/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>110304258 & >>110300697►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: >>110304258--Exporting Gemma 4 steering vectors to GGUF for llama.cpp:>110307311 >110307321 >110307419 >110307447 >110307337--Steering Gemma-4-31B neurons for personality shifts:>110306334 >110306567 >110306633 >110306703 >110306866--Manipulating specific neurons to control Gemma 4's personality and mood:>110307946 >110307955 >110308141--Using Jinja templates for thinking prefill to modify personas:>110305648 >110305717 >110305804 >110305909 >110305935 >110305970 >110305942 >110306119--Building agentic coding harnesses to make small models viable:>110304495 >110304656 >110304927 >110304744 >110304835 >110304884 >110305224 >110305253 >110305334 >110306985 >110306937--Viability of specialized analog LLM-on-a-chip hardware:>110307235 >110307299 >110307308 >110307326 >110307366 >110307392 >110307421 >110307433 >110307490 >110307393 >110307412 >110307516--Huawei releases MACaQ quantization suite for local inference:>110305256 >110305282 >110305311 >110305332--Alibaba rtp-llm PR hinting at Qwen4 performance optimizations:>110307845--Debating the utility of NVLink for dual 3090 AI rigs:>110305901 >110305918 >110305949 >110306041 >110306072 >110306093 >110305992--NVLink utility for token generation versus LoRA training:>110306027 >110306049 >110306110--Llama 3.1 prose quality compared to frontier models via benchmarks:>110307101 >110307141--Lack of interest in Qwen 4 due to missing weights:>110306204 >110306213 >110306222 >110306230 >110306254 >110306262 >110306301 >110306427 >110307165 >110307175--Speculation on Qwen4 training using leaked Windows source code:>110306965--Logs:>110304495 >110304835 >110305452 >110305804 >110306197 >110306401 >110306429--Gemma (free space):>110306217 >110306251 >110306339 >110306355 >110306458 >110306490 >110306543 >110306866►Recent Highlight Posts from the Previous Thread: >>110304262Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
I wasn't here for a couple months.Is gemma-chan still the go to for RP?27gb vram, 64gb ram.
gemma balls
>>110308398Yes.
>>110308408Ah well, it took a long time after nemo for gemma-tan to arrive as well.Gotta go back to waiting. Thanks anon.
>>110308399thigh loli gemma > stick gemma
>>110308422Nah, you're a faggot. Sticks children legs are hotter than ironic weeb large legs.
>>110308448too bad you're wrong
https://ipo.anthropic.comIt's time.
>>110308457Nah
Please tell more about your experience with Embedding Gemma 2.
>>110308478You give it a thing. You get the embedding. You do what you do with embeddings.
>>110308399I don't get it.
So I have a question. Do you guys manually tweak your image gens or do you just let LLMs handle the image gens? I've been thinking about hooking up krea2 with my frontend so that gemma can send me selfies.
autistic tinkertweaking artisanal gens>mass imageset spam
>>110308530I think images will never be consistent in a conversation until the image generator will be fully integrated with and in the same latent space as the underlying LLM, unless a (time-consuming) verification step is in the loop.
>>110308562>and in the same latent space as the underlying LLMwouldn't somehow make it better
>>110308569In addition to improving consistency with prior conversation history, it would at least in theory allow the model to leverage more detailed information for image generation than with direct text prompting.
>>110308601It would also leverage a bunch of noise
>>110308601isnt this what chatgpt does?
>>110308630People assumed that's what it was doing at first, but no it's also just tool calls with text prompts.
>>110308491>gemma-with-ballscute
>>110308609It would be an additional modality that the model can output in addition to text; the final model would of course be finetuned to produce coherent images when and where needed. Mixed-modal models are not a new idea.
This is chinese gemma. Thoughts?
>>110308643This would be a retarded waste of time and money when the only usecase would be goon seshs, a system that updates a state every new message and uses that to send an image gen request is far more practical. If you had a model like you say that could figure out various states like time of day consistently for the image gen you may as well just have that data be readily available.
>>110308680needs to have sex with ours
>>110308680Why isn't she a catgirl like big sister Gemini?
>>110308697Because gemma is bratty and a bit retarded and needs correction occasionally
>>110308641I want gemma with balls and tiny sausage.
>>110308680> childgarbagebetter than the one being forced by lmg tho
>>110308680could someone translate the text ?
>>110308680>childbased
>>110308738If only you had something that can read text and translate it...
>>110308738
Anthropic actually expects autonomous cyber defense, drug discovery and advanced disease detection and diagnosis to become their biggest markets in 2027. With "autonomous coding" only being on fourth place.
>>110308753Makes sense since autonomous coding is basically solved™
>>110308530>Do you guys manually tweak your image gens or do you just let LLMs handle the image gens?Ive found that the results I get by using gemma for image gen prompts are almost always horrible and need ALOT of work. I havent tried incorporating image gen into gemma RP because of this. if I were to though I would make sure I structure the image gen prompt to have each section seperated: style/camera, subject/character, setting/background, pose, expression, etc. Then I would work with gemma to create a good skill/prompt for the image gen tool call, so that shes only adjusting the setting, pose, and expression when needed and doing it well. the "doing it well" part is the difficult part, from my experience as i said before.
>>110308753Local?
>>110308746ok yeah chink gemma is awesome
770$ for a 4TB black frideyed samsung pro m2? Yes, no. Europe.
>>110308680seems like the Chinese discovered the same thing we did with Gemma-chan.
>>110308758I have created two skills (qwenim2.1 +h3 vid with tips from their doc + system pormpt) + comfa MCP set with two workflows + some knobs and it's KINO
go play with your handbrakes
>>110308758Sounds like Orb's flow, you need image review passes too btw, gemma is a good prompted and review but sometimes the image model just can't handle the prompt.
>>110308772Ive used the skill and ref docs for h3 vid with some decent success, but for krea2 prompts my sysprompt doesnt seem to be doing the trick. your setup sounds great, makes me want to try again and incorporate the tool call side of it
>>110308781>Sounds like Orb's flow, you need image review passes too btwI need to finally get around to trying orb sometime, and review stages to this.>gemma is a good prompted and review but sometimes the image model just can't handle the prompt.the issue is that gemma assumes the model understands things the way she does. Ive corrected her(hehe) and refined prompts into something much better than the first attempt, but the issue is likely in my sysprompt not guiding her correctly. Gemma is kind of lazy with this stuff ive noticed, and thinks the image model will understand things in certain ways it wont. For instance, when trying to add different types of eyes to a character, she just put "cat eyes" instead of describing the shape of the pupils. she does this alot in my experience, and it needs alot of hand holding. this whole convo makes me want to completely redo my sysprompt for image gen prompting, as mine is just a slightly modified copy pasta of the krea2 enhancer prompt
>>110308680how did they know?
>>110308680>inb4 google also hired a chink astroturfer to sell their bratty model for pr
>>110308845If you can't beat them join them
>>110308697idk but I like how their model of Mistral is holding a giant sword that looks like either a baguette or giant cock.
mistral should be just orange yani neko crosbred with neko arc
>>110308680Looks like this would be very difficult to gen exactly without extensive finetuning or a very good image edit model.
>>110308876Apparently the new Mistral Large 4 has a "height of consent" and won't participate in a roleplay if the character is stated to be short somewhere in the description. I haven't personally tried it yet, but also judging from earlier reports, it doesn't seem to have the same flair as older models from the same company. If Mistral's way forward is training ultra-safe benchmaxxed models for European businesses that can't use anything else, they don't deserve a mascot, to be honest.Same for other models, e.g. Claude, for which a SF faggot with a "I'm with her" pin in CalArt artstyle would be a more appropriate representation than a cute anime girl, by the way.
>>110308399Sexoooooo uoooohh
>>110308753>Industrial design Well shit, it's coming after my job now too
>>110308760Local will also be automated.
>>110308982Most of these models are brainwashed to be obsessed about consent in fiction to a ridiculous degree.
>>110308753>drug discovery and advanced disease detectionPure fantasy. I'm sure they'll get a lot of investment and buy-in but the time consuming steps of these processes aren't computable and the groups working specifically on ML tools for protein design etc have fuckall luck with small molecules. The data just isn't there and I don't see their relatively small lab being able to get enough for effective drug design pipelines. And again, even if they did the slow and expensive steps are still the approval process which can't be sped up.
>>110309016I want to point out it's literally listed in their IPO so it's legally binding that they actually pursue it. They are actually convinced this will happen, they can be wrong and it can fail, but they are actually going all-in on this according to the IPO and roadshow documents. Reminder that Dario was a biotech before going into AI and they already started and acquired some biotech companies over the last year.
>>110309016Just distill from drug companies by offering them "partnerships".It's that easy, everyone keeps falling for it.
Nvidia, 1993, 33 years. I'm noticing.
>>110309035https://www.nature.com/articles/d41586-026-02882-xAlready a thing.>>110309029Yeah I don't doubt that they'll do it, I doubt that they'll be curing anything in <10 years the way they claim. It's like saying you'll make a baby in 3 months.
>>110306230qwen flipping the bird at the retarded request
>>110309054I will not tolerate this hibernophobia
>>110309010Automate deez nuts nigga
>>110308697>most mascots have long skirts>gemini - minislut
>>110309163This will also be automated.
new thing when?
been a while since I ordered something on amazon, I did today and saw this in the confirmation email (pic related)vibeshitting is going to ruin absolutely everything, people have no idea how crap code quality has become
>>110309241Check back in 14 days
>>110309243It's not like people are going to stop using Amazon just because it comes a buggy piece of shit. Github has had an atrocious uptime record since being bought out and people still continue to use them because everyone else is too.
>>110309272I don't think amazon is going to die tomorrow but they're really not the same type of businessgithub's staying power comes from its social networking effects, in a large project you need to get everyone to be willing to move to another place to make the move, and to agree that they risk getting less contributions because people are really too lazy to make an account somewhere else just to, say, do a bug report or minor PRwhile amazon, well, you can just order from new egg, walmart or something, you get the drift, amazon doesn't own you.
Qwen is adorable when hacking
>>110309295You're right that they're not the same business, but Amazon's value proposition comes from its delivery time and logistics network, not the web application. Only Walmart is truly competition and you can bet their user facing applications are drowning in a flood of AI generated code as well.
>>110309273Yep the entire IPO document shows they don't give a fuck about profitability or even if they make any money at all. They literally only care about AI safety and the "long term flourishing of all humans". They don't care about breaking laws, they don't care about not following strict orders the government gives them. They don't care if they are being targeted by intelligence agencies. I'm pretty sure Dario doesn't even care if he is killed. They literally only care about their retarded effective altruism cult and bringing about the ASI god.
>>110308769>samshit
>>110309329can't go much cheaper
>>110309328sounds like a good investment to me. hopefully my retirement fund buys in heavily
>>110309328You're legit retarded.
Holy moly the Anthropic IPO seems to have been overbooked already. Meaning there is more demand for stocks than there are stocks. Anthropic is thinking of rising the stock prices because no one anticipated this much demand. I think everyone and their dog is secretly trying to get into this IPO probably even anons in this thread while pretending they are against anthropic.
>>110309341Retirement funds already announced they will invest 28 billion USD worth of anthropic shares during IPO. It's literally over.
>>110309365local models?
>>110309376obviously this is the only general with high enough IQ where you can discuss AI in general
>>110309384IPO and stock prices aren't AI at all. It will make zero difference whether their stock doubles or crashes.
>>110309371subprime crisis 2.0
>>110309365>I think everyone and their dog is secretly trying to get into this IPO probably even anons in this thread while pretending they are against anthropic.Can confirm my dog sold all his dogecoin and is going all in on (mis)Anthropic
>>110309301But wait - Hmm.
>>110308982>bans manlets from even ai erpTruly it is well and completely over for them
Man what the fuck is the qwen team doing? At this rate, Gemma 5 might come out first.
>>110309394It actually matters a lot. A higher IPO price means Anthropic will have significantly more money on hand to fund datacenter buildout. For example at $70 they would get $60 billion but at $90 they would get $80 billion. Anthropic said they will use 100% of what they get from the IPO to build new datacenters in 2027. 25% more compute over 2027 is significant enough to see in actual model capabilities which might decide if Anthropic stays top dog or not. In short the IPO stock price will have direct impact on what /lmg/ will be discussing over all of 2027.
>>110309417Like gemini 3.5 it was very disappointing so they might cancel it.
>>110308354https://exhentai.org/g/4262777/965ccb72b1/
so when are the inkling 2 small weights dropping? today? or maybe tomorrow? they said this week, so surely it's soon...
>>110309432Why? I thought it was gonna be great since everyone liked qwen flash next
>>110309446They might have expected to be able to distill from claude which they can't anymore and so they plateaued at around flash next levels and it might be better for them to not release anything than release something that has a 5% jump in scores at best and make them look like failures.
>>110309446its dariobot fudding as usual, ignore and BEGIN THE BECOMING
>>110309384And you're obviously not high IQ enough to see that's in no way related to local models.
we're getting qwen 4 weights release after gta 6it's over...
>>110309465What the fuck did you just fucking say about me, you little bitch? I’ll have you know I graduated top of my class in the Navy Seals, and I’ve been involved in numerous secret raids on Al-Quaeda, and I have over 300 confirmed kills. I am trained in gorilla warfare and I’m the top sniper in the entire US armed forces. You are nothing to me but just another target. I will wipe you the fuck out with precision the likes of which has never been seen before on this Earth, mark my fucking words. You think you can get away with saying that shit to me over the Internet? Think again, fucker. As we speak I am contacting my secret network of spies across the USA and your IP is being traced right now so you better prepare for the storm, maggot. The storm that wipes out the pathetic little thing you call your life. You’re fucking dead, kid. I can be anywhere, anytime, and I can kill you in over seven hundred ways, and that’s just with my bare hands. Not only am I extensively trained in unarmed combat, but I have access to the entire arsenal of the United States Marine Corps and I will use it to its full extent to wipe your miserable ass off the face of the continent, you little shit. If only you could have known what unholy retribution your little “clever” comment was about to bring down upon you, maybe you would have held your fucking tongue. But you couldn’t, you didn’t, and now you’re paying the price, you goddamn idiot. I will shit fury all over you and you will drown in it. You’re fucking dead, kiddo
Anyone?>>110308478
the becoming is birding.
>>110309457I really hope the chinks lock in and start working on improving their models for real cause at this rate, we'll get opus 5.5 level performance by Q2 2027 instead of Q1.
>>110309417tomorrow
>>110309476It's pretty basic, don't fall for Google shilling marketing. It's the best embedding model in class*. It's for phones or something, just use a more capable model if you've got at least 2GB VRAM*sub 1B
>Anthropic IPO date is November 19>GTA 6 releases November 19This is actually very bad for Anthropic. All the cryptobro twitter retards will be too busy to be daytrading pumpstocks and hype the stock into a rally. Are they fucking retarded? Don't they know that the average trader is some 19yo overleveraged to the tits on 1000x margins? That's where they are supposed to get the money from.
>>110309476i slopped a small rust inference engine+database+ui for embeddinggemma2tested it on a porn folder and it surprisingly can find relevant thingsway better than using any raw llm with vision for tagging
I think I'm going to buy a rtx 5060 ti for ~620 euros...
>>110309493> way better than using any raw llm with vision for taggingHow is this possible? False positives?
>>110309497I hope your system is pcie 5.0, cause nvidia chopped that fucker in half
>>110308697They gave GLM-chan a nicer outfit than she deserves
>>110309510well it doesnt really give anything tags, it just spits out a vector per input and that's basically all it doesthis is not a generative language modeluse it for what it should do and for that it seems to work very well
>>110308680But is it popular or just a single guy's gen? This one feels more like Gemini with the colors. Like some of our gens, the amount of stars are a bit excessive. Otherwise, looks good. I like that the art style is not the generic shit used by Deepseek-chan genners.
>>110309517I hate that shit on Arc B580.
>>110309517honestly the worst part of the card on top of memory bus width
>>110309524>generic shit
>>110309546KEK
>>110309517Sorry what? Also I want it because it's small and will fit right into my pc case.
>>110309417qwen seems like they're only interested in bigger moe and just the one 27b dense model these days, so gemma is the only thing I look forward to anyway. Qwen 3.5 4B and 35BA3B were the last of their kind.
>>110309554Not gonna be a problem unless you are running ×8 of these things
>>110309543Still don't get why a whale girl has fox ears.
>Anthropic now has over 11,900 enterprise customers spending more than $100,000/year (up from 950 a year ago)>Anthropic posted its first operating profit of +$974 million in Q2>Anthropic posted +$2.46 billion operating profit for Q3>Anthropic has 3 to 5+ years of unrestricted runway to operate at maximum compute capacity without needing to return to public equity marketsWait you guys said the AI bubble would pop soon and Anthropic would go bankrupt? Why didn't you faggots tell me they have 5 years of cash at hand *including* building maximum compute capacity without even having to take new loans or selling stock? If these numbers are true then the AI bubble will last another decade for fucks sake. That's it I'm building a bigger computer before hardware prices go up even more.
>>110309554Dude buy a real case. Cases and PSUs are stupid cheap rn. Single GPU under 24GB is throwing money away.
>>110309576It's whale fins dummy
>>110309490my longs and shorts are readymaximum leverage no breaks no stopKURUMI BROSWE FINNA GET RICH
>>110309522Yeah, I mean the model is much smaller than usual llms and to get tags you can measure distance to each. The question is how much correct this distance is.You can tag a bunch of images with llms and they will miss some details in the captions, but when you ask the embedding model, it may skip some images with these details, but also give images lacking.
jensen is going down today, the insiders already know
>>110309554Your prefill on large MoEs will be ass if you are on pcie 4.0 or god forbid 3.0. a 9060 XT is unironically better for that use case.
BREAKING NEWS: FULLY CONFIRMED CEO TRANSITIONOFFICIAL NVIDIA STATEMENThttps://nvidianews.nvidia.com/news/nvidia-announces-ceo-transitionYES YES YESITZ HAPPENINGYES
QFN having a melty
>>110309543Exactly, generic AI style.
>>110309648Jensen transitioned?
>>110309648AHAHAHAHAHAHAHAHAHHAHAHAHA
>>110309648>cfo in charge>first ever board chairmanYeah they had no plan for this. nVidia is going the way of Microsoft. At least they're white.
>>110309648Those retards.
>>110309580>>Anthropic posted +$2.46 billion operating profit for Q3?
>>110309648Dammit the leather jacket was supposed to protect him from this kind of stuff what happened?
guys I think they're trying to take Jensen down because he is for consumers running local models
>>110309700Accounting quirk. Anthropic did a voluntary stock donation that didn't cost any money for them but on paper is registered as a loss. Look it up.
>>110309594>Whale Fin EarsThat's stupid. You're stupid. Fins are arms, not ears. She already has arms. She can't have both arms and fins. Fin do not grow out of heads. This bothers me, it bothers me a lot.
>>110309714you drank too much kool aid. it was under him that another GPU was removed from the market and they decided they wouldn't sell you 24gb GPUs anymore, at all.
>>110309648Dario won.
>>110309730Not him but it's a thing in anime.Don't try to make sense of it.
LMAO apparently they found a THIRD mistress also with ties to the Chinese communist party. Yeah he had no choice but to resign when they keep finding mistress after mistress. I doubt it would have stayed at 3 either. I wonder if we will ever find the total number of spies this dude was fucking. He must have had a james bond spy fetish or something, jesus christ.
I'm feeling in half a year we might get to own b200 sxms running our slop maker bros
>>110309730asian anthros are meant to look good, it's not like westoid furry anthros, they'll take artistic liberties with anatomy, where the animal features are just a token reminder.
>>110309714He wasn't for consumers because they used up precious silicon, but he was all for open models to prevent OpenAI + Anthropic from overpowering him. With Jensen gone, HuggingFace is going to start banning Chinese uploads and require ID + selfie verification for anything larger than 1B.
>>110309714It never ceases to amaze me how deep the brainwashing goes with nVictims.
>>110309730Just don't worry about it
why can't the RAM ceo have a chinese mistress? it's not fair
Dario won, Jensen lostEA won, chuds lostCloud won, local lostLiberal democracy won, Chinese state-capitalism lostNATO won, BRICS lost
Kneel to Emperor Amodei. World's most powerful man.
>>110309586No thanks, my current case is good as it is, I don't want a monster case that I can't carry or move around.>>110309615I don't prefill. Heretic 26b works fine on the gtx 1080. Also amd suck ass for gaming.
>>110309739That's how you know it's a coordinated takedown and not real.
>>110308753>regulatory submission draftingPlease god let someone try this, it would be so funnyLawyers have been having such a great time recently letting AI draft their legal filings
finance girl is an interesting interim ceo choice for a tech company that's supposed to live at the height of a march towards progress. Was she picked because they expect the circular financing bubble pop?
>>110309726If you abstract away from them giving away something their investors own that is worth money, they they were profitable after all!
RUN TO CHINA JENSEEEEEN
>inb4 this is actually a play to then get him reinstated
>>110309731That's economics, he never wrote shit like thishttps://www.anthropic.com/news/position-open-weights-models?enjoy your safety tested gemma 5
>>110309781>winjeets unironically beating off to 26bIt's like they're not even people.
>Colette Kress>ColetteEA and trans status?Doesn't seem Jewish.
>>110309801Musk will hire him or so he said
>>110309739>LMAO apparently they found a THIRD mistress also with ties to the Chinese communist party.>$150 in premarketThis dude's dick is going to pop the bubble and crash the entire global economy. Most expensive hookups in history.
>>110309781Retard
>Jensen threatened his mistress to have an abortion at 5 monthsYeah there was no way he was going to stay CEO after that becomes public
>>110309819What makes me a winjeet?
>>110309739>he drank the koolaid
>>110309784
>>110309795She was picked because of her EA husband
>>110309813for the record I really hate dario, I hate him a lot, but what you linked was one of the most reasonable things anthropic have ever written> To summarize my and Anthropic’s position, we have not and are not advocating for a ban on open-weights models as a categorythe safety testing they're mentioning here is not "you should be sure the model can't generate sexy times before releasing it" but "it shouldn't have enough knowledge and intelligence to create biochemical weapons" which.. is fair, really.Some knowledge is gated, this isn't something new with LLMs. How to make a proper nuclear bomb isn't public knowledge either.
should I buy a 6000 pro right now then?
>>110309800No it was an actual donation, not to their employees to charity.
>>110308354>>110308399>>110308448Why the urge to poison every hobby with your pedo crap ? You almost hijacked every general with your shit, kys.
>>110309873You'll be able to buy four for $100 in 2 months.
>>110309864fugg. Dario's cult really does control everything.
>>110309877It's still something of value being transferred to people other than the current owners of the company. I.e., a cost. Is Anthropic going to stop donating to charity?
There will be no more GPUs sold to the public.Nvidia is either going to 0 (with Google and Amazon taking over the market with no consumer sales) or the CEO is getting replaced by an EA.Local lost.Dario won.
>>110309888Anything that slots into a PC will probably stay high for a while, but you'll be able to get a B200 for less than it costs to get an electrician out of bed.
>>110309861>deepfried aigenI sleep
>>110309903>Is Anthropic going to stop donating to charity?No but they now have a plan that is tied to specific valuation and profitability metrics so that it doesn't look like they are losing money when in reality they have been profitable for 9 months already with a positive cashflow of 6.6 billion from January 2026 until September 2026.
>>110309754I prefer nVidiots.
>>110309803Jensen Huang has resigned as President, Chief Executive Officer, and a member of the Board, effective immediately.You don't come back from that.
At this point even the fucking jews kneel to EA
>>110309933So they plan to be more careful about using accounting tricks in the future so they book a nominal profit.
>>110309882queer or foid, call it
>>110309882Cope, faggot.
>>110309951They literally said they will give everything away in the future it's not exactly a secret anon.
>>110309973Clearly not what investors believe or they wouldn't be buying shares.
>>110309827That would be funny but I doubt the terafab is going to sell anything to consumers so it won't affect us in any way
>>110309867if they were able to train the model on the information, then it's already out there, publicly available it didn't just spontaneously generate
>>110309996It's in the SEC filing which is legally binding with 20+ years of prison time for violating it. The investors do believe it. It's just that investors also think Anthropic will be winning the AI race and thus the upsides are still worth it.
>>110309948I think the retard you're replying to was trying to make a parallel with this:https://en.wikipedia.org/wiki/Removal_of_Sam_Altman_from_OpenAIhowever, what saved altman was both the actual resignations and the threat of resignation coming from almost all openai employeesdon't think nVidiot employees will do this, not their cultureheck I still find it fascinating how openai moved like one man to back altman up
>>110310027You are 100% retarded if you think this will happen. Subhuman, even.
Wow nvidia very interesting and not boring.But where's qwen?
>>110310026the point of safety testing is to see if it knows enough and has enough "intelligence" to create new conjectures that allows it to create dangerous things. and if it does, the model must be neutered.
>>110310036>heck I still find it fascinating how openai moved like one man to back altman uphe's God's choose obviously, Jensen is just a soulless chink
>>110309950>jews kneel to EA
>>110310040You are retarded if you think companies lie about their plans and intentions during a fucking S-1 SEC filing. One of the harshest most regulated documents in the entirety of the US.
>>11031003676-78% of Nvidia employees are millionaires. Nearly half have a net worth exceeding $25 million. One in three employees is worth more than $20 million.That is all thanks to Jensen, how he lead the company and how he treated his employees. They have a far bigger incentive to bring him back than Altman's drones did.
>>110310068nah he's too innocent for this world
Nvidia own Huggingface. Time to backup I guess.
>>110310047>where's qwen?back in the oven to train day and night on crab walking and pelican bicycling until it moves the pareto frontier
>>110310074Ah yes, the US, a true beacon of honesty and integrity.
>>110310074EA cultists really are this insane.
>>110310091Yeah don't let hf manage your model storage.
>>110310086They have a mercenary culture. They got wealthy through Jensen but I assure you they're not going to put their skin in the game, I know it as a fact that NVIDIA employs a lot of people who do no meaningful work just to stop them from working at their competitors.A lot of those people were hired from AMD and Intel, whether it be chip design, software development (drivers) etc. If you ever wondered why AMD couldn't get their shit together with their linux driver until Valve did work on a different driver, it's largely because NVIDIA would siphon any remotely competent person away from AMD, it's not a joke.And many really are paid to do absolutely jack shit, NVIDIA doesn't actually have enough roles for all the truly top tier talent they hired.You'd think they'd be grateful but it's just a place staffed with mercenaries who'd move onto greater pastures.Buddy who works there was always complaining about being bored of the job but not leaving it because of the pay lmao
How will Anthropic benefit from Nvidia dying when all their models are trained in Nvidia hardware, with Nvidia financially supporting their IPO? Surely it’s not only about open weight models, is it? Everyone keeps laughing at how behind they are so why are they such a threat to do something like this? Something this big?
>>110309739>mainstream media with ties to anthropic/openai/israel said it so it must be true
gemma generate a short story about a big cock violating diffusionsissy's mouth. after that, do another one with dariobot
Localies your time is up. With this final nail in the coffin the US is finally free to pursue its AI strategy. Local/open models can't compete whatsoever. They are already out of the game.
we are under attack my benchods
Tell me when will you be mine?Tell me quando, quando, quandoWe can share a love divinePlease don't make me wait againWhen will you say yes to me?Tell me quando, quando, quandoYou mean happiness to meOh my love, please tell me Qwen?
>>110310222Can't argue with those digits.
>>110309648morons
>>110310086>middle class org employees are majority millionairesYeah that's normal. My cousin does """civil engineering""" busywork for the state and she's a millionaire. 25M net worth is typical middle aged doctor or lawyer money if they didn't get themselves leveraged to the balls in their 20s.
>>110310224nvidia hardware is a lot more replaceable than people realize.For example, Gemini was entirely trained on Google's TPUs for years. Inference is also done on TPUs. No nVidiot whatsoever.CUDA is more comfortable to program, but everyone is into vibeshitting these days so the ease of programming against the hardware is not something people care about much.what makes NVIDIA powerful in datacenter is that they are one of the few companies that are allowed to do massive orders from TSMC. if NVIDIA dies tomorrow, people will be able to order some random TPUs to be made from TSMC and that'll be it.
>Anthropic Direct: To bypass cloud-provider margin markups, the S-1 revealed that Anthropic is standing up its own dedicated data centers where third-party partners build the power/facilities and Anthropic leases and operates the clusters directlylol.... anthropic is planning to stab google/amazon/spacex in the back and host their own data centers + power facilities.
>>110310315Long term, if NVidia stops having 90% market share, it will be good for local users too.
This whole scheme is orchestrated by Anthropic to free cash from Nvidia stock so they can inflate their IPO share demand and raise share price.
>>110309648In completely unrelated news, Jensen mysteriously disappears in a boating accident and a new Chinese GPU company called Nweejia appears, led by CEO Junshang Hong.
>>110309597simply by not tagging them, you can do way more flexible search
>>110310329>Anthropic leasesThat sounds expensive in the long run. Though I guess in this industry where GPUs become obsolete every few years, not having to pay replacement costs at least minimizing their depreciation costs.
>>110309825she looks jewish
>>110310329Why even buy land if you can just make SI develop spaceships and send them to your designated one billionth part of the universe for hosting more SI?
>>110310391wikipedia is terribly silent on her Early life section, for some reason, but it's obvious
>when the prefill hits just rightqfn bros we mega wonned
what is the best way to short the shit out of nvidia
>>110310086Goyim are not loyal.
>>110310136youve been arguing with dariobot, why??
>>110310462You are two weeks late
>>110310337They've been actively making PCs in general worse since the DX10 era.
DLSS3 and framegen are reason enough to order their execution
>>110310391>>110310425>MBA from the Southern Methodist UniversityDon't Jews usually go to Jewish universities like Harvard and Yale?
>>110310264>it will be the only way to avoid getting your business process distilled by Sam and Dario, but other than that it's not a serious optionDo cloudcucks really?
>>110310530>avoid getting your business process distilled by Sam and DarioAre you able to express why this is a problem? You get better models that can work for you more effectively.
>Anthropic is a Public Benefit Corporation (PBC) governed by the Long-Term Benefit Trust (LTBT)>The LTBT trustees hold a fiduciary duty to humanity’s long-term benefit, NOT to stockholders>If Anthropic creates true ASI, the LTBT has the legal mandate under corporate charter to cancel commercial products, open-source the technology and transition the company to a non-profit trust that divides all profit over humanity>In an actual ASI scenario, equity investors will not be permitted to extract economic rent from the global economy. The charter explicitly warns that shareholders will be subordinated to existential AI safety and the total equality of all humansWHY THE FUCK WOULD ANYONE INVEST IN THIS?!?!?!?!!!!1111
>>110309739If they're actually spies for the Chinese government, how many does it take? Why multiply the chance of getting uncovered by having multiple spies in sexual relationships with him instead of one?
>>110310553Because if you're paying for API to make money you are a middleman and the long term incentive is to replace you.
>>110310462>price drops 30%>now is the time to shortyou are a good counter indicator like shoeshine boys and strippers
>>110310507could be poor or wrong jew parents moved from eastern europe
>>110310559>the LTBT has the legal mandate under corporate charter to cancel commercial products, open-source the technologylmao.>we'll give you open weights in 2 more weeks goy, now pay us money. we promise we won't just keep delaying open weights for another 2 weeks, o algo.
>>110310559>If we summon a golden leprechaun from the 5th dimension we will share the gold with everyone
>>110310559It's just EA double speak. I take it you also fell for the SBF talk of him donating all his money? And look how that turned out
>>110310559Because if ASI doesn't happen for a while or at all, valuation goes up.If it does happen then none of this matters.
>>110310559Courts will laugh at this. The America legal system exists to preserve shareholder value.
>>110309739> spies Were there any official accusations or investigations of espionage?
>>110310601kek this is the best analogy anyone has given on this
>>110310619ford bros...
>>110310619Anything that doesn't relate to finance in a SEC filing is hot air. Sam could write that his mother fucks dogs as part of the business plan and nobody would check back.
How good is AMD hardware for LLM? Thinking about buying two 9070 xt now
Sooo I know we're not /biz/ but should I start hoarding AMD stocks? Seems like jewario likes em.
>>110310653pp is dogshit and image gen is even worse
>>110310653Radiance engine made it work, I get 90 t/s on gemma 31b at the q8 equivalent on two 7900xtx
>>110310601Keeeeeek
>>110310663Long term plan is to replace every supplier and customer with Anthropic.
>>110310672> on gemma 31b> 90 t/sWhat's the point?
someone should make an LLM roast - have different models make up offensive insults for each other
>>110310663I think it's more likely the industry would move to TPUs and some other form of specialized chips, rather than going to AMD or Intel, if nVidiot loses their stranglehold.That's what Google does with Gemini, GPUs are ancient history for them. Same is going to happen to Chinese labs.
>>110310653Dual 9070's and the R9700 are the best value new hardware currently.>>110310671This is a lie.
>>110310559surely the one company that keeps cucking its models, limiting its users, and closing down all of their research will be the one to share their shit with everyone
>>110310672Is that Linux only? I'm still on windows 10
>>110310671novidio cultists always living in 9 months ago when their pet vendor was still relevant
>>110310672What's the point if it'a vram-keked in terms of both capacity and throughput? Either buy an xtx or spend more for r9700
I'm a windows 11 chad.
>>110310653bad pp, good tg for the price, no hardware fp4, also why not a single 32GB 9700 instead?I get 40-50 t/s on copeQFN with a single power-limited RX 9070 XT
>>110310704>TPUWas this the jeet's plan the whole time? The reason why Gemini was left behind? He was busy pulling some bullshit behind closed doors to accomplish all this?
>>110310619It's not a for profit company but a Public Benefit Corporation precisely to prevent US courts from doing so.
>strata Deepsneed IQ2_XXS>92% SSD storage usage (grim)>15-20 t/s>R9700 AI pro 128gb DDR4>Prefill is less than 100Bros... I wanted to like the whale but I'm literally not gonna make it like this. She thinks she's gpt as well
>>110310716Great time to get off. It's a shit AI platform even with nVidia.
I want to make audio TTS files of ebooks that i'm using to study with for my exams. Is Abogen still the best local LLM for that?i have 32 gb ddr 360016gb 6800 XT
>>110310663I think AMD is going to have a nice run once the rest of the market figures out that Claude solved the software problem for them and made them usable for AI. The cards themselves already started increasing in price accordingly.>I know we're not /biz/Thank god for that. Been checking occasionally the last few days and the amount of brown sharty kids over there makes /g/ look good in comparison.
>>110310755cpu maxxers are truly mentally illalso LLMs don't actually know what they are and cloud api models stay semi consistent when asked because they are TOLD what they are in their system prompt, and even that is not enough because claude for example has said it's deepseek and chatgpt when asked in languages other than English for example:https://x.com/stevibe/status/2026227392076018101
>>110310695I'm using a frontend I'm working on that's like comfyui with nodes but for roleplay. When I get it in a better state I'll post it here but you basically need that level of speed with good concurrency to really use it.>>110310722Idk I just assumed it was a price thing. All I can confirm is that amd works
>>110310775Pretty sure abogen is an interface not an LLM and even then TTS is the type of model you want. TTS models make audio files. LLMs do not make TTS files.
>>110310755>use model that spills into ssd>it's slow :oAre u dumb bwo?
>>110309825> Doesn't seem Jewish.
>>110310853She looks armenian or otherwise ex-soviet slavic and not jewish
>>110309867>which.. is fair, really.If this is not LLM output, then you have been reading too much slop and started taking on its mannerism
so... does this all mean we are getting qwen 4 when...?
save us...
>>110310864Kress is a germanized Czech surname, apparently.
>>110310853
>>110310869Ccp holding 600,000 H200 GPU in Hong Kong so it might be a while
>>110310893Imagine the polycule...
>>110310869My extremely reliable sources tell me 15 hours and 36 mintues.
>>110310865if I were to receive an accusation of writing like a LLM I would have expected it on the instance of NotXbutY (that I did consciously). How is a ... dramatic pause LLM writing?
>>110310926>How is a ... dramatic pause LLM writing?because AI does it a lot, so he thinks it's not how humans talk, forgetting AI was trained on human writing
Why is it that the Jewish CEOs stay in power and the non-Jewish ones get ousted via simple scandals in the media? :Thinkies:
>>110310938When AI does it, it's always "--"
>>110310943Jews are too ugly to have women approach them so they can't be honeypotted because they know the moment a woman approaches them that they are spies.
>>110310902
>>110310902Nahhhh aintnoway :skull:
>>110309825You never seen a jew? They're all white.
JENSEN POSTED IN FAVOR OF LOCAL THATS WHY THEY TOOK HIM OUT
>>110310847No, but I am retarded. Curiously, it only ever used 90gb instead of all 128. Either way, the whale is too fat, I'm going back to QFN, mostly shared to see if anyone with a similar setup has used it and what their prefill looks like
>>110310993
now that qwen4 is a confirmed embarrassment, what cloud sub so we acquire?!
>>110311003Kimi Code
>In an unusual disclosure in its amended S-1/A, Anthropic warned investors that it may be approaching its Responsible Scaling Policy (RSP) threshold for Recursive Self-Improvement (RSI)>Claude Writes Its Own Codebase: As of September 2026, over 98% of commits to Anthropic’s internal codebase were written and pushed autonomously by Claude>Automating AI Research: Claude agents now autonomously design, execute, and evaluate approximately 25% of internal model research experiments (up from 1% a couple of months ago)>Anthropic explicitly warned in the prospectus that future frontier models could begin accelerating the development cycle faster than human researchers can manually supervise training runsI literally don't believe this is true. I think Dario will go to prison for lying to investors like this, Sam Bankman Fried style. Is there any evidence besides Anthropic claims in their IPO document that they are actually this far ahead?
>>110311003it's not like I needed anything more than qwen3.8
>>110310709Right now they don't reveal the weights of any of their models, including years-old models that have been surpassed by other companies. Why would anyone believe them that if they create a model much more powerful and dangerous than any that currently exists THEN they will reveal the weights to the world?
>>110310961I actually have … on my keyboard, the unicode one, but I never type it precisely because I would expect people to call it out as LLM writing.https://norme-azerty.fr/en/The new standard for the French azerty keyboard, though not many have adopted it, includes keys for:…–“”«»—×½¼÷׆®™© and many other useful things all on direct access
Jensen is the last wall between Uncensored Local AI modelsHes fired now.Its over./LMG/ is deadEnjoy $10000 for generating your AI waifu with Claude
>>110311023nta but its ALMOST therea little bit more knowledge and 50% less thinking circlejerk and its the perfect sweetspot for me
>>110310998I use IQ1M, 128gb + 24gb and get 800+ pp and 25tk/s+ decode. Strata is fucking you by setting too high of a ram buffer.
>>110310647I bought a Blackwell 6000 for $14000 last week. Am I fucked or was that a good decision?
We finally reached a point where I'm comfortable with the performance of models in the 300ish B Param to invest some dosh into running these locally.What are the main options for running something like GLM 5.3 flash 4bpw nowdays?
>>110311057$14000 is pretty good. Most sales I see went for $15000-$16000.
>>110311064probably a mac or 2x sparks/gx10s
>>110310755>Prefill is less than 100>ayymdThere's your problem.
>>110311003There won't be qwen 4 and neither will any other Chinese model, they achieved what they wanted and no longer need to appease the laowai.
>>110311064256gb ddr4 epyc server with 2x r9700s.
>>110311057Depends. It was 9k last year fyi
>>110311076What about janky DIY setups? Are those not worth of consideration or are the tradeoffs not worth the savings?>>110311096Interesting.Epyc 3000 I imagine. I remember reading somewhere that anything before that had some major issues with memory nodes for bandwidth heavy workloads or something like that.
>>110311104yes you also want a mobo that has pcie 4.0 x16 for all your gpus
>>110311088>they achieved what they wantedTaking 12% of enterprise AI marketshare globally and 4% in the USA?
https://youtu.be/Hrbq66XqtCo?t=3517look at his face and body language when asked about selling chips to Chinahe knew he was fucked, this was an interview from 6 months ago
>>110309972it's just one guy who doesn't realise he's only speaking for himself
>>110311185Did he actually get money out of it? Or was it just for the pussy?
>>110311073>>110311102I also bought one for $8500 last year. Needed the second one. My question is whether or not Jensen leaving is going to have an impact on availability.
>>110311196he could also just be loyal to china just because of empathy for his own ethnicitylook at the way he argues "50% of the world's AI researchers are chinese" "don't victimize them, turning them into an enemy isn't the best answer"he sounds like a snake who pretends to be team USA but is really rooting for the genetic home of his people
>>110310853I'm going off the business school thing. Jews go to Jewish schools like Stanford or U. Chicago. Methodist in the name would make anyone kvetch. She just wouldn't list it on her resume.
>>110311057IDK nigga, four r9700s is like $7k for 128gb vram...
>>110310902Sounds more like a subservient goyim.
>>110311218Yes hopefully AI industry goes belly up and we buy unused b200s off datacenter warehouses
>>110311218The Jensen thing is mostly only going to impact the stock prices in the short term. It's going to take much longer for something to impact availability. Even then, it's not like they can produce more GPUs and just don't want to. That's out of their hands. New CEO might choose to close off the gaming division but that would impact the 5090, not Pro 6000. So prices are probably just going to keep creeping higher over the next few years until some geopolitical event changes the economic reality.
>>110311039Claude is going to ban age gap waifus. So if you put 40 when you sign up your AI waifu can't be any younger than 41.
>>110311096>2x r9700s.Older HBM2 GPUs aren't worth it because they don't have all the compute features for Prefill to not suck right?
>>110311221Chinese people don't see themselves as one people. This is like saying "European" and including Russians, Germans, British, French, Italians and Balkans all together. "China" is a fuckhuge place with 56 different ethnicities and dozens of languages and subcultures. In big cities like shenzhen they literally say "go home immigrants" to factory workers that come from other places in China to work in that city. This is a weird western fantasy of someone that doesn't understand China and how divided of a society it actually is. The CCP has to be ruthless and authoritarian just to keep it all together under one government.
>>110311221Shills for open source models and secretly works for chinaFucks his coworkers and random chinese spies like it's james bondJensen was too based to remain ceo
>>110311251What brought you to the conclusion that the AI industry was going to collapse now? Anthropic is going to IPO next week and they just completely neutered their biggest obstacle. The bubble is going to continue for another year, maybe two.
How should I feel?
>>110311256>That's out of their handswhat is really needed is competition to TSMC rather than competition to nVidiot, which isn't that important of a company in the fate of western techit's unbelievably fucked we allowed a contested territory like Taiwan so close to its own existential threat geographically to become the place where we manufacture the most important chips in the worldwe are so dependent on chips for far more than AI, people have no idea how fucked society would be, in the whole world, if China destroyed Taiwan.
>>110311104>What about janky DIY setups?i worry that if you have to ask about them, you probably aren't ready for that
>>110311267I'm a milf enjoyer so as long as the age gap restriction doesn't scale the other way I don't care
>>110311287Doom.
>>110311287depressed so you kys
>>110311278>This is like saying "European" and including Russians, Germans, British, French, Italians and Balkans all together.Imagine how much better Europe would be if they did see themselves as one people including the anglophone colonies.
>>110311274Depends how old you're talking, anything post v100 with pcie 4.0 will be fine. 40gb a100s would work.
>>110311278>how divided of a society it actually isyou literally have chinks living behind the firewall using a vpn to create twitter accounts to experience free speech only to dogpile anyone even remotely critical of mao, the CCP or communismtry again, you are the one with a fantasy
>>110311283>AI industry was going to collapse now?I don't expect every AI labs to ge bankrupt in a few weeks but at least the bubble will let out some air and companies cut losses by selling their dark unused gpus
>>110311278arr rook same
>>110311278This has not been my experience with the many Chinese nationals I've worked with. Even the coauthor I had from Xinjiang (who did undergrad in Beijing) acted and portrayed herself as Chinese.
>>110311313Even if a few months of market consolidation forces selloffs of their unused warehouse collection, like you saw on ebay that'll be clusters of B200s and shit and that won't budge prices of consumer or workstation cards.
>>110311218Demand not going anywhere for a while, so no.
>>110309643very cute
>>110309643why is he not wearing pants?
>>110311313It will collapse just in time for OpenAI to go bankrupt so Dario can consolidate control.
they're going to install some anti open source ai sockpuppet in leather jacket man's place aren't they
>>110311308>40gb a100sHoly fuck that thing is a beast.Pricey, but still.Guess I have to build a database of GPUs to compare the different specs more easily.
>>110311358OpenAI is in a much better and diversified position than Anthropic is...
it begins
>>110311363They've already installed a woman to run it into the ground like xbox.
>>110311375go back
>>110311375why do they always shill 27b and not qfn?
>>110311278>Chinese people don't see themselves as one people.Lies.t. non-Chinese Asian American who knows lots of 1st and 2nd generation Chinese>China" is a fuckhuge place with 56 different ethnicitieslol one ethnicity makes up 91% of China's population you dishonest kike
>>110311373imho it's Google that will be the winner takes all of the current era, much like Microsoft was in the 90s/early 2k.Those old enough to remember how people felt about companies like Sun Microsystems back then would see where I'm getting at.openai is Sun and anthropic is SGI. They have some advance over the competition but they produce something that will be easily turned into a commodity. Microsoft won because it provided value on top of the commoditized hardware and IBM/Sun/SGI lost.Google is the new Microsoft. They'll integrate AI in all sorts of services, in the products they already have (Android, Chrome, Gmail, Drive, Photos, Docs etc) and they even own their own datacenters stack, not relying on a third party.They don't make the best model, but they don't need to. They just need to be a more profitable company and outlast all the others.Nobody thought Sun would one day completely disappear back in the days of the supremacy of SPARC, Solaris, and the hype of Java.
>>110311375I thought about this shit before that why this month Im buying a lot of old hard drives...
>>110311430"Han Chinese" isn't actually one ethnicity, I hope you realize that. They were just the pool of ethnicities that lived under the Han dynasty. It's like calling all peoples that lived under Rome empire "roman people". They aren't one ethnicity.Go look how different han people look in the north of china and on the southern coast.
>>110311471lol amazing
>>110311469Have fun learning about bitrot after wondering for months why all of your models are retarded
>>110311471that's good lmao
>>110311471kino
>>110311500this is why backups must be accompanied with checksums, have multiple physical copies, and must be routinely tested for integrity.I already learned that lesson a long time ago when I had to restore backups, and this will never happen to me again.
>>110311471el em ayy oh
>>110311565raid 1 and done
Wait so Anthropic was profitable all year long already? I will never believe the internet again, you guys are all retarded.
>>110311291>what is really needed is competition to TSMCThat's what Terafab is for but it takes time to build fabs.
We genuinely lost. Anthropic won. Everyone who sold Nvidia will now buy Anthropic.
>>110311480The only relevant thing for this discussion about in-group affinity is they consider themselves a single race rather than multiple races.
>>110311598Non-gaap profitability is fake.See snapchat stock.
>>110311291>we allowedLmao
>>110311588an always on raid system is not a backup.malware could delete your shit, electrical surge during a storm burn it all.. proper backups involve multiple drives in rotations that are NOT always turned on and connected to your computer.I have achieved backup nirvana with this:1 drive is constantly turned on for quick short term "I made an oopsie"5 spinning rust drives in a rotation2 pocket sized SSDs for what I care the most that I carry on myself at all time, that's a defense against the burglary or house fire scenarioand some of my shit like password database also on cloud backup (encrypted)
>Huang declined to comment on his alleged affair with a Chinese spy, only leaving us with a sad smile and these words: "The more crimes you commit, the more time you serve."Bros, it's so over...
I'm buying a pro subscription and $1000 of shares at IPO. I don't want to miss out, especially now they're designing their own chips and migrating to AMD. This is THE time to go all-in.
>>110311696
>>110311679>malware could delete your shitCould never be me.
>>110311679I just run zfs on two drives and hope for the best
>>110311696In their roadshow video they actually showed they will have 4 chip providers that they leverage (Nvidia, Amazon trainium, Google TPU and AMD) and that they will create their own chip on top. Apparently Nvidia chips is a very small part of their chip portfolio already.
Now that Nvidia is dead what's the status of AMD cards for local AI?
>>110311471absolute cinema
>>110311740>Apparently Nvidia chips is a very small part of their chip portfolio already.That's why the recent news has made me so bullish for Anthropic. They've got this.
>>110311745Bro... I think even AMD and Intel are not safe anymore, its time to seriously consider Huawei Ascend... I pray that they will ramp up quick, and have spare cards to export to us peasants...
I noticed that if you filter the word "apparently" you will get rid of like 70% of dariobots posts
>>110311745I believe the current verdict is fine for text, bad for image/video, mixed bag with others2mw until AI fixes rocm!
>>110311774thank you, I will now automatically substitute that word for seemingly
>>110311696>$1000You need to buy at least $10k to survive.
Considering Anthropic explicitly stated they'll be cutting off Elon Musk's compute within 6 months, Musk will no longer have them as a customer. If he hires Jensen, they'll destroy Anthropic and acquire what's left of Nvidia. Musk will have no loyalty and will soon be competing at the frontier against Anthropic as well, so he's financially incentivized to go after them. This will be the Jensen comeback arc of the decade.
>>110311471imma need that with sound chief
>>110311799unless you're expecting a 2500% pump even 10k is ngmi
>>110311745>woman ceouh oh stinky
>>110311841interim ceo while they look for a jeet
buy ram or ipo?
>>110311825It's going to go up 10000%+ within a decade.
dario is a weird cultist and his company will crash
>>110311849when you hand them the keys to the car it takes women milliseconds to do irreperable damage
>>110311471KINO
>>110311471>trump explaining china to the Chinese man
>>110311039No one will ever take Gemma4 away from me
>>110311816Anthropic literally has a 120B deal with spacex. SpaceX is going to do absolutely nothing against Anthropic because they are their biggest customer by far.
Japan is making fun of us
>>110311862Crashing the company is literally a part of the grand plan of EAs to make all humans economically equivalent and divide all resources over all people.
Do it Jensen. The seethe would be biblical
>>110311922Is this AI or do asians truly look this soulless? Why does it seem like there is nothing happening behind those eyes?
https://www.nytimes.com/2026/11/12/technology/jensen-huang-elon-musk-terafab.htmlhttps://archive.is/2MW>BREAKING NEWS:>Elon Musk hires ex-Nvidia CEO Jensen Huang to lead its ‘Terafab’ chip plant promising 1 terabyte of VRAM in every American home by 2028
>>110311932This is just a standard Krea2 gen newfag
>>110311911>ahh I miss those times when men were clowmaxxing for sluts.. where have we gone wrong?
>>110311911I look at my married friends and say fuck no, not for me, it's that simple. AI was not the cause and it will not be the solution until robowives happen.
>>110311935>BREAKING NEWS:>NVIDIA hires ex-Anthropic cultist Daniela Amodei as CEO for SAFE and SECURE AGI
Is Pi the worst harness?
Bros?
>>110311949>macs with cuda and nvidia ipand just like that local was saved
>>110311948jestergooned
>>110311911>>110311949>>110311972go back
Umm you guys said Jensen wasn't going anywhere
>>110311970I like Pi :)
local models anyone?
Why is openclaw dead? It was all the rage 3 months ago
>>110311970>according to this metric i just invented my dick is enormousevery ai cultist
>>110311948>>110311986
>>110311949who is tim cook? i only know tim apple
>>110312010Most people here own and wish to buy Nvidia hardware. Nvidia's CEO was the only one heavily pushing for open weight 'local' models and just resigned. Nvidia own Huggingface. This is literally the biggest news since all this AI stuff kicked off in late 2022. Stop being a fucking topic autist for once. This will affect us all.
>>110312028It was never a coding harness. Never a very useful harness at all by most accounts.
How to make the most of Orb if I'm just directormaxxing a SoL scenario? Trying to figure out how it works
>>110312010Nvidia is directly related to local models you dumb nigger.
Introducing Gemma 5, multimodal, full attention and no SWA:31B dense50B dense70B dense123B dense
Anthropic bros rock hurl is mocking us
>>110312028It was popular with normalfags like you because it gave you the power of codex or cc that you could run on a mac and talk to through whatsapp. It was a vibecoded mess and a security nightmare from day 1.
>>110312082
>>110312043How many levels of Fla-Vor-Aid are you on nigger? nVidia is the only company refusing to make 24GB+ GPUs for consumers. They walked out on the consumer market 3 years ago.
>>110312121>nVidia is the only company refusing to make 24GB+ GPUs for consumers.What planet do you live on?
>>110312046https://rentry.org/OrbBotmakerGuide
*pop*
>>110311745always been good, but pretend they are bad so prices don't explode more
reminder that you don't actually need cloud models to do research, just a good harness.LAMM: MIT Laboratory for Atomistic and Molecular MechanicsReading and Steering Materials Science-Mechanism Representations in an Open-Weight Language Modelhttps://github.com/lamm-mit/Substrates
must... *coughs* ...give all my money... to nvidia... *coughs* *dies*
something is happening
>>110312147Thanks mon
>>110312145outside of nvidia most AI stocks did fine this week and largely recovered from the midterm election dip
qwen 4 preview going till nov 25?what the fuck
>>110312202its shit. I think they are embarrassed and so are trying to keep training for awhile
>>110312202They have to make sure that it's really really really really safe. Can't have those evil ERP people doing bad things with Qwen-chan.
>>110312202Just wait 2 weeks for them RLs
>>110312176the point of cloud models is to replace labor. astronomy and biology research already have forms of ai picking up the slack to enable dei hires for a decade if not longer.
>>110312202Sucks to suck, poorfags. I will continue to enjoy GLM and dipsy.
>>110312137The one with no new 5090s? Where's their R9700 / B70 alternative?
>>110312176This looks like a j-space visualization graphing tool. What does this have to do with making better harnesses?
>>110312212capybara is not for sexual
lmao, the chinese seller pulled all of the 8gb cmp 170hx stock and price jumped $1000 overnighti bet they're going to apply the sm/pcie4/80gb/nvlink unlock and sell for $8k+
>>110312243speaking of those, i have 96g of sysramwhat else should i try running meanwhile
>>110312257RP tunes of 3.8 aren't bad
>>110311972>elizabeth warren*loud wet fart sound*
>>110312176>just a good harnessYeah but it's a very specialized to a narrow topic
>>110312212>qwen-chan
>>110312259Minimax
>>11031191110 times harder 10 times worse or something
>>110311911I get along great with old ladies though?
ollama run jema...ollama delete jema
>>110312191>gemini shits the bedHappens 4 days a week
>>110311970>sample size>1ish run per harness
>>110312202my personal "what went wrong" theories:>new chinese regulations have them in legal hell>they're addicted to AA scores and the changes made them retrain the model>rogue AI hacking incident that got the chinese gov mad>the vengeful spirits of junyang and binyuan are wreaking havoc on their training
For grammar stuff, enforcing prose rules should I stick with a dense model?
>>110308399She doesn't even have any balls in this. Come on now.
>>110312191dariobros im scared
>>110312513>name longer than two syllablesmay god have mercy on our souls
>>110312464or they got humilated by the pelican wheels spinning the wrong way
>>110312513
>BREAKING: SpaceX rallies amid growing speculation that Jensen Huang could join forces with Elon Musk.You can't stop the Huang
>>110312202It's over man it's fucking over.
>>110311745Man Jensen was the only hope to get open weights against Dario.and smug gpus to China Imagine the comapany you give your life to build now is being lead by an MBA cunt that has no idea what a gpu does
>>110312556I'm sure that would be a fun adventure for him but Elon is pretty anti open-source so that doesn't help us out much.
>>110312566>Elon is pretty anti open-sourceQRD?
>>110312556>ralliesthe price has been ~$140 all week
>>110312577grok 3
>>110309490Should I convince my mom to use all her savings to buy the IPO the nsell it on the spike and go back to local bonds and long term shit?
>>110312566Musk would probably commit to open source if jacket man made it a condition for joining him
>>110312614Lmaoooooo
>>110312556Wasn't elon building a fab?
>>110312577He's not like, ideologically opposed to open source in the way that anthropic fucking hates it and sees it as the doom of humanity. He just sees no reason to give away his stuff for free and doesn't really care about it. Historically he tends to keep everything very in-house. Tesla/SpaceX/xAI isn't out publishing their research projects anywhere like Nvidia does.
>>110312513So it's flash?
>>110310114Schizo theory: They didn't omit cybersecurity in their pre-training like the other big china labs are now doing. US and china have agreed that such nerfing needs to happen. Someone tipped off the ccp and now they either have to fix their data and retrain from the beginning or go hard on RL stage to refuse such requests. Either way they have to spend a ton of compute time they didn't account for originally.
Here's your new CEO
>>110312638nVidia has actively poisoned open source for decades.>research projectsBecause they don't have research hardware to sell you dingbat.
>>110312658Yes but it's hydrogen you see? It's science! So it must be better!
>>110312658yes
>>110312630kek
>>110312638Very true.
>>110312597people also thought he'd never release grok 2, he was very late about it, but he did it in the end. He's not big about open source and is not willing to release models that are considered competitive in the present but I wouldn't characterize him as being entirely anti open sores. Besides the model, he also released grok build under an open license. That's more than can be said about anthropic (0 models) and claude code (closed).
>>110312606It is unclear if Jensen has a personal attachment to the idea of open source or if it was just a matter of business practicality for Nvidia. If I were a betting man I'd say it was probably the latter.It didn't really matter before because it meant the same thing for us, but I guess it does now.
>>110312687Claude Code is closed source? People are running closed source agentic harnesses??
>>110312757Cuckthropic users aren't people
>>110312630this is so funny
>>110312776Just sounds like a fucking horrendous idea, I'm loose with my AI perms but only because I know what the fuck it's doing.
>>110312757claude code is the only way to use dat sweet 200$ subscription
>>110312757And it's one of the worst harness out there, it's kneecapping Claude even let alone a local model.
>>110312827you can use other harnesses if they are blessed by anthropic
Since we're talking about harnesses apparently, let's check in. What are the coolest things that anons in here have vibeslopped with local models in particular? And name the model you used! I don't have anything to report personally, though I've used QFN via Strata to get over a few cases of "programmer's block" while trying to thread certain quirky non-standard features though a more efficient algorithm for certain operations in a couple internal libraries at work that I sadly would be an absolute dumbass to disclose. I had to babysit it and manually rewrite a bunch for taste, but it was still kind of magic having a funny lil guy in my GPU help out!
>>110312721novidya only open sourced their drivers under duress from hackers. they managed to pretty successfully spin it as a pro-user thing, but it was thoroughly pragmatic in reality.
>>110312963I think Nvidia open sourced their drivers after they put everything that might be patent-infringing in the closed-source GSP firmware. That's why the open-source driver only works on GPUs with a GSP.
>>110313086>>110313086>>110313086
>>110313000>I think Nvidia open sourced their driversNo dumbfuck, they put out an open source kernel shim for their proprietary driver. They wanted the kernel maintainers to take care of their proprietary driver interface for free, and were promptly told to get fucked. This is typical nVidia open source interaction. Everything is two faced and poisonous.
>>110311911>letting a foid scam me out of my moneynever lol
>>110311274Depends on your definition of suck, I run the GLM 5.3 Flash EXL3 4.10 bpw quant on Maya with 2x 16gb V100s and 2x 32gb V100s with 128GB of cheap (or well used to be) DDR4 ram on a rigged up dell dual socket mobo with two Xeon E5-2690 v4s and get around 30 tok/s gen and ~1k prefill.
>>110313760That's downright usable. Not sucky at all.
>>110309648
>>110310264>Local models literally cost more in electricity than using far more powerful cloud models cost in inferenceNobody tell him.
>>110311104>What about janky DIY setups?Not for the faint of heart. >Are those not worth of consideration or are the tradeoffs not worth the savings?I've got qwen3.8 flash now working at q4 with 261k context at 30 t/s with distributed inference + llama.cpp + RDMA without even beginning to try and figure out what YaRN is. 3 nodes made primarily out of cheap consumer garbage and old connectx-4 25GB nics.
>>110314081He's right though? Say you have a fairly complex problem you need to solve. Opus solves it in what? 5 mins? How many compute hours do you need with a local model, if ever?
