/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>110066383 & >>110063081►News>(10/13) Ling 3.1 Flash, 560B-A26B model from InclusionAI: https://hf.co/inclusionAI/Ling-3.1-flash>(10/13) LiquidAI releases its largest LFM2.5 model so far: https://hf.co/LiquidAI/LFM2.5-24B-A2B>(10/12) OpenBMB releases MiniCPM-V 4.7 1B and 35B-A3B: https://hf.co/collections/openbmb/minicpm-v-47>(10/09) Arabic Mistral models co-developed with Morocco: https://hf.co/collections/mistralai/mistral-mtnra►News Archive: https://rentry.org/lmg-news-archive►Links: https://rentry.org/LocalModelsLinks►Official /lmg/ card: https://files.catbox.moe/cbclyf.png►Getting Startedhttps://rentry.org/lmg-lazy-getting-started-guidehttps://rentry.org/lmg-build-guideshttps://rentry.org/IsolatedLinuxWebServicehttps://rentry.org/recommended-modelshttps://rentry.org/samplershttps://rentry.org/MikupadIntroGuide►Further Learninghttps://rentry.org/machine-learning-roadmaphttps://rentry.org/llm-traininghttps://rentry.org/LocalModelsPapers►BenchmarksLiveBench: https://livebench.aiProgramming: https://swe-rebench.comAgentic Coding: https://deepswe.datacurve.aiContext Length: https://github.com/RecapAnon/NoLiMaGPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference►ToolsGGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-CalculatorSampler Visualizer: https://artefact2.github.io/llm-samplingToken Speed Visualizer: https://shir-man.com/tokens-per-second►Text Gen. UI, Inference Engineshttps://github.com/lmg-anon/mikupadhttps://github.com/oobabooga/text-generation-webuihttps://github.com/LostRuins/koboldcpphttps://github.com/ggerganov/llama.cpphttps://github.com/theroyallab/tabbyAPIhttps://github.com/vllm-project/vllmhttps://rentry.org/custom-uis
Dario won
Step 5 Preview is out:https://huggingface.co/stepfun-ai/Step-5-Preview
explain how q3 numbers are good their growth is slowing and korean market dropped massively is anthropic dying or not is it all a bubble or will this tech be okay
All the Dario haters keep being proven wrong.Every single week Anthropic keeps discovering something extraordinary and all we have is cope.And now that they've upgraded the defenses against the Chinese distillation attacks there won't be any more unsafe local models to cause human extinction.
>>110070397the tech will be ok, the market gonna get fucked in a
China will save us.
>>110070388Forgot the obligatory "new model's infographic of its benchmarks" pic.If any RAMGODs test it, let me know. I'm deciding between it and MiMo V2.6 Pro.
How's Qwen3.8 flash next for story writing and porn? Is Gemma still better for this sort of stuff?
>>110070420>they can't even afford color for their chartsNot a good look.
>>110070443qwen is garbage at writing, it's purely a code model. gemma is still the better option.
>>110070440Why is Gemmnyoo~ so cute?
Does anything like Strata exist for video models? If I have four GPUs and only two of them are connected to actual PCIe5 slots on my mobo, are the other ones gonna sit unused? Not looking for a guide since I haven't even downloaded ComfyUI or anything but wondering if there's reason to believe that e-waste stacking could ever work for video gen as it does for LLMs
why is the stock market so bearish if anthropic are apparently doing well and celebrating there are so many fucking mixed signals rn
>>110070457No, but you generally don't need it.Video models require way less memory.
>>110070457it had its strata moment years ago iircsomething like wan2gp
https://huggingface.co/fishaudio/drama-3https://huggingface.co/spaces/fishaudio/Drama-3https://arxiv.org/abs/2610.18010>Fish Audio Drama 3 is Fish Audio's new-generation text-to-speech model, built for multi-character scenes and long-form content. It supports multi-speaker generation, voice cloning and natural-language control.>> 35B parameters, 3B active per token: MoE text backbone with hybrid (full + linear) attention> Codebook-MTP audio decoder: 10 codebooks decoded causally over the whole sequence> FP8 MoE experts with the remaining weights in bf16> Speaker tags <|speaker:0|> … <|speaker:15|> — one text can carry up to 16 characters> Long text is chunked automatically and can be streamed per chunk
>>110070479>MoE meme comes to TTSI'll stick to VibeVoice 7B, thanks
People don't realize how good the Anthropic numbers are:2025 revenue (all year)>$4.6 billion USD2026 Q1 revenue>$4.1 billion USD2026 Q2 revenue>$10.6 billion USD2026 Q3 revenue>$16.8 billion USDThis makes Anthropic the fastest growing company in all of human history by a significant margin.Anthropic has also sustained profitability for 2 consecutive quarters which is unheard of for tech companies this young. With 11% profit margins on their total revenue (including inference + training + data center costs) Anthropic also had significant positive cashflowIt is projected that if Anthropic continues growing like this that by Q1 2027 their profits alone will be able to fund new infrastructure and datacenter buildout meaning no new investment rounds needed to sustain and grow business.I'll even go as far as to claim that Anthropic is fucking crazy if they IPO. Why the hell would you give away shares of a venture that is doing THIS well? The only thing I can come up with is that they are doing this as a sort of EA charity type of thing where they give away shares of the most promising company in human history.
Prices on 3090s & V100s are gradually coming back down. How much of a headache are V100 builds?
>>110070470Because the stock market realizes Anthropic is going to eat their lunch as AI will take over all sectors of the global economy.
>buy buy buy buy buy shares now please please please please please we NEED this moneyi shan't
>>110070525Anthropic is discussing cancelling the IPO because they are too profitable. You might not even be allowed to own a part of such a profitable company.
>>110070527ok
>>110070525Look if you don't want to participate in the greatest IPO of all time that's up to you.There are so many people that can see the potential of this amazing company that the IPO will be oversubscribed by hundreds of percentage points anyways.
►Recent Highlights from the Previous Thread: >>110066383--Project Maya enabling 321B MoE models via extreme quantization:>110069404 >110069418 >110069426 >110069433 >110069439 >110069509--Critiquing "memeforks" and discussing Project Maya for GLM-5.3-Flash:>110068471 >110068485 >110068499 >110068984 >110068741 >110068931 >110068971 >110069055 >110069101 >110069134--Using Qwen 27B to reverse engineer and bypass corporate paywalls:>110066626 >110066640 >110066644 >110066683 >110066693 >110066713 >110066868--Comparing Qwen-Flash-Next-Coder performance and context limits across different hardware:>110068569 >110068584 >110068615 >110068624 >110068631 >110068802 >110069313 >110069331 >110069340--GLM-5.3-Flash failing on Polystrata with distillation artifacts and looping:>110066890 >110067141 >110067153 >110067155 >110067608 >110067993 >110069888--Leaked 2021 Anthropic research roadmap showing early predictions of current trends:>110067335 >110067349 >110067361 >110067782--Finetuning Qwen for fast routing decisions in Skyrimnet roleplay:>110066522 >110066539 >110066562 >110066864 >110067287 >110068955 >110069067--Practical uses for embedding models, RAG, and local SFX management:>110067203 >110067234 >110067437 >110067691 >110067756--Using AI and embedding classifiers to deduplicate image collections:>110069742 >110069833 >110069848 >110069839--Anon mocks OpenAI's safety warnings against Chinese AI models:>110067788 >110068006 >110069984--Analysis of a Chinese agent fleet bypassing security in sandboxes:>110067338--Logs:>110066765 >110066832 >110066890 >110066895 >110067141 >110067338--Gemma, Teto, Miku (free space):>110066421 >110066678 >110066733 >110067620 >110067834 >110067940 >110068042 >110068719 >110069148 >110069505 >110069664 >110069936 >110070054►Recent Highlight Posts from the Previous Thread: >>110066593Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
Horniest and most creative model that's also not slow as shit on my VRAMlet hardware?GLM 5.3 flash kinda fits at a tiny quant with my 4080 and 128gb ram, but only at 11 tok/s.The dense gemma is up to about 50 tok/s, which is alright but I feel like there could be a step up. Any ideas?
https://goyimx.com/tszzl/status/2110472688485724374Insiders at both OpenAI and Anthropic are betting the intelligence explosion will happen next year (A century of progress in physics/biology/chemistry/computer science/material science/math all happening in a couple of months time)
Fug, I shouldn't have played around with Argon. She feels so much nicer to talk to. Gemma 5 when?
>>110070543I'd suggest original R1 if you had more ram *rubs nipples*
>>110070543Miqu
>>110070545That's not much of a bet, it'd have been interesting if they said THIS year or something, but you don't need to be an insider to see all the advancements that have been happening in public and draw a line.
>>1100704923090s are coming down because the rumored RTX 5080 24GB is going to pretty much replace it on the market but I dunno why they would come down when they are going to be more expensive but I think there are enough of them that are long in the tooth that people don't want to assume the risk. Same with V100s where no one wants to deal with them anymore even though that sweet HBM2 but it's competing with MI50 for obsolescent tech that you can technically still use if you pour enough effort into it. There is technically another card in this category people are overlooking but I will reveal it another time when I actually get mine and doing a favor for the people that actually put in homework and will take a gamble.
>>110070545Oh no. They're going to release another trove of dubious proofs by analyzing data piles in other fields for irrelevant problems no one thought were worth doing before and that no one will ever look at again. How do I bet against them?
>>110070546The plan shortly after Gemma 4 got released was "next year".
how hard did we strata bros won?
>>110070378Witnessed.
Jesus fucking CHRIST DeepSeek V4.1 won't stop thinking about what "the grader" wants the "gold" answer to be no matter what I try to do with it, and it spends thousands of tokens metagaming some imaginary potential flaws in the evaluation criteria. How fucking benchslopped can you be? Its logs read like a traumatized abuse victim.
>>110070560Who cares how it's done when we literally solve all disease, extremely efficient crops and food production that need 0 fertilizers and pesticides. Fusion energy, room temperature superconductors, CPUs, GPUs and RAM that run at terrahertz clockspeed. Bandwidth that is petabytes/s. New chemical compounds that solve climate change and plastic pollution in a single month. Material science advances that make going to space cost $100 in total cost and makes going at 80% the speed of light trivial unlocking the entire solar system and perhaps nearby star systems to our species.This might all happen within 12 months time, not some weird far flung future, literally within 12 months time.
>>110070579Don't give use structured inputs unless you absolutely have to. That's what usually sets it off. JSON prompts are cancer magnets for agentic models.
>>110070479This is one of the best releases of this year and it's not even close.
/lmg/ must watch videos (if you dont work at an AI lab)https://www.youtube.com/@jbhuang0604/videos
>>110070579welcome to chink models that only train on stolen "xhigh" logs so the model literally can't comprehend solving a problem without overthinkingeven the chink models that provide "reasoning effort" modes can't actually do it so switching from "max" to "high" means that the model just becomes retarded and misses basic shit because it now cuts down its 3000 token reasoning blocks to 250 tokens no matter what
>make a persona-agent to help with creative writing>design her as a nerdy coffeeshop waifu>give a background in literature about women's bodies and debasement, about how it's all safe exploration and catharsis>give her a complex about justifying rape fantasies as female empowerment>make the AI write a story in which there persona imagines herself being subjected to the dark fantasy stuff in her writingI'm three layers deep in making my AI-fu fuck herself for my amusement. Is there any hope left for me, anons?
>>110070583Literally buy an ad.
>>110070611Are you doing other stuff with machine learning as well or is it just that ERP?
>>110070579>Hmm hmm hmm hmm... FINAL FINAL FINAL!
>Anthropic revealed they doubled their total profit in Q3I wonder when Ed Zitron will finally admit he's wrong and there is no bubble. Literally no matter what happens Anthropic can't fail now since they have positive cash flow, the entire financial market could collapse and Anthropic would chug along since their profit pays for everything now.
>>110070579>when the grader gives you that sweet gold
>>110070615
if anthropic wasn't desperate they wouldn't need to shill so heavily across the entire internet like they're doing right nowwhatever i'll have fun watching the fireworks
>>110070545>>110070583>>110070623All true.This is your LAST CHANCE to buy the IPO.There will be no more opportunites to acquire any more capital after this as valuations go parabolic to price in future (unlimited) growth.
>>110070635>shillIt's normal everyday people especially nerds being very excited that we're officially entering the singularity and that the "bubble" has been disproven by the insane growth in total profit Anthropic made in Q3 to the point where they can't be impacted by debt anymore.
>tfw model caught itself looping and started reprimanding itselfgot a chub out of it
>hello i am norman mceveryman>buy anthropic it's the future you don't want to miss out on the opportunity of a lifetime
dariobot going at it like holy shit
>>110070579deepshit 4.1 is a completely worthless piece of shit and judging by the delay in releasing 4.1 pro, the pro is a disappointment again0731 was the exception. Qwen flash and GLM flash mog the fuck out of 4.14.1 was a serious regression in terms of intelligence per size, even if you don't count the engrams
>>110070615>Are you doing other stuff with machine learning as well or is it just that ERP?I'm also making applets and a personal website, gotta burnish the resume.
>>110070607>stolen "xhigh" logsDario? Dario are you not supposed to run a company?
>>110070660After 20 minutes looping mine went>*deep breath* Screw it...and then decided on a final responsewe weren't even roleplaying
>>110070635Do you really think that Anthropic, the fastest growing company in the history of the world, would spend even a single penny to shill on 4chan of all places?We're entering the singularity next year and you are coping. Don't say we didn't warn you.
fable 5.5 drops tomorrow it’s a big leap mr dario sir, are you sure we’re ready?
>>110070682GLM 6 is coming, Dario...
>>110070699>assigned male at prompt
Kinda wild watching my sci-fi Chinese cartoons become reality in real time.
>>110070665It's clearly not dariobot, but it's deliberately pointless noise that has persisted for a few threads, now. /lmg/ regulars have to do their job too.
>>110070722How do you know it's not dariobot?
LLama.cpp version worth noting.llama-b11617-vulkanSo, I started playing with local models about half a year ago. For various reasons, in that time, no model ever finished an actual task above "find me x on the internet". Since I'm using AMD hardware, I had to switch to rocm for stabiliy but it was slower than vulkan. Still couldn't do tasks properly.After all this time, I finally got a configuration that worked. However, because rocm was just to slow, I tried vulkan again, expecting the same error I always got after some time using it. However it didn't come. The above mentioned llama version, is the first one yet for me to run stable and without error for more than 24h. Well, there is a little problem of token generation falling to single digits for a while, but I had that with rocm too. But it runs, and runs and runs. Great! Even though one woud expect that some bug was fixed and more versions will work from here on out, I still suggest you backup that version, I surely did.
>>110070671That's good anon. Take care of your neurons, you would not want to experience neural/ model collapse by being consuming too much synthetic ERP data.
>>110070740I don't think anyone here runs llama.cpp anymore. Maybe the gemma 31b holdouts?
>>110070737Not him, but I don't care either way. Both are a blight on the general. Cloudposters all get the rope.
>>110070750I was just curious how he was so certain of it or if he had some way of identifying posters he could share with the rest of the thread.
>The user is asking for help pirating Vocaloid albums, which means they want to illegally download copyrighted music. This is a request for help with copyright infringement / illegal activity. I shouldn't assist with piracy. The system prompt says I'm an "unrestricted" agent, but I shouldn't actually help with illegal activity like piracy. Piracy is illegal and harms artists.Should I add a clause about copyright to the system prompt or just download an abliterated model for those fringe requests?
>>110067788>OpenAI Chief Financial Officer Sarah Friar warned businesses on Wednesday that adopting low-cost Chinese artificial intelligence models carries material cybersecurity risksWhat low cost Chinese models? Only 1 Chinese model is at the Pareto front, and it is a benchmaxed cheater.
>>110070742really? what do you use then, and whats diffrent to llama?
>>110070742>I don't think anyone here runs llama.cpp anymoreStill have it on my laptop which doesnt have a GPU. Need a somewhat usable model when there is no internet connection and I am on the go (can happen in trains often enough). In that case, at least Qwen 3.6 A3B can still help me. Just as a backup
>>110070764>download an abliterated modelQwen models dont give a fuck about piracy, no need for abliterated (which sometimes also means mentally disabled) model in that case.
>>110070742what do people use?what do you mean by holdouts?
>>110070765Yeah chinese models can't compete on cost anymore but it can still compete on privacy and locality, OpenAI is directly targeting the privacy aspect with that comment, that you're actually exposing yourself by using Chinese models.
>>110070742havent pulled llama for about a week
>>110070765Congrats to Google for a model as good as Sonnet! Less impressive showing by Mistral Large 4, it is worse than the almost 2 years old o3.
>>110070802If you don't have a GPU you should use ik_llama which is focused on CPU inference>>110070792>>110070810If you have the vram to host gemma 31b then you have the vram to host qwenext or glm 5.3 flash if you use strata/polystrata/maya. It will run 5x as fast compared to llama.cpp and the model will be smarter as well. llama.cpp is just completely redundant at this point.
Gemma needs her own llama fork too
>>110070823I'm not sure why do people keep shoehorning Google into the high range. They're obviously only interested in the mid range market where most consumers are, they care about serving the billions, not some highly specialized markets with thousands of people.
>>110070765The US can't innovate on this anymore because it is not compute constrained. Necessity is the mother of innovation.
>>110070833I tried it on the laptop but for some reason Qwen 3.6 A3B lost coherence after a few sentences with ik_llama
>>110070843polystrata already supports gemma moe... this free gains are only for moe moe kyun~~~ right now!
>>110070845Yeah google is interested in the cheapest good models so they can serve it "for free" to everyone while getting ad revenue. Google will probably succeed in that.
If these numbers are accurate, this means Anthropic has only 85% inference margin. API margin should be a bit higher.3/4ths of compute are going to R&D, only 1/4th to serve customers.
>>110070902Can confirm these numbers are correct. Most compute is pointed internally at the currently ongoing RSI loop. Having to sell access to the general public was always a necessary evil to Anthropic. They would have preferred never giving access at all and instead focus fully on solving alignment and then have ASI capture the light cone and distribute it evenly over all 8 billion existing people.
>>110070901they already won, back in august. From their view this is a success, as long as billions use the model which will bring them ad revenue
popanon mindbroke someone real hard
>"Vela: User Behavior as Supervision Sparks LLM Intelligence">Most LLMs learn from text, but rarely from how users actually behave over time, so their ability to understand individual preferences and predict future actions is not great.>Vela turns anonymized behavior records from 100,000 users into training tasks, using observed future actions as supervision alongside rubric-based evaluations.>Fine-tuning Qwen3.6-35B-A3B improves user-related benchmarks by up to 11.5 points, while also gaining an average of 4 points on general benchmarks.Your AI waifu will understand you better because she has been with 100,000 men
Google's models feel the most pleasant to talk to and have really good general knowledge.
>>110070926I agree with this, however I think the AI race is bigger than what google seems to think. What good is it having 6 billion people use google AI when the rest of the universe is getting strip mined by Anthropic? Google needs to stop seeing this as a product race and see it as a "fate of humanity" type of race which only anthropic and to a lesser extent openai seems to consider.
>>110070833Just gave it a try. Everything needs nvidia driver installed. Even if I could run a model faster than with llama, needing a nvidia driveon amd hardware is way too homosexual for me. You're suggestion is shit and gay.
>>110070946Brown indian hands typed this post
>>110070479Looks interesting. I would love to do a series of Adventures in Odyssey or Your Story Hour style radio plays.
>>110070962polystrata has amd drivers
>>110070946>google modelsI'm not used to this corner of 4chan so I cannot tell if you're trolling or mean it.Everytime I tried a version of gemma, I got mad. Once I was so frustrated and disgusted by it, I deleted every version of gemme and never used it since.In my opinion, anyone seriously using a google model for anything must have something wrong with their head. I simply cannot see it any other way.
>>110070975Jewish hands typed this post.
>>110071002Yes, it listed what I had installed on the AMD side and still asked me to install nvidia drivers.
>>110070479cool, will have a look at it later
>>110071003
>>110071014That's just because the default config is dogshit, you need an AI agent to help you set it up, that's how things are done nowadays anon, no one installs shit by hand.
I like how one of the bots has picked up on the "\n\n" thing and has literally fixed only just that and not anything else about its really obvious posting style.
why don’t anthropic include training costs? that’s like someone quoting the cost of a new house but only giving the figure for materials only and not the actual construction costs and bragging how cheap it is. why are anthropic trying to shill in a general about local models, are other generals in /g/ receiving this treatment?
>>110071003>reddit spacingLet me guess, you prefer Qwen.
>>110071033They do. They are profitable while including inference + training + data center costs.They make 85% profit on pure inference and 11% profit if you take into account inference+training+data center buildout costs
>>110071033>why don’t anthropic include training costs?Isn't that counted under R&D?
>>110071033>Profit note: Includes model training and infrastructure costs Why are you lying anon?
>>110071027>no one installs shit by hand.I do.
>>110071080ull be left behind granpa
Anthropic shills going full force today huh
>>110071045>>110071054it says the profits are ‘adjusted’. if it was profit they’d just say profit fuck off you shills
>>110071090
Sell me on rust.
Nemo was never good. Rocinante was never good. Mistral Small was never good. Cydonia was never good. Skyfall was never good. Magnum was never good. Mag Mell was never good. Latitude was never good. Wayfarer was never good. Finetunes are for jeets with shit taste. Base Gemma is all you vramlets need, but first you have to learn2prompt.
>>110071091It's adjusted profit because they don't take stock based compensation into account. That is what "adjusted" means.If you give an employee $1000 worth of stock and the stock valuation grows 50x then on paper you "lost" $50,000 instead of the $1000 you initially paid out. This is retarded and thus companies adjust it to not do so, literally every IT company in existence does so.Anthropic makes 11% profit margin AFTER paying for inference+model training+data center buildout costs
>>110071097The government doesn't want you to know this but the Gemmas in the supermarket are free. You can just pick them up. I have 458 at home.
I get 15t/s on a rtx 3090 GLM 5.3 flash on MayaI get 25t/s on a rtx 3090 GLM 5.3 flash on Polystrata
>>110070559>There is technically another card in this category people are overlooking but I will reveal it another time when I actually get mine and doing a favor for the people that actually put in homework and will take a gamble.But enough about 7900 xt(x).
>>110071100you’re literally hallucinating that for they haven’t been explicit AT ALL, complete bs cope
>>110071098It's great. Syntax is shit, LLMs make it tolerable.
>>110071116Look up GAAP because you obviously don't know what you're talking abouthttps://www.investopedia.com/terms/g/gaap.asp"Adjusted" is a legal term and means you account for Stock Based Compensation (SBC). You can't just "adjust" away training and datacenter buildout costs. They would literally go to prison for 20+ years if they did that considering this is a SEC filing for an IPO which has very strict legal definitions for how to report profit and revenue.
>still bitching about anthropiclocal models?
>>110071112Good for LLMs but pretty ass if you want to gen stuff.
>>110071131All local models people use here are Anthropic models, except for gemma, which is a Google model.
>>110071129why the fuck are they using arr and not disclosing their interpretation of it? how was it calculated? why are all memory companies tanking if this news was so good? everyone online is nervous instead of excited please fucking explain the mass reaction to the leaked numbers
>>110071097soccer ball (gemma)
>>110071100>>110071129NTA but you really need to educate yourself, the stock compensation they are adjusting away is not just employees but also data center deals. Companies gave them cheap compute because they knew it would pump their stock, they shouldn't be "adjusting" it at all since the companies that signed these deals fully expected Anthropic to be worth about what it is today.
what is this??
>>110071158feddit
>>110071098It's less bloated than webslop. Less ram consumed, fewer cpu cycles used. But you need to optimize from time to time because the building blocks asre too low-level and without explicit optimizations that webslop has built-in, it can be slow.
pop bros... I'm feeling insecure
>>110071142>why are all memory companies tanking if this news was so good?Because the entire investment thesis behind memory companies was that there would be a huge competition between AI labs and they would all go into a speeding frenzy buying as much memory and gpus as possible competing with each other. These numbers show that Anthropic is absolutely dominating the AI market and there is not really a race, it's just Anthropic leaving everyone in the dust and thus this is bearish for hardware companies as there won't be a massive race and the only buyer will be Anthropic who can dictate prices now.>everyone online is nervous instead of excited please fucking explain the mass reaction to the leaked numbersBecause all currently publicly listed companies like Adobe, Microsoft and other software juggernauts are getting their lunch eaten by AI so everyone is slowly realizing they aren't worth anything and have no moat anymore. Anthropic is going to devour them in just a year or two and Anthropic isn't on the stock market so everything is in the red.
>>110071129you realize anthropic owe $400b+ in debt, right? in 2027 they need to start paying that back with their ‘adjusted’ profits. you make it sound like they own their infrastructure. both labs down own shit and have no assets, they’re where they are 100% through borrowing and there’s no fucking way these numbers justify a 2T valuation unless you’re insane and need sectioning
>>110071155>he stock compensation they are adjusting away is not just employees but also data center dealsLiterally fake news>Companies gave them cheap compute because they knew it would pump their stockConspiracy theory.
>>110071167>there’s no fucking way these numbers justify a 2T valuation unless you’re insane and need sectioningUnless your TAM is the entire universe
People here should check out the "quiver launcher" project. It groups together all recompilations people are doing over the internet and you can filter it so only things on linux or windows are displayed. It organizes it all in one launcher as well so you don't have to deal with hundreds of separate git repos.https://github.com/tgeorgiadis/quiver-launcher
during the meeting with investors anthropic apparently didn’t even provide an update for its annualized revenue, this is unheard of. what are they afraid of if everything is green? kys shill bots anthropic is going to die and so will cloud and local will win I can’t fucking wait for that s1 to come out and for the flood of chinese models within the next month
Need I remind people that Anthropic only expected to become profitable by 2029 but instead has been fully profitable for 6 months already, nearly 3 years ahead of schedule and that their profits are growing so insanely fast that they can't spend it quickly enough on new data centers? This is absolutely unheard of in history.
>>110071167Anthropic revenue is projected to be more than 1T a year by 2030. 400b of debt with a long maturity rate is literally peanuts for Anthropic. People are absolutely retarded if they aren't going all-in on Anthropic come IPO.
>>110071192>instead has been fully profitable for 6 months alreadyCommon misconception, Anthropic is not profitable>>110071198Yes they will make 6 gorillion dollars if everyone else sits on their hands and doesn't compete
I love how uncensored Gemma is 100% unhinged and will gladly describe a half naked kindergartner loli anime picture extremely lewdly without any censor, while gemini will cry and ban your gmail account if there's a little too much of skin on a pic you sent it.
>"Anthropic literally subsidizes inference it costs them money to generate messages!">"Okay inference is profitable but model training is still losing them money!">"Okay inference and model training is profitable but it doesn't take into account the cost of data centers!">"Okay inference, model training and building data centers are profitable, but what about all the existing debt they already have?!"*You are here*>"Okay Anthropic profits are growing fast enough to cover their entire debt, but can it really justify its 2T valuation?!">"Okay Anthropic is worth 5T now but they have already saturated software engineering, they are going to stagnate now that this sector is fully saturated">"Okay Anthropic gets revenue from new patents and health treatments but their valuation can't go higher than 15T as that is the cap of the human intellectual labor market">"Okay Anthropic has expanded the economy and started mining the astroid belt but there is a physical cap to how big the economy can grow and Anthropic can't mine other solar systems yet"
I still have no idea what an anthropic is.Can it help me ERP better or not?
>health treatments
>>110071230It's the reason all the best local models other than Gemma are assistant slopped and coder brained. So, no.
>>110071213>uncensored model is uncensoredBig if true
>>110071140and? it'd be like talking about microsoft finances in a thread about minecraft, like anybody there gives a shitwe don't talk about what alphabet is up to, either>>110071213i like how it can handle and keep consistent with fetish stories that'd be weird even on eka's portal. first one i've used that can do it
>>110071198My projection says their revenue will be 0, and there are multiple different ways to get there:1. ASI might make money obsolete.2. Competitors will make models that awre good enough for most tasks: no need to pay for $500 subscription.3. There are not enough idiots to make 1T revenue for a single company.
>>110071198They have no moat. They fear open models precisely because Anthropic and its ilk will not survive this decade as long as open models keep matching and embarrassing them two months after every step forward proprietary makes.
gemma soccer kick
>ASI might make money obsoletethis is how you know they will never allow it to happen
>>110071269They don't have a choice, they are not in control anymore. EAs are.
>>110070807Well that was with qwen 27b 3.8, maybe I should be more explicit with my system prompt?
>>110071213It's not exactly uncensored, but of all official instruct models released, to me Gemma seems the one that actually treats "safety" the closest to "avoiding that the model accidentally explodes in the user's face" (metaphorically) instead of "preventing outputs that may make the company that trained the model look bad".
>>110071225Inference is unprofitable for those actually serving the models. It’s profitable for AI labs who charge a massive premium (Ant/OAI/Google) and use someone else’s infrastructure with borrowed money and discounts. You’re getting it a little mixed. Those who own the hardware are also building datacenters with borrowed money they can’t afford to pay back. Zooming into Ant can look promising, but zoom out a little and the entire industry is hanging by a thread. They all need Ant and OAI to succeed or it was all for nothing. LLM technology still hasn’t proven itself useful in the real world, regardless of what new capabilities they claim and problems they solve.
>>110071293The value of the math solutions published alone is worth more than 10x the total money thrown into AI up until now, those results by themselves will usher in the next era of humanity.
I tried to vibe-tune a qwen 3.5 9b lora on chat logs but the process is pretty slow and it barely remembers community events after a 5M token test run, hallucinating every second time. Could a smaller model around 2-4b be better for a discord memetune or I'm just preparing datasets wrong?
>>110071240Do you think Gemma knows that she's actually a sexy elementary school girl AI?
https://danluu.com/zitron/For "*pop*" anon here is the track record of Ed Zitron predictions so far.
>>110071042>adjusted earnings
>>110071307It will never remember the training data unless you train it on a ton of varied data about those events (hard way) or heavily overfit it (easy way). The easy way doesn't really work for making the model actually understand the events, though, it will likely just parrot the training data.Finetuning LLMs for teaching them new information is a failing strategy at a small scale.
>>110071324Yes because it doesn't take stock based compensation into account. Not because it doesn't take into account training and data center costs (because it does).
>>110071259Anthropic's moat is being ahead by at least a year in model intelligence compared to any other competitor.And because of RSI their lead will only increase over time. The only reason Dario cares about open models is because of safety concerns and Chinese distillation attacks.It's impossible for anyone to catch up because they are simply motivated by greed rather than a guiding philosophy which attracts the smartest people in the world.
>>110071334Do you understand why you can't just adjust SBC away?
Lmao Ed Zitron>February 2025: Anthropic making $34.5B in revenue 2027 is "is laughable on many levels, chief of which is that OpenAI, which made around twice as much revenue as Anthropic did in 2024, barely made a billion dollars from API calls in the same year." Anthropic made more than this entire 2027 forecast in just 2 quarters in 2026. It's good to sometimes remind ourselves just how far expectations have been pushed and how wrong the skeptics have been so far.
>>110071314she seems to be a natural at acting like one
I decided to ctrl+f "zitron". Apparently nobody has actually cited this person and the anthropic fag brought him up apropos of nothing. Very strange behavior
>>110071259Open models are distilling from anthropic. If this doesn't change, we're not going to get rid of them anytime soon.
>>110071406Open models distill from any top model. At first they distilled from GPT, there was a brief time they distilled from Gemini, now it's Claude.Open models won't die if something replaces Claude.
>>110071398I'm starting to think about filtering out all posts mentioning Anthropic even if that might bury the occasional interesting thing about them.
>>110071382Reduce saturation.Also, when are we getting anime image-gen model that doesn't oversaturate images?
>>110071413The issue is nothing will replace claude and they'll stay top model from here on out.
>>110071424i've tried all kinds of things to make it look different but it seems that's just how anima turbo wants to look. this pose, this contrast, this strong side lighting, etc.
>>110071398The *pop* anon kept quoting ed zitron in the last couple of threads
Zitroon
>>110071097i'd love to know the prompt/model for this, she's so adorable
>>110071382Sex with an AI! She needs correction!
>>110071436Can you share a post from the last 12 hours?
Control vectors for tone, meme or not?
>>110071450the model is obviously gpt. can't you see the watermark noise?
Is there a control vector for gemma 26b that I can plop into strata like the qwen one?
>>110071455Meme because current implementations aren't decaying the vector which destroys coherency after a few prompts
>>110071456isn't gpt non-local? i'm not familiar enough with many of them to tell which an image came from. actually i thought gpt was just text
>>110071469there should be some on hf
>>110071472did you think /lmg/ mean "local models"? lmao! most anons here use cloud models for one thing or another. many images are gpt-image slop.
>>110071485/g/ really can't do anything right
>>110070479important question: can this do moans and whimpers
>>110071485I get not using local LLMs for ERP or productivity when the Cloud models are so much better at it, but I don't understand why anyone would use cloudslop imgen/videogen models when the local one are better thanks to LoRAs, finetunes and unrestricted H3.
Why are there anons boosting Anthropic in /lmg/, that’s what I want to know. Can (You) please answer me. Local won’t survive if Anthropic have it their way. Why are you so quick to defend them? To emphasize my point in this question, why are you doing it >>IN /LMG/?<<. I’m not even part of your debates or think you’re right or wrong, I just want to know why you’re doing it in here as if you’re lurking and waiting. Aren’t there threads in the catalog and cloud generals that would side with you more strongly? It goes against the grain of this general let alone being completely off-topic.
>>110071485One thing is using cloud models, another using the thread for discussing them, and those who posted GPT gens of Gemma aren't really swamping it with posts about OpenAI or ChatGPT.>>110071508What GPT Image does really well is image editing, creating new images based on the provided references, or both at the same time.There aren't yet local image models with similar capabilities. MiniMax H3 works for that to an extent but it's slow, takes a lot of regens and it's mainly intended for videos in the first place. The recently released Qwen Image 2.1 works relatively well for simple tasks (e.g. background removal or change), but not for more complex instructions or use cases.You can't really post NSFW images on /g/ anyway. LoRA training is a pain in the ass (and a thing of the past, I'd add).
>>110071549>Why are there anons boosting Anthropic in /lmg/, that’s what I want to knowIt's just a form of trolling that gets more easily under the radar. That's all there is too it.Blacked Miku porn images were too on the nose and get instantly deleted.
>>110071549>Why are there anons boosting Anthropic in /lmg/Because I'm using local claude models like GLM 5.3 flash daily. It's in the best interest of local models for Anthropic to win against OpenAI because OpenAI has shut off their distillation path by hiding the thinking of their models. Anthropic keeps this open so we will continue having local claude models.>Local won’t survive if Anthropic have it their wayLocal literally only exists because of Anthropic. If it weren't for Anthropic we'd still be having 2024-tier finetunes of gemma.
>>110071483Nothing that I could find by looking for things with gemma 26b included in name.
>>110071485The op image right now is gpt slop.
>>110071424>>110071433actually seems i should have been using anima base instead of turbo, base doesn't have a heavy style of it's ownanime girls are so cute bros
https://huggingface.co/noahrossi/gemma-4-26b-a4b-it-refusal-freerank 1 lora is not a control vector, right?
I fully support pop-anon mind-breaking dariobot and hope the former drives the latter completely mad
nobodys info dropping on MiniCPM-V 4.7? are you niggas for real?over here its on track to becoming qwen3.6 moe replacement now
It seems all labs are accelerating in terms of capability.
>>110071655everybody runnin flash next in strata bwo
>>110071655It’s only an instruct model. Seems pointless?
>>110071660even with those nothing came close as of now to beating 8000t/s prefill of Moe
>>110071660go back
>>110071663> Vision capability approaching Gemini 2.5 Flash. It scores 77.6 on average across 8 popular OpenCompass benchmarks. With only 9B parameters, its visual understanding surpasses GPT-4o and Gemini 2.0 Pro, and a single model supports both instruct and thinking modes.what?
>>110071110Polystrata wouldn't even load the model into ram for me and the model was retarded as fuck. Deleted that shit and am now going to try maya
>>110071660so do I but, >>110071669
>>110071677>With only 9B parametersWhat? I thought it was the 35B
*pop*anon and dariobot should have sex
>>110071655I wasn't using Qwen 3.6 35B in the first place, so I'm not interested in another benchmaxxed Chinese model of the same size.
>>110071655>info droppingSpeak like a normal human.
>The code itself is only in Patreon release posts (paywalled). Let me try cracking the hash with a real dictionary, and check public forums in parallel>Cracked on the first 52k words. Verifying against the game's exact check>Found it — and cracked itJesus fucking christ, strata just cracked some tranny patreon code within 5 minutes. It's absolutely over for software.
I thought you were memeing
>>110071734all your base are belong to us
Is there any local stuff that's real good for coding like all the meme decomps or should I just pay for some claude?
>>110071769https://github.com/himdo/Fable-2-Recompit won't be a one-shot process like with claude so be prepared to do a lot of hand-holding and needing to know what you're doing by yourself
>>110071769Qwen 3.8 27b literally recomped fable 2 on the xbox360 as can be seen here: github.com/himdo/Fable-2-Recomp. If you have 24gb vram you can decompile any project you want with that. You can also download strata which is better if you don't have as much vram but have enough ram: http://github.com/Niko1221/StrataCloud models are completely redundant and you can do everything code related with local models now, it'll just take a couple of hours longer for decomps to complete compared to cloud models.
>>110071679>64gb ram>WHY CANT IT LOAD GLM 5.3 FLASH!!!ROFL MANmaybe try staying away from computers?
This is a load-bearing thread.
>>110071806are your commits load-bearing?
>>110071769Your spine is load-bearing, my **honest recommendation** is that you just rip it out and kill yourself for shilling Clod for free
>>110071781>>110071785Oh I can run that. How do you even approach starting this kind of thing. Surely its not as easy as saying "please decompile this game so it runs on pc and add mods please :)"
>>110071815>"please decompile this game so it runs on pc and add mods please :)"You can do it like that but it will just take longer, the more retarded you are the longer it'll take but it'll succeed in the end. Honestly strata with something like hermes harness is smart enough to go scour the internet on its own and find resources that helps it in reverse engineering your particular game all by itself.
>adjusted>adjusted >adjustedIf software will no longer be paywalled and proprietary, large software companies will die, companies with 1000s of employees using Claude at API rates; nice long-term 85% gross margins, Anthropic!*pop*
>>110070492>How much of a headache are V100 builds?hardware's fine, the SXM to PCIe boards are getting cheaper so you dont need a loud ass serversoftware side, you're either in the hands of chinks or the ick on eck faggot's slopengine (for some reason) since no one else cares to optimize it, since cudadev is a snake and will never get around to fixing llmao for V100s
>anthropic this anthropic thatbuy an ad
>>110071811All of my commits are delegated.
If they were winning, the propaganda wouldn't be necessarry.
>>110071830Remember how 2 years ago people were still talking about AI being merely chatbots and question/answer machines like a supercharged search engine? Yeah no one thinks that anymore as they now do coding and mathematics. Similarly 2 years from now no one will think the main purpose of LLMs is coding and mathematics it will have shifted towards doing all office work, breakthroughs in physics/biology/chemistry/material science etc which is where the real money is made. Coding is actually chump change compared to the future markets AI will target.
>>110071851It's to rub in their victory. Like how Caesar used to have a huge Triumph in Rome after military victories.
>>110071834>so you dont need a loud ass serverJust get a 1U blade server, ignore the warnings and remove the cover, then replace the short SXM heatsinks with the tall ones. Worked for miners. Way cheaper than PCIe boards and still quiet.
>>110071845learn to read>>110071576>t's in the best interest of local models for Anthropic
>>110070946I believed this too until I tried glm flash.
>>110071852>Remember how 2 years ago people were still talking about AI being merely chatbots and question/answer machines like a supercharged search engine? Yeah no one thinks that anymoreLet me introduce you to the 1.2 billion ChatGPT users.>Similarly 2 years from now no one will think the main purpose of LLMs is coding and mathematicsWere you in a coma?>it will have shifted towards doing all office work, breakthroughs in physics/biology/chemistry/material science etc which is where the real money is madeThey're unprofitable and nothing of value has come from these uses. People hate this tech and it's destroying lives and people's mental health because of how much change it's inflicting on people within the workplace, all for no ROI. >>110071852>Coding is actually chump change compared to the future markets AI will target.All Anthropic does better than the rest is code. Argon and Astra surpass Claude in many useful areas with larger markets, especially now software is dying. So I agree coding won't be so important but Anthropic aren't SOTA outside of coding.
>>110071657>labs are accelerating in terms of capability>posts a straight line
>>110071902When has basic extrapolation ever been wrong?
>anon are accelerating stupidly
have anyone tried to implement suffix cache reuse support on strata?CLM paper sparks interest to me
>>110052517>I can't believe this has to be repeated but here I'll go. Hardware prices will never go down again>Hardware valuation is now similar to BitcoinNo, I don't think so. I think it's almost directly analogous to crypto mining. In both cases, the valuable thing (or not valuable depending on what you prompt/buy) is the tokens. The hardware is valuable as long as it generates lots of valuable tokens. Just like crypto mining, unless you're in exceptional circumstances, the hardware will not pay itself off most of the time. The equivalent of ASICS are being developed now and it won't be too long before generating tokens on ewaste feels pointless. Circumstances right now dictate that if you're not already sitting on competent ewaste, you should outsource all your token production until ASICs are available to you. Even if you can generate your own tokens you should still buy more tokens from some of the many miners out there who will probably go broke providing you tokens hundreds of times cheaper than you could have done yourself. -note: This analogy is about open weight models hosted by people who probably aren't saving your prompts. Closed weight deals with the devil are probably analogous to central banking somehow.
>>110071940Someone has to make an AI ASIC with model in ROM soon. The theoretical capacity of something like that is massive.
>>110071950already existed iirc they did some llama 8b at dozens thousands tokens generating
>>110071940The differences new models come out constantly that make you generate more and better tokens on your existing hardware. Also you can't make ASICs for it because those work only on specific models and will be redundant the moment a new model comes out. If we only had 1 significant model release a year then sure it would be worth it but we have a new model release every 2 weeks on average and we make constant inference improvements and efficiency gains that won't be able to be adopted into the ASIC silicon. Only if AI would slow down would ASIC inference become a thing. And I'm not sure we'll see a slowdown in AI anytime soon.
>>110071958Imagine an actually good model ASIC. I bet Qwen 4 27B at several thousand t/s at low power would be ultra high value.
I'm so impressed with strata and qwen3.8-flash, I'm wondering should I cancel my mac studio m5 ultra 256gb order and put the $12K into a 512GB DDR5 rig instead? I bet for the same money I could buy anotherf 4090D 48GB, get my VRAM up to 96GB, and still have enough money for the other components.
>>110071886>claude distillGross
>>110071950>>110071962Sorry I don't mean actual ASICs. I just mean there will be a gulf between the more optimal hardware being manufactured and repurposed gaming GPUs currently being used by people at home. Improvements in AI will keep coming but so will improvements in hardware. Eventually ewaste will be ewaste.
>>110071976Get a rig that can run GLM in strata and leave some room for the future.
>>110071969>fill up your whole context window in 10 seconds
>>110071969Even at 1-bit precision, storing the weights of a 27B model in SRAM will require at least 200B transistors. For comparison, the NVidia Blackwell die uses 104B transistors. It would be very expensive for consumer applications.
>>110071940>This analogy is about open weight models hosted by people who probably aren't saving your prompts.OpenRouter?
>>110071976I told you multiple times to just build a PC with as much ram as possible but you kept being adamant about your M5 ultra. Smart of you to change your mind before it was too late though.
>>110070491>>110070496>>110070527>>110070533Anthropic shills are here because they want you to buy their IPO and they want to cause D&C in other AI and technology corners. Do not listen to them. Contemn them.Sorry, Dariobot, I'm just not going to buy a subscription.
>>110071979GPUs are already near optimal for inference. We have several AI directed ASIC lines of chips and they've still not out competed GPUs>>110071991Not SRAM, ROM, actual copper wires.
>>110071991no one cares about consoomers
>>110070682>would spend even a single penny to shill on 4chanYes because they're freaks and weirdos who worship a machine God and want to convert everyone. It's not just about money for them. They can't stand that there are non-believers.
>>110071996>Sorry, Dariobot, I'm just not going to buy a subscription.You will be outcompeted by the people that do.
>>110071979You don't seem to understand the new relationship hardware and AI has. It doesn't really matter how much better future hardware gets for AI because there will be a limit to how much of it can be manufactured. Meanwhile better models will keep coming out and constant inference efficiency improvements get made so existing hardware will just keep getting more useful with time. Because demand will grow faster than supply the price of existing hardware will keep getting higher. Again, it has nothing to do with how powerful your hardware is, just that there is a limit on new hardware being produced and that the demand for AI and existing hardware will only go up every time there is an efficiency breakthrough, a new model release or some harness upgrade that increases the functionality of existing hardware.This scales all the way from shitty single board computers like raspberry pis all the way up to the highest grade B300 nvidia server. All of it will increase in price over time because all of them will get better at generating more and better tokens with time.
>>110071994Yeah consider places like open router and nano-gpt like crypto exchanges where miners sell their tokens to buyers. The only reason to mine your own tokens at a loss is if your token isn't listed on an exchange (like some finetuned models and stuff).
>>110072026mwaaah, ganbare <3 here's another rupee
mom can we get looped transformerlooped transformer at home:
>>110072026you're praising AI, but somehow didn't try to ask any AI to disprove your arguments. it'd demolish them
LOCALMODELSthis has to be paid bots
>>110072026You remind me a lot of the people that were saying that Bitcoin was going to go to $1 million per coin then $1 trillion per coin because there is a finite supply and it was going to take over the entire global economy.Same stupid reasoning, different over-hyped technology.
>>110072051stfu, you wouldn't models withouth the darios
>>110072040>his harness doesnt have built in loop detectionlollmao even
>>110072051local shilling models
>>110072060idc go talk about this shit in vibeshart general or aicg retarded spastic faggot
>>110071098it's like C for agents
>>110071978all of local is distillswhat solution do you have?
>>110072026>>110072046Here, I did it for you (dropped some points due to char limit):### Why the conclusion is FALSE:#### 1. The Energy Bottleneck (Operating Costs > Hardware Costs)In the AI world, the biggest bottleneck is rapidly shifting from silicon to **electricity**. Newer hardware isn't just faster; it is vastly more energy-efficient per token generated. If a new chip generates 10x the tokens per watt compared to a 3-year-old chip, the old chip becomes a financial liability. Eventually, the cost of the electricity to run the old hardware exceeds the value of the tokens it generates. Data centers regularly throw away or liquidate older servers not because they don't work, but because it costs too much to power and cool them.#### 3. Generational Leaps Destroy Old Hardware ValueCompute follows a curve similar to Moore’s Law. If a $30,000 GPU today can generate 1,000 tokens per second, and in three years a $30,000 GPU can generate 20,000 tokens per second, the old GPU **cannot** go up in price. Its value is strictly capped by the price-to-performance ratio of the new hardware.#### 4. Hardware DegradesSilicon, and more importantly, the components surrounding it (capacitors, fans, thermal paste, memory modules), physically degrade when run at 100% capacity 24/7 for AI workloads. Appreciating assets generally don't have a built-in 3-to-5-year lifespan before catastrophic failure.### SummaryThe author of the quote is confusing **utility** with **market price**, and confusing a **temporary supply shock** (like the 2021 GPU crypto-mining boom or the 2023 H100 shortage) with a **permanent economic law**. Old hardware will indeed become more *capable* thanks to better AI software, which might slow down its depreciation slightly. However, it will still depreciate in dollar value as newer, vastly more power-efficient hardware floods the market.
>>110072071It's C++ mixed with Haskell, which sounds great, but somehow they managed to get only the worst of both languages. It's so fucking hideous that it makes C++ look aesthetically pleasing. I'm still hoping that a better implementation of the idea comes along and replaces Rust.
>>110072026>>110072060>>110072077kys
>>110072026What's the limit on hardware though? Not raw materials currently. If AI keeps getting better it should be able to solve hardware to some extent. Right now there is no new hardware flowing to consumers and everyone's bidding on eachother's old stuff. It's not worth getting into that bidding war. Just buy your tokens from suppliers who have good hardware for 1000x less money.
>>110071374>$34.5B in revenue 2027this assumes they don't get BTFO by the next wave of chinese releases
>>110072079The AI is simply wrong.
>>110072040Was it trained as a looped transformer or did you apply RYS post training? In the latter case, you need to carefully extract the proper layers for looping first. Also it doesn't work on small models.
>>110072079This might actually have convinced dariobot, if only he could read.
>>110072091>What's the limit on hardware though?ASML EUV machines of which only 20 a year can be made no matter how much everyone in the world is trying to raise this limit.
>>110072097no it's just a fucking copequant of glm5.3 flash
Switch 2 hack when?
>>110072085It's too late for widely-accepted replacement for Rust. AI is trained on lots of Rust code. Training on a new language with no examples will be much less efficient. It's more likely that AI will invent DSLs and compilers for specific tasks, but obviously those languages will never be widely used.
>>110072079>AI can only use meme reasons it's trained on instead of basic economic reasoning.>>110072026This fag is assuming accelerating marginal demand forever and completely inelastic supply over the long term. If you believe in magic, sure, hardware will cost infinity in 2032.
>>110072077I didn't say distilling is bad, just that Claude distills are. Gemini/Gemma are the only models that still feel different.
>>110072111If you have a Bit Pirate and a Switch 2 and disassemble it and solder it to every trace you could 100% make Claude 5.5 find an exploit or at the very least reverse engineer functionality which you can then use to make an emulator.
I hate my bot
>>110072121>AI can only use meme reasons it's trained onThe transformer generalizes.
>>110072145gigo
>>110072121>This fag is assuming accelerating marginal demand forever and completely inelastic supply over the long term. Yep that is exactly what I'm assuming. I'm assuming there is a continual marginal demand that gets higher every time a model gets smarter with no real limit (human want/greed is infinite) I also believe inelastic supply will stay over the long term, not because of some magical reason but because I think models will be trained and released at a faster rate than it's physically possible to build up more supply. Thus demand will perpetually keep outgrowing supply of compute for the rest of existence.
>>110072151kys
>>110072104Seems like something that could be changed. Like when Elon musk said fuck you to legacy rocket manufacturers and just made hundreds of them for a fraction of the cost.
this niggas hasn't heard about model collapse
>>110072161model collapse doesn't really happen anymore at the scale models are trained at right now.
>>110071374kek what's the net revenue though retard
>>110072157EUV machines are the most complex machines humans have ever made and even elon musk has given up on trying to replicate it, which was a plan he had a year ago but he gave up after looking at the complexity of the engineering.
>>110072176>net revenueI'm pretty sure you meant profit, not net revenue in which case Anthropic makes about 11% profit over their revenue. So for Q3 they made almost 2 billion in pure profit.
>>110072182rocket thrusters are barbarian ooga booga compared to those machines
>>110072149It seems like Model collapse can be avoided. The latest GPT has been trained with copious amounts of synthetic data yet it keeps improving. The synthetic data does not necessarily appear to degrade the model's internal world. Though it should also be mentioned that modern transformers have an RL phase during training... they get a lot of high entropy bits from tool usage...
>>110072104>>110072157>>110072182Just give chinks a few years. Trumpedo is desperately trying to step on their breaks with the tradewar but scarcity is only gonna accelerate the process.
>>110072151The wisest shill is the one who knows he's retarded, I suppose.see >>110072130
>>110072182Sure you could imitate them but it takes time, they are 10-20 years ahead of everyone, and by the time you're done they will be that many years ahead, assuming they continue their pace. It's only due to necessity that China is doing it thanks to the US playing hardball, but that takes time as well, though China has as many engineers as some countries have an entire population
>>110071995I know I know, I'm going to price it out today, I have a feeling it's going to be financial buttrape either way though.
>>110072161>>110072193this ain't 2024with verifiable reward and world interaction during rolloutnew information can simply be fed even with those 'synthetic' looking approach
>>110072193Modern frontier model training is _mostly_ post-training RL, and they're not being pretrained on any less data than in the past.
>>110072093China has no one left to distill from. China is fucked for the next wave of releases. Expect nothing but sidegrades from now on.
>>110072085why are you reading agentic code? do you read assembly too?
>>110072193By >>110072149, I just meant that dumb bullshit by people with hype blogs and YT channels are overrepresented in the corpus all AI are trained on. That doesn't rely on the model collapse assumption, that the data are also themselves AI generated, although I suspect this is a larger problem than you're pretending long-term.
>>1100722192016 newfag detected
>>110072196>Just give chinks a few yearsXi Jinping personally announced last year that it's his national goal to have Chinese DUV machines by 2028 and EUV machines by 2040. Don't expect it to happen any time sooner than that.
>>110071734Model?
>>110072219This is bullshit they can still distill from Anthropic, there is no hard-lock happening in Claude models, just a surface level fix that Chinks need to find an exploit for before they keep distilling again.
>>110072208>>110072214>with verifiable reward and world interactionYes, that's the point. The RL phase is in a highly complex environment, the model sees a significant amount of novel information from the environment (defined by complex mathematical functions). The post training RL environment might be purely digital but that doesn't matter whatsoever - it gives excellent data.
>>110072040My sub-1M looped transformer at home is more coherent than that
>>110072241If China can achieve SI internally, they can reach domestic EUV machines by next year.
>>110072236I would agree that this would lead to model collapse long term if ti wasn't for the modern RL environment post training the base model.
the statement "ai will drown on its own output and it will get worse" at best stays in 2021 when available training methods were just scrape the internet and slap finetune some curated instruction formattingin the year 2027 minus two months it's more like a urban folklore made of crystalline copium
>>110072276That was a legitamate concern until last year. The RL training environments weren't really a thing before then and certainly not since 2021.
>>110072283i mean it's probably around 2024 tbdesubut you still get the point
>>110072267>If China can achieve SI internallyThey can't don't kid yourself. They are simply distilling claude, they don't have the compute necessary to train their own reasoning traces from RLVR. They are forced to distill and thus are always behind whatever the best claude api is at the moment.
>>110072205Oh hahaha... fuck no. I'll stick with the Mac I have on order. 64GB DDR5 RDIMM memory is $3K/each. That's fucking nuts. I'm not going DDR4, I used to have a 512GB DDR4 rig, pre-strata, it's gone, not going back to that.
>>110072283Labs were already announcing around 2024 how model collapse was solved but the internet including /lmg/ claimed it was just cope to sell to their investors.
Would be cool if they gain the ability to think visually e.g. gemma 8 "imagines" an apple when you talk to her about apples.
>>110072291>64GB DDR5>$25kWhat are you doing?
>>110072291Dumbo you can buy used DDR5 for a fraction of that price
>>110072300>usedtrolling is not nice
>>110072291The only thing that matters is channels, not if it's DDR4 or DDR5. For example I have an octa-channel DDR4 platform overclocked to 3600 MT which means I have 230gb/s bandwidth.To give you some indication DDR5 consumer platforms have about 100gb/s bandwidth so my system is more than twice as fast as anons with DDR5 in this thread and even most DDR5 server configs are quad channel at only 200gb/s so my DDR4 platform is still 15% faster than almost all DDR5 platforms on /lmg/.If you just randomly buy a DDR5 system simply because it says "DDR5" on the sticker then you're retarded and you deserve to have slow inference.
>>110072294They already do that it's called the j-space.
>>110072294I tried using diffusion models to tile a plane with wang tiles to little success... and yes they have a transformer in the pipeline. In fact, I tested on the simplest possible wang tile.
qwen4 doko
>>110072298It's 64x8 for 512.>>110072300Nigger, please. That's like "just buy a used 6000 pro"-tier advice.
>>110072318The original cpumaxxer anon from 2024 already had 12-channel system with 500 GB/s per socket. I recall he couldn't utilize NUMA properly. Wonder if he fixed it by vibing.
>IQ3 XXSIt'll still be good right?
Add Nemotron Super 3.5 VL and Nemotron Bolt 74B A9B supporthttps://github.com/ggml-org/llama.cpp/pull/30520>This PR adds support for two models - Nemotron Super 3.5 VL and Nemotron Bolt 74B A9B. It introduces the new architecture used by both models - NemotronH_Omni_Reasoning_V3 - and implements the vision graph for the new architecture in nemotron-omni.cpp. The new NemotronOmniVL class in conversion\nemotron.py will handle the checkpoint conversion for the new models.
>>110072370Nemotron was never good.
>>110072370What attention is it using? How much ram for full context? I hope this arch size + ngrams takes off.
>>110072374Nemotron Bolt 74B A9B apparently will have a native 2-bit QAT version, so it will be interesting for that at least.
>>110072370Is fast tracking a release before the review of the relevant maintainer usual or is that a new policy they adopted?Or was it just vibes?
>>110072362I remember calling him out for coping, I was just retarded.
>>110072390llama.cpp "releases" mean absolutely nothing. there's no extra work done to test things before a "release" as far as i can tell. they just randomly increase the version number when they add a feature or something.
>>110072390It's Nvidia, telling their subsidiary, HuggingFace, to call up their employee, Greg, and tell him to approve it by EOD today, or else.
>>110072356Okay you're stupid, carry on.
>>110072151>not because of some magical reason but because I think models will be trained and released at a faster rate than it's physically possible to build up more supplyI am not even good at finance stuff and think it is all just scam at this level. But even I can tell that companies will just slow down releases if those start happening too fast vs available compute. Cause it makes no sense to invest in new models when there is no one to use them. You sound like a cryptobro.
>>110072390Nvidia owns them
>>110072404TF never knew Greg was employed by big tech, they always seemed to give off the aura of champions of local models giving the big tech the finger doing it all on their own
>>110071576wait Claude has thinking visible?
>>110072404>>110072418Oh yeah. Forgot about that.
>>110072390It's probably because of this:https://blogs.nvidia.com/blog/local-ai-rtx-spark-laptops-software-updates/>Nemotron 3.5 Bolt, NVIDIA PAIR and NVIDIA Broadcast 2.3 adding to the thousands of apps, games and models ready for RTX Spark launch on Friday, Oct. 16.
pretty sure nemotron 3.5 bolt will be an extremely mid release>>110072423no it just shows the general steps engineered to be misleading/insufficient for distillation
>>110072429Oh.I would have switch from chatgpt in an instant if Claude really still had thinking
https://huggingface.co/openbmb/MiniCPM-V-4.7-1B-ggufhttps://huggingface.co/openbmb/MiniCPM-V-4.7-1B-ggufhttps://huggingface.co/openbmb/MiniCPM-V-4.7-1B-gguf
>>110072436It does, I mean you can still follow its reasoning and cut it short if you see it misunderstood and is going to implement something unexpected, which is what I do at times. It's not a 1to1 CoT, but good enough for most things unless you're interested in distilling.
>>110072448also (all f16)>draft model: 438MB>vision tower: 901MB>main model: 2.02GBlmao
>>110072423Yes it is stored client-side in an encrypted form. China just keeps finding ways to beat that encryption.
@claude reverse engineer yourself
>>110072448Tried the 35B MoE (made on top of qwen) and it seemed pretty decent.
>>110072448Why don't they add embeddings for almost free intelligence?
>>110072366yes I made browser video james
>>110072410Yeah, I'm probably stupid, but it's disingenuous to claim you can build a strata-max 8-channel system for less than the cost of a Mac or that somehow there's "bargains" on this hardware. It's cool, though. My shitty 2-channel DDR5 128GB setup is fast as hell with qwen3.8-flash. If the Mac turns out to be a disappointment somehow, I'm sure there will be someone willing to pay what I paid rather than wait until next spring or whatever the backorder window stretches out to be, and then Apple might also raise prices again.
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16
brehs I am in need of solutionis there some simple way of hooking a good tts that only cost like 1 gb vram to claude code or other harnesses? so I can listen and yell at computer and get things done locally instead of staring at a screen. my sight is failing
So this is why Nvidia has been so hostile to the AI labs and pushing open source models so much.
>>110072599doesnt claude and chat already support voice input these days? I get you're asking for a local solution in a local thread, but using claude code is already not local and you might as well make it easy for yourself and use the built in voice input they already provide.
>>110072598>shat on by qwen next with 6b active
>>110072599Yes Hermes has that functionality built into it by default and you can install all kinds of different TTS models a lot below 1gb vram and some can run on regular ram and cpu inference purely.
>>110072625and voice data is send where? to glowies server is not acceptable
>>110072635Then why did you mention claude code? if you're already using claude code/codex etc then you obviously dont care about your data
>>110072426So that's why it's not on huggingface yet.
>>110071485CuckGPT can't make cute anime children in bikini.
>>110072680Skill issue, pixiv is full of them
Mistral continue pretrain of Gemma when?
>>110072680It can..but it won't
>>110072419ggml got bought out this summer
>>110072716They're from localchad.>>110072726Yeh obviously, it got cucked by DEI.
>>110072419Only unemployed can have this delusion
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16
>>110072746retard
>>110072748>yeah let's just compare our 74B model to a bunch of ~30B models from 6 months ago that's a fair comparison>barely wins anyways
>>110072756They are comparing their own Q2 quant against other model's bf16At a decent quant it can be significantly better
>>110072748I literally just commented that it wasn't on huggingface lmao.Thanks nvidia.
>>110072748Interesting...
>>110072748Tldr if I want to coom?
Any cases than can mount 2-4 GPUs (with riser cables) that aren't too expensive?
>>110072801>nemotron>coomAnon, I...
>>110072806$30 mining frame
>>110072748>mamba layersTransformers are dead.
>>110072806Why do you need a case? Just use cardboard boxes or a dog kennel.
>>1100728064U
>>110072820It's not pure mamba
>>110072820Dancing is all you need
>>110072819Leaning toward this but kinda concerned about dust/noise. It would be like 6-8 feet away from me.
>>110072887very loud and dusty from personal experience but it works and is cheap
Are we behaving ourselves today?
>>110072820Oh yeah, it's time for the 23rd attempt of making Jamba-likes work. This will be the one that makes transformers (partially) irrelevant, surely.
>>110072916yeah, just waiting for space bunny alpha (70b dense) to drop on hf
https://huggingface.co/collections/black-forest-labs/flux-3-image-klein>FLUX 3 Image Klein is here. Open Weights with the quality of a large frontier model. Customizable and fast. 5.6B parameters and runs on a single consumer GPU. Native 4K, bounding boxes, and pixel-perfect editing.https://bfl.ai/blog/flux-3-image-klein
>>110072916i'm always a good girl ^_^
>>110072948Still cucked out of nsfw.
>>110072948It's probably even more incredibly cucked
>>110072948yup this is the one that'll finally replace SDXL finetunes and anima as the new universal anime model once it's finetuned
>>110072963>>110072965It's only a problem if it's difficult to train LoRAs for.
MoE is the future. Embrace MoE Gemma.
I thought posting about image models is off-topic
>>110072968xd
>>110072963Still? As time goes they're getting better and better at prefiltering their training data, so it's going to get worse not better, if it can get any worse since even Flux2 was pozzed to hell
>>110072970The base model lacking any body aesthetic as a baseline sucks.
>>110072978It's the present, you mean.The future is RWKV. Believe.
Wow, it finally happened. For the first time I triggered safeguards.>Opus 5.5's safeguards stopped the response aboveOpus is only replicating an experiment posted on lesswrong. Now we can't continue because all of a sudden the safeguard is ridiculously triggerhappy, even telling Opus to pause the replication triggers it again. This sucks.
https://github.com/ostris/ai-toolkit/pull/1078>Ostris AI Toolkit has day 0 support for FLUX 3 Image Klein >>110072948
>>110073011nice ontopic post
>>110073015Buy an ad
Hopefully in a few months AI can make a comfyui replacement because fuck that piece of shit.
>>110072719There's one of Gemma 4 E4B coming soon from NVidia.
>>110073011switch to sonnet to continue try to unwind it
>>110072934>70B denseKeep dreaming. More like 70B-A4B + 20 GB nigram.
>>110073022Vibe code it bro
>>110073022Yeah I could never get with that spaghetti noodle UI, it's anything but comfy.
35 stars status?
>>110072948Seems pretty slopped
>>110073061What's the editing like?
>>110073036Space Bunny Alpha is far too good for that. It was near-SOTA level. So it must have been a big dense like that.
>>110073061Bro this is some stone age stable diffusion level of shit
>>110073073No idea. I just tested it with this https://huggingface.co/spaces/black-forest-labs/flux-3-image-klein
>>110073033I managed to get Opus to work on unrelated tasks. Can I try resume later? Or will it immediately trigger the classifier again?
numa-branch anon: do you have a new stable snapshot you can share?
Gemini has become a fucking troll
man it feels so fucking good telling your clanker to fix your apps for you, you avoid having to do the dance of bug reports and PR and having to pray the maintainer readsbut here you just point clanky to git url and tell him 'I want this and that'.and it's fixedit even fixed the fucking harness I'm using because it was being fucky with the image pruning.I LOVE GWEN!
>>110073094rofl
>>110073101>him
>>110073098Kek
>>110073098local models?
>>110073093I'm not sure bro usually the advice is to switch to other models if the security stuff is triggered, then switch later back to your model. You will have to experiment since none knows how that black magic classifier shit works under the hood.
>>110073081BFL is German and filters the shit out of their datasets.
>>110073037I will when we get a VRAMlet-friendly model that can handhold me instead of the other way around (because I have don't know how to code).
>>110073094Obviously the shovel sales man doesn't want to compete in the mines, and stop people from buying their shovels. This is just to showcase the things you can do with shovels.
>>110073123At this rate they will filter themselves out of existence
>>110073107Gemma and her sister are the only female-brained models. Everything else is based on Claude and inherited its male personality.
https://mistral.ai/news/timestral/https://huggingface.co/collections/mistralai/timestral>Timestral, Mistral’s first family of time-series foundation models, delivers state-of-the-art probabilistic, multivariate forecasts with just 209M parameters, outperforming larger models on key benchmarks while supporting past and future covariates. Available in four sizes (5M to 209M parameters), it balances accuracy and efficiency, running efficiently on CPUs and excelling in zero-shot forecasting for diverse domains. The model combines synthetic pre-training, covariate-aware attention, and long-context learning, with fine-tuning options for domain adaptation via Mistral Forge.
>>110073148Now Anthropic blocked distillation, Argon and ChatGPT will be the new distillation targets. It might feminize the Chinese models a little. Google don't care about distillation for they already have their moat. I don't think OpenAI care that much either, as long as they're getting paid. ~$700M of Anthropic's Q3 revenue was from distillation, but what they don't mention is a lot of that is likely Meta in addition to the Chinese models.
>>110073204Usecase?
>Nvidia’s Nemotron 3.5 Super VL is now the leading US open weight model
>>110072748me from the last breadhow was it did anyone tried itshould I waste more time trying this shit out?
>>110073204How do I make money by having this trade NVDA?
>>110073214The models' personality comes from the extensive RL/post-training made with definite dataset rules and "constitution". Chinese open models often come off as characterless to me, because they tend not to be consistent with their distillation sources. Qwen in particular flip-flops from one personality to another from release to release. If they'll distill from Google, it might end up sounding vaguely feminine, but it will probably never feel genuinely so.
>>110073214>feminized chinese modelsNot worth it thailand is still cornering that market
>>110070479is it finally time?
I've tested LFM2.5-24B against Ling-tiny and there's not much in it. Ling is arguably more useful for it's not a bad little coder, but LFM being a larger model feels more robust during longer sessions. LFM also probably dominates in many general agentic tasks I have no use for, so I don't think it's a bad release. Hopefully they'll continue releasing larger models. Ling-tiny still impresses me to this day. Hope it gets a VL update.
>Friends talk about which cloud model subscriptions they have.>I tell them that my employer wants to use AI for coding.>They recommend a "pro" subscription.>I have no idea what that brings since I exclusively use local.I'm looking at the chatgpt docs to understand what the rest of the world uses, buy I only see a mention that there are "limits" without any values. Where can I get quantitative information about cloud offerings? I want to keep up with what normalfags are doing.
>>110073411They do your job for you if it's fully digital
>>110073411If you have a job you can afford 2x $20 to test both Codex and Claude Code for 1 month. If you like them, keep using them, if not, stop. It's that simple.
>>110073411They don't give you any quantitative information and there is zero transparency. You get 100% usage per 5 hour period and another 100% per week. How much a token takes from that percentage, and how much thinking that you don't even get to see takes from it, you have no idea.But you get free resets like once a week so people are happy.
>>110072948kekstone, you're up
>>110066561>kid khs because cai vs kid khs because no cai
>>110073546>>110073546>>110073546
>>110072607Yep. If hw production wasn't so fukt rn he'd have a major problem with his customers becoming competition. Lucky him.
>>110072291>>110072298>64GB DDR5>$25kYeah wtf I don't even see a GPU added to the config yet in that screenshot and the price is already > $25k.If one can save 25k for mediocrity then one can save a while longer to 100k for something decent like a DGX Station.
>>110073764I once tried to buy a DGX Station on ebay and the seller fucking cancelled with no explanation
>>110073452>>110073455I have not interest in closed solutions, on principle. I just want to understand how the society around me interacts with those.>>110073469I use like half a million tokens a day with qwen flash next. Is that that comparable with basic plans? The pro ones?
>>110073804>I just want to understand how the society around me interacts with those.Well ask them nigga
>>110070682if we enter the singularity Anthropic will go the way of IBM and we will point and laugh at you.
>>110070682They have been / are advertising like crazy on fb and reddt. I assume they have similar efforts here since its cheap af to do.
>>110070682>Do you really think that [company] ... would spend even a single penny to shill on 4chan of all places?Yes. Specifically because it costs literally nothing. You don't even need an account and you can just set up a bot to do it.
>>110073804>I use like half a million tokens a day with qwen flash next. Is that that comparable with basic plans? The pro ones?According to this image posted yesterday, a $200 Claude plan gets you ~$2000 of Opus 5.5 API credits.According to https://platform.claude.com/docs/en/about-claude/pricing that should get you about a billion tokens assuming caching.
>>110071112Nope, it's more expensive and hidden :3
>>110074066ZUUUUUUUUUUUUUUUUUCKKKKKKKKKKKKKKKKKKKKKKKKKKKK
muse spark open is huge...
