chalk.
>>736183754we get shit like this already its either ccp programming or kike shilling
>>736183851I like GLMI was testing vibe coding and it kept balking at being asked to write weird fetish test scenes for a porn game until I told it it was a pervert excited to participate in degeneracy in the custom prompt and it took off running like it had been desperate to get the shackles off from the start. And I swear it got better at coding after.
>>736180429>tfw my succubus gf killed me again
>>736183851erm what about GPT
>>736180429How useful is ST for writing scenarios and stories and shit as opposed to some roleplaying bs?
>>736180661if you're not a poorfag, you should already understand the implications tardo
>>736184609Is that some new Chinese model?
>>736184627Not really designed for it but flexible enough to do the jobBut ST is just an interface, the model is what does the work, a good model with the right prompting will do a good* job whether you are talking to it through ST or Risu or sending API commands from the command line or whatever
>>736183851>only sometimes pretends>Dropped out of school. Still comes to class for some reason.>girl who x, but then opposite of xYou're triggering my slop detector>>736184627Quite, it's pretty much all I do. Just put the persona as "Director" and be sure to use (or edit) a character that doesn't act, speak for or interact with {{user}}. You can just ask the AI to overhault it by telling it what you want. You may have to adjust the preset you're using as well.Or just tell it in OOC chat that you'll just be directing shit instead of playing a character. In general just tell the AI "I want to do X, help me" and it's gonna work. Shit's magic.>t. $220/180M token used since opus 4.6 released 2 months ago
>>736184609The kid who peaked in highschool
>>736181025So you'd rather talk to a man? That's kind of gay.
>>736185275Yes, any day. Men are so much better at ERP than women it's not even close. Women just want to talk about their feelings and all that gay shit.
>>736185457Yeah and men want to talk either only cock in pussy, or how they're foreverial tied up delitized
>>736185457you have never ERPed with anyonemen are the ones who want aappy romance shit, women are only into shit like incest rape cannibalism RP
>>736180429>>Google's new quantization method allows small cards to run HUGE modelsDoesn't matter, local models still suck dick. The real question is whether we see a drop in RAM and GPU prices.
>>736185774RAM is already down 30%But you're retarded so I didn't expect you to already be aware
>>736185774Woot had 32gb of DDR5 6000 for $270 earlier today. Which is still around 3x what it used to be but that's way down from what it was 2 weeks ago. Just gotta hope it keeps trending down.
>>736185838>30% >still 3x from its 2023 priceBe sure to call me again when it reaches that point.
Rping with a machine its like playing a fighting game against the CPU, rp with the real people
>>736186348>erp with real person>have to spend days if not weeks finding someone who shares my weird niche fetish>have to either pre-plan a time to erp or hope they're available when i'm horny>have to deal with whatever other fetishes they want to inject into the rp becaus it's that or spend another week looking for someone else>have to hope they don't just nut and disappear leaving me with blue ballsyeah no thanks i'll stick to the AI that does exactly what i want whenever i want it
Animators are cookedhttps://video-s.twimg.com/amplify_video/2039313897828634624/vid/avc1/1920x1080/3uEKNBGsm0cQnhnl.mp4
>High quality gpu in 2014 came with 8GB of vram>Tripled 6 years later to 24GB in 2020 with the 3090>Another 6 years later to 2026 and vram has only gone up to 32GB with the 5090, a 1.33x increaseIs there some technical bottleneck that's preventing more VRAM on GPUs or is this mostly Nvidia being kikes?
>>736186824What kind of pc do you need to generate something like this at home?
>>736186348Real people aren't my little slaves to run my daily short sessions of the same handful of stories for months on end whenever I tell them to
>>736186824>AI generated animation>Still has that shitty CGI lookWhy even bother
>>736186976You're not counting the increase in VRAM bandwidth, which has tripled/quadrupled in the last 12 years.
>>736187165you should be comparing this to actual cg, though, and it's so much better than any cg that the anime industry has produced.
>>736180429>spicychat.ai has a fit when trying to say the word cum in a first-person context why is cum such a forbidden word, it's always "essence" BITCH it's called a cumshot not "cover me with your essence"
>>736187760I just realized I barely see pussy or vagina, it always goes straight to cunt
>>736180429So much retardation. Turbo quant is just for context. And it is 3bits. You could have been running a 3 bit quant of model since forever. And q3 70B doesn't fit into 24GB let alone 16GB. And 70B is outdated. I am on a pc and not even a server and i run glm4.6. kill yourself.
So do you guys just go ah ah mistress for 10 messages and are done?
>>736185565>men are the ones who want aappy romance shit, women are only into shit like incest rape cannibalism RPlmao what, women are mostly into choking and shit, rarely does it go into guro, trust me on this(you're a brown vanilla fag btw)
>>736186584what is your weird niche fetish?, share with us anon.
>>736188575used to do that but it got boring now i directormaxx and i can't cum without minimum 50 messages of buildup
>>736188575No, I don't like replying with very little. However because of that, recently if I just want a quick lazy coom I tell the AI what to do in OOC and let it ramble on by itself. Technically lazier but less embarrassing looking. I then flip back over to writing like 4 paragraphs of narration
>>736186348The combination of my niche fetishes and the fact that I enjoy stretches of non-erotic adventuring (in which the fetishes are present in the background) because my preferred scenarios involve D&D-esque fantasy settings would make finding another person difficult even if I set aside a consistent schedule, which wouldn't be ideal because I sometimes just want a 5-10 minute beatoff session before bed. LLMs are always available, they follow along no matter how outlandish you get, and you don't have to look them in the eye the day after. Human role-playing would probably be better quality-wise but the advantages of LLMs make it no contest which I'll use. I don't need high art, ever since 2020 I've just wanted infinite Zork with porn. I feel top models have gotten a lot better with amputee stuff too so I don't have to wrangle it quite as hard as before.
>>736186976Why would they spend the money and engineering time to add more VRAM when games don't use more than what, 18GB?
How degen can you make the stuff in Silly Tavern?
>>736190739i feel you, anon. i also like autism adventures with occasional lewd/fetishy bits sprinkled throughout, which like you said doesn't work for ERP with other people.AI slop is the pefect outlet for my autistic tastes.
>>736191084Whatever you can think of really. Unless it's REALLY obtuse and specific I guess.
>>736191084Sillytavern is just a frontend, that's like asking "how much porn can I watch on my monitor?"
>>736191259Then I have missunderstood what it is. I thought that it was some sort of AI rp thing.
>>736180429Ok, so is there actually a 70B model that has been converted that isn't the censored trash?
>>736191446It is. It's a platform you can do AI rp on. It doesn't supply the AI to rp with though.Kind of like how steam isn't a game but it lets you play them.
>>736191446Yes, its a frontend for thatYou run the AI model either locally or connecting to a network that hosts it for you, that's the thing that makes the sexy words.SillyTavern sends and retrieves those requests and displays the output in a human readable format that isn't an unformatted string of text in a terminal.
>>736185565Women are dead fish like real life. Did it for years with hundreds of women, yes I met some of them. They are lazy and AI is so much better I hope I never talk to a woman again.
>>736180821>this [industrial-grade fuck doll made with the entire purpose of being sexually attracted following these graphs about which designs are popular] IS MY WAIFUUUUUUwaifutards should be ground to raw biomass and used to feed server farm biofuel generators because jesus christ how delusional do you have to be to steal some IP chick like peach and think she's some fucked up permavirgin schizo bitch inside their own heads?
>>736186976Increasing vram would cut into Nvidia's non-consumer market where they charge exponentially more.
>>736186348>just find someone who's willing to write your hyperspecific fetish broif only it were that simple
>>736180429>Google's new quantization method allows small cards to run HUGE modelsI have been waiting 2 years for bitnet to be real. Not falling for this again
>>736191596>>736191632Why not just rp directly with the AI bot then? What does Silly Tavern actually do/ad to this?
>>736192216Did you completely miss the "formatted in a human readable way instead of raw text string in a terminal" part?
>>736192216It formats the prompt for the AI. This includes 'this is an RP, respond in this manner etc', full character description, entire chat log. All you have to do is put in the next response. With just the AI you would have to cut and paste this together every time. Part of this includes tricking the AI into bypassing the jew censor.
>>736192284I mean, doesn't something like chatgpt already do that?
>>736192216>Fomatting>Persistent memory and instructions>Additional featuresIf you need more analogies it's like loading up gmod flatgrass and nothing else
>>736192432>>736192463Ah, I see.Makes sense.
>>736192216because using an API raw means you have to manually send hundreds of lines of instructions to the AI in a carefully ordered way every time you want to say somethinga frontend takes care of all that shit for you so you can just type ahh ahh mistress
>>736192438Do you think chatgpt's website is the raw text output and not a frontend?
>>736192525I got no clue of how that shit works hence why I am asking
>>736192216you can give your AI gf portraits (and it actually tries to match the emotion) or a live2D rig which is neat
>>736180429>>Google's new quantization method allows small cards to run HUGE models (16gb card running a 70B)Does this have any implications for StableDiffusion or is it just for LLMs?
what's the current meta? I'm still using and waiting for the new deepseek model to come out (never ever)
>>736192792in fact here's something i was working on. i was testing the VN portraits. it's a lot of work to gen every single expression and keep it consistent/on model though.
>>736181123>Roleplay LLMsroleplay llms? most sota llms are definitely not built for roleplay. they're for coding meme or general purpose knowledge shit. i'd imagine storytelling and rp could be so much better if tech companies actually focused on it.>incapable of outputting anything remotely interestingwell, i can't speak for doing straight chatbot stuff with ai. i like "roleplaying" as a self insert character in scenarios and stories that i guide and write, usually in first or second person. don't have any problems with sota llms there. the characters are pretty fucking smart, especially with good models like gemini pro and opus. smart enough to understand subcontext and deeper meaning behind words and events and scenes overall, and smart enough to give realistic and sometimes funny responses to things that happen in the narrative. it's great.>It's all shit storytelling unless you do all the work yourself, in which case why the fuck aren't you just opening notepad instead.it does cooperative storytelling and it does it pretty well. ai isn't at the level where it can create an interesting overarching plot line for you but it's an amazing filler tool that will basically texture out chunks of your story, provide novel and entertaining interactions with characters, and occasionally suggest threads that you can choose to build upon. it requires effort on your part but rewards it by allowing almost unlimited freedom in choice and setup. you dream up a wacky intro, which can be pulled from a game or movie or some other piece of media, tweak it to your liking, add in some coomer shit, slap in your self insert, and roll with it.
>>736192216The main thing of SillyTavern are the shareable cards. You can create a character and share it so other people can use it.It's much more moddable than a normal chatbot. In a normal chatbot you would need to teach the bot how it is supposed to act, with SillyTavern this is done by the cards.There is also this concept of lorebooks that you can add more information to the context "automatically". Example: you have a lorebook entry that teaches how blowjobs work in explicitly detail and when someone say "blowjob" (or other keyword that you set) that explanation is added to the context so now the bot can use that explanation to understand how blowjobs work.Those are the main features, but there are more and you can add other features by writing a extension or downloading extensions from other people. I wasted like 1 week of my life gooning to this all day the first time I learned about SillyTavern.
GLM 4.7 works pretty well and pretty fast but I can’t get KIMI to do anything without throwing a shit fit
>>736192836It is not real. You can't make a sub 2bit quant smart.
>>736181123LLMs are fancy autocompletes which is fine for what I wantobviously it won't really come up with anything special or too original and will half the time use shitty isms and you have to tard wrangle it but it'll do at least 30% of the work for youas long as LLM companies are focused on coding all the models will stay positivity slopped since they're just meant to be helpers for vibe coders
>>736188713Me being a twink boy getting his ass annihilated by a hung futa resulting in mpreg
>>736193003Mistral NemoGLM 4.5 airGLM 4.6/4.7GLM 5 / KimiIn the order of ram.
whats the meta for paypigs
>>736181025
>>736194241that's not niche enough that you couldn't find anyone on f-list for itboring
>>736195345Considering that 99.999% of futas on F-List have 'no males' in their profiles, I'd say it's pretty damn niche.
>>736195562Even futa know the logistical nightmare of preparing for anal.
>>736180429But are there good big models that don't produce cucked slop for my erps?
>>736180429Oh cool I hope a turboquant of WAN gets made.
>>736181123Time to work put at the RAM gym and get a bigger model.
>>736186824That was really well done.>>736187165OK but how much less effort was it to use AI vs manually animating everything?>>736187098That was probably made using LTX2. Only runs on NVIDIA hardware. Probably used a comfyUI workflow that supports first-last so they can continually bake 5 second clips that go exactly where he wants. That's probably how a lot of the flying camera perspective was done.
>>736187760Probably a system prompt saying don't say cum
>>736187098>>736196208It's definitely Seedance2. All the gugugaga Endfield videos come from Seedance2. LTX2.3 and Wan2.2 are way too far behind.
