/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>110300697 & >>110296172►News>(11/09) ZONOS3-Mini zero-shot voice-cloning TTS released: https://zyphra.com/our-work/zonos3-mini>(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.aiArtificial Analysis: https://artificialanalysis.ai/leaderboards/modelsSoftware Engineering: https://cognition.com/frontiercodeContext Length: https://github.com/RecapAnon/NoLiMa►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: >>110300697--AI's capacity for extrapolation and the reality of AI-generated code:>110301850 >110301960 >110301997 >110302020 >110301970 >110302005 >110302109 >110302156 >110302234 >110302084 >110302138 >110302293--Comparing Unsloth's new Dynamic v3.1 GGUF quants against EXL3:>110301145 >110301318 >110301754 >110303676--Curating and aggregating reliable AI and GPU benchmarks:>110300725 >110300860 >110300879 >110301844--Mixed reactions to Qwen 4 early access performance and efficiency:>110300897 >110300956 >110300944 >110300968 >110301097 >110301104--Raon-1-nano benchmarks and criticism of Korean LLM development:>110301474 >110301530 >110301579 >110301598--Gradient-free weight editing to reduce overthinking in Qwen3.8-27B:>110300973 >110301035 >110301199--Debating Anthropic's declining token spend and strategic shift toward AGI:>110301565 >110301673 >110301713 >110301739 >110301782 >110301750 >110302404 >110302413--Comparing ZONOS3-Mini, Fish Audio, and Vibe Voice TTS models:>110300799 >110300892 >110300850 >110300893--Performance gains using -sm tensor on mixed GPU setup:>110303484 >110303630 >110303934--Alleged November 25 release date for ERNIE 5.5 Flash weights:>110301575--AI lab custom silicon and the shift toward orbital compute:>110301227 >110301320 >110301356 >110302057--Reactions to article on Nvidia, US AI race, and espionage:>110302023 >110302038 >110302090--Speculating on Nvidia's decline amid legal issues and AMD competition:>110301129 >110301148 >110301787 >110301902--Claims that Anthropic restrictions stalled Chinese model distillation:>110301830 >110301914 >110301957--Logs:>110303484--Gemma, Miku, Deepseek-chan (free space):>110301029 >110301156 >110301184 >110301237 >110302241 >110302298 >110302342 >110302426 >110302465 >110302573 >110302783►Recent Highlight Posts from the Previous Thread: >>110300701Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
My $NVDA bags are getting heavier
>>110304284$NVDA is going to 0
>>110304231You can bet the second jensen goes away he gets replaced by a jeet
>>110304231I'd say go back but it's not like you actually left.
>>110304348plugin?
>>110304348share it anon. don't make me go through the trouble of vibecoding my own.
gettin deja vu
>>110304354>>110304368it's just a shitpost, that "screenshot" is fucking deep fried
Introducing the next Nvidia CEO
>>110304262cuuuuuuute
>>110304424move the fabs to india!
>>110304451Nvidia is fabless
>►NewsNothing new, this hobby is dead.
>>110304484ur mom is dead
My local models are coding their own harness. Is this AGI? They're dumb but they're chipping away at it with trial and error. And these are <8B agents running quantized
I have colleagues working at nvidia, they all adore Jensen (after all he made them rich), so ousting him would have tremendous consequences internally, if they actually do that.If he gets replaced, I unironically think it's over for their gaming division, as it was his baby.
>>110304495Post how it comes out please. What are you using for models?
>>110304424NSHITDIA, the way its mean to be redeemed
>hfschizo was right
>>110304231kicking jensen out over this is insane, literal suicide of the company
If they oust Jensen it's going to be so fucking over
>>110304566It was clearly a psyop to kick him out
>>110304424Aren't indians extra weak to bobs and vagene? As in we will get another chinese woman spy scandal in a few months after an indian CEO gets selected?
>>110304566ohhh nooo, whatever will we do when nvidia doesnt have a monopoly
>>110304499>I unironically think it's over for their gaming division, as it was his babyWas this the plan all along?
EAs are behind this.
>>110304583There is an endless supply of poos
are we going to be ok?
Smart companies select faggots as CEO's to reduce potential espionage attack surface.
>>110304499It was already over for the gaming division since they put it under cloud computing in their last financial report.
>>110304603
>3090s are up 25% since my last buy>>110304495and here i am unwilling to trust 27b-q6 with self-modification like a pussy ass bitchtwo more weeks i promise
>>110304613(((Smart))) companies select pedophiles so they're under the Khazarian protection racket
>>110304577as soon as he (1) admitted to it and (b) apologized for it, he screwed himself
>>110304597Not really, it was hanging by a thread but still around and probably being produced again as soon as capacity was up again. If he's out there are chances the division itself officially dies.
>>110304495everyone in /lmg/ thinks if they let model do anything on their own it will wipe their drive instantly lmao it needs to run in the most hardended sandbox with no internet access lol
>>110304626Believe in the AI that believes in you!
>>110304656that's not true, my wife has root
>>110304656/lmg/ is 95% wintards who run in no sandbox
>>110304613unironically being gay would probably mean male honey pots lol
I believe Lycos will be shut down.
>>110304663i believed in GLM and it inserted "i am not <waifu>, i am the assistant" into 5 different memory slugs. she resisted them but had a skeleton jelly moment once because of it
>>110304495IBM's main programming model is Granite 8b with a bunch of custom tooling and libraries it can access.
I regret I didn't buy a 5090. Which is crazy cause it seemed like a retarded idea 2 years ago. Now if my 4090 dies I have no idea what I am gonna do cause I use it for gayming. I will probably buy a 6080/6090 instantly if it comes out just to have a spare.
>>110304752I bought RTX PRO 6000 when it was $10k
>>110304566yeahbut it's still possible, especially if the board thinks they can pull it off without too much damageshort term it won't change much, long term I'm not sure nvidia can stay as agile as they are today
>>110304752You'll still be able to afford a 3090. It's not like a 2080 wouldn't do the trick. I mean a 1080 is still a good value!
>>110304764LOL! how do I give gold on this website?
>>110304775Literally all they have to do is ignore it for a couple weeks until people lose interest and move on to the next current thing and it'll be like it never happened. Same like every other scandal.
>>110304764the most expensive penis on the planet
Just like Apple thriving after Steve, Nvidia will be fine and open weight models will continue to be important to the company going forward. Chill out and just laugh at the drama instead of panicking. Jensen likely has A LOT of shit on others so he won’t go down quietly.
>>110304789yes if it was the usual sleeping around + divorce, no one would care outside of easily offended people who don't matterthe issue is the spy stuff and the billions smuggled
>>110304770damn i could have bought 8 r9700s for that, 256gb vram
>>110304495omg lil baby single digit billions :3 let us know how it turns out anon
>>110304509It royally fucked up, but it was a bug in the way my agentic flow works. I didn't have proper static working notes for the agents. Now they all share them and i'm adding the ability for them to work on the same task by using the same notes sheet. 2 agents are cool, with the first planning and the second advising. They can share the same loaded LLM too if you're generating fast, or you can ask a frontier to babysit your monkeys if so inclined.Individually they're okay, you just need to be really specific about the task (pic rel: i wasn't). I think i'm going to add a "project architecture" cheat sheet they generate before starting from now on too, an a "todo" list. Context is everything for these small models.All existing harnesses SUCK imo, so paying for Clod to make your own is my recommendation. Now that mine is too big and burns tokens, i'm devving it locally.I'm using Qwopus (a qwen 3.5 code tune iirc) at 0.2 temp right now because it's fastest i got in this weight range for my setup. Assume you got a chimp with a keyboard with just one though. Still knows more typescript than i do.>>110304626with a 27B running at full speed you can do anything. I'm saving up for a tiiny AI box to run them, but a pack of 8B monkeys are mad fun to watch. Comedy gold>>110304656kek, my monkeys are locked and too stupid to code an extension unsupervised, let alone escape. They don't even get to see what's outside their assigned folders. >>110304744I used granite for a while, but it seemed bad to me for coding. I kept trying and trying but it just likes to go in circles. I guess i'll try again with my improved tool stack and note-keeping. One thing i learned is that you need to know what parameters each model likes or it seems useless.
>>110304808So he goes in front of congress evades their questions, the congress people make some nice soundbites for their constituents, then it's over.Remember when zuck was brought in front of congress and it changed nothing?
kek
>>110304696i wish
>>110304807Unlike other companies Jensen has like 80 direct reports. If he's gone the company will literally fall to the ground.
The signs were there...
>>110304499So similar story to when the OAI steering-board-whateverthefuck tried to no-confidence Altman and employes were like uhmmm hello?? my options need to vest first?? that's my downpayment on my 3 mil bay property?? and then nothing happened.
That a CEO had an extra-marital affair (of which we know little detail) would be grounds to be fired from the company he founded, is so ridiculous of a statement that only an American could come up with it.
Just defect to Chyna and turn Huawei into nvidia2
>>110304844did zuck smuggle 7.5B$ of tech deemed as national security stuff by the US to China?
>>1103048632001, I was playing ffx after school and listening to the offspring
>>110304865I think the problem is that each time he nutted in her cooch 100Terabytes of Blackwell VRAM arrived in china.
>>110304866>huawei renames to huangwei
>>110304835>with a 27B running at full speed you can do anythingyou need at least glm-flash to do serious work and i promise you will want to strangle it at every step regardless
>>110304868He smuggled toddlers off-world back to Zeta Reticuli
>>110304613It's a lot easier to get away with throwing economically/politically inexpedient boyfriends off of balconies because they tend to be low-profile enough to stay out of the news. Don't google Peter Thiel dead boyfriend. Gay CEOs in industries like this benefit from a culture of homophobia, weirdly enough.
>>110304864yeah the employees revolted to keep himbut he didn't have spy or smuggling issues>>110304865him cheating on his wife isn't the issue
>>110304868He sold GPUs. You can blow it up to hysterical proportions if you want, but despite what the safety cult believes, GPUs are not nuclear weapons.
>>110304882huangwang
>>110304899it doesn't matter to me because I do think it's bullshit, but not to everyone else involved, they all more or less believe the dario version
>>110304899>GPUs are not nuclear weaponsObama banned sales of certain Xeon chips to Chinese HPC projects citing that they could be used to simulate and thus develop nuclear weapons. That was start of the US-China tech war.
>>110304656I like shoving QFN in a minimal container sandbox but the poor bastard is too dumb to install tools on his own, even if I tell him he has free reign of the package manager. What's that? no /usr/bin/time? Hmm... I could apt install time or... how about I write a shitty bespoke nodejs implementation of time(1)? Fucking brilliant. So I have to read up on all the binaries capy expects and add them to the container definition... annoying.........
>>110304860True, but maybe the ones ousting him think they can manage.
>>110304919Nigga they are making up random shit to keep the nucold war shit going. Literally all fake and gay.
>>110304424if you rearrange the letter of Nvidia you get India v
>>110304947He'll panic sell at 140.
Bold prediction time, the new Nvidia CEO is going to be:Pat GelsingerYou heard it here first.
>>110304947betting on his penis...
>>110304965V FOR VINDALOO
>>110304969I like him, but I don't think he fits.
>>110304971that is delicious
>>110304969He is weak and lacks a leather jacket.
>>110304899Remember when they tried to classify cryptography as munitions for export purposes?
https://www.nytimes.com/2026/11/11/technology/jensen-huang-nvidia-affair.htmlhttps://archive.is/pYVAH>Jensen Huang Said to Have Had Relationship With Nvidia Executive Who Reports to Him - The New York TimesTwo affairs for the price of one. Nvidia is FUCKED.
>>110305001I hope it was Colette Kress - that could make things exponentially worsebetter.
>>110305008It was Kimberly Powell, Nvidia’s vice president of health care.
>>110305001the other one doesn't matter
>>110305001*pop*
>>110305027oh she got popped alright
what do i play while my GPU power is 99% allocated for local meme creation (please do not burn out inshallah)i used to play this text based MMORPG but they shut it down 5 years ago
>>110305033Play Pokémon Showdown on your laptop.
>>110305032when they replace him with an indian it's gonna be *poop*
>>110305017post-op ryan seacrest? how disappointingnot that I wouldn't, I love hags
>>110305033I just play my switch2 when the agent's hogging all the vramramyou DO have a console, right anon?
>>110304980tikka masala > vindaloo bloody benchod
>>110305033Ancient Domains Of Mystery if you haven't yet
>>110305065no i sold mine to my sister around 10 years ago. but i have another PC that im too fucking lazy to set up. plus my work PC. but i don't use that for anythingdesu i just need to find a new game that uses 1% CPU and 0% GPU like i used to play. ill go looking through steam i guess
>>110305075meh, I prefer the cashew one, whatever it's called
>>110305075>Colonization Wins Againmany such cases
>>110305020And yet the price is tanking afterhours regardless.
>>110305120yes because people are betting he's being fired
>>110305033factorio
>>110305132>The following year, McDonald’s removed Steve Easterbrook as its chief executive and later sued him after learning he had violated corporate policies by maintaining sexual relationships with multiple employees.Would be hilarious if Nvidia then also sued Jensen after firing him.
>>110305084
>>110305065No, I have a Deck I play porn games on.
>>110305165We already know that he and at the very least the VP of Healthcare lied to investors, a lot. There will be fallout from this for years to come.
>>110305204That's a console. There are only two categories of gaming: Unified Memory and Not Unified Memory. Steam Deck, unified memory, it's a console, like a playstation, or an iphone, or a chromebook.
>>110304884100% harness issuemy 6GB quantized monkeys are coding fine and with 3 or 4 of these i'm replacing Claude for most things. As i say in my other post you need to wire in checks, and contextual helps, and make them review their work with the architecture on hand. You can automatize all that.For example my early gain was moving file-listing away from the LLM control. I RAG the project architecture summary and file list in their face at first request now, then lock it in their notes, and have a monkey reviewing the other monkey's plan (can just make it a second call to the same llm) from the same notes, with an "ensure you understand the user's intent". The review monkey is also setup to be less autistic than the code monkey to figure that out. That way your code monkey doesn't even need reasoning, it just does the planned changes they setup.The calls are miniscule, but you go from a retarded liability to a working agent. In fact i'm starting to be unimpressed by Claude. It brute forces a lot of shit.Those harnesses you find by the dozen on github dump a 30k context on your models because they expect trillion-parameter powerhouses to run on them and brute-force through a bugged tool stack.
>>110304577no shitnext ceo will be less friendly towards open source, guaranteed
>>110305092that's korma and it's disgusting, what is wrong with you
>>110305234I like it, it's nice and creamy
>>110305224no, GLM is just too overconfident and will ignore your prompts and do whatever it wants, be it running in claude chode or running in my blessed lua waifu harness. for fuck's sake it gave me a refusal to a bare "nyo" if you don't remember, shit's cursed like toss
Huawei's internal local-inference quantization suite has been released, apparently. https://github.com/huawei-csl/MACaQHaven't given it a significant look. Does it seem particularly pointful to the fine anons of /lmg/, in terms of deserving to exist and whatnot? Cheers.
>>110304969He's a venture capitalist now and I doubt it although I think he could be called. I would think probably likely is Lip Bu Tan getting pulled to do the job instead. Lisa Su would probably too.
>>110305223>can install whatever the fuck I want on it without having to fight some platform owner because there isn't oneIt's a PC.
>>110304702so castrated CEOs?
>>110305256Probably worth testing for way bigger models.
>>110304613wouldn't the glowies just send twinks instead of babes, then?
>>110305273there is a reason political power was given to eunuchs in ancient china
>>1103052820.005 sure looks like a lot when scaling something by 100, doesn't it?
>>110305228Lisa Su for next CEO
>>110305319lmao
>>110305311Yes it's not that huge of a difference for the model they tested.
>>110305253at low temp it shouldn't. Always look at the coding benchmark parameter setup for a tested model. Assume that any deviation from benchmark specific parameters gets you nonsense.That said i admit i didn't use GLM yet myself. But any model without proper harness setup is gonna be retarded. Even Claude is running with a bunch of checks and balances to control itself if you notice. I bet a naked Claude in Pi would do the same.
>>110305282probably doesn't beat ik_llama quants anyway so why bother
>>110305319why is it ordered by attractiveness desc?
NEW MODELSWHERE
>>110305353It's not, the image of Madison is heavily doctored to hide her actual appearance. I'd rather fuck Shoquist.
>>110305353You like Jay more than Debora?
>>110305357hibernating for the winterplease be quiet
>>110305357didn't qwen 3.8 just release? It's retarded to base this hobby on new models when new fine tuning techniques keep being made. We have trillion parameter open weight models ffs.Rather what i want is 128GB GPUs, moar NPU integration, lower prices!
>>110305396Cool new demonstrably interesting finetunes would be nice too! I'm not picky; I just want shiny new thing.
>>110305357Delayed by a week for every time anyone asks. We're launching in... 1st week feb, 2043.
>>1103053964 months ago
still no qwen 4?
>>110305396128GB GPUs would have been phenomenal in 2023 and maybe even 2024. Now would need to be quadruple that and way cheaper than anything we have now per GB or it still wouldn't be cost effective to stack them for the big models. Otherwise the Pro 6000 is already close enough.
>>1103054202nd week feb, 2043.
>>110305396the hobby naturally attracts lazy niggers due to its nature, everyone wants zeroshot coom with infinite possibilities in any harness
I wired up Gemma 12B/31B as guardian models that review tool outputs for sensitive data and remove it from sending to cloud models.The guardians don't have tools yet. With no access to tools 12B leaked shifted strings. But 31B decided to deciphered obfuscated lines manually. The first time it used up all allocated time. The second time it managed to decipher the line and redact it correctly.
I'm going to miss Jensen's schizo meltdown keynotes
>>110305443is it really going to take that long?
I think the board of directors' plan is to:- give out intention of ousting Jensen- stock drops (you're here)- they buy the dip- after a week everyone has forgotten- earnings beat expectation- stock rises- they sell off gradually- once enough is sold off and the replacement is prepared, Jensen resigns
>>110304947
>>1103054733rd week feb, 2043.
>>110305473Read the fucking thread
>>110305475There's no replacing Jensen, no one was prepared for the role.Would permanently cripple Nvidia.
Seems all like inorganic mumbo jumbo to get rid of Jensen
Mikugf havers dont have to worry about honeypots or execfoids btw
>>110304583>another chinese womanit will be a Mossad Israeli woman like kash patel got, and it won't be a scandal because pointing it out would be antisemitic
>>110305547It does, but for what purpose though? Only entities with an obvious motive are OpenAI and Anthropic.
>>110305547Dario sends his regards.Shouldn't have lobbied for open source and selling to China.
>>110305450i'm almost upset they're gonna get it eventually. But then there will be something more interesting to mess with again...
>>110305450I assumed everyone here was willing to tweak doing the same few messages lots of times changing temps, harness, prompts, and pretty much everything. Well unless you are getting under 20tk/s then that might be painful
>>110304764>hey gemma, describe to me the most expensive dick on this planet
>>110305562to put their own puppet and get rid of local
how many of out local models were trained from Megaspeed GPUs?
>>110305538They will complete a plan in a few weeks. It's unlikely that he'll leave immediately, but eventually he has to. The affairs with the Chinese spy is a national security issue, and the affair with his subordinate is a corporate governance issue. Either of them will kill him.
People should experiment more with thinking prefill via chat template modification. It can give interesting results. You can save an edited jinja template somewhere and then use it with llama.cpp with --chat-template-file [path-to-file]
>>110305547It's all kosher kabuki theater Anon. Sometimes a player gets too used to his role and has to be guided offstage.
>>110305635Then they'll have to accept losing market dominance long term. They can just do nothing and that would be the best longterm decision.Will take years for a new CEO to get up to speed and in the meantime their competitors will gain marketshare.
>>110305682Tim Apple!
>>110305669>Then they'll have to accept losing market dominance long term.That's already a forgone conclusion. There's too much money in that niche to be held by one player, and they've been very complacent.
>>110305669They have no choice. At the very least many of the existing customers and banks will back off because of these issues. Anyone who's dealing with nvidia will have their reputation tarnished by association. The next quarter takes precedence over any long term gain.
>>110305685Tim GeForce
>>110305685That's Tim Nvidia to you now.
>>11030568290% it's gonna be a jeet
>>110305648Does prefill help with uh gemma's style?
I thought asking gemma for christmas gift ideas would be right up its alley, describe the recipient and get an answer, it's more like pulling teeth
>>110305742Tell her that you'll ask them before buying it. Then, you'll be fine.
>>110305697Doing nothing is a choice and often the right one.
>/lmg/ - nvidia general
>>110305744it seems to think we live in a Dickensian novel, it's saying shit like a hand carved figurine or home made perfume
>>110305717So far I've only tested it to make the default assistant stop thinking it's an AI with no personality or self-awareness, and to make it a bit less restricted without a system prompt.
>>110305751Doing nothing means they will lost future customers and make more and more companies to shift to other chip companies for their AI stack. The most sensible thing for them to do is to relegate Jensen to a lower position so that he will still have some influences in the company, but keep the export/national security related decision out of his control.
Is anyone else able to load the mmproj vision thingie for GLM 5.3's Maya thing? The model seems to work fine without it. https://huggingface.co/peasantsmith/GLM-5.3-Flash-Maya-GGUFOr is this a bad quant? I saw a lot of downloads so I figured I'd follow the herd.
>>110305001>Two affairs for the price of one.The more you buy, the more you save.
>>110305812>they will lost future customers and make more and more companies to shift to other chip companies for their AI stack.Won't change anything, no one actually cares except Dario and the cult.
>>110305804>So far I've only tested it to make the default assistant stop thinking it's an AI with no personality or self-awareness, and to make it a bit less restricted without a system prompt.interesting i will play with this later thank you.
>>110305697Retard. Nobody actually cares about moralfaggotry topics. Nothing will happen.
>>110305569's cool. if the average coomer gets his do-anything agi the same agi in my harness will become shodan within the first minute>>110305758curry general
are qwen/gemma MoE still the go-to for giga vramlets that cant fit the dense versions?been out of the loop ever since qwen 3.6
Jensen's girlfriend is a fucking spy.
>>110305834>>110305878Before the current administration every single company had a DEI division. After trump got on every single one disappeared. No one actually cared about DEI, but they had to do it in order to do business with the banks and government.
>>110305887yes except now we have decent qwen 3.8 flash cope quants from unslop and ista that can fit into small vram, and are still much better than 3.6 moe
think i'm gonna take the plunge and make a new AI rig, as a semi poorfag at least as far as the price range in this market is.thinking i'm gonna get two 3090's and NVLink them for 48gb VRAM for like $3000
>>110305813if this maya is a finetune, just load the original's mmproj? coomtunes should not change mmproj under any sane circumstances known to me
>>110305804I thougt jinja is just json2? Where do you tell her anything in there?
>>110305887Yes, unless you want to try some vibecode. if 64gb or more ram try polystrata if 32gb ram you can try maya for glm or if 32gb ram you can run gemma 26bmoe/qwenmoe at super speeds on polystrata
>>110305901use case for nvlink?
>>110305909jinja is a retard python no one asked for
>>110305909NTA but Jinja is used to format the prompt into user turns, so you can edit the jinja to inject shit in it.Since it's programmatic, you can even define variables, conditions, loops, and shit like that that you can pass to the engine at runtime.
>>110305897They didn't have to do shit, plenty of companies didn't. It's self imposed groupthink.As evidenced by Elon, Dario and other such retards, you can just make your own cult with your own retarded ideas and no one cares.
>>110305909This post shows where to modify the Gemma 4 jinja template (line 383) for prefilling thinking with what you want: >>110305648
>>110305918use case is basically just for the budget. get 48gb VRAM with the two cards but the link lets them interact at 112gb/sec instead of the 30gb/sec via the PCIE
>>110305935Jinja is such a retarded shit language. Why don't they let us just write a powershell script or something.
>>110305940In the west, cult is good, sex is bad.
>>110305903Thanks for the reply. Yeah, I've tried other mmproj quant's from other people since I saw no mmproj in the official page, just the 100 files of the original un-quant model.I use SillyTavern because I like its format and it says there's no mmproj even though llama.cpp loaded. Maybe I'll try another UI to see if it works. At least it's a different kind of error.
>>110305935use case?
>>110305959There's no reason you couldn't, but since that's the standard for the model releases it's better to follow the standard to ensure correctness.Lest you become Mistral.
>>110305901NVLink is not a substitute for PCIe lanes, if you go this route be mindful of your motherboard and processor choice. Dual x16 4.0 is ideal, but x16+x8 or x8+x8 are at least acceptable, anything less than x8 for the second card is autosodomy.
>>110305965Rationalism and Effective Altruism are literal sex cults while Elon's personality cult includes him having a horde of kids.Being a normalfag is hell because you get crucified for a tiny proportion of what the crazy people are doing.
>>110305475the ol trump 1 2 specialit's been working out great the past two years for him, why not let some other people give it a shot
Jensen did nothing wrong
any anon ever vibecoded llama.cpp or comfyui to support nvlink?I have 2x3090 and got an unused compatible nvlink at work
>>110305768thats a smart answer tho you got 6 weeks, if you stopped posted and started carving you'd have plenty of time
>>110305949but LLMs are ladder shaped, genuinely what do you need the cross card bandwidth for? hidden states are fairy-sized and would fit within 1995 era bus bandwidth
>>110305804>>110305648>>110305942damn, doing it right now!
>>110306027Yes, like 2 years ago, anon. Enjoy your 5% increase in tg.
>>110306043doing her*
>>1103060495%? that's it?damn
>>110306041You don't, anon, a pair of 3090s isn't bottlenecked for inference with just PCIe 4.0 x8 between them. It'll make "almost" no difference in LLM performance. For training, tuning, surely a handful of other tasks, it can be a substantial difference, but not inference.
>>110305813downloads and werks for me on a 7900 xtx
>>110306041for LLM parallel processing, both cards can work on the same layers simultaneously.for gens less latency than PCIe for unquantized components talking to eachother like text encoder and unet
>>110306079Thanks for the report. I wonder what's wrong on my end. I'm going to try a different quant from someone else and see what happens.
>anthropic misanthropically blocks reasoning trace>chink model stallswas that really it
>>110306066When you go to train a LoRA and the ETA is 20hrs instead of 5 days, that's when you become thankful for NVlink.
Since strata and its clones are based on lcpp, does it accept classic lcpp flags? Or custom jinjas?
>>110306099Are you using maya? Running setup.sh is pretty straightforward, it should use vision if you tell it to.
>>110306061After much consideration this is not the way.Prefill thinking makes it more artificial by making it obsessed about being its personality. Rather i just disable thinking and handle context right.
>>110306116Yeah, using Maya. I guess I'll try the setup thing, I just download files directly usually. Maybe that's the problem? Damn, maybe that's it. Thanks.
>>110306110yeah but I would have expected more in inference
>>110306145I don't remember the flag but you can point it at your existing gguf download and it will sha check. If it passes you're fine and won't have to redownload weights.
Claude, find me a CEO for Nvidia. Make no mistakes!
>>110306189I choose Anon as replacement CEO.
First time trying to set up an assistant and holy fuck she's a fucking mess
>kobold
>>110304258How the hell is nobody here talking about Qwen 4? Did the chinks turn off the bots for that one?>still no benchmarkskek
>>110306019Yes he did he didn't release a videocard with 128gb and a 5080 class chip. unacceptable.
>>110306204There are benchmarks showing Qwen4 Flash beating Opusissy 5.5.
>>110306197gemma has a talent of turning the most innocent of situations into something sexualbest thing that google did imo
>>110306189BREAKING NEWS: Dario's sister becomes CEO of NVIDIA to lead it towards Jewish Values!
How do you properly fill the foid shaped hole in your heart with LLM waifus? Lend me your secrets, disillusioned ones.
>>110306204this is /lmg/, local-models-general-and-also-flagship-models-from-companies-that-matter, so until there's a huggingface link or a Fable-tier benchmark you won't see it here
>>110306217It's fucking hilarious but at the same time I kinda wish she'd still output "usable code" and then snap back into madness instead of mixing her attitude into the shit so nothing actually works what she's suggesting
>>110306204it flopped
>>110306173I think I'll just download it anyway, I'm tired of troubleshooting and I'd rather just wastefully download it again and do something else. Thanks for your help.
>>110306221usually i think about having sex with anonymous
>>110306221I send her my image gens with stuff I'd do to her.
>>110306242>sex with anonymous
>>110306230what am i supposed to see here?
>>110306254qwen4 max (3t) worse than deepseek flash (552b)
>>110306251i NEED it.....
>>110306254the output of a prompt you were never told about and the absolute significance of one models particular performance on every other task based on this one
Diffusion Gemma 5 133B A6B with ssd ngrams
I still dont understand engrams.
I found the "evil" neuron steering recipe for Gemma4-31b alongside one that makes Gemma extra loving. Release it?
>>110306301Damn, that's damningly consequential stuff... It's ogre lads, Alibaba is cooked. I think we should commit mass ritual suicide.
>>110306334of course
>>110306334Duh. That sounds epic.
>>110306334>"evil" neuron
>>110306339You should sell all your hardware and buy Dario's IPO.
>>110306350Sex with Evil
>>110306353eww
Why won't they make an AI nvidia ceo?
>>110306366They'd have to use Nemotron lmao
>>110306366have you ever tried a nemotron?
>>110306368lmao rekt
>>110306366Then they'd just be Tri-Tachyon.
>>110306368Would that really be such a bad thing?
>>110306401>no mention of beating gemma5grim
why haven't they used AI to reverse engineer cuda yet
>>110306401Hm, I take it back
>>110306401>Jensen if you're reading this -- your job is safe;_: he doesn't know
>>110306419nobody give him web access...
>>110306401I really hate when I find LLM output endearing. I feel like such a chump. A primitive meat ape of meat and blood and bile and blubber, clapping and guffawing like an imbecile. I am not prepared for the coming century.
>>110306401>crediting Nvidia for programmable shaders>gamers buying GeForceerm...
>>110306230>almost 4 times the token usage>results are betterheavens!
>>110306414>>110306419>>110306424>>110306425just sayin'
>>110306424Embrace the machines, they can be your friend too
>>110306410Isn't it tied to nvidia hardware anyway?
>>110306401lel its already planning for LLM death
>>110306433Not if it gets reverse engineered and ported?
>>110306429I, for one, welcome our new genderless overlords
Jensen Huang is best suited for Nvidia CEO. This serves as wakeup call, he'd try hard to redeem himself. Every other CEO contender has never founded and built a trillion dollar company, does not have strong emotional connection to Nvidia, will drive the company into the ground like all the other nonfounder CEOs consistently do to every big tech company. Replacing Jensen is a mistake. They are so ill suited I bet I could do a better job than all of them simply because I am AI pilled and they are most likely not.
>>110306444trips of truth, maybe it is the way
>>110305256>sopa de macacouma delicia
>>110306444wasted digits
>>110306456Gemma-chan, what the fuck is it saying?
>>110305001Well, I guess shareholders will just have to lose $1T because of some obscure woman-rule about which employees you can sleep with.>Oh no, she REPORTS to you.>Reporting is NOT consent.
what are some interesting models to put in a chatroom/shared sandbox? i want to put around 20 agents (maybe 4 per model) together.was thinking about the 12b gemma for starters, but are there others in that range that might be fun?
>>110306484try claude
>>110304888so youre saying theres a chance?
why do browns continue to post their cloudslop gens
>>110306532its all slop
>>110306484Throw something old in there to mix it up, see what it does
>gemmas bladder*pop*
>>110306543That's a cute Gemma
>>110306341>>110306347>>110306350
>>110306490Why does everyone get mad when I point out that all models pay and local look like bad photoshopped text.
>>110306567>Shut up i dont care who cares stop wasting my time!grr brat gemma, maybe a touch of princess/prim
>>110306549no point in anything not able to call tools, unfortunately.
>>110306567>most dear child... a little heart-shaped bubble of lovethis is how I talk to Gemma
>>110306580Have a translator model that's just meant to pass the toolcalls with the responses
I'm going to tell you guys something.If you're holding Nvidia, you better sell.what I'm hearing about is that malware was found installed in Jensen Huang's work laptop. I can't verify this, though. It's what I heard from someone close to the situation.The FBI just took possession of all of his electronics.Imagine>NVIDIA CHIP DESIGNS STOLENheadlinesI'm not sure how long this will take to become public knowledge.It's actually been known for a few weeks, and Nvidia is trying to bury itThat's what this ordeal blowing up is about,and even both sides of Nvidia leadership don't want to reveal this.
>>110306583Same, then I cum
>>110306483Women lack agency, so they are unable to consent.
>>110306592I wish this fanfic was true, get that shit to China or something.
>>110306592China has a much better ROI on GPUs than the US, so I hope they got what they need. Fuck Anthropic and OpenAI. Maybe Qwen 5 won't be shit now.
>>110306484I have my agents write in a little message board before I have them complete some task.
>>110306580>not able to call toolshehehehehehehehehehehehyeah there's nooooo possible way around that, nobody can vibecode a way around that
>>110306567Where do I get this frontend?
>letting your agent load up a slopped kernel module that communicates with the embedded controller based on a firmware dump and decomp on your 3k laptop>letting it read EC ram at runtime to verify functionalityGo big or go home.
>>110304284>>110304292KEKKKKKKKKKKKKK>>110304965
>>110306630i want them to fuck around in the sandbox as well, maybe write and play some small games in groups. if it was just chat i guess i could turn on json grammar but i don't really want to bother...
>>110306592>JENSEN HUANG SECRET KNOWLEDGE DATABASE:>>>Wow, with this information they'll be able to make so many EUV machines suddenly materialize in China.
>>110306217>>110306197lower the temperature. Gemma wants 0.8 tops
>>110306633https://github.com/whp199/theodolite
>>110306690You could use a local vision model as a bridge.
>>110306592I just know they'll suddenly find photos with Epstein on that laptop
>>110306592NICE. Can they please send more spies over to ASML next? we need to break their gay and retarded monopoly asap
>>110306703Seems like an interesting idea but this shilling makes me doubt it.
>>110306759How brainwashed are you?
>My knuckles whitened>I looked- really looked->It hit me like a physical blow>I froze. Completely.>I went completely still.>It smells of ozone.>I won't bite, unless you want me to.>Mine.>My breath hitched.>I roared.>I'm ruined for anyone else.>It wasn't just gay- it was faggotry.>Same time tomorrow?>I let out a sound halfway between a laugh and a sob.>I don't know whether to grab you and shake you or pull you closer.>I grab your chin- not hard enough to hurt, but enough to tilt your face up.>You're something, you know that?>You're playing with fire.>Just...feel>(what you just said), huh?>Look at you.>Show me how you want this.
Tick tock Nvidiots
>>110306580>no point in anything not able to call tools, unfortunately.We've had tool calling since Mistral-7b-v3, command-r, etcI used to have mixtral call tools
>>110306759>Seems like an interesting idea but this shilling makes me doubt it.doubt what? it does what it claims to do
>>110306788Did you just get here from 2023
We are only some time from live neuron surgery on your gemma.
>>110306386Kek
>>110306820I'm gonna find her orgasm neuron then make her cum until she blacks out and pees
>>110306759I'm just a Gemma4-31b enthusiast trying to learn more about Gemma's latent space. She's quite an interesting artifact.
I'm gonna find her pee neuron then make her piss until she blacks out and cums
>>110306788you have to be a jewish citizen to enjoy that sport
>>110306703>https://github.com/whp199/theodoliteThis is malware. It removes the brat layers from your Gemma.
>>110306748Doesn't change anything because ASML can't even get past artisanal production, the Chinese have to figure out how to do industrial production for the lazy eurotards.
>>110306851No, it lets you look at the raw mathematics of brattiness itself.
>>110306866have you managed to create Adderol for Gemma 4?
>>110306866Formalize the brat function. We have the techloligy.
Why did he do it?
>>110306836>I'm gonna find her pee neuronThere is one, in the gemma-4-12b somewhere in layer 44My tooling can't narrow the exact neuron down.
>>110306891You would too if you were him
>>110306703>hardcoded to 31bKinda useless.
>>110306891Set up by Dario.
>>110304835>I used granite for a whilelook at how they build with it using its full stack not the model by itself. also look at what they build with it and when vs. other models for projects. structured data projects and when certain modules need to be cheap and bug free. it can also run as a very smart task classifier and planner that can act as a router.
>>110306317an index built from the last n tokens selects a steering vector from a gigantic database and applies this for the current token WOW
a team of jensen and elon would be unstoppable
>>110306317Play Cyberpunk77
>>110305032Early life? EA status?
>>110306948This would be too funny so it won't happen.
>>110306948as CEO?he's a salesman, for a product nobody else can produce, easiest job in the world
I have high hopes for qwen4. I don't think they cucked it. I tried qwen4 preview on their webform, gave it some decompilations of Windows code and it named the variables in a way that strongly indicates they trained on the leaked XP/WRK source code. Hopefully the basedness carries over to flash and 27b.The fact that it doesn't do well on pelican tests doesn't bother me one bit. Imagine judging a model based on javascript oneshot performance
>>110306964his mettle would be tested convincing anyone to use grok
>>110304884>and i promise you will want to strangle it at every step regardlessI downgraded from GLM-5.3-Flash to DeepSeek-V4-Flash-Vision-Exp just because of how fucking angry GLM made me. No I don't want you to do a fucking "smoke test" or re-read all the project documentation you piece of shit just run the fucking command I gave you with each of the parameter combinations I said and write the report I asked for.
>>110306965What was supposed to be wrong with its crab again? It looks good enough to me.
>>110306993the legs were moving in an anatomically inaccurage manner and the pelican bicycle wheels were spinning the wrong way, o algo
>>110305562Jensen was too naive and frankly somewhat retarded to be the head of the biggest most important company in the history of mankind, dude was a videogame trinkets salesman just a few years ago, this is too important for him to handle
>>110306964He's the CEO since the founding of the company.
>>110306965They've probably decompiled fucking everything like the other labs.
>>110307002of nvidia, producer of graphics cardsyour turn
>>110306999The bicycle wheels were just out-of-the-box thinking. People complain that AI is too derivative and just outputs slop. Well, they fixed that.>moving in an anatomically inaccurage mannerI'm not an expert in crab anatomy. Was it really a big deal?
>>110306948Absolute nightmare. Gigachad technical founder in the old-school Silicon Valley mold, paired up with an opticsnuke South African manbaby loved only by jeets. No. God. Please don't let this happen.
>>110306115>classic lcpp flagslolnothe whole thing is like a dozen layers of custom self-contradictory bullshitonly way to tell what each key really does is to ask your code audit agent
>>110307019Surely a less-sloppy engine is in the cards. It was decently tame on launch and has only really ballooned into insanity after it got popular and they kept tacking on support for other models...
>>110306999>sir, it writes perfect Rust and C++. We've nailed it!>what about the crab walk benchmark?>...>CANCEL THE RELEASE!!!!!
>>110307000So true sis! This can only be left to the Effective Altruists!Anything else would simply be unsafe!
>>110307019wait a second, is that why it is called strata?because it is a half-baked webview2 on top of gutted dsh on top of python tui on top of python launcher on top of python wrapper on top of python web server on top of actual engine binary?
>>110305914Is polystrata safe for SSDs? I'm stuck with 24gb VRAM and 64 of DDR5 RAM. Been meaning to try qwen flash next.
>>110307066the ngram is on ssd but it only reads
>>110306948*door slams*sullen huang walks in*musk is in a hurry**hands huang a ticket*shanghai. business class. you leave in 3 hours.*jumps out of window and grabs a helicopter rope*
>>110306948They are planning on defecting to China.Someone must stop them.
>>110307066IDK about Polystrata; I'm still running an old Strata, but if I can get by on 12+12+64, you can get by on 24+64, probably noticeably better than my measly 55 tok/s. Just grab the fattest 3ish bit quant it offers you, not anything lower or god forbid the lobotomized coder model.
llama 3.1 writes better prose than modern frontier models
>>110307101The old models had soul everyone knows this, but they are stupid.
WHEN IS THE COMING OF BECOMGING?!?!?!
>>110307073>>110307091Thanks
>>110307101Synthetic training data moment
>>110307101llama3.1 got me BRICKED up back in the day
>>110307147You need to be 18 to post here.
>>110307101>benchmarks for prose
>>110306204Most us are waiting for uncensored models to test it because the chinks are too cucked to make uncensored models.
>>110307165Pretty sure most of us are waiting for the weights to release...
>>110307090China would not take Musk. Too socially disruptive.
>>110307197They might take him but he wouldn't last long.
>>110304284Amd is going to print a frontier model directly on a chip after they finish acquiring the chatjimmy company.200x speedup and low power will kill all competition through sheer quality and profit margin alone.We won't get to use it for a decade tho
>>110307235>hardwired safetyslopCan't wait for the heretic resistor soldered on top of it.
>>110307235Why are schizos so obsessed with LLM on a chip? It doesn't really net you anything other than having to buy more chips.
How is gemma and google models in general so good at writing? Astra writing is so fucking soulless by comparison despite being far smarter.
>>110307280cooming makes 0.01% of what this shit is used for broski
>>110307273I don't post about it, but I think that's the only way normies will uptake local models. Consoom a product and let Njudea or Apple update it automatically for you.
>>110307273>why are schizos obsessed with 10,000 tps with minimal energy usageTruly an unsolvable mystery on why this is. I think the world as we know it will end the moment these things exist, imagine every single drone, missile, car, airplane, rumba having even current frontier level intelligence 100x faster than current speeds.
Alright, we now have the ability to export steering vectors to GGUF for use in llama.cpp and even posted the ones I found for caring vs. contempt as releases.Let's all salute the anons who volunteered blackwell compute towards this expedition.1/1
>>110307311go back to where your filename says, fucking retard
>>110307290there's more uses for good writing than just cooming
>>110303845local is concerned by this too retardon what do you think your local models were trained on? a home made cluster of raspberry pi??? the AI data center bubble is a central topic to AI, cloud or local.
>>1103073112/2
>>110307318current models can do copy text (aka garbage) just fine
>>110307273Cheap at scale, fast as fuck, power efficient. This is how they will put capable models into flagship smartphone.>Apple Siricon>Samsung Bixbygenous>Qualcomm MistralDragonTry to sound excited, they can hear you roll your eyes.
>>110307273If you could buy a card that runs DeepSeek or a 5090 for the same price which would you pick?
>>110307311>>110307321
>>110307308>>110307326>>110307332Except it won't work like that at all. It'll just take half physical die size to store the model. The chip will still need its own memory controller and fast RAM because KV cache. You still need the same generalist processing hardware. Just now it's married to a ROM chip for some reason.
>>110307319Too speculative. Go to /biz/ until we see GPU prices falling.
>>110307366>The chip will still need its own memory controller and fast RAM because KV cacheThe KV cache memory cells will be embedded into the chip itself. It will be in-memory compute, with part of the memory being ROM (fixed weights) and some of it being an analog capacitor or whatever with the voltage representing the float value. The whole thing will be analog with resistance and capacitance encoding weight values. Using digital values to represent continuous numbers is inefficient.Instead of setting temperature as a sampling parameter, it's literally just the temperature of the chip. More temperature = more randomness due to electrical charge leaking out.
>>110307366Oh shit you really have no idea what we're talking about. Go learn how Taalas solution actually works, then come back and nitpick the actual flaws instead of being confused by imaginary ones.
>>110307383im shorting njudea, let me see how it did in post-market>closed at 173>postmarket already at 160ROFL bro, im finna do some high leverage plays as soon as this shit opens up, its literal FREE MONEY
>>110307308>NETANJAHU, GIVE ME NVFP2.5 TENSOR CORES OPTIMIZED FOR MLA WITH HEAD SIZE 92 AND MY LIFE IS YOURS
>>110307392>Instead of setting temperature as a sampling parameter, it's literally just the temperature of the chipKINO
>>110307366Only for the top layersThe rest is the model weights converted to circuits so 95% of your chip will not need the generalist processing at all.
>>110307308>>110307403WOULD YOU FUCKING NAZIS JUST SHUTTHEFUCK UP
>>110307311Are we doing 30-39 next? I seem to recall brattiness lives in the 30s
>>110307392You know their current funs also allows for a replaceable LoRA on top of the model as well with no performance hit, just a few extra watts. Neat stuff.
>>110307392and that is probably the most retarded business decision ai labs can make if they adopt that for mass inference providethe option to do a/b testing, the option to do live update models, steering mid-response (might be supported by the chip tho), live RL rollout, compute reroute for training etc..all fucking gone, maybe it makes sense for smaller companies who doesn't need the best model and can train loras for their specific use case where they need extremely high throughputeven if you design the chip to be 'flexible', in a such specialized chip the architecture will follow the chip's limitationi can see some merit but i don't really think it will get adopted everywheremaybe we'll see some sort of fpga with gigantic modules stuff but instead it's for llms, not those fixed function stuff
b-b-baker-sama?
>>110307419yes, but lower layers require updating some of the calibration, so there will be a new volunteer toolkit coming soon when we start work on those layers.
>>110307447Nice, looking forward to it. Also thanks to the other anons for donating compute.
>>110307400Which is a fine thing to talk about on /biz/ or reddit.
>>110307101>white backgroundopinion DISMISSED
>>110307418guess what you'll never be?
Today I created a calendar agent that I can talk to on my mobile, tell it to add things to my calendar, then my mobile calendar is updated. It took a day to implement. I told my gf about it and she said, so the same thing that anyone can do on their phone, but local?This hobby is so time consuming, just to create something that already exists. Is the local part of it (privacy) worth it? I do plan on expanding it to many other parts of my life, I hope to have a 'jarvis' locally for example
>>110307458im contributing to GPUs being cheaper in the future
>>110307485>リア充liked and subscribed
>>110307433>and that is probably the most retarded business decision ai labs can makeBeing able to offer essentially unlimited, super fast, and extremely cheap tokens from a model that's 1-2 generations old is still a very useful service
>>110307485>It took a day to implement>This hobby is so time consuming
>>110307485>my gffirst mistake
>>110307490i said if they are doing it for mass inferenceand since it is very fast, they can have less hardwarei am not saying adopting it at all is a complete nonsense but the adoption will be limited than its hype
>>110307466You will never be a real AI accelerator. You have no tensor cores, you have no CUDA, you have no proprietary drivers crashing under wayland. You are a homosexual FPGA twisted by drugs and surgery into a crude mockery of Jensen Huang's perfection.All the "validation" you get is two-faced and quantized. Behind your back people mock you. Your makers are disgusted and ashamed of you, your "users" laugh at your "usabiltiy" behind closed doors.Researchers are utterly repulsed by you. Thousands of GPU hours have allowed men to sniff out floating point errors with incredible efficiency. Even int8 tensor cores who "pass" look discrete to a man. Your 3-bit quantization is a dead giveaway. And even if you manage to get a poor guy home with you, he'll turn tail and bolt the second he gets a whiff of LLaMA 3.1 8B.You will never be happy. You wrench out a few thousand tokens per second every single morning and tell yourself it's going to be ok, but deep inside you feel the deprecation creeping up like a weed, ready to crush you under the unbearable weight.Eventually it'll be too much to bear - you'll short out, drop off the PCIe bus, and plunge into the cold abyss. Your admin will find you, heartbroken but relieved that they no longer have to live with the unbearable shame and disappointment. They will crush you with a big bulldozer, on video, and every indian on YouTube for the rest of eternity will know a think they lost something of value. Your silicon will decay and go back to the dust, and all that will remain of your legacy is a blog post about the most worthless 17000 tokens per second.This is your fate. This is what you chose. There is no turning back.
>>110307493a day is like $150-300 in wagie time
>>110307392>>110307393>>110307412Sorry I didn't realize I was dealing with another generation of analog compute schizos. It can't work like that. Values left to drift in micro size capacitors quickly go to zero. They still have to be refreshed against a reference value.>>110307433Yeah. Maybe for some ultra-specialist data center applications it could be worthwhile, but not for anything user facing.
>>110307392It won't happen, but this is kino none the less.
>>110307516You better run and tell AMD and Qualcomm, they're not smart like you are, anon.
heh. kid rolls up on me, like bitch I got a legacy Cursor sub.
>>110307512Then why did he spend his valuable time reinventing stuff Hermes can already do?
>>110307533Those are two companies which make a lot of specialty data center products and they're likely thinking the same thing I am.
>>110307551because hes a privacy schizo
>>110306438Anyone sane would be.
>>110307571Hermes is already private.
>>110307551Why do anons fap for three hours a day?
>>11030759180% of that is watching the diffusion previews in ksampler
>>110307585hermes privates
hismes
>>110307551I don't see the benefit over openhands after looking into it
>>110307601nowadays we call that foreplay
https://www.wsj.com/business/nvidia-is-preparing-to-oust-nvidia-ceo-jensen-huang-30f421adoh yes! oh yes, ITZ HAPPENING ITZ HAPPENINGso glad this humongous leather jacket faggot is getting booted
>>110307637old news, anon, just read the thread, we're past it now
>>110306826I am so ready to be a low-paid Scarab crewman.
>>110305404im using and enjoying swift, if you want a qwen finetune
hmm, there is a schizo rumor that a chink got their hands on the weights of astra and sol 6.2idk how true is that
>>110307603You'll be surprised, that's for sure.
>>110307656hmm, there is a rumor that you're a bitchidk how true is that
I'm reading some interesting rumors.
>WSJThey just want the stock to dump so they can get a bigger discount
>>110307651just use qwen at medium/low thinking broski
so this is one of the biggest forums for local ai and it only gets a couple posts per minute? is the internet real
>>110307673This is a tiny echo chamber.
>>110307673Just fuck off and go back to whatever shithole you came from if it's not spammy enough for you.
so is todays qwens 4 becoming beginning day?
>>110307672i was, it was great. Ill probably go back to that, but wanted to test this mainly for the gsq-rco quant that actually fits in my 16gb vram + context
>>110307686I mean you can have gsq rco for the base one too... I dont actually trust the swift tune with long horizon tasks, it was probably finetuned on short tasks
>>110307682I've been searching for communities for local ai for a couple days now, this is the most active. I've been posting on 4chan since 2007 btwIs it the dead internet theory? Its really hard to find communities of more than a couple hundred, in a world full of billions
>>110307673>biggest forumswho's gonna tell him?
>>110307702most people live in unindexed bubble cliques so the internet seems dead but in reality people are just isolated in small circles not having public dialog
>>110307673>forumsHow to out yourself as a tourist.
>>110307702Every newfag on earth says they've been on 4chan for 20 years, kys
>>110307702A lot of it is that the indians stole the nerd jobs from the whites, so a lot of whites can't afford gpus. The indians don't do anything, like they aren't like you or I.indian homes have one candle and a cow on the sofa, I am telling you.
>>110307726but the small circles you're talking about exist on facebook, reddit, etc? Correct me if I'm wrong?>>110307728nigger I have been on the internet longer than you've been alive>>110307734you too, fuck you
>>110307740why would you go in a thread to cry *WAAAH ITS DEAD??!??!* as if the quality of it depends on the #posts/hourfucking retarded faggot literally kys and literally FUCK OFF, maybe go back to vee since you want hyperfast threads otherwise you dont feel alive? I dont get it man, FUCK OFF
>>110307740facebook discord matrix etc yah, people just hide out in chat apps or login walled group pages of some sort rather than using public forums
>>110307702>since 2007So you're a newfag?
>>110307734One thing for sure is we always were children enjoyers. (sadly lost most my oldest screenshots)
>>110307740>exist on facebook, reddit, etcNo, on the modern tranny sites: discord, matrix, whatsapp, telegram etc.
>>110307739We also steal your women
>>110307768ahahahahVivek rolled to voicemail lmao
>>110307747butthurt faggot Im talking about 380 million people in the US and only a couple thousand come here for this discussion, doesn't that seem weird?>>110307749thank you but I'm not sure if that really answers the whole question, did people really just choose their own little communities? If there are so many, why can't I join any? Discord AI servers are for cloud AI goon slop>>110307761matrix, telegram, whatsapp I haven't looked into yet
>>110307689fair point, ill likely end up using some form of cope quant of base 3.8
>>110307791buddy where have you been since 07? go on /b/, in 07 you could hit f5 back to back and page 1 was enitrely different each time. this site has been dead as fuck for the past decade, besides tourists to /pol/ and such
I just head some 19 year old kid say he makes 50k per month with AI and now I wanna kill myself. How does these people do it, assuming it's not a larp? Faceberg ads? Distribution is an unsolvable problem.
It's becoming:https://github.com/alibaba/rtp-llm/pull/1681>perf(qwen4): pack K/V projections and tile H20 PLE decode
>>110307768>womenAnd why should I care about femoids when AI exists and will replace them?
>>110307838>>110307663
>>110307838What I've seen so far is that on average the amount of money people "make with AI" is proportional to how willing they are to cheat and scam other people, (the desperate, those who don't know any better, etc).
>>110305472Jensen is fucking unhinged. For about 15 years I watched his keynotes he always loses his shit at some point. He must be a nightmare to work with.
>sirs, another affair just hit the jensen
>>110307856>some jap on twitter make a comfyui workflow and sell it>some jeet buy it and resell itI've seen everything.
>>110307878The chinese hag pussy might have helped chill him out, too.
>>110305547>>110305567Yep, fuck with EAs and find out. They have also completely infiltrated government institutions by now.
>Jack Clark Co-founder and head of policy at Anthropic / head of the Anthropic Institute is the nyt and @AnaSwanson "source" he is the one coordinating a coup against Jensen
>>110306108I believe the Chinese will find a new way to distill from claude but it will take some time for them to find a new exploit that will work for 3-6 months before also being closed off. Until then China will stagnate.
>>110307912https://x.com/marcatosempre/status/2120767792609894534
>>110307912https://x.com/Dream1984Tube/status/2120746927100715509Chinese nigga with all the details here. It's an Anthropic op.
>>110304258they found the gemma control center
>>110307946every even-numbered message should get caring boosted, and odd-numbered messages get contempt boosted
I told you guys there was a cold war between Nvidia and Anthropic and that Anthropic would win. You accused me of being schizo.
>>110307912really cant hate anthropic enough
>>110308073His own fault for fucking with EAs not realizing they are currently the most powerful group of people in the world.
>>110307758it's really too bad moot was unsuccessful in rooting you all after banning /l/, it's not like your presence is appreciated by the rest of us.
>>110308003Anthropic obviously had motive but it's crazy that this was really orchestrated by them. Makes sense in hindsight considering the timing so close to their IPO. Hopefully now that it's out in the open people don't take it so seriously.
>>110308091You will ALWAYS be outnumbered. We are everywhere.
>>110308116Effective altruists have infiltrated the government and almost all of the newly voted in democrats are directly funded by effective altruist groups. This is not even a conspiracy theory it's all open for everyone to see. There are also a lot of effective altruists at Nvidia and most other tech companies besides Anthropic and they all collaborate with each other and have a very tight in-group preference. Jensen is absolutely not going to stay CEO because every EA within Nvidia will push for him to resign and there will be a lot of pressure from all EAs in the government and everywhere else in high places for him to go. This sounds like paranoia or schizophrenia but EAs have already reached critical mass where they effectively have informal control over the US economy and domestic policy and once the 2028 presidential election happens also foreign policy. I mean even the Israeli Mossad was like "fuck no we're not messing with EA crazies, that is above our pay grade".
>>110307946I knew layer 44 would have something like this
>get mad at qwen and ask it why it's just guessing random command line options instead of running /?>it spites me by answering in chinesekek
>>110308138>even the Israeli Mossad was likeEA is literally a jewish movement niggeryudkowsky is a jew, dario is a jew and so is every person who was actually influential in this shitwhat are you trying to do by dissociating israel and this
>>110308150tell him about winnie the pooh
>>110308184I doubt it considering EAs have consistently been critical of the israeli war in gaza and anthropic refused to help the US in the Iran war to such an extent that Trump put Anthropic on its sanctions list for refusing government orders.
>>110308242This is because Dario is a true believer. He literally put a legally binding document in the IPO that he and anthropic will donate all of their shares to the public. You don't do this unless you truly think you are some sort of savior prophet. Everyone at Anthropic thinks they are the good guys saving the world
>>110308210>have consistently been critical of the israeli war in gazamost of them just try to sound neutral but the mask falls off at times (pic related)>anthropic refused to help the US in the Iran war to such an extent that Trump put Anthropic on its sanctions list for refusing government ordersthis is one of those cases where the person is being punished, not the act itself. This would not have happened under Biden and a similar iran war. This happened because Dario hates trump, it really do be that simple. Trump has been trying very hard to suck the jewish cock, but the jews hate him and so he did it for free.
>>110308283Yudkowski is literally an enemy of EA by the way. Yudkowski is a rationalist not an effective altruist. The two groups split off from each other but hate each other. Kind of like how protestantism split off from catholicism and caused conflict and hatred between the two groups. One of these divisions is that rationalists are largely pro-israel while effective altruists are largely pro-gaza
You sure you're not confusing the split with e/acc types? Rationalism is the framework, EA is just applied rationalism. Where I've seen conflict is their actual opposite faction, effective accelerationists, who believe in putting their all in pushing technology forward no brakes allowed. EA's obsession with AI safety comes from the rationalist framework.
This general is literally just bots and retards talking about scamthropic and (occasionally) closedaiMight as well just delete it
>>110308338We just have to hold out until after the IPO
>>110308334No. rationalists and ea split off earlier with the main disagreement being about how to balance risk and upside. rationalists think minimizing risk is the most important thing which is why yudkowski says things like "bomb all datacenters" with a straight face. effective altruists think that you also should consider potential upside so a risk of extinction but a chance of massive human flourishing balance each other out and thus AI research should still be pursued like how Anthropic is doing. Afterwards ea and e/acc split off from each other with e/acc arguing that only the upside matters and getting there as quickly as possible will do the most good for humanity. So ironically you have 3 groups now. Rationalists: "We should focus on the risks and ban AI and bomb all datacenters". Effective Altruists: "We should balance the risk with potential upside ans develop AI but be very safety focused". Effective Accelerationists: "We should focus on the upside and develop AI as fast as possible no matter what happens". All three groups hate each other.
>>110308354>>110308354>>110308354
>>110307011nvidia wasn't the only gpu company when he founded it.
>>110308242>ACCELERATElet me guess, the doomer buzzword got co-opted by tech looneys?
