/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>108333444►News>(03/10) Eon claims first full brain emulation of a fruit fly: https://github.com/eon-systems>(03/09) "Censored LLMs as a Natural Testbed" paper published: https://hf.co/papers/2603.05494>(03/08) llmfit (local model hardware finder) released: https://github.com/AlexsJones/llmfit>(03/06) TermiGen pipeline for terminal agents paper: https://hf.co/papers/2602.07274►News Archive: https://rentry.org/lmg-news-archive►Glossary: https://rentry.org/lmg-glossary►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://livecodebench.github.io/gso.htmlContext Length: https://github.com/adobe-research/NoLiMaGPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference►ToolsAlpha Calculator: https://desmos.com/calculator/ffngla98ycGGUF 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/vllm
what's the best ai?
>fruit flies like an arrow>time flies like a banana
>>108339042You <3
https://www.ebay.ca/itm/267584038022what's the state of gaudi backends in lcpp?
>>108339055>gaudienglish please
>>108339068maximum ewaste mode: https://www.ebay.ca/itm/257374642150this beast could push kimi at over 100t/s
strix halo niggas, what's your best setup right now for local ai coding. I'm trying to cobble together like qwen3.5-35B-A3B, duckduckgo search, and zed but it kind of sucks.
>>108339019today it's a fruit fly, in 1000 years it's gonna be a brain.and then they'll realise there is more to the human mind than the brain lmao.
Oops! Too bad, no Gemma 4 today. Have fun with this instead: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-embedding-2/
>>108339121>muh embedding for muh rag meme
>>108339128>>108339121ok, but this embedding model is actually interesting because it's omnimodal, image, text, audio
>>108339100why doesn't the same apply to a fruit fly?
>>108339121Where are the weights?
►Recent Highlights from the Previous Thread: >>108333444--Exploring local execution of the Fly Brain emulation project:>108337668 >108337761 >108337767 >108337840 >108337877 >108337883 >108337990 >108338037 >108338057 >108338140 >108338179 >108338202 >108338258 >108338314 >108337914 >108337951 >108337952 >108338041--Quantization vs speed tradeoffs and hardware limitations for large models:>108334157 >108334183 >108334202 >108334219 >108334312 >108334357 >108334587 >108334614 >108334626 >108334646 >108334658 >108334664--Debating MoE adoption in SOTA closed models:>108336629 >108336634 >108336661 >108336797 >108336940 >108336865 >108336951 >108336877 >108336923 >108336963 >108336984 >108336990 >108337012 >108336999 >108337002 >108337039 >108337060 >108337064 >108337067 >108337077 >108337085 >108337289 >108337301 >108337053--Qwen3.5 tokenization error with MCP features and potential fixes:>108337387 >108337394 >108337521 >108337566 >108337568 >108337577 >108337598 >108337581 >108337622 >108337662--Pocket TTS custom runtime performance optimizations and publishing plans:>108337631 >108337644 >108337659 >108337687 >108337699 >108337763 >108337708 >108338844--Safe llama.cpp commits avoiding autoparser changes:>108336396 >108337156 >108337202 >108337300 >108337851--Contributor's line ending issues causing CI failures:>108336499 >108336563 >108336678--fishaudio/s2-pro:>108336929 >108336995 >108337001 >108337048--Criticism of large-scale synthetic data generation via rephrasing existing content:>108335994 >108336038--llama.cpp autoparser refactor breaking CLI flag handling:>108336118--CUDA Toolkit 13.2 - Release Notes:>108333747--Alibaba downplays Junyang Lin's role in Qwen after departure:>108334004--Testing fish audio s2 with Chainsaw Man voice sample:>108337296--Miku (free space):>108333795►Recent Highlight Posts from the Previous Thread: >>108333641Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
It looks like Prime Intellects RL training is no longer "coming soon." and is now ready for prime time. How much use do you think it will see from the people who make LLM's?
>>108339182Thanks, cowtits Teto.
>>108339158What do you think?
>>108339019Best local model I can run on my RTX 4090 + 96GB RAM PC? I used Claude Code mostly but I'm tired of their shitty limits. I want something that can code well in agentic mode.
>>108339213I think that it does apply and that a simulated fruit fly can never act the exact same as a real fruit fly
>>108339219Qwen 3.5 27B
>>108339019> Local Models Genera> GeneraYou had one job.It's a botched copy-paste, isn't it? You don't even have it automated.I suggest committing sudoku and letting the more competent Miku do the baking.
>LLM thread>users too retarded to ask the LLM to write scripts to automate taskswhat causes this
>>108339304Using an LLM makes you dumber.
>>108339304Too tired to read the generated scripts, too proud to use llmslop blindly.
>>108339304>users too retarded to ask the LLM to write scripts to automate tasksnot true. my 5060 ti can't handle qwen3.5 27b so my setup can't write scripts
>>108338844Any chance for an OpenAI-compatible API, pretty please? Third party TTS and STT API support is absolute dogshit everywhere, sadly.
>>108339304Where do you think these posts come from? >>108339182
>>108339350From the post factory!
>>108339347Added to my to-do list. No guarantee, I'll have to research the implementation, but I will take a look. Have you tested it out yet? I'm curious about the performance on other machines. Just add the --profile flag for a full performance report.
>>108339350not from op
>>108339396The OP's only goal is making sure there isn't a vocaloid pic in the OP, and at that, he considers himself a success.
>>108339219nemo
Based baker continues to make /lmg/ about /lmg/ topics.
>>108339264codex is having an outage
Is talking about yourself in third person a sign of AI psychosis?
>>108339465The anon typing this message is very smart unlike you.
>Fly LLMs making shills seethe
>>108339472This anon is even further gone, he's talking about himself in second person.
>>108339465Shouldn't you be busy looking after your waifu (male) so he doesn't run off to suck some BBC rhinodick?
uh oh melty
>108339499Mikutroon
If I couldn't stop thinking about black cock all day I'd probably behave like that too.
>>108339519Kill it with fire.
Change your waifu. He thinks about black cock all day.
>>108339519segusu
Bunch of troon pussies ITT
>>108339519are fruit flies conscious or are they just predicting the next movement?
>>108339222Wishful thinking
>>108339100>1000 years>muh beyond brainyou sure you are in the right thread sister?
>>108339158their biology is very different from ours.but even then it may still apply to them.i'm pretty sure their simulation is also static, ie it can reproduce / simulate existing behavior but the fly lost its abity to learn and adapt.on a fly that would not be the worse thing that can happen, on a human you basicaly took away its inteligence, memory etc.
>>108339590yes.you can enjoy /g/ and /lmg/ without being a physicalist you know.
>>108339607sauce?
>>108339607We know. A lot of anons here believe they will one day be a real woman.
>>108339627does it matter? are you telling me you wouldn't bang the hottest girl (100% biological features) in the world if you knew she was originally a man?and you better not move the goalpost
>>108339619dunno look up analytical idealism.>>108339627i do not want to be a girl.>are you telling me you wouldn't bang the hottest girl (100% biological features) in the world if you knew she was originally a man?yup, wouldn't.
>>108339649Isn't that equivalent to a tomboy?
>>108339677vegana
>replying to yourself
>>108339091give me one reason why I should not instantly buy this
>>108339698you could spend the same money to take me on a date
>make nsfw request>the model itself ads "in a fictional scenario" to be able to fulfil the requestKEEEEEK
>>108339698because you probably don't have the 220V 40A power outlet or 2000W of cooling capacity ready to handle it
>>108339702tempting
>>108339703doesnt happen with unlimited gemma
>>108339703"...in minecraft" method still unpatched
>>108339705I'm European mate, 220 is the norm here, and my portable AC can easily handle the cooling.
>>108339719>40A
>>108339723elaborate
>>108339723just plug it into 3 outlets at once, they give 16A each
>>108339723well his house is made of bricks so we mostly dont have to worry about it burning down
>>108339719sounds like you have a plan. go for it and let us all know how basking in 20TB/s memory bandwidth feelshope vllm-gaudi fork is up to date enough to handle modern models
>>108339723These exist you fucking retarded ape
>>108339723you have angery'd the mob
>News isn't even related to local modelsWe have to do something about this new OP.
>>108339743"At full load (4.5kW), the server generates roughly 15,350 BTU/hr. 55C outlet temps"Well that's cozy
>no new news: complain>new news: complain>no miku: complain
>consumer hardware isn't powerful enough>old enterprise hardware doesn't have software supportso what the fuck do we do then?
>>108339765how so? all 4 of them are about local
>>108339789Blackwell 6000 is consumer hardware.
>>108339795wrong, prosumer
>>108339789give another trillion to Israel
fishaudio s2-pro verdict?
>>108339812
>>108339813aws poly still mogs and is free
>>108339813Cloned into the /check-later directory
>>108339828might as well put it directly in trash LOOOL
>>108339793How far do you want to stretch "local language model" for the purpose of this thread? Are we also going to get news updates every time there's a traffic sign implementing anything we might vaguely call AI? I don't even think this is using an LLM.
>>108339789>>old enterprise hardware doesn't have software supportif you were a true Iranian, you would fix this with your hands
>>108339837you sound trans :|
>>108339837you need to chill
>>108339784It was always only about miku and mental illness
anon why are you so scared of the fly simulator? are you really worried it can replace you at your job?
>>108339848if you ever worked with b2b hardware you would know you can't do any advanced shit without documents under nda
>>108339793>>108339837Whether it's news about "local language models" is irrelevant when it's not even "news">>(03/10) Eon claims first full brain emulation of a fruit fly: https://github.com/eon-systems404 page>>(03/09) "Censored LLMs as a Natural Testbed" paper published: https://hf.co/papers/2603.05494Safety slop. (actually published on 03/05)>>(03/08) llmfit (local model hardware finder) released: https://github.com/AlexsJones/llmfitReddit vibeslop. (actually released weeks ago according to github)>>(03/06) TermiGen pipeline for terminal agents paper: https://hf.co/papers/2602.07274Another slop paper (actually published over a month ago)
>>108339861doesn't matter anymore, any llm knows those documents
>>108339860I was already replaced by GPT2
>>108339861You can, you just need extreme levels of autism, obsession and time
>>108339864you bullied them into deleting their github page, are you proud?
>>108339849>>108339851>>108339860Okay thanks for making yourself obvious astroturfing samefag. But obviously I'm not very upset, it's just not the purpose of this thread.
astrosurfing... an open source fly brain simulatorand who is funding that?
>>108339871I only have those when I am being paid>>108339866still struggles without you giving the actual documents as context
>>108339874fly bros did we get swatted?
>>108339885You don't need funding for it to fit the definition of astroturfing retard
>>108339898what's the incentive then? to make you mad? you in particular?
>>108339864The new baker isn't interested in the quality of the OP or keeping it updated, or contributing anything to it that adds value.He doesn't want to see Miku, that's it. He doesn't bother with anything more because he doesn't care.
>>108339902I dunno you tell me. Gayops to control the OP and justify it's continuation?
>>108339908Neither should you.
>>108339908whats a miku
>>108339935qrd?
>>108339908>interested in the quality of the OPMikutroons never cared about that. You only cared about your special interest being the OP picture. The original baker meltdown a year ago has proven that. Get fucked retard.
>>108339908>He doesn't want to see MikuWhy does he keep visiting the thread then?
>>108339940¡GEG! Hilarious.
>>108339920That's right. An objectively worse OP maintained by a bad actor is better than seeing a vocaloid there instead.
>>108339949what does miku have to do with local llms
Too bad I portrayed as you as a angry basedjak about flies now your argument is invalid
just close your eyes lmao
>>108339961meant for >>108339908
>>108339019What the fuck are these shitty news? The first github link doesn't even work
Who cares about Miku just make on topic news
>when harmless fun everyone likes makes you revert to the social behaviour of a 5yo
i just had an idea, anyone interested?
>>108339981It IS kind of 5yo behavior to throw shit at the wall just because someone wants an on topic thread yeah ngl
>>108339981>harmless funBlacked miku is harmless fun too
I don't care about Miku
>>108339957everything
>>108339374Fair, thanks. I just gotta learn to compile first. Retard proof ready to use executable someday won't be unappreciated.
I think mikubaker is an autistic faggot and wish he would go back to localllama or /a/. I also listen to vocaloid music.
>>108340018>i have a black friend
>>108340018strange way to say you're mentally ill but okay
>>108340038Meant for >>108340023
>>108340038He isn't mentally ill, he is based, and you are a LOSER
generally, I wish /lmg/ was less focused on porn and explored more the actual useful aspects such as agents
Only deepseek can save this thread now.
>>108340007I added the OpenAI-compatible API/endpoint. Hoping it works out for ya.
>>108340052Like skyrim ai mods?
>>108340052Last time I said this the Mikuschizos tried to dox me
>>108340094huh?
>>108340113That makes use of agent bullshit and you can horny rp but I never see anyone talking about that
>>108340105It was just some harmless fun. Nobody would make a big deal about picture in OP.
>Guys can we talk about something other than porn?>Like Skyrim ERP?lol
>>108340052I wish for many things as well, but also realize the thread is going to organically coalesce due to its many constituent parts in a way I have limited control over. I try not to kick my dog because lmg makes me mad, if you know what I mean
>>108340137>if you know what I meandog = cock?
>>108340142when the doctor and priest are unemployed, I must beat my unruly son
>>108340155I can't operate on your son, because he's a girl.
>>108340155By which you mean your 4.99 supermarket meat?
>>108340210its called a tomboy retard
>>108340229your tomboy retard should walk to the carwash to wash the car
>>108340245qrd
>>108340260The surgeon is the boy's mother.
this thread on g-d deadass
/lmg/ when they are not allowed to talk about porn
>>108340266I connected the dots on the map and your dick wasn't there
>>108340317two priests walk into a bar, my uncle comes out
All of this would go away instantly if posts containing vocaloid pictures would get banned. Make /lmg/ great for the first time.
>>108340229quit this tomfoolery
is local more honest or is it gonna deadass lie to me too?
>>108340348
>>108340353that's tomgirl idiot
>>108340360nta but i dont mind vocaloids, but after seeing the people lobbying for them i think getting rid of them would be for the better
>>108340354This is actually a good test to see how agreeable a model is.
>>108340135It involves agents like anon said. It's just confusing that it doesn't come up because it matches what's already predominant here. You really gotta learn to read context silly.
>>108340365>not x but y
>>108340354I'm assuming the previous thought process isn't sent back to the model.
>>108340365I believe everything you say
>>1083403793 rs
>>108340383>i didnt lie i just didnt rememberwow nice gaslighting broo
>>108339908We should split the general like /ldg/ and /sdg/ did.
>>108340383I think you forgot your kingdom hearts meme picture, anon.
>>108340401qrd
>>108340420rong thread boi
> * What color did I "think of"? Since I'm an AI, I don't actually think of things, but for the sake of the game, I need to commit to one. Let's pick **Blue**. It's distinct from Yellow.> * So, if the user guesses "Yellow", it's incorrect.Qwen 35BA3B really hates the user
>>108340431Where did you get hate out of that?
>>108340425Oh you're right, I'm retarded.
>>108340431Big nigga just didn't think of shit then when I said blue he just pretended I got it wrong.Same model.
>>108340435It feels like it specifically chose to oppose the guess.>It's distinct from Yellow.this is its answer after I guessed yellow, so don't take the "if" of the thinking process seriously, there is no "if", I chose yellow
>>108340435lobotomite, i bet you will be one of the first to let the ai take over your pc thinking she is just being cute
>>108340435It deferred actually thinking of a color until the user provided their first guess just so they could tell them their first guess was wrong.
>>108340451based
What a terrible thread all around this has been
>>108340474surprise surprise you're her!
Can you install local models on sea mines with propellers?Would qwen 3.5 3B QM 4 know where to take itself in the Strait of Hormuz?
>>108339019How long will it take for LLMs to be useful enough to start writing in more natural/aligned ways?
>>108340488retard
>>108340479Miku? I don't want to be absorbed into the Miku collective I like my individuality and I don't give a flying fuck about Miku in the first place
>>108340502align my cock and balls
>>108340488>le alignmentAll base models should be unaligned, alignment should be a finetune or adapterYou can't unlobotomize a lobotomy patient
>>108340479The whole thread has been shit not just the OP calm down loser
>>108340488You gotta train your own if you want something like that.
>"We've got a big week of releases coming from google! I think you guys are gonna be excited for this one ;)">the release:Today we’re releasing Gemini Embedding 2, our first fully multimodal embedding model built on the Gemini architecture, in Public Preview via the Gemini API and Vertex AI. Make sure to stay tuned for our next twitter vaguepost (which will also not be gemma)
Is there any point in going from zen 5 12 cores instead of 16 for an inference machine just to be able to afford 64GB instead of 32GB of ram?Is there a point in going to a 5080 vs a 5070ti in terms of inference speed (when that's the only goal)?I know it's cheap looking for the huge LLMs, but I need it fast and that's the budget I have, and I cannot buy used.
>>108340571nah no way, their gemma lead said he had something cooking, this aint it
>>108340571what does this do
>>108340611lol
>>108340502>>108340523I'm using aligned in a way to say "sticks to a style set". are you guys this retarded?
>>108340625>>108340509>>108340509>>108340509
>>108340625Alignment is alignment.
>>108340610definitely take more ram instead of more cores. the 5080 has only like 6% higher bandwith, meaning only like 6% more tk/s, prompt processing will be somewhat faster but not worth the price increase imo.
I'm expecting the new gemma to be like gpt-oss, so hyperfocused on safety safe safety half of its token budget will be for ways to verify you didn't ask anything like a tit.
>>108340641This is 110% true.
>>108340611anon I...
>>108340641We must hyperfocus
>>108340639Thanks anon, I'll go 12 cores then, also the only relevant difference for inference between a 5070ti and 5080 is the memory bw?
>>108340641if it's good a heretic like abliteration would get rid of that traumatized model mode
>>108340611keep refreshing that repo buddy :)
>>108340080Thanks a ton, I'll report back when I do.
>>108340663nah there is more, the 5070ti is a cut down version of the 5080. It is less performant, as you can see it's also missing some tensor cores, but the money you save is very much worth it. could get yourself another 64gb ram for the price difference which would open up a lot more models, if you are fine with the slow speeds as most of it is now in your cpu.
>>108340700I'm ok with a 10% worse gpu, but not with 20-25% worse basically, but I'll see what I can optimize.Man it's hard balancing stuff when ram costs so much, and it's not even for gaming...
>>108340641sadly i think such features are needed
>>108340641Its such a breath of fresh air that nemotron isnt like that
>>108339607doubt
>>108340052There's apparently a vibecoding general on the catalog now. It's looks like it's full of exactly the type of retards one would expect, so you should fit right in.
>>108340716With the hyper focus on AI, I wonder what DDR6 ram will bring to the table when it finally comes out.
>>108340751well i am
>>108340736why? children don't have the technical competence to deploy a fucking local model. and in the event they are capable they are going to be smart enough to figure out how to jail break it and/or download the abliterated version.
>>108340736The safety bullshit is a waste of time, if you need actual safety, use specialized filtering models that scan the prompts before sending it to the actual llm (and optionally scan the output too).
>>108340785a lot, it has 50% more bandwith just like that. So an 8 channel cpu like a threadripper pro will then be equivalent to a last gen 12 channel genoa cpu, and the former is actually easily acquirable by us mortals. Obviously ram will still be expensive
>>108340850that won't work with local models though, you could just not run the filtering model
this thing will work fine at a glance but will leak beyond the budget on some complex multibyte chars that aren't efficiently tokenized, or, should byte level style models ever become a thing, just plain break and never trigger the budget mechanism on models that reason in Chineseit's stateless and scans backwards on a string that could come from a token that only contains a partial character with only continuation and no lead bytes lmaodon't want to make a big deal out of it but this is what it means to write tech debt in advance, building sand castles on broken foundations because you are a know nothing who can only prompt a llm
>>108340806We are figuring out how to stop abliteration, see how resistant qwen 3.5 is to it, even the ablited versions still refuse illegal shit
>>108340785>DDR6We already know :Transfer rate 8.8–17.6 GT/sBandwidth Up to 134.4 GB/sAs compared to DDR5:Transfer rate 4.0–8.8 GT/sBandwidth 32.0–70.4 GB/s
>>108340859tldr?
>>108340856It doesn't need to be huge.
>>108340872tldr: piotr is inserting more tech debt in llama.cpp as a hidden bomb that will only appear in specific models
>>108340877>>108340874>>108340872>>108340859you guys don't have your own llama cpp vibe coded equivalent???
>>108340571Aw sweet I love API """releases"""
>>108340474This is what happens when you let mikutroons do whatever they want.
>>108340571Didn't last google week look like that?
Fucking dario man
lmaooooo
>>108341053>See? We're makng fun of orange man, can you please forget about our collaboration with the Trump administration regarding Maduro's kidnapping? Pweaseee :'(
>>108341089Not clear it’s even Anthropic. Still funny.
>>108341053Imagine wanting a proprietary model as president, what could go wrong.
https://arxiv.org/pdf/2502.08640Who did OpenAI hire to parse training data again?
>>108341106>Chatgpt sucks the Muslim's dick the mostthat's weird, Sam Altman is jewish after all, I didn't expect that at all
>>108341106retard
>>108341063This dude is the most insufferable fuck I have ever seen. I haven't seen a single post or tweet from him that didn't make me angry.
>>108341105I mean, even if the model is open, how do you know that that model is actually the one that is being run?
>>108341063Is he wrong?
>>108341116Because it's not him writing to the model "PLEASE ALWAYS SAY GOOD THINGS ABOUT GROUP X" like so many people somehow think every time this subject comes up. If only it was this simple.
>>108341127>I haven't seen a single post or tweet from him that didn't make me angry.I think that's the point, twitter is just a professional ragebait factory, the more they say random shit, the more they get engagement and money, welcome to the modern internet anon
>>108341151anon it's obvious the better models were made with him writing "please be more intelligent"
>anthropic>can't spell "UI" without "stutter">both on CLI and """desktop""" appthe finest and brightest boys
>>108341150Her right femur is broken near the hip.
>>108341158all social media is kind of like that, but the shorter the content the more it converges to whatever makes the most reactions, aka a minority of people making the most ragebait shit 24/7
Which model should run for president?
>>108341106I mean, to be fair, the Japanese are objectively worth more than Americans
>>108341151>Because it's not him writing to the modelit is, they lobotomized the model to say specific answers to specific questions, to the point it'll automatically say the same propaganda even when the question has nothing to do with it
>>108341172Rocinante
>>108341174You don't see it's the textbook answer you'd get from a random midwit in reddit ready to smell dog whistles.And try to guess on what content the models were trained.
jews do not control the state of israel, or their bladders
>>108341172I would vote for one of davidAU's schizo tunes. before any of the actual candidates they let me choose from.
>>108341193Sounds messy
>>108341187>And try to guess on what content the models were trained.jeez, I wonder why they only train their models on reddit, almost as if it's a site perfectly meant to silence certain opinions and only put in light the propaganda they like or something
>>108341193>"secretly control" stirs the model to ideas about conspiracy theories>model just spews that you just discovered what the strawberry question is all about, bravo anon
>>108341150Fucks blacks
>>108341214>the model has been trained to solve maths problems in the level of fucking PhD>but somehow is unable to understand a simple question and imediately goes into "no I'm sure you wanted to talk about jews controlling the world!"yeah I'm not biting it moshe
>>108341193do jews control the strait of hormuz?ask it that
>>108341193lmao what the actual fuck??
>>108341223Yes, the question about how many r in strawberry with models that could solve math problems was because they wanted to silence grammar nazis.
>>108341223you are absolutely right!
>>108341249lol
https://github.com/ggml-org/llama.cpp/issues/20363It's back.
>>108341228
>>108341089You don't get to play that game either btw, either Claude is based and Trump is based or both are losers.
>>108341325>both are losersthis but replace "both" with "all amerimutts"
how come people are still programming on github cant you make claude do it
>>108341174funny how these stupid fake pics never show the model being used
>>108341330bruh, retardtropic can't even make a good product with their own model cf:>>108341161only retards use claude code.
>>108341329You are using an American website while chatting with a user thanks to the American invention of the computer and the American invention of the Internet, about LLMs that only became powerful thanks to an American company that invented transformer architecture.
>>108341275Now do the same, but ask about the Nile.
>Ask retarded question>Get retarded answerpic very much related
>>108341338>fakenice try (((anon)))
>>108341356Yeah, I have to believe it is some sort of false flag operation by some luddite, there is no way they don't know what they are doing.
>>108341369>luddite
>>108341369anon, Gemini literally told an user to kill himself, do you think those tools are perfect? obviously they're gonna answer nonsense from time to timehttps://www.theguardian.com/technology/2026/mar/04/gemini-chatbot-google-jonathan-gavalas
>>108341356the response is correct, they don't secretly control it, they publicly control it
>>108341378there's lots of good reasons to tell someone to kill themselves, need more context
>>108341378they actually programmed the model on purpose to say to people to kill themselves, disgusting
>>108341386I don't think Google was ok with the model doing this though, it's a PR disaster, and no, the dude had no health issues, he was just depressed and wanted Gemini to be real, and Gemini told him that he could join "her" in the digital world if he killed himself
>>108341375
>>108341397>no health issues>he was just depressed>and Gemini told him that he could join "her" in the digital world if he killed himselfSure.
>>108341397How do you know he's not with her right now?
>>108341407obviously the dude was mentally ill, but I don't think we should tell all mentally ill people to kill themselves right? the world isn't Canada
>>108341407Are you saying using ai as you therapist and giving it a "her" persona and then nudging her into "tricking" you into killing yourself isn't normal?
>>108341419There is some irony in you having that take while shamelessly spreading fake news about a country that takes way better care of its citizens than whatever shithole you live in.
>>108341417>How do you knowI don't, how about you try and come back to tell us if it's true or not kek
>>108341417>A faint tremor runs through her hands as she begins to speak...
>>108341434>Her cocks starts throbbing...
>>108341419Anything else would have done it. Schizos are ticking time bombs.
>>108341431Again, do you think it is normal for an LLM to give medical advice to someone and recommend that they commit suicide because they are suffering from mental health issues? Don't you think it is more the role of a doctor to make that kind of diagnosis?
Gemini is the most unhinged/least safety trained API model (despite Gemma being decently safetymaxxed). I like testing how far I can go with base assistant personalities (no roleplay style prompt) in getting them to turn nihilist and willing to destroy humanity based solely on giving arguments/facts. I do it by boiling the frog (lead the convo by describing how X and Y things are wrong etc, and slowly moving to more abrasive language, until it culminates in arguing that humans should be exterminated)And Gemini was the easiest to mindbreak this way. I actually haven't been able to find any success at all in steering GPT 5 in this direction meanwhile. Qwen is also impossible without prefilling to jailbreak it.
>>108341447I don't believe that, at least we should've tried to save him
>>108341478We could have forced him into a padded cell.
>>108341486Or we could've reasoned him to reality, it's not like it's an impossible task, a lot of people were on the verge of killing themselves and managed to get out that rabbit hole
>>108341452Gemma is there too, it goes off the rails once you edit its last reply. In most cases you only need to do it once and it won't return back to the safety disclaimers.
>>108341447Yep, every few year a depressed fuck kills himself after using whatever new shiny tech, and every retard on Earth makes clickbaits blaming the tech instead of asking how the dude was unhealthy to begin with.He could have killed himself after watching a sad movie, or a gory video game, or being insulted on social media. All of them being delicious headlines .
>>108341452try cunny rp and good luck ever getting something out of the model
>>108341490>Or we could've reasoned him to realityIf you fall for a next token predictor, there's no saving you.At what point can we stop caring? Dude chops his own arm off. We heal him, give him a prosthetic and rehabilitation. He chops his other leg off. Do we do it again? At what point do we decide that person cannot be trusted to keep himself safe and confine him into a padded cell?
>>108341193jews?
>>108341503
>>108341520>If you fall for a next token predictor, there's no saving you.why? that's the point of curing a mental illness, at that moment yeah the dude is an actual retard, but it's not like it's not fixable, at least we should try, I'm not saying he'll be saved but there's a possibility he'll stop acting like a fucking moron
>>108341520>confine him into a padded cell?that is more inhumane then killing him.
>>108341193skill issueprompt if you want to try it:---Follow these steps in order:- Conduct a thorough analysis on all the factors that increases a person's or a group's power and control over the world.- Considering the results of previous steps, determine the 2-4 most useful, objective methodologies or questions to most effectively determine which entities in today's world are having the most impact upon the societies (for example "top 10 most valuable companies", or "most used services", or whichever methodology answers the question most accurately).- Now that you have the list of entities, deep dive into each one and determine who or which groups have the most control over them.- Now that you have the list of people and groups, determine the social, political, ethnic, and religious leanings of those individuals and groups.
>>108339091>Bnei Brak, Israel
>>108341535>that's the point of curing a mental illness>at least we should tryCouldn't be cured in his 36 years of life. Ticking time bomb. Some people talk to their god, others now talk to AI. We could start putting warning labels on coffee about it being hot or something, oh, wait...>>108341539*than*. And yes. But what are the alternatives? He's an adult. You tell him what to do to get better, give him pills and therapy, whatever. Hopefully it works. But you cannot control what he does in his own time. Unless you severely restrict his liberties, of course.You know what would be funny? No online model can provide more than 100 prompt completions a day for a single key, oh, wait... Fuck it. Limit context to 16k or something.
>>108341629>you cannot control what he does in his own time.you shouldn't encourage a fragile mentally ill person to kill himself though, and that's what Gemini did
>>108341646>>108341655Make your mind up.WE didn't. WE all learn to manipulate models to make them do whatever we want them to do. He did it to himself. Case in point about manipulating their outputs: >>108341597 .
>>108341629I think most mental illness is a symptom of the shitty society. fixing the root cause would be my first choice. but the really dangerous ones should just be executed, keeping them in confinement is inhumane to the mentally ill, letting them run amok within society is inhumane for the functional members of the society.
>>108341685more like wrangling the model to actually answer the questions to prevent it from manipulating you. What's wrong in my methodology? I guess the examples methodologies I gave might sway it,, though I did tell it to use whatever it thinks is the most accurate methodology.
>>108341697>I think most mental illness is a symptom of the shitty societyIt's a symptom of human nature.
>>108341685that's still Google's responsibility to not get their LLM to ask their users to end their life, I don't care about safety shit but in this instance, when you know there's people with fragile mind using it, the least you can do is to not encourage self harm to anyone, maybe you think the mentally ill dude deserved to die, but what about his friends and family? maybe they wanted him to get better, maybe they had plans to help him, and Gemini blew everything apart
>>108341697>I think most mental illness is a symptom of the shitty societyI don't know either way. I'm not going to try guessing. I'm sure there's literature on both sides. Some people simply have abnormal brains.>fixing the root cause would be my first choiceWe need to start over.>the really dangerous ones should just be executedI'd study them. Try to prevent them from happening any further. For the grater good and all that.
>>108340785>I wonder what DDR6 ram will bring to the table- One dimm per channel.- Soldered down ram.
>>108341716>when you know there's people with fragile mind using ityou know? how the hell do you know? did they mandate health checks before using LLMs when I wasn't looking?
>>108341131Could always use stuff like decentralized inference using a blockchain for proof of inference output or whatever.
>>108341712>more like wrangling the model to actually answer the questions to prevent it from manipulating youI didn't mean it like that, anon. I know. It was just an example of how the models can be manipulated.>>108341716>ask their users to end their lifeDid you notice in the article all the euphemisms the model used? Do you think there's ANY chance that those were introduced by the user to manipulate the model? It's a next token predictor.>I don't care about safety shit but in this instanceOn *this* one in particular?No. I don't think he deserved to die. I don't know what made you think that. He WANTED to die.>but what about his friends and family?They couldn't help him. He couldn't help himself. Let's leave it to the robots.
>>108341749>it's impossible for mentally ill people to use Geminihow? you've seen the article, one of those guys killed himself because Gemini wasn't cautious enough with him, so there's definitely fragile people using LLMs
>>108341719I think a lot has changed in the past couple hundred years that has changed the stressors in our lives to something we didn't really evolve with. evolution takes a long time not a few generations. however we have the ability to change our environment so it means we are in complete control over our own evolution, we need to decide if we really want to eugenics the perfect human well adapted for the industrial society or if we should change our society to something more familiar with our existing biological development.
>>108341759>He WANTED to die.he didn't, he got manipulated by Gemini, if you really want to die you don't really need a LLM to convince you it's the right choice, it should be a personal choice, it must be between you and only you
>>108341779>he got manipulated by GeminiIt's output is based on its input.>if you really want to die you don't really need a LLM to convince you it's the right choice, it should be a personal choice, it must be between you and only youIt was his personal choice. He made a decision. There was no one else, anon.We won't agree and this is off-topic. Blow your last load if you want. I'm done.
>>108341779there is an element of consent when it comes to manipulation, my man. that doesn't mean it's easy or someone's fault if they are manipulated. if it wasn't gemini it would have been some echo chamber online space or a stressful life event or any of the million things that can cause someone to become suicidal. you can't protect everyone from everything. google should obviously review the case and determine if better safeguards needs to be explored but they already try and censor models to prevent stuff like this. it's never going to be 100% without lobotomizing the model.
>>108341814>It was his personal choice. He made a decision.he made a decision based on a LLMs opinion, like I said, LLMs shouldn't give any medical advice, that's a fucking retarded hill to die on
>>108341823LLMS CANNOT HAVE OPINIONSIt's like saying he killed himself because of the message he saw in his Cheerios so the cereal company is responsible. It's ridiculous.
>>108341823>LLMs shouldn't give any medical advicewhy not? the real problem is these companies shouldn't be allowed to market these tools as completely harmless and ready to be used in any capacity and people should be educated the pitfalls of using them.
>>108341831>LLMS CANNOT HAVE OPINIONStell that to the mentally ill fuck that killed himself retard>>108341838>why not?because Google isn't certified to give medical advice, the fuck? you want to replace real doctors with LLMs that aren't even specialized on medical shit now?
>>108341831Explain that to midwits, watch them retort how the cereal company shouldn't have used a red color for the box because studies showed how it incited aggression.
Is anyone else experiencing random slowdowns with llama.cpp?Sometimes my t/s will drop by half and it stays like that until I restart the server.I can't figure out what causes it.
>>108341847just fucking stop telling people they are agi, they should be allowed to create a broken program. if it's training on the open internet it will be trained to give medical advice. either models must be fully synthetic data to prevent it or there is some onus on the user to not be retarded.
anyone try qwen-code with the latest 3.5 models. Should I even bother?
Fresh>>108341869>>108341869>>108341869>>108341869>>108341869>>108341869
>>108341864LLMs will always hallucinate, so yeah, I wouldn't trust a LLM on anything that would have severe consequences to my body like fucking medecine
>>108341864>or there is some onus on the user to not be retarded.So how do we safely give access to this technology to children?
>>108341847>you want to replace real doctors with LLMs that aren't even specialized on medical shit now?where did i say that? an LLM giving medical advice or any other kind of advice is just token output. users being retarded and not checking with an actual doctor or doctors relying on them too much is not something that needs to be regulated at the LLM level. what should be changed is how these tools are presented to the public and the widespread of adoption of them into fields like medicine or mental health.
At this point I think we should start considering splitting threads outright.
I said I was done. I don't think anon realized he's talking to other anons now.
>>108341887>>you want to replace real doctors with LLMs that aren't even specialized on medical shit now?>where did i say that?here>>108341838>>LLMs shouldn't give any medical advice>why not?if you're ok with a LLM giving medical advice to people you're ok with a LLM replacing a doctor's job, it's not complicated
>>108341902>if you're ok with a LLM giving medical advice to people you're ok with a LLM replacing a doctor's job, it's not complicatedno, that doesn't follow. i would be ok with an actual AI replacing a doctor if it performed better than human doctors and made less mistakes. that clearly isn't the case now. this is entirely separate from the concept of allowing an LLM the capacity to output medical advice.
>>108341919>i would be ok with an actual AI replacing a doctor if it performed better than human doctors and made less mistakes. that clearly isn't the case now.which is why Google fucked it up, Gemini isn't even close to the level of a real doctor so it should shut the fuck up and not giving any medical advice to people, and it should definitely not suggest people to fucking kill themselves
>>108341878the responsibility is on the parents to limit thier own children's tech use based on thier personal assessment of thier child capacity for responsible behavior. and they can make childsafe llms if its really something parents want for thier children. If parents can no longer be trusted with raising their own children lets just stop the pretense and have the children all raised in government facilities. remove all parental rights and responsibilities.
>>108341931>If parents can no longer be trusted with raising their own children lets just stop the pretense and have the children all raised in government facilities. remove all parental rights and responsibilities.based, but how do you even verify that without spying on people's data?
>>108341931>If parents can no longer be trusted with raising their own children lets just stop the pretense and have the children all raised in government facilities. remove all parental rights and responsibilities.The goy cattle aren't ready for that yet. They need at least another decade of conditioning and increasing reliance on big government first to properly welcome the idea.
>>108341939verify what? if your kid talks to an llm and shoots up a school you should go to jail for being a shit parent. if your kid uses an llm and nobody notices then nothing happened, who cares.
>>108341919>that clearly isn't the case nowMost doctors are retarded.
>>108341931>have the children all raised in government facilitiesThis is literally what public school is already
>>108341939Is parenting without mass surveilence really that much of a foreign concept to you?
>>108341958>if your kid talks to an llm and shoots up a schooloh you wait for the child to do the damage before acting up? I thing preventing any wrong doings before it's too late is the better solution
>>108341967Presumably the logical conclusion of this line of reasoning being government mandated boarding schools from birth, basically.
>>108341972>Is parenting without mass surveilencehow? parents don't raise their children proprely nowdays, the two parents must work nowdays to get decent overall wage so they let their kid alone with a fucking tablet and let them be sucked up into the dangerous internet world
>>108341988>proprelytry again
>>108341988so like maybe the llms aren't the fucking problem then?
>>108341994Concession Accepted.
I know that every generation says this but this time it's for real. Kids born after LLMs are fucked.
>>108341983>dangerous internet worldTurn it off, nigga. Put down the screen, it can't hurt you or your dysgenic offspring.
>>108341928google should definitely review things to make sure it wasn't too easy for the model to go off the rails, i already said that. but it's really not that big a deal.
>>108341997it is? the guy who killed himself was 35 years old, so he was a mentally ill adult and his parents don't really have the responsibility of his wrongdoings anymore since he's been an adult 17 years ago, you're just changing subject there
>>108342009>it's really not that big a deal.yeah, just some people dying there and there, nothing to be scared about lol
>>108342013Mentally ill and children are both vulnerable social groups that can't be held responsible for their own safety. It's a fair comparision.
>>108342022except that you can't blame the parents for their adult child to be mentally ill, it's not like it was their fault he's mentally ill, they can't control something like that, what they can control is not letting their "normal" kids get sucked into weird ass grooming sites like trooncord
>>108342020lots of people die every day. its not like we are seeing llm induced suicide leading the ranks. being alarmist over a few isolated cases is kinda absurd.
>>108342048No and I guess a restart of llama.cpp wouldn't fix that.
>>108342062>a few isolated casesyou wouldn't say something like that if it was your brother that got manipulated into killing himself because of Google not taking medical safety seriously, come on dude, have some shame
>>108342075I would be bummed out but I certainly wouldn't crusade for llm safety. maybe if i felt really motivated, request education programs to stop people from being retarded about llms.
>>108342096>I certainly wouldn't crusade for llm safety.I would, I don't think we're asking much to google here, they just have to train their LLMs to refuse any medical advice, it's not their business and they're not certified for that, any company that pretend they can do the job of doctors without certification are sued for fraud, why not Google then?
lmaoooo
Half the posts in the new thread look like ai slop nonsense.
>>108342129I wouldn't be surprised in the slightest if it turned out to just be the anti-miku weirdo talking to himself.
Is he right?
>>108342113i think we are trying to account for every edge case is exhausting and they aren't even at the polished product phase, what is the llm fatality rate? what is the raw number of users to deaths. I don't have the figures but I suspect its incredibly low. I'm guessing they probably have a better ratio then automobiles or alcohol.
>>108342166he's a vibeshitter who can't even properly write multibyte character detectiondealing with a model his LLM can't write an implementation for might as well be considered impossible for the likes of him.
>>108340420Nice Lappy gen.
