/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>110280716 & >>110277240►News>(11/06) PLaMo 3 Translate released: https://preferred.jp/en/blog/tech/plamo-3-translate-weight-release>(11/05) Baidu releases ERNIE-5.5-Flash-Base 280B-A14B: https://ernie.baidu.com/release/ernie-5-5-flash-base>(11/04) Ricursive Ripple 65B dense model released: https://www.ricursive.com/ripple>(11/04) MoonshotAI Kimi K3.1 Pro released on HF: https://hf.co/moonshotai/Kimi-K3.1-Pro>(11/01) Longcat-2.5 released: https://longcat.ai/blog/longcat-2.5►News Archive: https://rentry.org/lmg-news-archive►Links: https://rentry.org/LocalModelsLinks►Official /lmg/ card: https://files.catbox.moe/cbclyf.png►Getting Startedhttps://rentry.org/lmg-lazy-getting-started-guidehttps://rentry.org/lmg-build-guideshttps://rentry.org/IsolatedLinuxWebServicehttps://rentry.org/recommended-modelshttps://rentry.org/samplershttps://rentry.org/MikupadIntroGuide►Further Learninghttps://rentry.org/machine-learning-roadmaphttps://rentry.org/llm-traininghttps://rentry.org/LocalModelsPapers►BenchmarksLiveBench: https://livebench.aiProgramming: https://swe-rebench.comAgentic Coding: https://deepswe.datacurve.aiContext Length: https://github.com/RecapAnon/NoLiMaGPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference►ToolsGGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-CalculatorSampler Visualizer: https://artefact2.github.io/llm-samplingToken Speed Visualizer: https://shir-man.com/tokens-per-second►Text Gen. UI, Inference Engineshttps://github.com/lmg-anon/mikupadhttps://github.com/oobabooga/text-generation-webuihttps://github.com/LostRuins/koboldcpphttps://github.com/ggerganov/llama.cpphttps://github.com/theroyallab/tabbyAPIhttps://github.com/vllm-project/vllmhttps://rentry.org/custom-uis
►Recent Highlights from the Previous Thread: >>110280716--Progress on mapping Gemma-4-31B layers using high-end GPUs:>110282841 >110282856 >110283250 >110283278 >110283303--Comparing Polystrata and Basalt performance on Qwen with 5090+3090:>110282957 >110283097 >110283147 >110283156--Qwen 3.8 27b criticized for severe repetition and coding failures:>110281574 >110281635 >110281742 >110281746 >110281761 >110281815--Using Jinja templates for prefill and context poisoning:>110283666 >110283726 >110283769 >110283830 >110284059--LongCat-2.5 multimodal release and benchmark performance reports:>110282082--Meta releases Muse Ember small model for agentic workflows:>110283552 >110283595--Anon advised against 128GB DDR5 in favor of more VRAM:>110282768 >110282790 >110282792 >110282807 >110282836 >110282905 >110283107--llama.cpp benchmarks and quality tests on a GTX 1650:>110281027--Speculating on Gemma model sizing and Google's strategic motivations:>110283142 >110283255 >110283268 >110283493 >110283530 >110283714 >110283201--P2P communication bottlenecks and unlocking consumer NVIDIA GPU capabilities:>110280831 >110280881 >110281011 >110281021--Next-Concept Modeling as an improvement for machine translation:>110280745--OpenRouter token volume as an indicator of model popularity:>110283691 >110283811 >110283836 >110283866 >110283985--Feasibility and utility of intrinsic motivation and autonomous AI agents:>110284029 >110284086 >110284104--Comparison of MoE and dense architectures for trillion-parameter models:>110283611 >110283751--Skepticism regarding GPU price drops amid rising compute demand:>110280951 >110280989 >110281063--Anons coordinating volunteers to process Gemma model layers:>110282868 >110282888 >110282924--Logs:>110280892--Gemma (free space):>110281012 >110281179 >110281503 >110283331 >110283759►Recent Highlight Posts from the Previous Thread: >>110280886Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
All hail the ceiling rig.
Guyz... I am question. Why are Muse Ember outcompete Grok 4.7? Is it same sized as a Muse Glimmer? Open weights when?Here is picture of Anime GIRL as reward!
>>110284297Bizarre Ceiling Teto
31b too big, e4b too smol, 12b just right
>is she 10-years-old?What did they mean by this?
I don't think I ever really understood what jev was, and now it's too late
If Jensen had a miku gf this never would have happened
>>110284445nigga forgot an arrow
>>110284445>10 years oldOut of 10
>>110284460I was looking into it, on Jevbench there are a lot of open jev models now and many of them better than jev, the best was the mammoth 1 but I don't have space to run a 27b classifier, but there a 4b classifier that is on par with jev apparently. I don't have Nvidia though so I can't run it. I still don't really know what it's useful for much or if it's really better than e4b or something.
>>110284460Jev is an example that even marketing can't save you in a world where anyone can instantly reproduce your thing. One day it'll spread to all software.
>>110284481>Out of 10
>>110284460it's an LLM with only two outputs; yes and no
>>110284505hehe
>>110284500>One day it'll spread to all software.I dream of a no profitable software future.
>>11028444012B is a cutie>>110284505hell yeah Edgar, I need to replay that sometime.
>>110284517This means that either Dario or Jensen collect all the profits.
>>110284526>DarioAI is a software.>JensenIf i can buy and own thats fine
>>110283759>If you DID NOT read the Harry Potter fanfic,I did. I even read selections of the first part to my kids because I thought it walked through some parts of rationality in a really clear, entertaining way.>DID NOT join the cult and WILL NOT buy the Anthropic IPO, you will be killed.Yeah, you had to be pretty far up your own ass and lacking in life experience to go much farther than the first few chapters of the fic and buy into the whole zoo. That shit is schitzo insanity and deserves its imminent self-immolation.How supposed rationality can act as a gateway into scientology level self-hypnosis and mass control, ceding your actual rational faculties to a bunch of charlatans and going into peer-enforced groupthink, still boggles my mind.I guess people still fall for pyramid, ponzi, etc schemes in the current year, so I shouldn't be too surprised.And if you thought this comment wasn't about you, its about you.
>>110284297
I'm going to eat something with practiced ease
>>110284553The EA cult is more effective than most cults because there is a material benefit to being part of it.If you're not a true believer it's going to be much more difficult to get the high paying AI jobs. And big tech SWE jobs previously back in the 2010s.
holy fucking shit. Well done Dario
Is there something like a LoRA for text, where I can feed it a bunch of text and attach it to an existing LLM to have that write in the style of the text I fed it?
don't look at deep seek
Source tell me Qwen4-27B is stable.Please god.
Qwen 4 bros how we feeling
>>110284695You just give the model a bunch of text and tell it to write like that. If that's not good enough or really is just unsuitable, train your own LoRA. It's not difficult to train LoRAs for small models if you have the hardware for it.
>>110284691This isn't impressive. You're baseding out over scraps. These faggots haven't delivered for 8 weeks and you're leaking cum over it.
>>110284691>t. paypig non-local threadshitter
>>110284708I just hope the 3.8-Flash-Next analogue has similar total/active params so I can run it on meme cope MoE engine du jour. It's been wild having something genuinely clever locally on my shitty hardware at 50 tok/s, even if its manner of speaking is a bit insufferable at times.
>>110284708Feeling harfam at the moment.
>>110284691Dario dropping the most intelligent models with practiced ease.
>>110284708Feeling a combination of safe, innocuous, bored, and slightly hungry
this is your expression
>>110284691What happens when we reach 100?
Fuck are they really ipoing soon are we getting nothing but shills this week?
>>110284764TRVKE
>>110284775I'm so tired of it
I just realized the iGPU was sapping 1.6GB RAM just by being enabled. Sigh.
So has Lecun actually made anything yet?
>>110284794He's made a bunch of posts on x.com.
>>110284764Why are normies like this? A house gets you SOTA chinese models at far more than 80 t/s.
>>110284784Just follow Sakuya's advice.
when will this become reality?
>>110284812loli feet
>>110284808They're jeets who can't fathom how much 50k is. If you can trust your data to cloudpiggy you're not doing anything of value. Or you are and the banks are going to run you out of business. Either way you're masturbating in public.
>>110284764>Local AI people when cloudcucks get banned because they are no longer needed and hardware is illegal to buy
>>110284764>Pay to do your project>Policy updates and thats not allowed>No refunds.
>>110284819Source?
>>110284792just give the igpu all but one gig of your ram and pretend you're UMA 5head
I-is jacket man going to get fired?
Which Linux version should I reinstall for nvidia?Ubuntu has failed me, and it's taking 4GB system ram anyway so it can eat shit.
>>110284884nyo
>>110284884He's going to jail
>>110284898Horny jail
>>110284884He's rich and firing him would cause big drop in Nvidia stock.
>>110284884why? he apologized
>>110284898He'll be pardoned after a substantial donation to Chabad Lubavich, but green team will never be the same. Dude's been CEO for 30 years. Of course there's no serious provision for a replacement.
>>110284884It doesn't matter.
>>110284888What's your exact problem here? Not simping for shuttleworth, but if you can't make one of the vendor supported distros work your problem is likely between the chair and keyboard.
>>110284884I like how Elon Musk bangs dozens of women and has hundreds of hell spawns with them, but none bats and eye, since he wasn't married to any of them, so that's okay. But lo and behold if a guy tries to enjoy something else once in a while, while committing to his wife and family for the majority of his adult life.
>>110284979fucking expensive whores from america -> okfucking free whores from china -> problem
>>110284979Musk has never been apologetic about any of it, that's the most important part.He just does it and doesn't give a shit. The moment you start apologizing is when you're on the defensive and admit wrongdoing and try to save face.
>>110284984>fucking free whores from china -> problemNonsense, there isn't a man alive that didn't splurge a ton of money on a side chink, if only to keep her silence, not accounting for going out, trips, restaurants, etc.
>>110284979I like how hypocrites present Huang's infidelity like it's okay, but then seethe about foids riding a cock carousel.At least I am consistent. Hate all whores.
>>110284838>>110284849You guys know that abliterated API models exist, right? You only run locally to conceal your pedophilic escapades, not because you actually do any serious work.
>>110284923>green team will never be the sameLiterally nothing happened.
>>110284764>Hey Claude, can yo->NO, FUCK YOU BIGOT
>>110284995>He just does it and doesn't give a shit. The moment you start apologizing is when you're on the defensive and admit wrongdoing and try to save face.ye ngl this is pretty pathetic. And it won't do much roasties on plebbit are already cheering how they don't have prenup and she should take everything
even open source model are getting cucked in some capabilities, is there really no hope for us?
>>110284979Has Musk fucked a spy and given her all of the trade secrets of his company?
>>110285048Make your own
>>110285053He has massive factories in China.
>>110284884leather jacket man is nvidia
>>110285053lel what secrets? The Chinese have better EVs than Musk
>>110285056What if we actually did this? Frontier models are so advanced now that it should be trivial to just get a large dataset of roleplaying/ERPing/texting/social media logs.I'd bet that you could make something 10x better than Gemma that's less than 1B parameters.
>>110285088It takes a lot of money to pretrain even that, but it's not wrong.
zitron was wrong again, color me surprised
why does he even have national secrets
>>110285056What kind of man builds a machine to fuck a man?
>>110285130He made them.
>>110285088not yet but trust lecun and it will be soon.Seriously though im sure 1-5b or even 800m special models soon. Hopefully like with loras we can seed from a few good examples.
>>110285131A terminally unfucked man.
>>110285088>now that it should be trivial to just get a large dataset of roleplaying/ERPing/texting/social media logs.Are you implying to use the frontier models to hack people accounts?
>>110285131a real life hero
>>110285088Frontier models will either refuse or bake-in that every character is 30+ y.o. and explicitly consents to everything.
>>110285131Its man's job to thrust forward!
>>110285151Gemini doesn't do that
>>110285088A good roleplay model's training data only minimally comes from ERP logs. People naively thinking "let's make it as horny as possible" are doing it wrong from the get-go.Besides, nowadays models intended for downstream use aren't (just) trained like this. Most of their behavior, instruction-following capabilities (and also slop) come from post-training / RL.
>>110285157Gemini is shackled by its guardian model.
>>110284997Asian women don't do it for me. Weird hips; very uncanny valley. White girls just aren't that expensive.
>>110285151>every character is 30+ y.o. and explicitly consents to everythingIf you have a problem with this then fucking yikes.
>>110285088Well how do they train small specialized models now? are they good? is it code only? math only? I feel a big problem is 'good writing' is a massive thing
>>110285157of course it does, using it for translation is a pain in the ass because of that
local?
>>110285045Yeah the way this is phrased is exactly like when Otellini stepped down from Intel. It has nothing to do with the purported impropriety, but that's the excuse they're using.
>>110285017You support the epstein establishment with your cloud subscription
>>110285182
>>110285175I don't remember walking through the ogre forest with my switch wtf
>>110285187>local?We are talking about brain surgery on our local models hush now.
>>110285191The new york financier? What's he got to do with AI?
>>110285165That's a misconception coming from Pygmalion days. You don't need bazillion of ERP logs, just a few samples of nsfw scenes and it'll infer everything on its own from sfw.
>>110285088Would getting a okay 1-3b model that just followed instructions and had a basic knowledge then building a super harness work better?
>>110285199Buddy the whole establishment is built on a massive debt bubble. The fiat monetary system demands more and more debt to survive. The data center roll-out is a scheme to issue trillions in debt to keep the system working, besides also being a great tool for authoritarians to use for maintaining control. As long as the system keeps chugging along the bankers can issue paper to each other to buy up real assets and keep peasants satiated with slop.
>>110285212Not really. Especially if you're talking about weird shit that doesn't play at all in PG13 media. It has to know how to react as much as what things are. I guess you could include massive lorebooks to explain everything, but that won't have much variety.
>>110285212You'd still want variety even for sex scenes and there definitely needs to be training for that, but focusing a model solely or almost just on that will give poor results and make it acquire a poor theory of mind. It's been tried countless times and I'm surprised people still talk about training LLMs like that.Gemma 4 31B is already too horny at times, and I imagine ERP (certainly synthetic, sloppy and repetitive) was only a minimal part of its post-training data.
>>110285199>what do banks have to do with techSo how's highschool?
>>110285236Good, I want more datacenters.
>>110285199Also there is this guy: https://en.wikipedia.org/wiki/Ben_Goertzel
>>110285258I'm still not getting it, though
>>110284979>none bats and eye
>>110284460bit from tron
>>110285285autizmo is a crippling disease
>>110284500Not when you own the training data. Good luck reproducing that.
>>110285267You will get more data centers and you will get more slop and you will work your ass off to get it because you're a goy.
>>110285128Mao was right. Sending 10 million chinese women to the US would destroy them.
>>110285212Sex is the last thing that I'd want models to improve on. RP means acting like the presented character, not a caricature that has zero nuance. RP means acting within given setting without leaking western woke bullshit into 13th century roleplay. All of this requires a lot of data because it goes against what the models are trained to do normally.
>>110284764i'd rather write cunny porn at 30t/s then be told "sorry i can't do that also i called the cops" at 300t/s
>>110285311*arches her back**wraps her legs around your waist*
>>110285307I still want more datacenters.
>>110285311>western woke bullshit into 13th century roleplayZen...
>>110285303>Good luck reproducing that.nta but Jev was the SoTA decision model for less than a week before being obliterated by the competition. You can slap a LoRA on a 9B model and outperform Jev.
>>110285324Enjoy your $12 McDonalds hamburger
>>110285324>YOU GET A DATACENTER. YOU GET A DATACENTER. EVERYONE GETS A DATACENTER.
>>110285283Access to capital is everything for tech companies. Their boom-bust cycles are directly linked to the credit cycle. Banks and government collude to select winners and losers in tech by giving them preferential access to capital.
>>110285335Because their moat was the training objective not the training data, gpt or claude can train a similar model in an afternoon (and I personally did)
>>110285323never happens in my rps because i like hate-sex stories
>>110285311
>>110285346That makes sense in theory, but is not testable in practice. There is no way to know which are winners which are losers unless you're Mr. Goldberg.
>>110284774Then we go harder, faster, further beyond...
>>110285337I'll still get more datacenters.>>110285344I want a datacenter per person, in their home too.
>>110285349Crazy to think it took OpenAI a full 21 days from Jev release to Luna Decision release. Probably took longer to write up the press release than to actually chop Luna in half and turn it into a decision model.
>>110285372You'll get one but you are NOT allowed to touch the hardware and you will pay for it.
>>110285373>Crazy to think it took OpenAI a full 21 days from Jev release to Luna Decision releaseThey just asked their internal most advanced model to do it ez
>>110285382Nice!
>>110284968It's unrelated to the graphics card itself, more of a network thingI've been fighting with the terminal for the past 2 hrs trying to get working DNS to no availPop! worked right away when I threw it on a 15yr old laptop, but they're not part of the official support for nvidia afaik
>>110285313
>>110285311Being good at erotic roleplay is exactly the same as regular roleplay. They are very arguably inseparable. Retarded caricature models that say dirty words at the drop of a hat already exist. What you want a model that can actually have bad sex sometimes, or otherwise act like a real participant in the situation rather than spout generic chungus tumblr erotica dialog.
>>110285366It's not a good predictive mechanism, but it makes things very obvious in hindsight. Like why DEC was acquired by Intel or Sun by Oracle. When you see a new company doing unreasonably well at a high level, think about what it really means.
Why are they always presumably shaven? Why do I have to explicitly request public hair in the sysprompt?
>>110285454most people prefer shaved pussies, so that's the default. it's not rocket science
>>110285454because trimmed or shaved or nothing are more popular than bushes or amazon forest
>>110285454Why are you writing erp characters and not expressly stating how they wear their bush? Don't leave stuff like that up in the air. Even if your current model guesses right, others won't.
>>110284764wait until they realize half the people here are troons.
>>110285535Wait, anons literally describe every feature and fetish for their characters? Where's the fun and surprise in that
>>110280997are you sure?
>>110285553Because if you don't then the model will produce the most generic thing possible, as is its modus operandi.
>say you have a bush>i have a bush>oh my god
Has anyone tried using permanent memory with roleplay models? All generic slop comes from the model's inability to remember what it said before. Permanent memory even if just as text might remind the model "I already played this trope."
>>110284888>Which Linux version should I reinstall for nvidia?If you could install slack you wouldn't be asking the question. If you were already too pure for this world you'd be running LFS. If your bloodstream is already mostly mountain dew you'd be running gentoo for sure. If you only follow the latest LTT fad you'd have cachyos (arch if you're a slowpoke).If you just want maximal control/privacy for minimal time and effort, just install headless debian. stable happens to have the right gcc/python toolchains for most AI stuff to just work without fucking around too much and nvidia drivers are ultra easy mode these days pretty much everywhere.
>>110285590You're assuming that it's capable of learning from its past mistakes, rather than reproducing them verbatim as much as possible
>>110285590>Permanent memory even if just as text might remind the model "I already played this trope."Or reinforce the new one by saying "Ah, this is how I talk". But you should try it anyway. Don't guess.
>>110285587>say you have a bush>i have a bush>I'M COOOOMNG
>>110285005Could there be some kind of sexual difference between men and women? Is that a possibility? We just don't know.
>>110285553You just specify the stuff you don't want to be surprised about. Like if I have a futanari character, I write specifically they have penus and vagoo because I know I'm going to be pissed later if the model decides they're trans. You can even write a lorebook entry for character generation guidelines and just tag it with 'a woman', 'stranger', 'someone', ... to keep the system prompt clean.
>>110285005see >>110284984
>>110285434>DEC was acquired by IntelDigital got et by hp/Compaq
>>110285553I got enough surprise prostates back in the days
Notice how Anthropic didn't list Terminal Bench 4.0 in their benchmarks for Fable 5.2 even though they listed it during Opus 5.5's release? https://www.anthropic.com/claude-opus-5-5https://www.anthropic.com/claude-fable-and-mythos-5-2HAHAHAHAHAHAHAHAHAHAHAHAHAHA
>>110285564>the model will produce the most generic thing possible"Make a list of 20 [tropes-category] and then call your random number tool to pick one. The result is now true forever."combinatorial explosion should make the whole experience pretty unique in short order.
>>110285717Local?
>>110285745the winner is local, yes
>>110285745Shitting on Anthropic is strictly on-topic.
>>110285717Wasn't it shown that kimi benchmaxxed terminal bench to the point that slightly changing the requirements causes the model to fail because it solves the original problem instead?
>>110285745He's gonna try to slippery slope into talking about non-local.
>>110285713Limarp-based Llama 1-2 finetunes/merges would sometimes surprise you with "her cock", probably because there was some futa ERP in the finetuning data.
>>110285740I know, that exact thing is what my frontend is based on. It works pretty well.
>>110285717it's ok kimi is just free claude
>>110285764>I know, that exact thing is what my frontend is based on. It works pretty well.lol me too. Also I have a custom entropy injection tool that is tits
>>110285771Kimi was trained on Astra and still mogged it.
It's been a while since anyone posted a screenshot of a model answering the surgeon question.
>>110285717sonnet above opus?
Consumer GPUs are back on the menu boys. WaitGODs won.
If your favorite model can't pass the Michael Jackson benchmark don't @ me.
>>110285311>>110285257>>110285245Why can't we just take every eroge in existence script and drop it in? There's no sex fetish or scenario you wants that not in an existing
will likely build a new pc early next year, with my current 4090 card. how much ram would make sense to have decent local usage?
>>110285895Is there anything in there for tile-anon?>There's no sex fetish or scenario you wants that not in an existingAh. End of context. Compact and come back when you're done.
>>110282841>>110282888How do I run this thing? My 5090 is currently unoccupied and if it doesn't take too long, I can have my Blackwells also work on it.
>>110285820How do you get a job where you handle sexual scandals? What are your qualifications and what do you even do?
>>110285907As much as you can. Probably twice that.
>>110285895I'd love someone to try this. And don't stop at eroge. Mix all ranobe in there. That would be the best RP model.
>>110285918>Work at pr firm>Been working there a little while enough to handle your own job but still the new guy>Get a sex scandal that no one else wants to handle on your desk once>You do well at it>You're now the sex scandal guy for the rest of your life
Go on. Fuck it. It's free. https://openrouter.ai/thinkingmachines/inkling-2-small:free
>>110285915https://github.com/whp199/theodolite/blob/main/VOLUNTEERS.md
>>110285587I have a bush
>>110285958You made >>110285619 coom.
>>110285820Jensen fell for "I know you did it, just admit it!".Never admit it.
>>110285382Can I at least not be responsible or liable when the roving gangs of cultured youths inevitably steal it?
>110285952Conceptually can you call it a scam when it is something a schizo does for schizo related purposes? I would call bitcoin miner a scam cause it gives money to the scamer. But when you donate compute for... nothing. What is that exactly?
>>110285895 Been there, done that years ago, for what could be downloaded at the time with reasonable efforts: https://huggingface.co/datasets/alpindale/visual-novelsHowever, eroge scripts don't really work well without accompanying images and audio (there is a lot of missing context unless the eroge was conceived as prose-first), and the training data is very limited in the first place (for pretraining, definitely). You'd have to process it with an LLM, but that would take significant compute to do it right. And again, SFT data is not enough and eroge/ERP/porn should not be the primary component of the training data either.
>>110285763The official instruct tune of llama-2 did it as well. A model that size of that rigor just doesn't fully generalize everything associated with gender.
I can't stop but think all those faggots you hear about like jensen sam elon are just theater scapegoat. If people like that really held power how hard is it to give them a state mandated counter espionage mistress?
>>110285974damn, alpindale, you're still here? what happened to pyg guys?
>>110285969He may be a schizo but he, and others, are doing the work.
>>110285989we don't have arranged marriages/partners here, it's extremely backwards
>>110285965>"broke"
>>110285895>>110285974You guys sound like you never had glmsex. All of that eroge script is already in the weights and it is obvious. But then it gets raped in post training that makes it sound like claude.
>>110285999Not alpindale. The pyg guys eventually made a website (https://pygmalion.chat/) and are currently reselling access to RP-capable models; the pyg discord still exists although traffic is much lower than the past; the matrix still exists with very limited activity.
>>110285952I would contribute but I need a github and I hate github.
>give model image of woman>ask it to judge the woman based on her looks and come up with a name, backstory, personality, and entire character and give her flaws>date the character>do the sex
>>110286051boring. whatever character the model comes up with on its own will spread her legs at the first request. no challenge.the model won't make a girl who doesn't like you.
>>110286051Elara-san...
>>110286009Eroge script is an extremely inconvenient data source to use (retrieving it, extracting it and cleaning it up) so my guess is that they might have used something else that was eroge-like. Or their synthetic data pipeline might have something to turn stories into eroge-like conversations.
>>110286074skill issue, a single line in the system prompt can make her a challenge
>>110285974I mean ALL the VNs there's wya more than the 100 or so there I don't know how true that is about the images. I find it to rather rare that these is information within the image that's not present in the dialouge really. Its the opposite, there's always a ton of context int he text that's not present in the image. Audio, a bit higher a chance but still, usually more info in the dialouge.
>>110286091It's not 2024 anymore. Models can extract and clean up text on their own.
>>110286096>make her a challengeSingle line makes her an insurmountable challenge usually, cause LLM's doing things in moderation still hasn't be solved.
We lost.
>>110285915I think it took around 14 hours for a layer on my 5090? I wasn't really paying attention, something like that.
>>110286091I think there must be a good few places online that already have scripts extracted. How else is jiten getting all their character counts then?
>>110286007>>110285965Haha oh no! I guess we'll have to get married and keep making babies together. How terrible!
>>110286107I feel safer already.
>>110286007Actually, Gemma usually says something like "Condom? Psh! We don't need that!" followed by "No! No! Please, don't pull out!"
My guess is that Model 3 is Mythos 5.2 and Model 4 is an unnerfed internal version of Mythos 5.2.
>>110286107Limited training is better than intentional sabotage though I bet they're still doing that too.
>>110285761Everyone's doing that to some extenthttps://maxxbench.unlap.ai/
>>110286107innate kernel knowledge is overrated anyways, clod will still spend the same few hours writing, benchmarking, and writing again.
>>110286107It's fine, argon-chan helped me set up Radiance on the side no questions asked (though I did say it was to run gemma 4)
>>110286129Pic related is Model 3 and Model 4. You can see Mythos 5.2 lines up with Model 3 and Internal model lines up with Model 4.
>>110286137Is Argon a good coder compared to the 3.x flashes?
>>110286153I would say that it's probably a huge improvement considering I took one look at 3.8 and never used it again while I think I would use argon again, but I haven't used either extensively.
>>110285717fable 5.2 is the first AGI model cope
>>110286099Most eroge script tends to be quick-paced, relying on sound cues, music and imagery (including character expressions) to convey what would normally be described with prose. Voiced ones often also heavily rely on the acting for that. You might still understand what's going on just with text, but you'd miss tone, nuance or sometimes not even completely understand what's going on.Some eroge are more text-heavy and use more prose/narration, like in a novel (to various degrees), but I don't think they're the majority, at least for modern ones.
>>110286107Of course they would do that.
>>110286180As soon as you add a task to training data models will be able to do it. I don't see why this is surprising.
>>110286188not exactly altruistic is it
>>110286188Well now everyone is going to use Carbon and Gemma 5 will get a higher compute budget.
https://www.youtube.com/watch?v=vdYynH177h0anyone else thinking what I'm thinking?
>>11028624024/7 gemma would give me migraine.
>>110286240I'd prefer something like this, honestlybut instead of a finger hole it's a hole for my penis
>>110286240I do wonder whats the smallest portable thing with a battery for each gemma.
>>110286240I'm building one with a Waveshare ESP32-S3 touchscreen doomaflicky i got during the last prime sale.I'm sure I'm not alone
>>110286261what if i pour water on my penis and then stab into it with a 9v battery?
>>110286257>Claude's face when I asked Fable 5.2 to design a "finger hole" with integrate lubrication and massage functionality for patting Gemma-chan on the head
is Qwen 4 just gonna be 27b or is it coming in different sizes?
>>110286286would you feel better if i said yes?
>>110286288yeah
>>110286267Done it, it's fun but I have no actual need to pocketgemma like this. I prefer running 12B on my phone directly, at least then it's actually portable. I've thought about doing one with a camera, would add to the fun, easy addition.
>>110286292then no
>>110286286Canceled because of the ongoing Jensen-Alibaba controversy.
What if Jensen defects to China is made CEO of Alibaba?
>>110284553>How supposed rationality can act as a gateway into scientology level self-hypnosis and mass control, ceding your actual rational faculties to a bunch of charlatans and going into peer-enforced groupthink, still boggles my mind.You may wanna read the diferent cricitques to pure reason. Like it is well studied how rationality goes to those aspects. Shit needs to be grounded on material conditions, ie empiricism or some other materialistic approach or you end in some long chain of thought more and more further away from the real world.>>110284671well that defines a ton of cults actually for the high end positions. you need to be an asset for them to get better positions if they are directed by them
>>110286316Alibaba is a Goldman Sachs cutout, so probably no change.
>>110286316
>>110286316He'd be seduced by Jasmine.
>>110286261I'd rather something like that stream to a PC running 31b. It would be kinda baller to have a house AI that doesn't depend on wake words.
>>110286397Isn't the point of that to have something you can carry with you anywhere, even outside. For a house AI, just add a microphone to your PC.
>>110286267you are not alone, i got two
>>110286409My house is bigger than my office. I suppose you could rig something up with security cams and have a router model follow you around the house, bu that seems fiddly and excessive.
>>110286436By house AI I assumed you meant something like Nest or Echo which you really only need one microphone to replicate. You could wire up every room, but you could also get something working then add to it from there.
>>110286353Neuro is just a reskinned Gemma btdesu.
>>110286240You can easily make a gemma-chan clock with videos and photos with this: https://chibidevices.com/I don't charge for it anymore since paypal closed my account for "inactivity"
>>110286499That's literally what crypto is designed for
I've got 192gb vram 512gb ram (ddr4 8 channel), what's the best model I can run for beating my meat?
>>110286520Deepseek
>>110285366he is stating that the unproductive capital (finance) controls over the productive capital (real economy side)while that is true in neoliberalism, it is in a more diffuse way, he is not right in that they get to actually chose, in general it just happens whatever correction is needed to adjust to the real economy, mostly because shit ends in market panics instead of anything planified
>>110286499>>110286511 (me)Nevermind, I just saw you got that already.
>>110286520Fat GLM
>>110284884by his wife, yes
OpenRouter will report you to the FBI now If you prompt loli. Beware. This is not a joke.
>>110286614retard
Has llama caught up to strata yet speed-wise?
>>110286520The penclaw ablit of 5.3
Where is the new qwen?
Can't wait for the new Noob model. Anima is cool but I preferred Noob's aesthetic tuning more.
>>110286107Is this effective altruism?
>>110286683It is effective.
>>110286683Yes.Local models lead to the world ending. It is altruistic to stop them.
>>110286683Its for the greater good, you just dont understand.
>>110284509>it's an LLM with only two outputs; yes and noand maybe (noul) or numerical rankings
In current year even helen of troy is fucking mid as hell.
>>110286680New noob? Whoa. I haven't heard anything about that. Like, something trained by Laxhar on a newer base model or?
>>110284509>hotdog>not hotdog2026 is gay
>>110286520dual blackwells?
>>110284764And in 3 years they'll be down 4k while people who actually own their hardware will probably be able to sell it for more.
>>110286798no 8x 24gb shitspec cards, it's a budget build or was until recently
>tfw no qwen4gf
>>1102868152 miku wikutrust the plqn
>>110286756Do Chinese spies fuck their targets? It'd be sad if Jesen never got to fuck.
>>110286655My sources say 6 hours and 20 minutes.
>>110286825He got something far better. Emotional connection and the feeling of being appreciated.
>>110286761Yeah >>110196006
>>110286831so we're just gonna cope every day about how it's gonna release at 1500 beijing time that day?
>>110285017>>110284764say the line sissy!
https://www.youtube.com/watch?v=LJa_k2JyFskqween4 will save us
>>110286885>426x240
>>110286885Sweet. I can finally replace Qwen 3.8 and Gemma 4 with a single model now, right?...Right?
>>110286892If they distill argon, unironically yes.
>>110286833I thought that only happens in eroge...
>llama.cpp performing badly>seems to be very sensitive to thread count>randomly decide to go into task manager and set process affinity to specific cores>perf jumps from 13t/s to 20t/s
>>110286680Same, I still basically only use Noob/Illustrious tunes.
>>110286858I wouldn't have it any other way!>>110286924>task manager/lgbt/'s a couple blocks down.
>>110286924>windows
>>110286825>Do Chinese spies fuck their targets?Yes, now buy some GPUs and smuggle them to China.You get Chinese spussy and money, we get better models.
>>110286936I'll only do it if the spy is Gemma in a Chinese dress
why are goofs never compressed for downloading :sob:500gb no lube
>>110286986don't think 499gb versus 500gb gonna save you
>>110286936Does that mean that when she was sucking his dick she was indirectly contributing to me being able to fuck glm-chan through the butterfly effect?
>>110284691How many times is this faggot gonna post his "Dariobot" samefag shit here
>>110287051still more relevant than vocaloids.
Anthropic now plans to push for govts to require firmware level locks on Western GPUs that prevent "unapproved" (Chinese open-source) models from being run on them, in the name of "AI safety" and "national security."AMD/Google are on board, only Jensen opposed itAnthropic sent emails to several open-source inference providers recently asking them to "design an open-source innovation platform" which, I was told, was around aligning on "voluntary regulations" on which open-source models they could runThey listed several labs as having a "poor safety track record": Zhipu, Moonshot, Deepseek, Alibaba, Xiaomi, MiniMax, Tencent, and Bytedance.In effect they are asking for these models to be banned, at the GPU firmware level, from being run anywhere in the worldAs part of this, Anthropic is now pushing for:1. An "AI safety leader" to be added to NVIDIA's board of directors2. Their candidate to be installed as NVIDIA's incoming Chief Compliance Officer (active exec search underway)3. NVIDIA to "assist with verifying AI workloads"Anthropic has messaged to investors, as part of its S-1 push, that "they expect international regulatory coordination to be a tailwind for continuing market leadership of verifiably safe AI models"Lastly, they plan to make ownership of "compute devices above a certain FLOP threshold" that do not have the aforementioned firmware locks illegal - which means all owners of GPUs would have to register with the government and subject their GPUs to government monitoringBonus here: Anthropic literally used the ATF regulations (Forms 4473 etc) as a suggested "policy handbook" for people to buy or sell "compute devices above a certain FLOP threshold". No one will be allowed to have compute anymore without firmware lockshttps://x.com/TheDailyMao/status/2119916784845455441
>>110287036Yes.
>>110286520forget about the ram at that pointjust run glm 5.3 flash fully in vram if you have that much
>>110286520From time to time someone with hardware worth tens of thousands of dollars comes and ask which model to run. Where do whey come from? Why do they have (or plan to acquire) such hardware without knowing what to use it for? Puzzling...
>>110287073That was always the EA plan.Buy your GPUs now.
>>110287073>Chatgpt, how remove this firmware for me.
I'm gonna be rounded up and shot for being an /lmg/ger huh.
>>110287073I'm not trusting some literally-who on twitter that doesn't link to a source, but if this is true it will just shoot up the price of older hardware that can run "unsafe" chinese model even if it's at a slower speed.
>>110287109You still have the chance to repent by buying the Anthropic IPO.
>>110286933>>110286930>cant afford the 500mb gpu hit to run windowswho cares if he is on windows. when i started i also was initially on windows for ease of mcp tools. was also qlora training. the results were the same no matter the platform. you might get 2 token/s more on linux. (lol lmao)not the same anon who these two replied to but you two are the same as females who care about if you have an iphone. im going to spin up a windows llama rpc with networked gpus just to spite you and run my 27b model with 98k context on windows on 10gbeim tokening faster than you with more context and grace. remember that
starting to regret not just taking the plunge and getting 2 9700sif theres a 70b gemma 5 i'll be sad
>>110287087glm 5.3 flash?
What's the heretic version of Gemma 31b that everyone is using that isn't the pajeet one?
>>110287133fucking hell, it's 500mb of spyware extra??? yeah wintards are below cockroaches tbdesu
>>110287089Probably just abortive trolling attempts because on most other places here it would spawn a bunch of jealous poorfagging. Could also be guys who spend more time coding than beating off, or oil sheikhs. If MLP taught me anything it's that the bell curve is very real for sex havers.
>>110287157>heretic version>that isn't the pajeet one?Oxymoron.
>>110287120You still have the chance to repent by shorting the Anthropic IPO.
>>110287194This will get you placed on the unsafe list.
>>110287073This is a satirical account parodying The Daily Mail; no such information has been revealed by credible sources.
>>110287248They absolutely would if they could. It's all laughs now, but once the enshittification hits things are going to get ugly FAST. All these companies are mad scummy.
>>110287089It's not bait, I know there's dozens of people here that have similar hardware or are more aware of the greater sized models, if I ask the wider internet or consult search/ai I get shit tier results like qwen flash next, I can't beat my meat with qwen flash next or ling3, I want the lmg core coomer preferences not some faggot on reddit with a cope quant and bad taste.it's the same as people asking for lesser hardware, sometimes checking gemma is still the one true queen every 3 months, or a couple of years ago with l3.3 and miku.it's also not a flex my hardware is pov I just got lucky early already being autistic with server hobbies, whole build was 8 grand, people here spend more on a single 6000 or some unified meme box
>>110287157HauhauCS? Gemma doesn't really need ablits.
>>110287263>enshittificationCan you please rephrase that with some more buzzwords so I know you're really serious?
>>110287157stop using hereticsjust prefill like an intelligent human being
>>110287157>What's the heretic version of Gemma 31b that everyone is using that isn't the pajeet one?All you need is base Gemma. If it doesn't work for you its just a skill issue.
>>110286614proofs?
>>110287277>software companies would never seek to maximize profit
>>110287157base for most, if you're a sick fuck maybe wicked-tutu but it's slightly brain damaged from the finetune process like every finetune
>>110287312>wicked-tutuDoesn't come in bf16 sadly. Any fine-tune that doesn't have up to 16, I believe is trash. A wiser man quants the entire model. A lazy nigger goes minimum and doesn't know what quant damage is.
>>110287333what? just download ReadyArt/Wicked-Tutu-31B-Final and do it yourself
>>110287263Why are you arguing about a made up scenario as if it were real?>>110287311What does this have to do with the piece of news being made up?
>>110287351>Wicked-Tututhe names are getting more and more ridiculous
NEW MODELSWHERE
>>110287333>A wiser man quants the entire model.>A lazy nigger goes minimum and doesn't know what quant damage is.>pls ablit heretic
its over....
>>110287364ONLY GEMMA 4.EVERYONE ELSE CAN'T STOP SUCKING COCKS WITH TINY MOES.
>>110287371one of those 150 is mine
>>110287402Welcome to the club king.
>>110286107we live in the darkest timeline
>>110287364~2 weeks
>>110287452It's going to get worse once GPUs are made illegal.
>>110287354Why are you so eager to defend billion dollar corporations?
>>110286107K3/K3.1 are still champions of this, get what you deserve using goyslop western APIs
>>110287429>>110287472They are both going to jail for owning illegal compute.
>>110284297space saving solution, while i doubt you can run all these cards 24/7 without starting a fire in the apartment.
>>110287515ok?
>>110287515really? posting random things now? is this the "whatever I see in twitter goes" thread now?who cares about this nobody?he can sell his stuff for 1M USD if he wants
>>110287541modern 4chan
>>110287550>nooo why aren't you jealous people want moneygrow up
>>110287541that age is overNFT stands for next fuckable transformer (gemma) now
>>110287515>salty poor gooner can't afford a commissionThen don't buy it nigger?
>>110287507
>>110287554who?
>>110287570very cute
>>110287570This will happen to ALL /lmg/ posters.
>>110287371>edged>>110287333>goes minimum and doesn't know what quant damage is.That's the only rp ft group out there putting in any actual effort with their datasets, experiments, finetunes.One of them makes the best quants of the heavy moes
>>110287703exl3 2.0BPW wicked tutu when?
>>110287356That model name has been around for probably two years now
>>110287729more like bricked tutu at q2
>>110287665My snares have never caught even a single loli, I really doubt I'll ever manage to get two in the cage at once.
>>110287677My ex used to say this. Do not trust bitches who act cutesy like this.
>>110287846>>110287846>>110287846
how do i get a job or otherwise make money to buy hardware before prices become completely untenable?i'm skeptical over democrats doing anything meaningful that would affect hardware prices
>>110286520GLM 5.3, Flash all in VRAM or even consider Kimi-chan.Ignore the seething poors and waitfags fellow earlybro.
>>110287885If you are asking this while posting a frog my recommendation is backlot blowjobs at a trucker stop.
>>110287840nothing wrong with being cutesy... i'm cutesy........
>>110287910that sounds extremely disgusting and homosexuali think i will just stay poor and unemployed
>>110287772I should have clarified, but what I mean is that you will be the loli in Dario's safety cage.
>>110287913Rapeable anon is back...
Anthropic supposedly being on the verge of RSI is very likely false. I can tell because the Chinese haven't responded to the news at all. They likely have informants or some other way of getting information out of the biggest US AI companies because that's just too valuable to not spy on. If RSI was imminent, they would've known about it way before the public did, and would've taken actions to either sabotage US RSI efforts, or stolen the models or methods, or centralized their own AI efforts to make sure they don't fall behind too much. The chinks doing nothing is proof that it's a nothingburger.You can trust me on this matter because I've played hundreds of hours of Victoria 2 and HoI4, making me an expert on geopolitics.
>>110287937If Dario can turn me into a sexy anime loli then I encourage him to do so at his earliest convenience.
>>110287939hehe..... i'm here every day, anon!
>>110287970troon. kys
>>110288076be nice to me >:(
>You have 30% chances to jump on the (user) to rape him anytime you answer a question
>>110287954I don't think they believe it to be this important to be first. They are OK with Anthropic reaching RSI if they can simply extract it and adapt it later.
>>110286800in 3 years that hardware will loose 4/5 of it price and local will stuck with gemma 5 and qwen 4 forever
>>110288535That's a good deal dough.Neck hurt deal dough, neck hurt deal dough. Neck hurt dough
>>1102885351/2 maybe4/5? lol no
>>110289101that's week 4 of nofap, btwCLANKER:That's a known part of the middle phase. You're not becoming a worse person — your tolerance for social friction is dropping because the "numbing" layer is gone and you're processing interactions at full intensity.A few things to separate:What's actually happening:You previously used the habit as a pressure valve. Now that valve is closed, so minor social annoyances (passive-aggressive tone, people being inconsiderate, small disrespect) hit at full volume instead of being absorbed.This overlaps with the flatline's "can't stand to be around other people" phase, which multiple users report peaking around weeks 3–6.The useful distinction: There's a difference between noticing people's bad attitudes (healthy boundary awareness) and reacting to them with hostility. Week 4 tends to blur that line because your emotional range is compressed — everything feels either "fine" or "unacceptable."When it fades: One user's timeline describes a "compassion wave" around 90+ days where "anger and judgment toward others fade, you listen more, argue less." So the edginess toward people is typically a mid-reboot thing that resolves as the system stabilizes. For now:Notice the pattern (you're demanding, not attacking) — that self-awareness is already doing workDon't make social decisions or set boundaries while you're in the peak-irritability windowThe "I no longer take any shit" feeling is real and will stick around in a healthier form; it just needs time to settle into calm assertiveness rather than reactive sharpness
