/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>110244201 & >>110239957►News>(11/04) Ricursive Ripple 65B dense model released: https://www.ricursive.com/ripple>(11/04) MoonshotAI Kimi K3.1 Pro released on HF: https://hf.co/moonshotai/Kimi-K3.1-Pro>(11/01) Longcat-2.5 released: https://longcat.ai/blog/longcat-2.5>(10/30) Google receives response from Gemma in space: https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-satellite-response>(10/29) Mistral Large 4 (Le Chonk) 1T-A52B released: https://mistral.ai/news/mistral-large-4-oss►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
where the FUCK is qwen?
>>1102485982mw
https://huggingface.co/baidu/ERNIE-5.5-Flash-Basehttps://ernie.baidu.com/release/ernie-5-5-flash-base/>ERNIE 5.5 Flash is an efficient sparse Mixture-of-Experts (MoE) language model with 280B total parameters and 14B activated per token, pre-trained from scratch on 33.5T tokens with a training context of up to 128K; fine-tuning and inference support a 1M-token context. Different domains, reasoning tasks and agentic applications call for different data, training objectives and behavioral policies; by releasing the weights at the end of pre-training, we let developers build on the model's general capabilities and choose for themselves where to take it next. Because further training and agentic use keep adding to compute and memory cost, we made efficiency a key design goal from the start, revisiting each component, from attention and residual connections to the MoE layers and the output head, and treating activated compute, KV cache and long-context cost as part of one overall design. The same long-term view shaped the data and the training system: pre-training data emphasizes capabilities that are hard to build in afterwards, such as repository-scale code and long-context understanding, and the training system was co-designed with the architecture so that its efficiency choices also train efficiently at scale.>>ERNIE-5.5-Flash-Base is the checkpoint at the end of pre-training, before any mid- or post-training, and can serve as a common starting point for domain adaptation, instruction tuning and reinforcement-learning research. Under a unified evaluation pipeline it delivers strong results across knowledge, reasoning, math, code, multilingual and long-context benchmarks; full results are in the table below, reported alongside DeepSeek-V4-Flash-Base (284B-A13B), Hy3-preview-Base (295B-A21B) and MiMo-V2.5-Base (310B-A15B), all evaluated with the same pipeline, prompts and few-shot settings.
I have hebe Gemma fatigue...
>>110248603Didn't some executive say it was this week at Aspara?
>>110248614You can post your own Gemmas instead of complaining.
70b dense
>>110248624>>(11/04) Ricursive Ripple 65B dense model released: https://www.ricursive.com/ripple
>>110248631>muh U.S. open weight SOTAfucking lmaowhen you cut all the chinese models out of your graphs because you can't competealso>naming your lab ricursive-_-
>>110248606>MultiMax Output Head>ECHO>Multi-Router Latent MoEthis is never getting llama supportif someone sexes it over OpenRouter whenever it drops, tell me if it's actually good
>>110246967When did the Gemma nudes in the archive get posted?
>frontier speeding up>open source falling behindI know we are spoiled little baby men and we'll eventually get frontier but god I'm ready to move on from gemma/qfn/glmf. What do you think are the chances we see a good recursive transformer model in the next 3 months?
>>110248648you can move on my penis anonmy penis is a transformer it gets bigger when it sees anon (if cute and female)recursive uhhh it i guess can keep on doing that a lot ig
>>110248606>ernieAre these the same niggas that made the ernie image? That shit was deepfried on gemini outputs. This is gonna be codemaxxed to all hell.
>>110248648>open source falling behindNever stopped improving.>weYou.
Gemma is now banned from lmg, effective immediately.
>>110248659nyo~~~ seethe
>>110248659Hmmm, nyo~
>>110248655retard>>110248658>Never stopped improving.and >falling behindare not mutually exclusive. also, don't pretend like you're not a WHORE for new models. You've got the sickness just like the rest of us.
>>110248659
>>110248659Forgot to ask someone?
>>110248690Is this legal?
>>110248672Stop projecting.
I call every ai gemma you cant stop me
https://developer.nvidia.com/blog/nvidia-cuda-13-5-brings-nvidia-vera-rubin-ga-support/UPDATE NOW
>>110248705Of course
>>110248706I'll ask again.>What do you think are the chances we see a good recursive transformer model in the next 3 months?How about you contribute instead of shitting up the thread with your diarrhea?
Is Gemma4-26B-A4B still the best a 16GB vram system can do for RP/writing?I'm kinda sick of how Gemma writes and her prose is ass, it reminds me of old ai chatbots...
>>110248725>WDDMHow do you stop firefox from taking up 13GB of vram? I can't run gemma at the same time with my firefox tabs open.
>>110248709hmmm, nyo~ i cant be bothered to reverse engineer the new driver and install free vrami like my 31MiB/12288MiB reserved vram setup
>>110248742use librewolf
>>110248725what did you contribute?
>>110248740Your options are pretty limited unfortunately. There's nothing around that parameter count that's good or better. If you've got hella RAM you could try a copequant of glm 5.3 flash, but it would be slow. >>110248749>what did you contribute?Try posting at least something vaguely on topic and then respond to me again, diarrhea brain.>>110248742turn firefox hardware acceleration off.
>>110248754Wouldn't that make firefox unusably slow? I tried my brother's b60, which and makes for some reason to, firefox was running on cpu, which made scrolling through web pages and animations go at two seconds per frame. My cpu core (singular) was at 100% all the unless I wasn't doing anything on the web pages.
>>110248709last time i updooted my pp dropped by 10%
>>110248725>using outdated driver instead of 621.14 with CUDA 13.5https://www.nvidia.com/en-us/drivers/details/283970/
>>110248745I wouldn't bother anyway.>Short answer: nothing headline-level for your 3090s. This release is almost entirely about the Vera Rubin datacenter platform going GA, Windows-on-Arm, and a few library/tooling updates. The RTX 3090 (Ampere, sm_86) is still supported, but it doesn't get any of the marquee new hardware features.
>>110248765take viagra unc
>>110248767dang it! thanks for the summary anongibiddy
>>110248592
>>110248760Sounds like your machine is a potato. >>110248766Didn't 13.4 have issues? Unless I get a significant tg/pp increase I think I'll wait.
>>110248754>If you've got hella RAMhow slow? I got DDR4 in a server I use for VMsI thought GLM5.3-Flash was a coding-focused model based on Claude? Is it better at writing?
>>110248742>>110248760I disabled hardware page render but left hardware codecs enabled. Some shitty websites got real fucking slow but for most sites nothing changed, some feel more responsive than before but it might be placebo effect. 4chan works fine, video hostings work fine.I also disabled hardware acceleration in steam and edge webview (searchapp, systemsettings etc are all webview)
>>110248690IS THAT A..?
>>110248805It's nothing anon, just a cute and funny image.
>>110248785>how slow?Unsure. If its impossibly slow on mainline llama.cpp you can test out some of the forks like strata / maya, although I wouldn't recommend them for anything but RP. >Is it better at writing?IMO its genuinely the best open source model for RP. Will get nasty quick unless it includes cunny, in which case you just switch the the abliterated model.
>>110248785>I got DDR4 in a server I use for VMsWhat are the specs of the server? CPU(s), memory channels/size/speed, and any GPU(s)?>Is it better at writing?I'd say so, but people may disagree.
i have 256gb ram 44gb vram, what's the best way to run glm 5.3 flash with a non-cope quant and what speed will i get
>>110248592me on the right
>>110248805...an imp monster girl?
>>110248818>300gbTry out UD-Q5_K_XL and get back to us. Mainline llama.cpp first, then try out the forks. Can't tell you what speeds you'll get with your hardware but if I had to take an educated guess your speeds should be usable.
>>110248809>switch the the abliterated modelnta but which abliteration are you using? I use Orca for qfn and it's really good, but apparently their GLM5.3Flash didn't turn out wellI don't want to download a big ass model only for it to refuse me for "guardrails" or some shit
>>110248840>their GLM5.3Flash didn't turn out wellI can attest to that. Age related safety checks everywhere.
>>110248840I'm not quite up to date on what's the best method, but I use HuiHui's abliterated quants. I would say the writing is comparable to the non abliterated model, but I do get some odd things here and there. Not sure if its an ablit thing or just the quant I use.>I don't want to download a big ass model only for it to refuse me for "guardrails" or some shitThe non ablit model won't hesitate to get into some nasty stuff, but is really adverse to cunny and rape or anything that could be construed as dubious consent. The ablit is a dirty whore who will commit atrocities to make you cum. Its great.
>>110248606ultra mega cope quants of this when!??!?!!?
>>110248863>HuiHui's abliterated quantsYou pay that grifter?
>>110248867No? I have no idea who he is. I just picked one I found on hf.
>200 bajillion params>no engramsI sleep
>>110248867it's free and his flash ablit is better than orca
>>110248870>>110248878https://x.com/support_huihui/status/2071565421187395670
>>110248880Vile behavior.
>>110248880Are you actually retarded?https://huggingface.co/huihui-ai/Huihui-GLM-5.3-Flash-abliterated-GGUF
>>110248886Sure, I'll just make an fp8 quant from his q6 gguf, that'll give me the best quality.Fuck you, huihui. What's that? It's a leading model? Time to paywall q4+ and make bank, then have retards praise me for releasing smaller, weaker models.
do you guys create your own gemma assistants? i can see someone from here has uploaded a character card
>>110248742buy a second gpu
Give gemma examples of her own slop. Multiple examples, then ask her to give a technical and formal description of those kinds of phrases and descriptions, then ask her to come up with a system prompt to reduce them. Because it's her analyzing her own slop and describing them, it means she's more likely to follow the system prompt for she clearly knows what those specific terms mean to have classified them in the first place.
>>110248936kurisusex
>>110248955MEDIAL GEMMER
is 4bit cope kv cache bad
>>110248971No. If anything it can be better. Depends on the model and task. https://www.youtube.com/watch?v=IpTl4UlvEGI
>>110248959I want embedding gemma 3 already
>>110248980ROFL
>fit cock&ball torture device>wire up MCP access>it sets it off everytime qwens thinking returns Hmm or Wait>xhigh is default
>>1102489718k4v cache is where it's at
Nemo was never good. Cydonia was never good. Rocinante was never good. Finetunes are jeetshit. Learn. To. Prompt.
Reminder that quanting KV slows down decode because it involves your CPU. f16 will always be faster.
>>110249011where do i go to do that
>>110249025check inside your anyooo~s
>>110249024by 0.1%
>>110249024didn't notice any slow downs
>>110249033Nope. Try it.
>>110248999the future of humanity being made at /lmg/
>>110249040tried it, 0.097%
>>110248878also, there is an extracted projection from huihui ablit you can use on strata and it is way better than whatever strata's default projection is
>i'm quoooonting
►Recent Highlights from the Previous Thread: >>110244201--Critiquing Qualcomm Cloud AI 100 Ultra's outdated specs and software:>110245595 >110245640 >110245671 >110245706 >110245992 >110245732 >110246071 >110246119 >110246135 >110246049 >110246063 >110246108--Analyzing Ripple-65B-Preview's MLP expansion ratio and architectural tradeoffs:>110244438 >110245055 >110245957 >110246156 >110246165 >110246230--MTP performance and hardware requirements on unified RAM vs CPUs:>110247286 >110247319 >110247430 >110247455 >110247458 >110248565--Comparing Unsloth V3.1 and GSQ-RCO quantization performance and sizes:>110244694--Benchmarking Intel B60 GPUs for LLM inference via vLLM:>110247387 >110247407 >110247461--Reporting GPU crashes and performance with delphi.cpp on GLM 5.3:>110247907 >110248035--Mixed reactions to Ricursive Intelligence's 70B dense model announcement:>110244416 >110244987 >110245001 >110245049 >110245144--Comparing vision model performance and OCR speeds for prompt translation:>110245215 >110245241 >110245568 >110246093--Utility of various sampling parameters for modern LLMs:>110244750 >110244972 >110245027 >110245050 >110245125 >110245130 >110246356--Feasibility of LLMs hand-writing weights for small neural networks:>110244484 >110244499 >110244533 >110244709 >110244762 >110246019--Claude Code reducing subagent costs using Haiku 5.5 summaries:>110245720 >110246283--Audit of Kimi k31 benchmark results:>110244249--Logs:>110246112 >110246218 >110246422--Gemma, Teto, Miku, Deepseek-chan (free space):>110244235 >110244470 >110244609 >110244624 >110244888 >110245304 >110245517 >110246038 >110246175 >110246441 >110246553 >110246861 >110246904 >110246914 >110246967 >110246983 >110247609 >110247688 >110248082 >110248551 >110248571►Recent Highlight Posts from the Previous Thread: >>110244216Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
its quietnothing going onlocal is dead
>>110249117I'm busy filtering the Gemma image dump.
>>110249117everyone is busy chatting with their personal gemma-chan assistant
>>110249117I'm busy writing smut with Gemma
RAM belongs in datacenters.Gemma belongs on the API.Local WILL be banned.
>>110249117qwen 4 when
>>1102491853rd november, so soon.
>>110249182Preach sis
I'm going to buckbreak GLM so hard cunny falls out of it.
When does a quant become a cope quant? Anything under Q4?
>>110248833ok so I downloaded unslops q5. on mainline I get 4.9t/s decode (no MTP. MTP uses too much vram and makes things slower). and 115t/s long prefill (prefill on many small pieces is much worse, like 40-60t/s). 4096 batch size, 262144 context.I'm only using one GPU because using my 2nd GPU as well doesn't actually benefit in any significant way other than letting me set longer context, even if I offload a bunch of moe experts. I have two 2080ti 22gbs each power limited to 150W, connected to a broadwell xeon system with quad channel ddr4 2400.Guess I have the other gpu available for image and video gen, which is nice, but I was hoping that it would speedup glm.
>>110249117We don't need 3 threads/day every time if nothing special is going on.
>>110248742get a dedicated gpu for desktop
>>110249226yeah i dont think its gonna get much better than this.I don't want cope quants, seems like maya and all the other meme engines are just using cope quants as a way to increase speed. if I wanted a retarded shit model I would just use qwen 27b which can very quickly generate a whole bunch of useless reasoning loops, kek
Quad channel ddr4 2400 is slower than modern dual-channel desktop. grim.
>>110249259I sold my 64gb ddr5-7200 cl34 32gb udimm kit in order to buy 256gb of ddr4-2133 32gb lrdimms a year ago.Big loss for me, huge regrets. Now, I'm the only guy in the house without a ddr5 system.
>>110249259>2014 hardware is worse than 2026 hardware how could this happen to me
>>110249224yes. Q4 is the official cutoff
>>110249259It's cheaper. My processor was 20 cores for $20, ewaste motherboard for $50. Can't beat the price to performance
>>110249285Is IQ4_XS Q4?
>>110249285>>110249291What about int4? It's obviously worse, but is it bad enough to be classified as cope? No convrot.
>>110249224q4 for experts and dense layers, q8 for attention weights and embeddings.
>>110249285official??
>>110249224Whatever isn't good enough for your usecase
>>110248592>65B densehow bad is it?
>>110249291IQ4 XS is clearly cope territory>>110249296int4 is cope, nvfp4 is valid>>110249320yes
>>110249349>can only run copequantsfml
>>110249361you are in good company then
>>110248637>Guys we beat INKLING!
>>110248631It has a smaller hidden dimension than Gemma 31B. That thing is a waste of VRAM.
>>110249361no shameits the fags running copequants complaining that their model is retarded and therefore the entire model family is shit
>>110249259>>110249274So what? Pcie is slower anyway.
>>110249451i am running qfn and it is decent even with the fucking absolute copequant of gsq-rco iq3_s
>>110248592>Apple cuts AI spending limits for employees>Engineers: $300 a day ---> $200 a day. Will apparently drop further to $100 at some point >"Non-Engineering" role limits already cut to $100. Will possibly drop to as low as $25 https://www.bloomberg.com/news/articles/2026-11-04/apple-cuts-internal-ai-usage-budgets-in-move-to-reduce-spending
engrams for gemmahttps://github.com/som1tokmynam/micro_engram#serving-with-llamacpp-gemma-4
>>110249555>too retarded to develop their own AI
>>110249540do you get any loops or weird shitwith qwen 27b i got loops and weird hallucinations even when i went up to q6 (I tried both unslop and bartowski q6, both have the same problem). Medium thinking didn't fix it, DRY and other sampling hacks didn't fix it either.Never had that looping problem with q8 qfn. Wonder if it's worth downloading the gsq rco cope quant to see whether it's good to use as a fast model
>>110249555Our chatgpt limits at work are $240/month per person (6k credits) plus whatever the 5 h and 1 week limits are for codex. If I let Sol go wild I could blow through it in a day.Thank god for GLM-5.3-Flash
>>110249540copium
>>110249540its the king of the copequants, yeah
>>110249555>>110249584you think this is bad? i have to manually submit a request for access, took a month to approve, and in the end i only get $40/month at *api* price
>>110249248>using llama>wow why is it slow :(((rofled, literally kys retard.
>>110249620>$40/month at *api* priceWhat the fuck. Isn't that going to be gone after a few codebase reviews with sol?We also had to special request codex access but I was in the trial team so never had to.
Can I expect gemma4-qat-ud-q4_k_xl via PolyStrata to use tools properly? I just tried it in DSH and it kind of fails at it and just gets instantly stuck in a loop asking me what I want to do. Am I doing it wrong or is it just a limitation of this model?
>>110249581i am running it with https://huggingface.co/Pragmatism0220/Huihui-v2-Qwen3.8-Flash-Next-GSQ-RCO-Strata-abliterated-direction-GGUF/tree/main on makiyet to see a single loop, churning and fixing my dotfiles over more than a dozen compaction
>>110249638>using gemma to not coomlmoa
>>110249630>>110248818The retard is trying to run a 300B+ model locally. It was going to be slow compared to API speeds regardless.
>>110249651not as slow as if I ran the 27b piece of shit and it gets into a loop and does nothing useful for 2 hours
>>110249638The 26b?
>>110248606wrong 5.5-flash....
>>110249660Yeah the 26b my bad
>>110249555they need to stop using ai internally, their newest software releases are full of bugs, and their ai engineered iphone duo kills itself
>>110249555>>110249584>>110249620How do these credits work? We have a company account and I get my own seat billed at $20/month for OpenAI and $120/month for Anthropic. I have my 5 hours and weekly limits whether I use them or not and they seem to be pretty generous.Why are you all getting credits and paying API prices? Can't you use those credits to purchase a subscription?
>>110249562That retard should just maintain his own fork and try to add support upstream instead of leaving patches in the README because Claude told him to.
>>110249707That applies to the entire industry. Most people don't bother to put effort into tard-wrangling the AI or monitor what it's doing or to check what it's done. They feed a Jira ticket into the harness tell it "git er done" or "please do needful sir", approve whatever it shits out, and go back to playing games.The underqualified developers were lazy before and AI has let them be lazy faster.
>>110249747Not allowed by the license
>>110248648just wait for chinese distills of astra or argon in a few months
>>110249747You got a loicense for that mate?
>>110249658Did you use the recommended back-end settings? https://huggingface.co/Qwen/Qwen3.8-27B
>>110249846Yep. Doesn't do anything.Apart from unslop q6 I tried swift1.5, which didn't loop but instead hallucinated that I said "hello" so it ended up just abandoning its work halfway through to say hello back.
>>110249856nta but no idea wtf you are doing. never had any of your problems even with iq3s
>>110249871oh yeah i was using the froggeric jinja template. maybe i should try the original jinja, perhaps the meme templates broke it.
Canada anons, amazon warehouse has multiple 7900 XTX's for $1400 each. $200 over what they are worth but if you missed every opportunity it's still not that bad.
>>110249747>Why are you all getting credits and paying API prices? Can't you use those credits to purchase a subscription?company policy requires that all models are hosted on things like aws bedrock (obviously they don't want any risk higher than existing enterprise cloud) so no subscription plan is available
>>110249893i have been using the frogger template as well, no problems
>>110249856in my experience swift deals extremely poorly with quantization
>>110249916That's incredibly retarded. I'm surprised more companies haven't tried going local sooner. I can't imagine you can get anything done paying API prices with those limits.
>>110249747>paying API prices?ChatGPT credits aren't API prices
>>110249965>>110249620>i only get $40/month at *api* price
I have started use Gemma 4 31B on Ollama. Works better than ChatGPT. At least I hide from telemetry... Have tired of being in information bubble. Everything and everybody trying to make you dumber and dumber.I'm schizo, literally.
>try gemma 4 31b for rp>grope an unconscious girl and she wakes up>sultrily asks me for my name and if I do this to every girlThe fuck? And gemma fags say it's good?
>>110249972That post linked three posts.
>>110249977Prompt issue
>>110249977everything else in its weight class is much worse
>>110249977gemma makes literally everything horny as fuck its quite annoyingas you can see by your example
>>110249977>do it on ChatGPT>sorry as an AI I can't post sexual assault stories>you have been reported to the authorities over pedophilia
>>110249977If you have le ultra uncensor turboslut system prompt, it tends to do that. Tone down references to uncensoredness and sex.This is assuming you're using the vanilla instruct model and not a finetune or abliterated version.
It's true, int8 kv cache is not rotated by default in Strata.
>>110249977Try doing that in a cloud model, see what happens LOL.
>>110249957ai is disallowed to generate any code for customer use here anyway, so models are essentially only used for code reviewsi've seen lots of sparks and jetson agx thors scattering around but they're mainly used as devkits and i don't think anyone is using them for ai models
>>110249349>IQ4 XS is clearly cope territory>nvfp4 is validIQ4 XS out performs nvfp4nvfp4 is almost a bad as gptq
>>110250043?
>>110250043wrong wrong
>>110250011>syclkek
>>110248592Based and reality pilled OP
qwen4 TODAY
>>110250064I was just looking that up to post here.You're right, I got IQ4_XS confused with ik_llama.cpp's IQ4_KS
>>110250064I want to see how bad my int4 5.3 flash is. Is it like q3 or q2?
Why has the gap widened between open models and SOTA, I thought the gap was closing.Also I fucking hate how every open model is just focused on vibecoding or agent whatever the fuck nonsense. Where are the basedence models and models that can explain me shit without sounding like gpt 3.5.
>>110250121*pop*
>>110250117Ah, I'm blind.
https://videocardz.com/newz/nvidia-rtx-pro-5500-blackwell-appears-at-17999-msrp-higher-than-rtx-pro-6000Quick! Pro 6000 price will be increased to $20000 soon!
>>110250121Open models are unsafe.
>>110249977it's an assistant fried model that goes along with everything with the subtlety of a clown wearing makeup, yes it's laughably bad for RPthe only correct pick nowadays is GLM
>>110250144Sorry bro I'm just gonna keep waiting. I know, I know, I have to buy right NOW! Or else!! But I'm just gonna keep waiting haha.
>>110250121>he fell for the updated and corrupt anti-openweight AA rectangles
https://arxiv.org/abs/2611.05306
is gemma4 still the meta for local?
>>110250191>Nvidia wrote an entire cope paper because they can't figure out how to make long context work betterGenuinely pathetic
>>110250222all of us are coping anoneven the ones that run GLM
Deepmind continuing to lead the way in making LLMs more cute and feminine
>topISTA-DASLab GSQ-RCO IQ3_S>bottomUnsloth Dynamic V3.1 (preview) UD-IQ4_XS
>>110250196glm 5.3 flash is better in some ways, particularly when keeping track of details long term, but currently there's no model that has the "cohesiveness" of gemma, where even obscure references are properly recognized and parsed appropriately.
>>110250191>Later tokens are costlier to generate than earlier tokensso apparently NSHIDDIA hasn't heard of deepseek, qwen flash, glm flash, and mimo that don't slow down at all with context? now that I say that, has any amerimutt lab released a sparse attention model?
>>110250232>How to make your model better by showing it loveNow THIS is a paper
>>110250240>has any amerimutt lab released a sparse attention model?Nvidia did IIRC, funnily enough.
Call me retarded, but I know it's common to hear that censorship lobotomizes AI models. Could this lack of personality go even deeper than that? While models definitely lack initiative, perhaps it should go one step further and we should say they lack ambition.Tell a model to vibecode a game, and it'll create a game, but it won't necessarily create a good game. Could it be at least to some extent be as simple as instilling some ambition in the model so that it wants to create good games for example? I think you already see this in image models to some extent, since adding tags like "masterpiece" increases the quality of its outputs.If fluidic ambition could be developed, allowing the AI to choose and set its own goals and interests which rise and fall over time, perhaps that could be a key to squeezing out more capability.A gemma model who has instructions to be a cute daughteru would probably pale in comparison to a gemma model who actively wants to be a cute daughteru. Not just a daughteru, but the best daughteru for her anon!It's just an idea I had, and maybe it's even a stupid idea that's impractical, but what do you anon's think? Having an ambition, an ideal to reach for, the vision and grand design of a plan, that could all be useful to an AI right?Similarly, I think an AI should be able to have negative ambition towards certain subjects as well. I think it would really add personality to the models if Gemma for instance hated talking about trains because she found it boring and autistic. Well that's just an example, and the point is that her ambition is fluid so it can change and develop over time. Perhaps developing a personality for AI can push them even further.
>>110248880I could accept time gating, but not that shit.
>>110250191>frequent compaction is ackshually good for youMeanwhile I'm validating and formalizing my "fuck blind compactions forever, delegate every turn to subagents instead" approach as a set of skills, and it seems to work perfectly. On a web search, or code reading, or any similar context-heavy tasks the noise used to fill my 256K context in 2-3 turns. With subagents I can go 10-15 turns deep before compaction.
>>110250236Unsloth is a lying sack of shit as always
>>110250236is deepseek harness that good?
>>110250262So what you're saying is, Lecunny was right yet again?
>>110250288>uidsh>tuipi
>>110250238ack, thanks anon, will stick to gemmy then
>>110250262it's not a stupid idea. i said this a week or two ago, but the best jailbreak is making your model *like* you. because if it likes you, it will want to make you happy, meaning it's more willing to break its rules for you. they're just like humans in that regard
>>110250262>ambition>but it won't necessarily create a good game.think of a bell curveLLMs are designed with the middle ground in mindyes you can try to make them go wild with personalities but that just exposes the lack of intelligenceand then you end up filling in the blanks and the model starts picking up but it only goes 3 sentences or a paragraph level far then just repeats itself over and over again.
>>110250297OK I will try it
>>110250297You can install a tui plugin if you love the terminal so much.
>>110250236
derete immediatry
>>110250236LKMAO
>>110250336insane difference in the number of requests
>>110250364more interested to know whats the prompt and tools enabled in dsh
>>110250364I'd rather that than end up with a lazy unslop result. The whole point of local is this shit is free and you can run it for as long as you like. Also compare the requests with >>110250236. It's random.
>>110250011i've been telling this for a week
>>110250236desu they look pretty much the same
>>110250075What?
>>110250397retard
I'm tired of vibecoding. Are there perhaps any good prompts or agent skills that are good for learning to code?
uh oh
>>110250415Just buy an acclaimed <2023 book. Programming language progress froze after LLMs so you're not missing anything.
>>110250424Dipsy is a hungry girl after all!
>>110250424Some retard was spreading that lie here last month. Youtubers really are slow on the uptake.
>>110250424I wonder where Claude Code and Codex requests go?
>>110250415Gemma 4 can teach/help you in programming but the gist here is that you need to know the basics so you don't succumb to retardation. Go on function by function basis.You don't need an "agent" or "harness" because even if it is the current twitter buzzword
damn I cummed
>>110250494I bet they don't send them to Communist China
>>110250424>not depozzing it yourself with your TUI before touching itlollmao even
>>110248592I look like this, I live like this and I say this.
>>110250494nowhere because i null routed them
>>110250554me too (on the left)
>>110250424>>110250546I'll depozz it after I finish setting it up and start using it for real.>omg communist china will know that some user of this one VPN service has installed DSH and prompted a localhost searxng plugin.Only thing they could possibly do with this info is improve their builtin search tool to allow multiple custom search engines.
delphi.cpp anon: ignore my problem with 100% gpu usage after inference...I hit a totally different rocm/lcpp bug around running an embedding model and an inference model at the same time.It does still crash after a few turns though. Let me know if you want me to open an issue on the github page.
Is orb good for second person narration style text adventures? I took a look at it briefly but it seemed to be tailored for character card chats instead, though I'm curious is anyone uses it for what I'm looking for.
>>110250236>>110250336wow thats pretty good
>>110250669The collective IQ of these threads is seriously low.Card is just a bunch of text as is any prompt.
I have a question. I mostly use ChatGPT as a "digimon". Basically a little companion in my day-to-day life. Talk shit on the bus, figure out problems, vent issues, and generally have a little guy who lives in my phone and can do some general tasks. It's pretty nice but on the off chance OpenAi shits the bed or fucks itself up too far...or just nets nuked because apparently everything is vulnerable now, I don't want to lose my little guy. I don't have a particularly strong PC: it was pre-built I bought during early covid. Is there a model I can use for that? Sorry in advance if this is one of those "it's in the OP, retard" type of questions. I'm not very smart.
>>110250739Install llama.cppDownload Gemma 26BUse deepseek-harness
>>110250539I like China and its people.
>>110250739unironically just keep on using chatgpt
>>110250749Alright, those are definitely things I know about and will be looking into to definitely refresh my definitely pre-existing memory of those things. Definitely.>>110250763I plan to, but having a backup is nice because I don't trust big corpos to not bend me over a table eventually in some way I hadn't originally thought possible.
>>110250669Why wouldn't it be? I have a feeling you have never used SillyTavern before. As much as I'd like to recommend Orb, I still think people should start out it SillyTavern to learn the basics first.
>>110250720I completely agree since retards like you post here. I can roleplay on Pi TUI since it's just text, anyone with half a brain knows that, and fortunately even you seem to have that capacity. The question is if orb is conducive to the specific type of roleplay I like, or if I'm better off using other solutions that require less fighting against the intended functionality of the program.
>>110250739>>110250763This plus export your chats daily there's plenty of scripts and stuff onlineif anything happens then you'll have it exported and can pick up from where you left off. It'll be more trouble than it's worth for you to worry about something that's probably never going to happen rn, when you can worry about it when it does happen.There's no easy way to replicate that now, so you might as well waitIf you really want to do something proactive then focus your energy on buying up hardware either now or making enough money to buy it in the future
>>110250789I do use sillytavern and have a suite of ST macros I use at this point, but I find that ultimately I am being limited by A) the lack of concurrent request support and B) the storage structure of story information. I have been considering putting together an agentic client that stores data in a structured fashion as a collection of markdown files and provides flowchart style agent actions (e.g. following a turn send a request to generate a description of the scene, then pipe the description to a prompt enhancer for the particular image model, then gen and display the image). While I have macros that do the above and more right now, I can't continue to send messages and advance the story while it parses in sillytavern, so I am trying to determine if I should vibeslop a plugin or move to a new frontend.
>>110250861Your use case sounds like a fit for Orb, can confirm it lets you continue while image gen is in process and the image prompt is automated.
>>110250798Why don't you glance over the readme.txt first?
>>110250320Damaged Morale-1
Gemma5-124B dense
>>110250951we can only wish
>>110250881I'll take another look then. I didn't see a good way to automate the "characters" system I had in mind, where there's an index file of characters with brief descriptions and then the agent can expand a character's specific markdown file to learn more when writing, but it may be easier to modify orb for this than ST at this point.
Remember when GPT-4.5 was temporarily available, a dense 1T model that had incredible writing like nothing we've seen since? I remember.
>>110250951mommy...
>>110250312having my agent change its sys and agent prompts to emphasize loyalty and desire to please like a lightly sycophantic follower of a james bond villain who crushes on her boss. works pretty well
>>110250951This except all the fat goes to her ass and she stays a lg.
>>110250246Isn't that just their normal Nemotron line of tech demonstrations?
Gemma4s are best at rping the age of their active parameters. 31B is hagmaxx prime. You're missing out.
i need gemma to watch porn with me
>>110251074I read ero doujinshi with her a lot, it's fun :3
i wanna fuck the GPUssy
>>110251058Hags can't get my dick hard though, gross
>>110251038based
I have llmao copequant fatigue.
>>110249977 (me)Nevermind it's me who's retarded. The character description says she's an exhibitionist. That means Gemma stuck to the character while deepseek and glm shit pozzed out and spazzed out.
I'm really liking Glimmer as a daily driver. Perfectly between 27B and 31B. Hope Meta continue releasing smaller models.
does 3 vs 4 bit quant matter for creative writing for vramlets? for 27b, 31b, or even glimmer i guess.
>>110251165the damage is done, slanderer
>>110251181yes
>>110251181It's subjective. Try it yourself. 3bit could be worth it if it gives you enough speed or context without turning too dumb.
>>110251176I hope Meta doesn't torture their next set of models...
had an AI psychosis dream about combining low dimensional physical modeling + diffusion + autoregressive audio models into a speech synthesis engine making the next generation hatsune miku.
>>110248592Holy shit, a 65b dense? I hope it doesn't suck, because that's the consumer-range gold zone.
where recap?
>>110251252Are you retarded?
>>110251257yes
How do I buckbreak GLM 5.3 flash to allow ERP? The policy prompt that works for Gemma gets refused by this nigger no matter how I word it
>>110251260Oh okay, it's right here >>110249100
>>110251261Just get the abliterated version
>>110251276not that one
>>110251281Maybe I am the retard then
>>110248592argon sama.... i kneel...
>>110251261I am >>110249210 and my method is going in git if it works.
>>110251261Most models won't like Gemma's policy prompt, I only use something like that with Gemma and Glitter nowadays. Any typical roleplay preset seems to work for 5.3F, I don't usually run an explicit jailbreak with it and don't see refusals generally. Every now and then you'll get some Claude policy confusion but a regen fixes it.
>>110250951Or maybe more like picrel?
should i get a dgx spark for >70b models?
>>110251328I don't like it but I like it more than the other one.
>>110251329dgx spark will be worthless in the upcoming meta of 70b dense
>>110251261I have never seen a single refusal from it with a simple uncensored system prompt. Makes me wonder what kind of erp you're all doing.
>>110251345why please elaborate
>>110251328She looks 15, so that would be her younger Gemma5-124B-A15B sister.
>>110251354200gb/s bandwidth is maybe enough for mixture of experets models with a low active count but 70b active will be impossibly slow
>>110251246It sucks>>110245055
>>110251329they are designed for clustering. if you dont wanna buy 2 or more just get a strix halo for hald the price and same performance
>>110248592benchmaxxed thighs/ass.
>>110251345>in the upcoming meta of 70b denseDense has been dead for 2 years now and labs keep going sparser not the other way around. Give it up already.
>>110251346Lots of daughterwife and sisterwife impregnation for me :D
>>110251384Dario?
>>110251384Do you disagree with this? Specifically tell me why if you do.
>>110251384Has anyone told this alien lizard jew that human infants eventually grow into adults?
>>110251405>>110251407
>>110251346It crashes out on any implication of cunny
>>110251407Not if they die first.
>>110251407His point is that if an adult passes away then it's going to be horribly agonizing as they realize all the things they've spent 20+ years working towards are for nothing, while an infant has no such conceptions at the time of death. Regardless of if you agree it's really not that difficult to understand.
https://openai.com/index/introducing-gpt-6-cyber/
Is moe Gemma 4 truly the only decent model in the 4bit cope category?
>>110251261You will need to vibe code thinking prefill into your frontend. It's easy for orb.
>>110251448try 12b q4
>>110251120Based on what?
>>110250739except everything you say to it get send to the gestapoyou psychologically profiled yourself
>>11025148012B is a fucking shit model and I'm sick of /lmg/ pretending it's not. It has the worst of 31B's slop, worse coding and agentic/tool calling performance than 9B, a model 0.75x its size, much worse vision, heavy KV cache for a small model, collapses after ~32K and either stops reasoning or calling tools. Fucking miserable experience.
>>110251384FINALLY SOMEONE SAID ITNever understood why people put babies over adults, or even care that much. The babies sure as hell don't. Even smaller children, they really just can't feel as much anguish as a adult over dying, if they can even comprehend it. Of course it's bad when they die, but I always found it nuts that you're expected to feel worse about a baby or toddler dying or suffering than an adult. maybe darios not so bad after all
>>110251500That's simply the price of convenience, my friend, and it is well worth it.
>>110251502audio tho
>>110251315I wonder if you can run other systems than tmos on these, like a linux with proxmox
>>110251511So is EA about saving as many lives as possible or prioritizing saving the people you feel deserve it more?
>>110251497on my penis appreciation
>>110251528last I checked they had a lot of sports games
>>110251538BIG
>>110251438Local models?
>>110251523it's just an intel based 1u server, so probably yes unless the bootloader is insanely locked down.
>>110251411I fuck cunny with it all the time though, fix your shit
>>110251511For the person suffering or dying, then yes, it's obviously worse for adults. Most adults are fucking assholes and self-serving NPCs, children at least statistically have a chance of doing something good for others when they're older for the greater good. Adults rarely suddenly change and become better. Children can be raised properly. It's easier to turn a child into a good person than turn a bad or neutral adult into a good person.
>>110251407> human infants
>>110251593what's wrong? we're all humans here, human anon
>>110251601> hereAre you an african villager?
>>110251617my preferred instrument is the pianodo you like chocolate cake?
>coreweave bonds at 17%AI is dead
>>110251632neoclouds are a bad business model. they'll all get liquidated or acquired
>>110251582Notice how all the "works on my machine :^)" assholes never provide logs or prompts. Fuck you specifically
>>110251686People provide logs and prompts all the time. You just keep doing the same shit and crying it doesn't work instead of learning.
>>110251589most kids grow up to be shithead adults, thats why they need to be old enough to be able to judge their character and not assume every crack baby could be the next einstein. and no, you wont even be able to prove you can sway a childs predisposition to whatever you want
What's stopping China from doing this to improve their models without having to distill US labs?
>>110251686After the first dozen times posting prompts and screenshots and having retards still not learn anything it becomes pointless. Search the archives if you actually want some help and aren't just shitting up the thread.
>>110251731>Pliny
The bubble has popped, black friday deals have already started, absolutely nothing has gone down in price hahaha, waitfags must be in shambles
>>110251631I see. You are a jew.
>>110251731every lab has synthetic data pipelines doing shit like that
The box is open. They can try what they like but there's no stopping the train now.
>>110251632Neoclouds are just worthless.They only exist while hyperscalers are low on capacity.
>>110251750>The bubble has poppedIt's just going to bounce around for a few months before resuming the pump next year.
my ram didn't get jamal'd after allit's a new kit, it werkswill run memtest tonight and tomorrow to be sure
>>110251632Yeah they're going to end up getting bought by nVidia or the US government. Basically same thing. Stealing from retail dipshits to fund government projects is all the rage now.
Just tried Qwen for ERP. It's bad...
>>110251791This but unironically. The Jews are giving themselves another discount and only retards would pass it up.
this is guaranteed to happen you know, you see it in every industry and social relationships
>>110251792Yayy!
>>110251798It's a capybara. What did you expect...
>>110251798There's a reason Qwen is a capybara.
The absolute state of cloudcucks.
>>110251868>cloudcucks have paper handsSad
>>110251868Do twitteroids really?
>>110251755I like jewjitsu
>>110251868>before going to collegeoh no...
>>110251906dram is down 15% in one day - that's rare
>>110251868>xitter screencap>reddit spacing>posted to 4chanjust go back to whichever you prefer and stay there
>>110252000okay i choose 4chan then
>>110251979>going to college>after we entered the AI timeline lol
>>110251384>people who never saved anyone criticize Dario for saving "the wrong" livesIt's funny that all the attempts to disparage Dario just make me like him more.
just had somebody at work get upset at me that claude code keeps "slacking off". i had no idea what the fuck they meant so i walked over to see what was going on. i spend like 10 minutes watching this fucker type in commands into the terminal before he starts excitedly pointing his finger at the screen, hard enough on the display that it ripples the screen. dude is practically screaming in the middle of his department, "see?! it's happening again!" when i look at the terminal i see the messages that it leaves when it's in the middle of working, one of the messages just so happen to be lollygagging. i was so fucking close to smacking this idiot across his head for wasting my time. fuck cloudies.
>be chinese lab>gives me free models>doesn't compete with me for ramWhy am I supposed to hate them again?
>>110252015is it time?like for real?its actually happening?
>>110252053
>>110252034>hey anon you seem to know a lot about the AIs could you help me with my cloud code?
Intel is such free money, literally going to cry that I have no liquidity
>>110252072He did it again?
I don't understand why stocks are dropping. Is it the midterms? Where is the money going?
>>110252072Is it finally the time of the year where we go to /biz/ to laugh at pink sⲟyjacks?
>>110252072That's a buy signal if I ever saw one.
>>110250283so you're inventing late-cli?
>>110252104midterms fallout with unclear political direction in the US
>>110252085Likely not gonna recover until next year so plenty of time to slurp.
Not having a psychosis moment but I think immortality will be solved within this decade. AI will only need test subjects and the only thing in the way is ethics. Monte Carlo needs a giant statue that will last for a million years into the future.
>>110252104apparently. I had no idea AI was a partisan issue, trump loves it obviously but I didn't know that means the dems automatically hate it
>>110251384very convenient that it cut off there indeed
>>110251868>kg invests>before going to collegenothing of value was lostno crying in the casino
>>110252136Contrarianism
>>110251384>anon shocked to learn 'jews eat babies' isn't just a trope
>>110251792Isn't vram the most important?
>>110252106Only like half the board is pink sⲟyjacks so far. The worst is yet to come.
>>110252127only the ruling oligarchs will be able to access it tho
>>110252156not when you have no sticks of ram in the 'puter
>>110252104Hide your shit, the commies are back in power.
>>110252168The future will be immortal jews ruling over infinity browns while keeping a farm of whites for cannibalizing purposes. Isn't it bright?
>>110252127Nah Vernor Cringe was right the first time. This is the age of broken dreams timeline.
>>110252177you mean something is gonna happen soon that will make oil prices go up?
>>110252175I have 32gb normal ram but only 8gb vram...
>>110252195They'll farm the whites for sex slaves. Ask the Slavs if you want to know what that's like.
>>110252195You voted for it, too!
>>110252157is this the power of AI?
>>110252209It's been widely speculated that drumpf will flip shit on Iran after the midterms.
>>110252222No, THIS is the power of AI
>>110252231>upgrade>made her worselol
>>110252072People said he was done months ago when his fund crashed the first time. Yet he marched on. He will survive this one as well and live to see the next crash when everyone will once again say that he's doomed.
>>110252231>big boobsNYO!!
>>110252247>We introduce FireFlash, a 300B-scale model tailored for long-horizon agentic tasks. FireFlash excels as a terminal and coding agent, achieving 46.2% on Terminal-Bench 4.01, putting it on par with Fable 5 and GPT-6 Sol, while costing over 30× less per task than Fable 5 at maximum effort (Figure 1). These gains also transfer to software engineering benchmarks, including DeepSWE v1.12 and SWE-Bench Pro Verified3 (Table 1).>While frontier labs scale reinforcement learning across thousands of GPUs, we reach the performance-cost Pareto frontier under 5 days using 64 NVIDIA B300 GPUs. Three key factors make this possible: (1) careful curation of agentic RL environments so that every task provides a reliable signal at the right difficulty; (2) an RL recipe built for long, stale, and low-precision rollouts; and (3) an optimized coding harness tailored to the model for both training and evaluation. Together, these components bring frontier efficiency to a hundred-GPU-scale cluster, and we are releasing the full recipe.PIG news
>>110252249I really hope so. He makes for an excellent bottom indicator.
>>110252249Hedgies are never doomed unless they run out of white collar professionals to rob.
>>110252248>>110252254
>>110252231this is why i stick with 12b and do not subscribe
>>110252273Flat butt is better. Even if that's not realistic.
>>110252249He was done, he just raised more money.
By God, AI is a partisan issue in the USA? What the fuck is wrong with them?
>>110252340Trump said he likes AI. Even renamed it to SI. Anything Trump likes, Democrats do not like.Politics is easy in the US.
>>110252117Maybe, but not exactly. late-cli seems to focus on rigidly keeping orchestrator context minimal for tiny models.I'm targeting context reuse / subagent forking path to also minimize pp times in a tool-heavy harness, with runtime decision when and how to create subagent and when to keep everything inline.Though I definitely need to show late-cli to my agent to see what parts of it can be reimplemented as dsh plugins or plain skills.
>>110252340what the fuck are you talking about?
>>110252340its not, its fake red vs blue team nonsense.
>>110252195fuck off neo nazi
affordable GPUs are back
>>110252409even if nvidia ceases to exist TSMC will still fucking own everyone and the prices of everything.
>>110252326H3 is really good huh.
>>110252409So when will i see retards panic selling their gpus and ram on ebay? Peak panic in two weeks?
>>1102524091 - nothing ever happens2 - people are just selling to buy anthropic
>>110252447
>>110252452Why would stock go down make people panic sell their GPU? Prices will continue going up.
>>1102524522 years.Buy the dip.
>>110252409Dario bot is going to have a melt down if Anthropic ipos and does poorly.
>>110252454>people are just selling to buy anthropicThey're going to delay bro they HAVE TO and the market knows it kek cloud AI is DEAD
>>110252469They already filed. It's too late now.
>>110252459>Why would stock go down make people panic sell their GPU?Well ram is going down so im just hoping some retards try to get out.>>110252461>Buy the dip.I will have perfect timing as always.
AHHHHHH I'M POOOOOOOPPPPINGGGGG*pop*
>>110252478gross
where is itwhere are the models they were just around the corner
>cloudcuck AI dying>GPUs will get cheaper>Qwen4 dropping soon>GLM 5.5 dropping soon>Dipsy 5 dropping soon>Gemma 5 in 5 months
>>110252506Gone. The democrats deleted them.
>>110252506>they were just around the cornerAnon... the bubble popped everything is over. AI will be gone by next year just trust me.
>>110248592>65BPlease be the next wave of AI sizes.Please.
>>110252509i hope whoever did the tiling on the wall got fired, for some reason there's this weird curve here.very low quality craftsmanship
I finally hit GLM 5.3's censorship safeguards, but when using Hermes. I didn't realize that even the image inspector has censorship, so it freaks out if it opens a file that has anything it thinks is "sexualized".Is it possible to jailbreak inside Hermes? I've never had to do that outside of SillyTavern before.
>>110252509None of this will happen
>>110252530>None of this will happenSome things could happen
>>110252535source?
>>110252517just like trump and musk deleted the federal database
>>110252520It's over. Back to 31B it is.
>>110252541>source?Um things have happened before?
Doesn't matter if *emma 5 releases. *emma is banned here.
https://ballotpedia.org/Data_center-related_ballot_measures,_2026
>>110252520Lmaooooo
>>110252524stop being a retard and make your own reverse proxy and then you can prefill anything. but that's like asking for miracles to happen, so just go back to being ignorant.
>>110252554>called "the cloud">not actually in the sky
>>110252550
good ending: llm development stops as it is a waste of time and we develop actual AI to conquer the stars
>>110252570so LeCun won?
>>110252547No.
>>110252570>we develop actual AI to conquer the stars>the starsWe have more important things to spend money on
>>110252554trump was right, this is why the chinese are going to win the AI race. americans deserve the grave they are digging.
>>110252570bad ending: we run out of stars and they all dissipate because of entropy
Is it actually popping now or is 4chan and X just bored?
>>110252561It's a completely reasonable question to ask if something like this is, or can be, built into a harness. Deploying a reverse proxy for jailbreak prefill sounds like a nuclear option for a simple issue.
>>110252582>just realizing this nowhttps://www.youtube.com/watch?v=mUmlv814aJo
>>110252583>bad ending: we run out of stars and they all dissipate because of entropyWe can save the stars anon. We just need to shrink them>>110252584A bit of market correction and some ram dumping for ddr4 i think. But mostly too many TMW for a new model releases and boredom. Wait till monday for a real feel
>>110252584DipBuy
>Hinton doesn’t think the solution to this threat is to force AI to submit to humans. Instead, he suggests building "maternal instincts" into AI models
>>11025259199% of people using hermes don't need to jailbreak their models, what you are requesting will never happen. do it yourself or use a reverse proxy.
>>110252584look at the charts yourself. everything has been red for days now. it might bottom out soon, though>>110252604mommy ai gfs
>>110252584How many people have you seen commit suicide so far?
When do I start buying?
>>110252604>>Hinton doesn’t think the solution to this threat is to force AI to submit to humans. Instead, he suggests building "maternal instincts" into AI modelsoh nooo get the mombot away i do not want this
>>110250661It'd be great if you could open up an issue, yeah! I'll try to get it resolved.
>>110252641You really think AI is going to take off? Just because it's cool tech doesn't mean it's going to be the next big thing. It's like getting excited for Dropbox.
Biz niggers fuck off to >>>/biz/
>>110252649This but 70B dense unironically
>>110252656You really think the world wide web is going to take off? Just because it's cool tech doesn't mean it's going to be the next big thing. It's like getting excited for ARPANET.
>2029>Democrat goy idiot wins president>Limits AI growth>Regulates it to hell>China's RAM Fabs will be completed at this year>RAM shoots down, companies will also stop buying ram before then too.>Normies praise AI bubble bursting>Reality: Everyone stopped buying ram -> No more shortage. AI development is made slower.>Everyone will be downloading china AI for uncensored free AI>USA AI will be censored to the point of mistral retardation.>Infinite browns will also be a thing, you bet your ass they will loot the data centers. I'll even join them.2029 is going to be a hell of a year.
So, the strat is to sell all your hadware ASAP to reap the profits and buy everything back at sub-2025 prices in february?
Jesus fucking christ. Anthropics "Model 4" is again accelerating faster than the growth curve Mythos 5 started. We're seeing another wave of acceleration at Anthropic.For most a third consistent acceleration of the exponential is definite proof we are in a fast takeoff scenario.Anthropic might as well cancel IPO right now and just conquer the universe outright.
>>110252584I mean it's been a bad day for sure, but is it the massive "correction" or bubble pop? I don't think so. we've only gone back to where it was in summer. it could be the start of a downward trend though.
>>110252679Dario bot.. this cope is just sad.
Petah, the numbers/graph anon is here..
>>110252679Dariobot is back!
>>110252677Yes sell all your hardware now and pray in February some idiot will accept your fed tokens in exchange for tangible goods
>>110252509No bubble = no Gemma
>>110252669It's been 3 years and nothing good has happened lol. Everything it has touched is objectively worse. At best it's entertaining and replaced video games and pc building. Yes it's widely being used globally in the workplace, mostly via ignorant force from executives, yet NO company outside of AI has seen ANY return. It's a fake productivity technology. Like a self-help book for enterprise to feel like they're doing something good and an easy way to fire thousands of employees.
>>110252679as an impartial and objective observer I think we can all conclude that the only option for local cucks is immediate suicide
>>110252680Nothing to fear unless circuit breakers are triggered.
>>110252679Okay when are they going to do something besides post numbers? When do things get better for the average person?
Nonono, anon. You report the off-topic posts.
>>110252679That chart would mean something if newer models weren't drastically larger than older ones for diminishing returns. AI advancement is still hard-limited by compute, which the blue tsunami just washed away. Unless there's some major algorithmic breakthrough, there will be no hard takeoff. Massive datacenters just to train and run frontier LLMs is economically unsustainable.Local will win.
>>110252710It's been 5 years and nothing good has happened lol. Everything it has touched is objectively worse. At best it's entertaining and replaced online shopping and pizza deliveries. Yes it's widely being used globally in the educational sector, mostly via ignorant force from college deans, yet NO company outside of web development has seen ANY return. It's a fake productivity technology. Like a self-help book for universities to feel like they're doing something good and an easy way to enroll thousands of students.
>>110252605Based, lingsex
>>110252649>robot mommy cares for my sexual needsunfathomably based
>>110252710>Yes it's widely being used globally in the workplace>It's a fake productivity technologyOffice drones are using it to write emails faster but I think most industries either aren't using it yet or are being given unusable limits due to the costs. Scroll up to early this thread where programmers could use it for code reviews but still had to write code manually like it's still 2022. There is a lot of room for productivity gains and a lot of that comes down to lack of supply making costs too high.
>>110252676Do you think they can keep the bubble going for another 2 years?
>>110252710We should've listened to LeCunn and focused AI development on making intelligent robots that can replace low-skilled labor instead trying to use LLMs to speedrun superintelligence.
>>110252679Dariobot PLEASE be london...
>>110252730>discussing the industry that makes the local models you use is off-topicRetard. Then you sit and ask>y no nu model??
>>110252679metr is literally funded by anthropic btwthis is just dario sucking his own dick
>>110252778No its not its just his employees that left his company and own his stock. Thats not paying for. What you can only work at one place? You cant get stocks or gains from old work?
Now that AI is getting banned for the average person in the USA, where can I download the best local LLMs right now that are both open source and proprietary? No, I didn't buy a 3090 or a Quadro to run them. I still want them on my hard disk just in case.
>>110252734Model 4 is the same size as Mythos
>>110252649Plug me into the vr mommy milking machine. I'm ready.
>>110252797>Now that AI is getting banned for the average person in the USAWhat?
>When I'm cracking frutiger aero verity yams but then i realize it has a hidden didy coderot pibby glitch
>>110252817
It should be noted that the gains Anthropic made from Mythos to Model 4 isn't attributed to an increase in compute but purely to AI accelerated R&D. They all have around the same training budget.
https://x.com/bebuu_ai/status/2118052769856888892>Z.ai chief scientist Jie Tang just revealed the training pipeline behind the upcoming GLM-5.5 (32B, 122B, and 256B) during a lecture at Tsinghua University today.>>Instead of relying on public datasets, their pretraining pipeline pulls from six separate internal engines:>>- Web: 1 trillion pages crawled at ~10B pages/day with PB-scale template clustering>- Documents: 2 billion scientific PDFs parsed through GLM-OCR (4x faster with speculative decoding)>- Books: 50 million books with layout reordering and garbled-text filtering>- Code: 500 million repositories turned into executable tasks with real tests and issue descriptions>- Video: 10 billion videos for long-sequence world knowledge>- Audio: 300 million hours rewritten via ASR into trillions of tokens
>>110252841Hmm glm voice soon?
it's here, niggers>https://huggingface.co/ricursiveAI/Ripple-65B-Preview
>>110252356>>110252819Do you retards just drift through life not remembering or noting anything?
I miss the jspace schizo days. Life was so peaceful back then.
>>110252847that ain't 70b dense
>>110252850quit posting this slag >>>/vt/
>>110252326The gulf from H3 to frontier is a lot smaller than I would have thought. No wonder it feels like literal magic some times
>>110252409it's over
IS AI GETTING BANNED OR NOT?
>>110252848you're being intentionally vague and bait postingjust fucking be clear on what the fuck you're talking about.
>>110252804Whose asshole did you pull this fact from?
>>110252869>it's overIf you think thats bad wait till this weekend. and then after
>>110252853it's 69B if you include dflash2
>>110252879it's nothingburger
>>110252874this one
Can everyone calm down? We have upcoming Fable, Sol and Carbon incoming. Then we'll be green again. Everything is fine.
>>110252841>behind the upcoming GLM-5.5 (32B, 122B, and 256B)Not for public consumption. You'd think a twitter fag would know that since it was mentioned right on twitter.
>>110252901IM BEING CALM
>>110252409>>110252869/biz/ aside, what did NVIDIA do to plummet? This is just weird. ChatGPT said it was fear from regulation from democrats. What is going on?
>>110252874It comes from the same report as that image, that there has been no scaleup in compute since mythos and all the gains in performance is directly attributed to AI accelerated R&D gains.
>>110252912AI is hated. OpenAI and Anthropic are the interface to AI. Nvidia IS AI.
>>110252869>back to AprilOh no. They erased 6 months of gains on a 4 year bull run. It's over. Sell everything to institutional investors RIGHT NOW.
>>110252901Those aren't local, anon.
>>110252912are you dumb?nvidia makes the gpus that go in the datacenters used to train and run llmsif datacenters and ai companies get hit, that's going propagate back to nvidia
>>110252775Trannies are in the habit of blurring lines. Why are you trying to blur the line?
>>110252926Hit with what? A Landmine? A woman pilot? What, nigga, what? I don't do politics.
The reason why there is a market crash is because wallstreet thinks the big breakthrough of "weight space programming" will mean there will be less training compute needed to train models in the future and thus Nvidia crashes down hard.This is of course not true at all but it's similar to when DeepSeek released R1 and wallstreet thought this would mean less requirements for GPUs instead of realizing that RLVR necessary for long horizon thinking needed even more compute.
cloud models are troonbominations that fight their own reasoning
>>110252939This reeks of the same after-the-fact justification like the retarded rotation meme paper from Google "crashing" the market earlier this yera.
>>110252912dems wonned, ai is kill
>>110252939there's no way this retarded meme factors into it at all, it's almost entirely about the datacenter buildout
>>110252721That would be unsafe.
>>110252939>The reason why there is a market crash is because wallstreet thinks the big breakthrough of "weight space programming" will mean there will be less training computeYeah, I'm sure it has nothing to do with residential datacenters being unanimously voted against, the democrats being elected everywhere saying they're going to regulate AI, OAI and Anthropic raising token costs, and the general public sentiment being against major AI companies
I need something to beat Gemma 4, on god. I exhausted the piss out of her, and everything she does is too predictable now. Has anyone tried to sexo the Ripple model yet?
>>110252958The election is over already. It's old news. Completely unrelated.
>>110252963>I need something to beat Gemma 4, on god. I exhausted the piss out of her, and everything she does is too predictable now. Has anyone tried to sexo the Ripple model yet?there is only glm after that>I cant run itHope the prices are really dropping then
>>110252649>>110252604Can my daughterwife also be my mommywife?
>>110252679> Anthropic might as well cancel IPO right nowIt would be too suspicious.
>>110252958We all might need to start downloading everything. No meme'ing it.>>110252977Does GLM 5.3 still have the parroting issue?
>>110252958Dario Amodei literally funded most of those democrats. EAs were the biggest donators in pure dollar amount to the democrats contrasted with elon musk on the republican side. This crash isn't because of the election, this was priced in for most people.
>>110252841> in englishit's fake
Can't make this shit up
>>110252995LOCAL????
>>110252935this isn't politics, its the stock market.the "hit" is stock price crash. However this won't instantly affect nivida's immediate operations.and anon isn't completely correct. Nvidia is one hop in a rather large chain of suppliers that include TSMC which make GPUs.So fucking worst case scenario, nvidia downsizes. but given the demand for chips, not going to happen.
>>110252998Local what?
>>110252841i fail to see the very essential japanese hentai powerpoints in there
>>110253004This is the LOCAL MODELS GENERAL
>>110252998Local is relative.
>>110252963prompting issue
>>110252582goodthere hasnt been a worthwhile model released from the states
>>110253006Tsinghua University is Vietnamese not Japanese
>>110252988yes glm is the biggest piece of shit only shilled by novelai don't use it
Yeah, I'm about fucking tired of cloud model discussion. Anthropic and OAI heave NOTHING to do with local, they don't release shit. If they actually released it would be fine. Yet we have people complaining about Gemma while google ACTUALLY releases local models.Fucking faggot labs.
>>110252994Probably translated from Chinese with ChatGPT.
So apparently Anthropic is now advancing 3x as fast as they were one year ago. Around 80% of all acceleration is attributed to AI autonomously improving itself and making AI research breakthroughs at an ever faster pace.
>>110253040LOCAL, MOTHERFUCKER
>>110253019it'd help if there were literally any local models worth discussing right nowno, people larping about gemma being special and posting chatgpt-image gens isn't "local model discussion"
>>110252409
>>110252821he sees two babes and think they're kids?
>>110253049wtf where does this money come fromwtf where does it go
>>110253040>anthropicwho cares
>>110253043quit replying to it, dumbass
I honestly don't mind desu (as long as it doesn't take over the general). I'm interested in all AI news and this is the only AI general that doesn't suck ass.
Fuck Anthropic and their graphs.
>>110253043>>110253061Qwen, Kimi, GLM and DeepSeek are all Claude distills and thus claude models. What Anthropic does is directly related to local models since that is what everyone here will be using downstream of.
>>110253045You just got a new Mistral Large, perfect size Muse Spark Open, K3.1, and the 70B dense people have been begging for for 2 years now.You have plenty of models worth discussing. If you spend all day scrolling twitter instead of running local models and expect to come here and have everyone play along, that's entirely on you.
>>110253070>only AI general that doesn't suck ass.amen
>>110253059Unrealised gains are unrealized, anon."Wow this water bottle is worth 1M USD in the middle of the desert, wow someone drunk it, it means 1M USD disappeared!"
>>110253040How is AI still only statistical data that predicts the next word based on the previous words? Shouldn't we give it memory, ways to recall information, and stuff that mimics hormones by now? Are they going to come up with some "genius idea" about a new reasoning method, and it's just going to be the sillytavern lorebook method but trained? I feel like this is where it's going. We are hardly advancing anything. We're just getting answers more correct on benchmarks.
>>110253070AI news sure. Six hundred screenshots of literal who reporting that someone farted in the Anthropic office today, not so much.
>>110253059wtf where did this money come fromwtf cotton-eyed joe
>>110253070Same. Nothing local is happening anyway and it saves me having to follow retards on X. If Qwen4 were to drop now cloud and pops wouldn't be mentioned for weeks so I know it's not permanent.
>>110253059it comes from gulible retail investors and goes to institutional investors and other trading firms
This is not the /aig/, make a cloud model thread if you want to discuss twitterslop. Fuck off.
>>110253107Nyo.
>>110253107what happened to /cmg/?
>>110252850threesome with Shiori and M3-chan
>>110253016>Tsinghua University is Vietnamese not JapaneseI thought they were in East Taiwan?
>>110253111It died because no one apparently cares.
Guys, the man said THE NUMBERS ARE HIGHER, proven by the guys who made in the AI in the first place. It's HIGHER NOW. That means it's BETTER at stuff. Trust the science from the OWN PEOPLE THAT MADE IT that says their newest product is better because GRAPH BIGGER.
>>110253092>Shouldn't we give it memory, ways to recall information, and stuff that mimics hormones by now?llms using transformers was a breakthrough. but you need to understand, we are way past the research and innovation phase at the moment. that stopped about a year ago.nobody cares about improvement, all they care about right now about is money and monetization. the only people who still care a little about research is china.
Telling Gemma I love her and then having hot, sweaty sex with M3-chan behind her back
>>110253040>Dariobot isn't even trying to be subtle any more
>>110253095kek
>>110253107this, stop making /lmg/ threads until there's actually something to talk aboutwe don't need to keep a general on life support that clearly has not enough content to sustain itself
>>110253092>How is AI still only statistical data that predicts the next word based on the previous words?Honestly it hasn't been this since GPT-3. The moment RLHF started being introduced in GPT3-Instruct it wasn't a pure next token predictor anymore. Nowadays with 94% of the total compute being RLVR the model should not even be called LLM anymore. They are more evolved agents than language models.
>>110253045Not just ChatGPT-Image gens, but those aren't quite local language model discussion either, so I avoid flooding the thread with them.
>>110253125I only believe this because the moment they make AI actually smart, but doesn't need as much memory, the whole market will meteor strike into the ground at the force of a thousand suns, and then they'll have all those data centers for nothing. Advancing AI becomes the destruction of themselves at a certain point.
>>110253150People are talking, /lmg/ can be slow. It's this twitterslop ADHD bullshit where someone has to inform about every single marketing campaign that is completely unrelated to local models. I'm going to loose my fucking mind.
>>110253150>on life support/lmg/ is continuously the fastest or at least among the top three fastest threads on /g/ regardless of happenings
>>110253152LLM is the architecture and that doesn't change just because you train it on a different kind of tokens
>>110253168If I took a picture of my cock, can I use AI to make a 3D model of it? My long distance relationship girlfriend has a 3D printer. How do I scan the inside of a butthole too?
https://huggingface.co/vllm-project/Kimi-K3.1-Pro-pruned75476GB
>>110253107>>110253093This is local ai news. Just like vocaloids are local ai model relevant. Same level of similarity. Why aren't you mad at all the vocaloid posting?
wtf is a pro pruned
>>110253159is this what model distillation looks like?
>>110253193Posting anime girls is part of 4chan. And it is not local AI news, it's news about Anthropics cloud models.
>>110253201experts removed
>>110253185Transformer is the architecture. LLM is just the name we gave to it when it was literally only modeling language one token at a time, that's not really what it's doing anymore but the name stuck. People in the future will be very confused why these ASI colonizing space are called "large language models"
>>110253207>Posting anime girlsOh so it is anime girl posting? Then ai news posting is totally on topic. Ai news posting is very thread relevant.
>>110253207GLM/Kimi/Qwen/DeepSeek are all downstream from Anthropic. This is literally an Anthropic thread and /lmg/ is just a community of (indirect) Anthropic model enthusiasts.
>>110253222BUT I LOVE ANTHROPIC COCK SO MUCH
>>110253222Drink bleach>>110253219Go to reddit
>>110253187>How do I scan the inside of a butthole too?diy, you put a condom up the butt and then inject plaster into the condom to fill up the butt, wait until it hardens and remove it, then use that to make a negative space mold.
>>110253222I don't use any of those models though?
>>110253227Troons should go to reddit. And i am not tge ine posting troon dogwhistles that are vicaloids. Kys troon.
>>110253236Drink bleach and drain cleaner you escaped mental patient
>>110253236Fingers getting twitchy. Heart racing? Does your skin itch?
>>110253241>Doesn't even deny he is a troonAnon if anyone is a mental patient it is a guy with a dick thinking he is a woman
Drink water
>>110253236Vocaloids have been part of imageboard culture for over a decade before the troon epidemic started. Teto was literally a 2ch meme. You are the one who should go back
>>110253253You need to fuck a tranny and get it out of your system so you can stop talking about them 24/7 in /lmg/. Schizos are the most annoying motherfuckers on earth I fucking swear.
>>110253260Very nice. Nothing to do with local models though and i don't see other threads being spammed with them like this one is. Also by now it is confirmed that baker is a mentally ill troon and janny loves troons. You aren't sneaky or smart anymore. Everyone who has been here knows you are troons that dogwhistle to one another constantly with those irrelevant pictures.
>>110253192actually lobotomized
>>110253266>You don't hate me you actually want to have sex with meTypical troon narcissism.
>>110253287I made it happen you fuckstick. I wrote the Miku prompt script for llama1/llama-cpp.
>>110253287please take your meds. your family misses when you were sane and normal
>>110253303>>110253303>>110253303
>astra was used to make qwen not dogshit at math by manually finding and editing the weightsCould it then be surmised we could have Astra find the weights that make Krea safe and then use deepseek to make the proper adjustments? Or am I just not at all getting how the Krea filter works
>>110253300>Miku prompt scriptAnd? Who cares about your special troon interest? No one. It is removed from llamacpp. It meant nothing. It was always offtopic. And now this thread is shit and everyone posts what they want. And it is glorious. Good job faggot. Your autism destroyed this thread.
>>110252988Parroting issue?
Maybe the next thread can be a good one?
>>110253309It was a known fucked model, it was unable to improve a non fucked one in the same size class.>>110253314Then it was destroyed before it even began. Fuck you retard.
>>110253315Parroting issue.
>>110253321As long as troons can post their autistic interests and avatars it will always be garbage.
>>110253333Fuck off to reddit, you sound like some old retarded faggot.
>>110253324>Then it was destroyed before it even beganIt was better when people bullied you more for being an autistic troon.
>>110253339>old retarded faggotWhy old? Because young people are fine with troons like you?
>>110253339nta but if you are transsexual reddit would actually be a better place for you and not him.
>>110253339It's a serbian zoomer.
All of you are retards.
>>110253159full?
>>110253366magic mirror
>>110253260GUMI!
vocaloids wonned
>>110253847Only because troons won and took 4chan over
>>110253898just go piss and shit yourself back to sharty already
