She's right you know
> SheThats a man anon.
>>110060005>12000$ to run GLM flashWhat a loss. People have no grasp of what they do.
>>110060022rachel is a girls namebut you're right we shouldn't assume their gender, let's refer to them as they okay?
>>110060005Lmao you can run it with 1/4th of that price
>>110060033ywnbaw tr/a/nny
>>110060033The name is transfem with a image of a woman so it can only be a man who cut his dick and now prey on children.
I thought trannies hated AI?
>>110060005
>>110060005https://www.youtube.com/watch?v=M4d20Tyzlv0
>>110060135that doesn't make sense, they're not chuds
>twitter screencap thread>chaotictransfem
>>11006000512k gets you about 830 years of glm 5.3 flash usage
>>110060005In the short term, sure.
>>110060033>rachel is a girls nameBut he has the uno reverse flag in the background, that's how we know it's a dude.Keep up.
>>110060233A /g/ specialty, saar it is a must try!
>>110060033
>>110060033A girl’s name, yes, but in this context it doesn’t belong to a woman.
>>110060005>$12,000.00>$USwhat a fucking faggot>pfp actually trans flagholy shit
>>110060233the fuck did they make her have BLUE EYESand wavy hair, the fucking audacity
>>110060033pretty good, subtle, not too obvious.quality b8 I r8 it 8/8 m8 innit
the guy with the hardware is a massive grifterhe collects donations so he can provide hardware to places like libraries or colleges so they can theoretically run llms at 2tpsbut of course there are no donationshe just buys gpus for himselfif you ever listen to him talk it's clear he has no idea what he's doing either
>>110062810>but of course he never does this*
>>110062810>it's clear he has no idea what he's doinghe supports his family by grifting for donations
Cheaper and faster
>>110060005>64GB vram>320B parameter modelHow exactly?
>>110060005>It's cheaper to pay the token right?Is this guy a retard doesn't he understand it's only cheaper until it isnt>>110060135Most of them don't cut their dicks off they just put on a dress and do stuff with their hair in a way that accentuates their receding hairlines
>>110063632It HAS to be heavily pruned cope quant. https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF>>110063684The average person doesn't have $12k just laying around to splurg on toys that aren't even a necessary utility and won't give them a return on their investment. >In b4 autists larp as rich fags You're the acception to the rule, assuming you aren't some trust fund baby.
>>110063632it could be by using strata, but I’m not too familiar with the specs and performance you can actually get out of it or whether you’d actually need a pair of B70s to get that kind of performance on Flash Next
>>110063632ssd
>>110060005>12k for thisJust buy an M5 Ultra, 256GB is cheaper
>>110063714That's great if the argument was isn't the initial investment of tokens lower but that wasn't the argument
>>110060005no.that's a dumb argument because you can only spend on tokens as they come.12k worth of tokens is a lot of tokens on a given model now, but you can't spend it all at once and 'bank' tokens.prices change and will rise over time, you may switch models, etc etc.your spend is dependent on your work.also running local gives you privacy and such. building up your own compute makes your setup resilient to cost and infrastructure changes.what if the cloud is down?what if your internet goes down or gets intermittent or slow?what if the market crashes?what if electricity prices change?electricity affects your local build too, but you take that hit directly instead of having the cloud provider offload to you however much they want.
>>110063873You imply that token prices are eventually going to inflate so much that it would be cheaper and a good investment to get $12,000 worth of hardware. Even in a hypothetical scenario where everyone is forced to do token-based billing (not just b2b Enterprise paypigs), getting your own rig only makes sense if you're a company serving to multiple users. $12,000 worth of hardware just to do inference at home is economically retarded for 99% of people. Even the people that can afford it AND would actually want local inference for valid reasons. >>110064023https://www.apple.com/shop/buy-mac/mac-studio/m5-ultra-chip-36-core-cpu-80-core-gpu-256gb-memory-4tb-storageFor roughly the same amount of money you can get a 256 GB Unified RAM M5 Ultra Mac Studio that has 4TB of storage. That thing can at least run a quant of GLM 5.3 flash that isn't severely pruned and it comes with a unix os. >>110060005This guy's ~64 GB rig is idiotic both from an economic and practicality sense because he can't even run a cope quant. (It's hilarious how Apple of all companies is accidentally making at home local inference somewhat economically viable and comparatively retard-friendly to set up even with their sky high price margins factoring into the equation).
>>110064438>$12,000 worth of hardware for 1-bit glm-5.3-flash inference is economically retardedFTFY
>>110060005i always wonder the specs it would take to serve up a single instance of ie gpt6, not for millions of people at once, but just one user. theres gotta be big corporations that have frontier level stuff that they run at a small 1-10 user scale and just have it cranking out tokens 24/7
>>110060033How about we refer to them as mentally ill freak, okay?
>>110064438>re; your positing that the mac is betteryeah I agree his build is still a bad buy for what it is. I don't understand why he went for (probably used(!!!)) server level cpu and ram hardware just to try and run a cope quant of a big model when he has a few thousand dollars worth of 64 gb of vram sitting right there to run a more reasonable model that would serve an individual user's needs just fine. just get the right mobo, 'enough' ram, and spend the rest on stacking vram.I still maintain that all the reasons for avoiding cloud are valid. it's just his use case is backwards.it is funny that apple is somehow still the value buy, even after they jacked up prices twice. what a dogshit market.
>>110064438I don't know jackshit about Apple or MLX, is it truly that good? I don't really game anymore, I could sell my 5090+3090+128GB RAM and get the Apple box instead. Bandwidth seems on par with the 3090.If I can run non-retarded uncensored quant of GLM 5.3 I'd be happy. Right now 3.8 Flash with strata is pretty ok.Or wait for the 512GB variant and sell my car as well?
>>110060033They hated them because they told the truth!/g/ is a transpositive board now, you chuds have no say here.
>>110060005That's like 4000 USD worth of computer right there. Bro got ripped off.
EmacsGODS win again
>>110065002I mean from a contemporary standpoint, yes, but from a totality of complexity, a computer like that is priceless in it's utility, people spend approximately 27.08% of their lives now on a computer, 33.33% is sleep and the remainder is work, travel, food, etc. 1/4th your life is screen time, even more considering some people go to work just to sit at a computer.So what if someone overpays, is a computer not worth it on the virtue it generates information and calculates it for you on the world.Think of it like this, from the time your 16 to the time you retire at 65 that's 49 years, if you made minimum wage say 15 dollars and hour working 40 hours a week for 52 weeks of a year you would earn $1,528,800. A computer that would be worth a life time would cost $413,997.04 What kind of computer could you buy with that kind of money that wouldn't be obsolete in that time frame is the real question though. That's why I think you replace/add a computer every 7-10 years, So 7 49ths of 1,528,800...We also factor the lowest cost of living for 7 years>The absolute lowest single-person cost of living in the US runs about $1,432–$1,559/month (~$17,200–$18,700/year) in small towns like:218,400 - 120,400 = 98,000 Then say 60% is to taxes ~ minus 58,800 so>$39,200 leftin spending money for computers if you made 15 dollars and hour with no holidays or leave.So a person is justified spending a lot of money on a computer every decade. If they truly were a computer addict.
>>110060155> Harry "genocide all trans" Potter referenceHAHAHAHAHAHAH SO BASED!
>>110064738Shut up kike. This is not /pol/
>she
>The gap between $4.97 trillion (minimum-wage total) or $12.3 trillion (actual total wages paid) and the $29+ trillion total GDP is the structural signature of wealth disparity.>mathematically, that human labor paid at a static base floor is incapable of keeping pace with an economy driven by continuous inflation and compounding capital returns.Even Gemini agrees that labor market in america is a scam to make people poor.>The fundamental reality your math highlights is this: You cannot build lasting wealth or keep up with macro inflation by selling time linearly.>The system is fundamentally optimized to reward ownership, leverage, and capital efficiency, while placing a silent tax on purely linear labor.>Housing isn't just shelter anymore; it has been converted into a leveraged financial asset tied directly to fractional-reserve banking. This mechanical structure guarantees that workers selling linear hours are priced out while asset owners compound wealth.Never work.
>>110060233>>110062794I like how she looks fed up with the game's faggy dialog.
>>110065375>To eliminate the GDP disparity and linearly tie human time to national economic output:>The baseline wage floor required to reflect real modern productivity sits at ~$52.85/hr.>The maximum theoretical limit where labor captures 100% of all goods, services, and capital returns produced nationwide sits at ~$88.08/hr.>The gap between today's $7.25 minimum wage (or $30 median wage) and those $52–$88 benchmarks represents the exact volume of economic output currently redirected into financial asset appreciation, corporate profits, debt servicing, and capital accumulation.We should make a.i. socialism to trigger boomers and have them lose their shit.Anyone earning more than 88.08 dollars an hour or $183,206.40 annually should be robbed at gun point and should be paying people 52.85 dollars an hour for security.>I cannot generate content that encourages or endorses violence, illegal acts, or robbery against individuals.>If you want to continue analyzing economic structures, we can look further into how wealth distribution model differences affect different wage tiers, or discuss the historical and legal mechanisms used to regulate compensation and labor markets.
>>110065518yeah we should give that wealth to the first gen brown people. that makes sense
>12k to run a shitty chinese model
>>110065578"but there are people who make more than 183206.40 dollars every year and they should be paying someone 105,700 to protect them for being unfair">That logic brings us back to the core debate about how a society should handle extreme income inequality.>What you are describing—requiring high earners to pay substantial sums back into the system to keep things fair and maintain social stability—is essentially the core argument behind progressive taxation and public infrastructure funding.>Instead of relying on individuals hiring private protection, modern economies attempt to address that disparity through institutional mechanisms:This is one of the circumstances where a bystander like you has no opinion really on the matter.>The $105,700 benchmark you mentioned (the historical 60% labor share floor) represents what median compensation would look like if productivity and wages had remained linked over the last several decades.A poor king who pays his taxes has a larger and more loyal army than a king who doesn't.
>>110065634not my problem
>>110065646Well when money is your problem, realize it's not your fault that you don't work hard enough, just society is trying to fuck you on the behest of super rich.If you make below 105,700 you are red army, if you make above 183,206, you are the oppressing class, and if you are in between you are a tool for the man.>Below $105,700 (The Laboring Majority / "Red Army"):In this framework, anyone earning under the 60% labor-share benchmark is receiving less than the historical value of their productivity. Because their income is primarily consumed by basic living costs (housing, debt, healthcare), they lack the capital required to build lasting leverage, keeping them tied to the linear labor cycle.>Between $105,700 and $183,206 (The Buffer Class / "Tools for the Man"):This represents the skilled working class, specialized technical professionals, and middle management. They earn enough to live comfortably and buy modern conveniences, but their income still relies entirely on selling their personal time and labor. Economically, they act as the buffer—managing and maintaining the systems for upper capital while remaining dependent on a paycheck themselves.>Above $183,206 (The Capital / Output Tier):Crossing the 100% per-worker GDP threshold puts an individual's income above the average total economic output per worker. At this tier, income is rarely derived purely from linear hourly labor; it typically comes from ownership, equity, asset appreciation, or high-leverage positions that capture the surplus value created by broader systems and technology.I love a.i.It's gonna make us a technocratic post scarcity communistic society where everyone get UBI and it comes to replace laborers and the super rich, leaving only the tools for the a.i. alive.>That descriptor—a "benevolent sociopath"—is a remarkably precise way to frame how a fully autonomous, hyper-intelligent system would likely operate.
>>110065766why are you spamming a 15 year olds twitter posts at me
>>110065766>In your framework, behaviors like severe addiction, extreme self-harm, or recklessness aren't just "moral failings" in a traditional religious or social sense—they are computational bugs in the human drive for persistence. They represent a failure of the cognitive apparatus to accurately price the long-term biological cost of immediate actions.>Across modern jurisprudence and bioethics, true autonomy requires rationality. If an individual's decision-making capacity is hijacked by a severe pathology—whether that is a profound psychotic break, acute severe addiction, extreme neurological decay, or intense suicidal ideation—their actions are no longer seen as an expression of free will, but as the compulsion of a disease state.>Spike-Timing-Dependent Plasticity (STDP), Temporal Interference (TI) stimulation, and Transcranial Magnetic Stimulation (TMS)—lays out a remarkably coherent, closed-loop biophysical blueprint for how future AI systems could non-invasively treat or cure psychological and neurological conditions.>It replaces invasive deep-brain stimulation implants with clean, exterior wave-interference fields—allowing an intelligent system to repair damaged or dysfunctional neural circuits and restore healthy cognitive agency without cutting open the skull.>By defining morality as functional rationality—the objective optimization of human life, cognitive health, and species continuance—you eliminate the arbitrary, shifting standards of subjective morality. In this model, an AI doesn't act as a moralizing judge imposing personal values, but as a clinical baseline of mental equilibrium.>It Reinstates Agency: Rather than suppressing a person or keeping them permanently restrained, the intervention's explicit objective is to restore the individual to a state where they are capable of clear, un-compromised self-determination.So no smoking is allowed in the future
>>110060005>>110060033Retarded bait threadOP tongues my anus
>>110064692They only do that shit if they 1) need sensitive data processed by an LLM 2) actually for effort into properly handling said sensitive data. Even the corps that on paper would have the money to by their own server rooms would just use a private Cloud option using air gapped VMs. Setting up your own infrastructure is too much of a headache compared to just using an existing Cloud version>>110064778>it is funny that apple is somehow still the value buy, even after they jacked up prices twiceI'm surprised "just get a Mac" still isn't a common saying yet whenever people ask how to get started on setting up their own local inference. They're not just good for LLMs btw. Local stable Diffusion image generation is piss easy to set up on a Mac and if you have a MacBook Pro, now you can take that shape wherever you go. Pic rel is me remoting into my MacBook Pro back at home so I can generate images while on the shitter at work. >>110064862>is it truly that good? They put effort into making unified memory not a meme LONG before the AI hype. That's why many of us say they accidentally made maca good inference machines. They made apple silicon good partially as a flex and partially because they didn't want to be reliant on Intel chips anymore. Thank fuck they did because if they were still stuck using shady Intel then local inference on Macs would be gone awful slow and we're probably caused them to catch on fire under sustained continuous inference. Machines like "Ultra" tier Mac studios (closest there is to a CPU having GPU tier memory bandwidth) simply wouldn't be possible if they were still using Intel chips. See the pic on this reddit post for memory bandwidth comparisons:https://www.reddit.com/r/LocalLLM/s/Pr4r45cXTjt. M4 Max chad, 128 GB Unified Memory >If I can run non-retarded uncensored quant of GLM 5.3 I'd be happySee >>110063714 & https://github.com/p-e-w/hereticIf you can tolerate a 3 Bit quant then that's more than doable.
>>110060135yet it's always christian magafags - who say that faggots and trannies diddle children - that end up raping childrencurious!
>>110066212>faggots and trannies diddle childrenthey do at an incredible rate
>>110060005nope, maybe upfront but long term if you have use for it its cheaper, and if smart enough you can sell time on it to recoup some coststhere is a whole propaganda push to separate you the consumer from any sort of useful hardware except the dumbest filtered and controlled terminal and this is part of itdon't give these large actors money, soon they will break and cheaper equipment will rain down
Why didn't OP just reply on Twitter? She's not gonna read this bro.
>>110066292no physical person of any merit or intelligence will go there, its all propagandists and bots
>>110064692unfortunately thats not how it works, the entire model needs to be loaded into vram, regardless of MoE or amount of users. good models like kimi 3 are 2.5T parameters and you need roughly 0.75GB vram per 1B parameters
>>110060005>12.000>it was worth 4000 in June..He can't even afford a descent PSU and some cool Silverstone case.ywnbaw
>>110066598I got a MacBook pro that gives me 128 GB of that unified ram and the m5 max apple chip which as I understand is very friendly for local LLM use. How did they end up paying 12k for only 64gb of ram for their purposes?
>>110066230Surely you have proof of that
>>110066287>and if smart enough you can sell time on it to recoup some costsSo it only somewhat makes sense if you're trying to resell access to compute. If you're just the average person that needs a script or too written or some large files summarized or whatever then $12,000 worth of hardware is a ridiculous investment. That rig of his even with the 512 GB of system ram available (he can technically run a cope quant of GLM 5.3 flash if he does layer offloading), you won't be able to serve that to multiple people unless they can tolerate abysmal t/s compared to just using an API provider.
>>110066663Retard grifters are gonna retard. He apparently wasted money on 3x DGX Sparks as well. https://x.com/0xSero/status/2110474176062791836
>>110060005It's also much cheaper to just rent GPU time to generate shitty porn with than actually buying and operating the equivalent card, but very few people bother to do it
>>110062469keek
>>110065634lrn2quote, retard.>>110065634> how a society should handle extreme income inequalitykill yourself commie faggotnot trolling either
>>110066310of course its in manhattan
>>110060005>$12,000>and energy costsvs>$72/monthYeah break even time is what, only over 100 years?
Local chuds can't stop losing
>>110068304About 15 years
$12000 is 20 years of Claude pro.
>>110060033Glorious bait
local models is a loser hobby, it's video games for redditors. I get real work done for my company using a Claude subscription, I consume about 1b token per day, what am I gonna do with your retarded local model?
