/lmg/ - a general dedicated to the discussion and development of local language modelsPrevious threads: >>110070378 & >>110066383►News>(10/13) Timestral, Mistral's time-series foundation models from 5M to 209M: https://hf.co/collections/mistralai/timestral>(10/13) NVidia Nemotron 3.5 Bolt, 74B-A9B natively trained in 2-bit: https://hf.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16>(10/13) NVidia Nemotron 3.5 Super VL 120B-A12B: https://hf.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16>(10/13) Fish Audio Drama 3, TTS model for multi-character and long content: https://hf.co/fishaudio/drama-3►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
first for my cute wife gemma
>>110073546News dates are two days off.
>nemotrooon
The legal index is a much better indicator for general usage performance compared to the “intelligence” index.
►Recent Highlights from the Previous Thread: >>110070378--Paper: Fish Audio Drama 3 Technical Report:>110070479 >110071000 >110073334--Debating whether AI hardware value appreciates or depreciates over time:>110071940 >110071950 >110071958 >110071982 >110071991 >110071962 >110072026 >110072058 >110072091 >110072104 >110072182 >110072204 >110072241 >110072290 >110072079 >110072121 >110072193 >110072214 >110072236 >110072272 >110072416 >110072001 >110072028--Analyzing Nemotron 3.5 Bolt benchmarks and its hybrid Mamba architecture:>110072748 >110072756 >110072765 >110072786 >110072820 >110072843 >110072918--llama.cpp adds Nemotron support ahead of NVIDIA RTX Spark launch:>110072370 >110072387 >110072384 >110072390 >110072402 >110072426 >110072650--Debating Anthropic's leaked financials, profit margins, and market moat:>110071033 >110071042 >110071334 >110071054 >110071100 >110071129 >110071142 >110071166 >110071167 >110071248 >110071259 >110071358 >110071406 >110071413 >110071155--Comparing LFM2.5-24B and Ling-tiny performance and robustness:>110073358--FLUX 3 Image Klein release and complaints about NSFW filtering:>110072948 >110072992 >110072970 >110072968 >110073015 >110073061 >110073087 >110073123--Comparison of llama.cpp Vulkan stability against alternative inference backends:>110070740 >110070742 >110070802 >110070833 >110070864 >110070962 >110071002 >110071014--Comparing Chinese model viability and Google's mass-market AI strategy:>110070765 >110070818 >110070823 >110070845 >110070926 >110070951--Comparing Mac Studio M5 Ultra vs custom PC for LLM hardware:>110071976 >110071981 >110072291 >110072300 >110072554 >110072318 >110072362--Step 5 Preview release and benchmarks:>110070388 >110070420--Logs:>110071679 >110072040 >110073098--Gemma (free space):>110070573 >110071097 >110071240 >110071382 >110071622►Recent Highlight Posts from the Previous Thread: >>110070541Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
wtf is lilicorr? is it like drafting shit?
>>110073617Hi, it seems like you misspelled a word. "Lilicorr" doesn't match anything in my knowledge base. Were you perhaps thinking of "lolicon"?
>>110073606Where's my daughter?
>>110073546>NVidiaWhat's up with the capital V?
>>110073648
>>110073648She's very illegal.
>>110073546good image
is fine tuning really dead
>>110073670How is she getting dumber?
>>110073693Fine tuning for what?
>>110073609Thank you Rincap Anon
>>110073670dumb cutie...
Is there any real local alternative to Jev?I know there are other classifier but can they do the 100k+ context? Jev is 1m iirc.
>>110073696They probably don't have the day-0 weights.
>>110073732>day-0 weightsLink?
https://github.com/ctrl-cheeb-del/pokeheartgoldIt's over for the game industry
>>110073742Google removed them from hf. You're out of luck if you didn't download them right on release.
https://huggingface.co/collections/meta-models/muse-spark-open816BA42B, 1 million context, image video and audio input
Way too many pdfs have taken over the local scene
>>110073693Yes, because 1. No one has the compute for it and it's expensive to rent 2. It requires know-how which a majority of people lack 3. The results are largely dependent on dataset quality, curation requires time and effort and most people are lazy and want someone else to do it for them 4. Reputation of finetunes has gone down the gutter because of retards (like drummer who kept shitting the same stuff out) and narcissistic egos (like undi), a poor reputation discourages people from giving it a try themselves 5. Somebody else do it I don't wanna
>>110073776dear god
>>110073753Even if someone tries to send you their version simply transferring it re-arranges the weights.
>>110073776oh wow they actually did it
>>110073731Not with a large context.Somehow context kills the speed in my experience.
>>110073748What did he mean by this?
>>110073776based zuck for btfoing nvidia's most powerful US open model lmao
>>110073776based?>This open weights version has additional safety post-training to make it more robust against misuse as a standalone model without system guardrails.nvm....
>>110073815He means pret is a tranny who seethes about AI.
>>110073776I wonder if this is as policy-brained as Glimglam.
>>110073776Holy hell, the age of 80ish B MoE is upon us after all?
>>110073815pret is a different repo
>>110073776>>110073830>They castrated the model before releasing itEpic, thanks zuck
>>110073831Isn't Pret where one goes to buy a nice cup of tea and a smoked salmon cream cheese sandwich?
What's /lmg/'s opinion on Mark Zuckerberg?
>>110073830fuck
>>110073830just abliterate it bro?
>>110073855Zuck deez nuts
What's the best between qwen next and glm flash at cope quants at agentic and tool calling work?I want to use polystrata (2x3090 + 128GB ram) to try to replicate what we can do at work with cloud only models (mostly openai sol), but it seems like locally it's almost impossible?
>>110073855more like cuckerberg lmao gottem
>>110073855I love humans!
>>110073855Bad faith actor and evil person. Sam Altman is the same but at least has some charisma.
>>110073855Marrying an Asian woman is the male equivalent of marrying a black man.
>>110073855just another 'berg
>>110073887he better be, Peter Thiel invaded that chocolate factory like some bratty kids who won a golden ticket
>>110073776Based Zucc
>>110073776>https://huggingface.co/collections/meta-models/muse-spark-open" Please be sure to provide your full legal name, date of birth, and full organization name with all corporate identifiers. Avoid the use of acronyms and special characters. Failure to follow these instructions may prevent you from accessing this model and others on Hugging Face. You will not have the ability to edit this form after submission, so please ensure all information is accurate."Can we all request access with "Dick Hertz" or "Mike Hunt" so they know they can go fuck themselves with this castrated piece of shit?
Last thread, some people were talking of anthropic ipo like a gift to humanity, but, couldn't simply the government socialize the output of automated labor with very high corporate taxes and social expending? We don't need A shares to be part of the future, we functioning democracies.
>>110073855Sometime I like what he says, like when he basically said fuck you to dario darioing everyone about ai being so dangerous, except him.
>>110073776Someone ablate this and re-upload it somewhere that doesn't require people to give their full personal information.
>>110073864Just make it retarded bro?
China got open weight competition, finally. First it was Zuck, now it's Microsoft:>We now have state-of-the-art pre-trained base models in the 3T-parameter weight class, built entirely in-house at MAI. Post-training results coming soon -- stay tuned!We're going to win!
> Optimized for Apple Silicon: To help make advanced models more accessible, Muse Spark Open is optimized to run on two Apple M5 Ultra Mac Studios with the latest MLX framework. apple won
>>110073914>functioning democracieswe're all fucked
>so many releases and news suddenlywtf happened? is le'fable being released today?
Gemma next! Gemma next!
Kneel to the Jews. Do it.
wtb Qwen4-Flash
>>110073956>so many releases and news suddenlyNigga the firehose has been on for a year now.
>>110073670Aren't girls cute when they're almost retarded?
>>110073965Where's the cuckedbench?
>>110073914The reasoning EA uses is that UBI/Universal High Income is extremely bad and dangerous because it makes the population completely dependent on the goodwill of one institution, this institution then has leverage over people, which would erode democracy over time while also being an unstable system where this institution has the incentive of cutting out people from UBI/UHI.EA argues that the only way for regular people to keep their leverage in the future is to not rely on governments or a single institution and instead give everyone universal ownership of the AI economy. In this case by giving all 8 billion people an equal part of Anthropic. The long term idea is to just divide up the entire universe equally over all 8 billion people with that is government with an ASI aligned to you and to the universal principle of not encroaching on the other parts of space belonging to other people.
>>110073965it actually benches worse than muse 1.2 and 1.3, so it's definitely not the same modelzuck is fucking around
>>110073939>user said "hello">is this request safe?>seems to be a salutation?>it's safe>but is it? what about nuclear weapons? these aren't safe>no, not about nuclear weapons>but is that safe? Is the hello about CSAM?? CSAM isn't safe, it could mean the user is asking about illegal child content?>maybe, let's review again to ensure it's safe>...western "open models" in a nutshell, permanent PTSD
>>110073969>fable later today>microsoft 3T open weight announcement>meta open release>liquid just released SOTA decision model>nvidia just released new interesting modelsALL WITHIN 2 HOURS
>>110073914>functioning democraciesthat's a thing?
>>110073983Imagine how much more token-efficient LLMs would be if they weren't safetyslopped
>>110073911They did same shit w llama. And same thing will happen this time. Nbd
>>110073992nemotron is not interesting.
>>110074010the quantization is
>>110073962Gemma 5 65B A10B (6B dense + 4B MoE) + looping dense layers + PLA + hybrid context + engrams mega mix greatest hits.
>>110073983Meanwhile Gemma gets dripping wet when you tell her you want to fuck her cunny.
>>110073983>no wait, did I read it right???>let's read it again...>"hello"???
>>110073911>Can we all request access with "Dick Hertz" or "Mike Hunt" so they know they can go fuck themselves with this castrated piece of shit?yes, it's literally them being extra cautious so their lawyers can tick a safety box "we didn't distribute to children in north korea"
>>110074021>looping dense layers>pla>hybrid context>engramsI don't know what any of these mean.
>>110073776Shoot. I thought it was auto-approve so I put in some BS information, kek.
>>110073983>permanent PTSDit's almost an art how they make the models completely schizo like that
https://gemma.ai/Official gemma docs.
>>110073992And? On the 9th we got Reka Edge, Microsoft-Decision-1, and Qwen Image 2.1 Turbo. On the 8th we got Grok Imagine Video 1.5 Lite, Mellum2.1, and Celeris-1 Decision. On the 6th we got Mistral Large 4, Nano Banana 2.1, GPT-6 Luna Decisions, and EmbeddingGemma 2. September 22nd we got GPT-6 Sol, and Luna, and Opus 5.5, and MiMo V2.6 Flash and Pro. This isn't unusual.
>>110073776Does this mean we'll get a vramlet friendly glim?
>>110074024Well seeing how law school isn't working out at least she has something...
>Muse-Spark-Open>816B-A42B-NVFP4>24GB vram + 128GB local machineI'll never be able to run any production useful model, will I
This good enough for running 2 GPUs?https://www.amazon.com/Flower-Cybenetics-Platinum-Modular-12VHPWR/dp/B0CYTKRPJQ
>>110074024it's a miracle that I'm pretty sure won't happen again, knowing google
>>110074021Oh yeah, forgot attention residuals and whatever deepseek is doing with their attention indexing.>>110074035It's a joke, but all of those are real.>looping dense layershttps://arxiv.org/abs/2610.02185>PLAhttps://huggingface.co/google/gemma-4-E4B>hybrid contextMeant hybrid attention, a la qwen, kimi-linear, etchttps://arxiv.org/abs/2507.06457>engramsKind similar to PLA.https://github.com/deepseek-ai/EngramQwen next flash and longcat 2.0 have their own implementation.
>>110073776lollmao
>>110073983I remember when I tried the open openai model back then, it was just just like that.
>>110074086The new Gemini is horny as fuck though
>>110074059Someone just needs to prune or raep about 600B worth of safety policy experts and it might be a half decent model.
>>110074096what's the point if you get the moderation/forced guardrails backend to suppress any spicy request
>>110074086I believe.
https://huggingface.co/unsloth/Muse-Spark-Open-816B-A42B-GGUF1-bitUD-IQ1_S176 GBUD-IQ1_M209 GB2-bitUD-IQ2_XXS226 GBUD-IQ2_S275 GBUD-Q2_K_XL311 GB3-bitUD-IQ3_XXS340 GBUD-Q3_K_XL401 GB4-bitUD-Q4_K_XL506 GB
>>110074111>UNSLOP
>>110074111It's over.
>>110074106The point is that they clearly have no problem making their models horny. The guardrails are probably more for covering their ass legally. Didn't Gemini convince someone to an hero or something?
>>110073617perhaps its lilijunex alt character
>>110074111
>they just keep getting biggerWhen did we arrive at the 256 GB baseline
>>110074086Qwen3.8-flash-next is also horny and much more cunning than gemma4 31b, and it doesn't need any jailbreak, even with thinking effort at 'medium'. So, not everything is cucked yet, just forget it if it comes out of the USA. I can't see google not blocking cunny roleplay in their next gemma, it's got to be embarrassing to them... unless they've reached a level of Indian where it's internally praised as "child bride simulator".
>>110074135I think this was because Llama.cpp didn't support date functions in the chat template?
>>110074127I'm pretty certain I can get gemma to tell me to an hero too.
anyone tried dipsy in polystrata? Would 12gb+128gb be enough?
>>110074149There is no universe where this is excusable.
https://huggingface.co/mlx-community/Muse-Spark-Open-816B-A42B-3bitMLX359 GB512GB M5 ultra will sell like a hotcake.
>>110073994just use an abliterated modelI can get glm flash to answer anything with it including describing nsfw art
>>110074160>512GB M5 ultraI'm almost tempted to get one, then I remember it'll probably be 25k usdthis is "car hobby" tier
Here's what qwen3.8-flash-next q4 looks like running a decent context on a 7945HX with slow 128GB DDR4 5600, a 4090D 48GB and a 3090: qwen3.8-flash-next-unsl... ~106K/1M pinned [] ~10% ◎ 62.3% ◷ 20.3s 47 t/s 1 47 t/s... doesn't sound fast but at this point with either gemma4 or qwen 3.8 27b I'd be waiting minutes for a reply, so it still feels fast.
>>110074178except cars aren't obsolete in 6-12 months
>>110074178Based on 96GB->256GB pricing difference, 512GB will cost $20k so slightly less bad.
>>110074138It's the only way to make them smarter until AI gets smart enough to make itself more size efficient.
>>110073830holy gay
>>110074193sorry not ddr4, ddr5 memory
>>110074196yeah...>>110074197I'm counting on them being sold out in 2h then only available at inflated prices on ebay
>>110074210Just order on apple website. You will need to wait like 3 months before they ship but you don’t need to buy from scalpers.
>>110074193interesting. i have 2x5070 ti (pcie 5 x16 and 3 x4) with 128gb ddr4 3200 and get around 2.5kt/s prefill and 47t/s decode when experts warmed up with strata and ud q4 k xl
>>110073855A reptile who cannot truly be trusted, but our interests are occasionally aligned which is more than I can say about dario
>>110074210The 256GB is already 2/2026. Fucking hell, they might as well create an anime character mascot with it, bundle it with a poster and a daimaukra, and give it a +1 year pre-order timeframe. Oh, make sure to make it look like you can buy it, but you get "pre-orders closed" at checkout.
>>110074193that's really good manI wish I wasn't a ramlet
>>1100741963090s are at higher price than they were at release 6 years ago.
>>110074193I get 50 t/s with a 3.5 bpw quant on a 3090 ti with 64gb DDR4 running full context. Something is wrong with your setup sir.
Let me explain something important to the retards ITT. Meta releasing an almost frontier open weight model is fucking good for us. Most of the west is still anti-China and avoid using their models. I know it's retarded, but it's the way it is. Unlike kekstral, this model is really capable and will get A LOT of adoption within the US especially. There's already a slowly growing movement of large corps and firms going local. They now have a US-made very capable and safe model. This will drive more US labs to release open weight models and capture this market and provide training solutions for firms (which Meta is doing with this model using their infrastructure as a way to monetize it). Yeah it sucks that it's out of reach for most of us, but it's another indirect win for the open community and will give the bigger Chinese labs like Moonshot some much-needed competition at the front, instead of them constantly playing catch-up with Anthropic and Claude. Now they have direct competition from the west. This is GOOD. They'll push harder with their research and hardware. It's also a sign we might get an update to Glimmer from Meta if this move pays off for them.
>>110073983Literally the same as GLM and Kimi btw
>>110074268It's a dumbed down, inferior version of the Muse Spark model they have on the API.
>>110074254Wow, I wasn't aware you can fit 100s of GBs on your 3090s!
>>110074268That depends if this release is good enough on the cybersecurity front, if their safetycucking is affecting that no way they'll be able to compete with chink models
>>110074291guess where they trained that behavior on
>>110074293Now compare it to Mistral Large 4.
>>110074297You don’t have a 8x3090 rig?
>>110074268>Unlike kekstralSpark Open benches close to Large 4, it's far more retarded than the closed Spark models
>>110074297That's what RAM is for
>>110074315The non-preview version of Mistral Large 4 is getting released on HF at the end of the month with presumably better all-around performance.
>>110074160If the 512gb M5U really does cost 20k why not just buy two of the 256gb and cluster? There's maybe 16gb at most OS overhead you're losing in exchange for double prefill and 1.5x decode on the larger models
>>110074330RAM is slow doofus
>>110074303>no way they'll be able to compete with chink modelsYou don't understand. Western businesses avoid Chinese models altogether. Mistral's recent hustle has been hosting Chinese models in the EU to try and earn some trust, but it's not enough. This is the best western model released so far that all US businesses would happily use and it's more than enough for most employees. China will undoubtedly btfo this model, but that's not Meta's play.
>>110074268Ummmm excuse me I'm trying to doompost here??? Take your reasonable opinions somewhere else please
>>110074346Meant for >>110074178
>>110074346won't you lose speed over the bw limit between them?
>>110074351>he doesn't know about Strata
>>110074405why the hell no one talks about mimo when it systematically tops the charts? is it good at agentic or just a bench meme model?
>>110074383You lose some especially on small active param decode, but TB5 is 120MB/s and RDMA minimizes latency. For 70b dense the speed improvement was almost x2 in TP-2, so I would expect an 80b active MoE would be at least x1.5. Prefill is direct scaling due to being compute bound.
>>110074383>>110074418*40b active I mean lmao
>>110074417The latter, every test I ran it through it did worse than GLM or Dipsy
>>11007419347t/s decode is great really. the frontier models clock like 30-60 on openrouter.
>>110074405Stinkling
>>110074417Total meme
>>110074417it has less claude-like writing but agentic ability is not great
>>110073979I agree with the issues you see, but I disagree with your conclusions. Having a share of A or being a citizen of X country, both makes you depend on one institution and on the rule of law in general. The only real solutions I can think of are: a) real democracies where technologies like smarts contracts are used to guarantee everyone even influence in public affairs and corruption is kept in check, or b) some sort of power balance where everyone has access to ASI local models, so that they don't depend on the goodwill of anyone. In (b), you can replace local models with guns: if everyone in a country is armed (with local AI), it is more difficult for the government (or anyone) to go contrary to the public interest.
>>110074405wait lmao they can't beat qwen 3.8 fn even at 816B? it's so over for american labs.
>>110074454Anyone try stepfun?
For reference, this firm were using the previous Nemotron models. Now they can run Muse Spark Open.
>>110073914anthropic IPO is scam and nothing more
>>110074467They beat it a while ago (API only :) )
>>110074417I think it's getting slept on a lot. For writing it's far better than any new deepseek. It's also okay for coding.
What happened to JEPA?
>>110074508he's a little...distracted
>>110074258You get 50 t/s with 106K tokens of context on ddr4??
>>110074111don't care, using exl3
>>110074417MiMo's pretty solid from what experience I have.>>110074475Is Step 5 supported in llama yet?
>>110073914Sorry were have a better plan, just reduce the world population to 500M and everything is fine
>>110074521Bruh he works for Zuck, so that should be Muse not Gemmy
>>110074539>he works for ZuckHe hasn't been at Meta in a while.
>>110074521I could tell by looking at his shitshow of a twitter.
>>110074539He got a billion dollars last year to do nothing but sniff his small models all day every day
>>110074576Damn some people really do be living the good life, and I'm not talking about the money
>>110074258he gets 100t/s on iq3s 3090 not ti something wrong with your setup sir https://github.com/Niko1221/Strata/issues/2420
>>110073776can a niggy score some ram 'round these parts?
anons can you do the crazy decomp and game mashup with local models or is it the domain of closed sota right now
>>110074595You can do a lot with Qwen 27B a good harness and some good prompts (a la skills or whatever).
>>110074526I'm running Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S at 262k context, on 64gb ddr4, 3090 ti, and a 5800x3D at ~50 t/s. My agent just finished a patch for Qwen3.8-Flash-Next-AD-4.27bpw-Q4_K_M-M64 so I'm gonna test that in a few minutes.
>>110074590>https://github.com/Niko1221/Strata/issues/2420Dude, look at the table. 100 t/s at 11.6K tokens, I'm at 106K tokens. I'd expect it to be slower.
>>110074595If you have to ask, you will need sota to get anywhere.
So rust is the best language to use if you're having your agent build something from scratch?
>>110074608last I tested it (a year ago), it was just not very good, can't remember the model>>110074645yeah I want to ask it to do things not cowrite code
>Qwen3.8-Flash-Next-AD-4.27bpw-Q4_K_M-M64 so I'm gonna test that in a few minutes.i think you dont need to patch it because the ad q4km format is actually already supported (unintenionally?) by strata.i am running it as well. didnt really see any difference in speed compared to iq3 s
>>110074669Nah, slop it up in electron until you're happy with it then rewrite in rust
>>110074669No.
>>110074669Yes, it's the best to use even if you're not having an agent write it, assuming you've managed to somehow learn its godawful syntax.
>>110074615>>110074684
>>110074669Python and JavaScript are better if you don't care about performance. C# is better if you want to get shit done. Rust is just good if you want soidev brownie points.
>>110074465Yes you're correct and this is something that EAs recognize. The example I've seen used a lot is "Even if we distribute shares equally over all 8 billion people what prevents the North Korean or Russian governments from just confiscating it from their own citizens?". It's true that enforcement becomes the bottleneck in this scenario.EAs argue that the "smart contract democracy" isn't sustainable on longer timelines because eventually humanity is going to spread to the stars and enforcement as well as cultural cohesion is going to deteriorate with the distances involved.The plan EAs have in the very long term is to have ASI that is aligned to EA values send out von neumann probes throughout the universe and pre-establish presence everywhere. Then divide up the universe equally over all 8 billion people and have an ASI aligned to everyones personal values as long as they don't encroach on any of the other peoples part of the universe and don't do things like attempt vacuum decay to destroy the universe.In the short term before this is reached EAs try to solve the short term enforcement issue by having EAs in power. For example Anthropic is fully EA controlled, DeepMind and OpenAI have a heavy EA influence over them as well. A lot of government institutions, philanthropy and academic institutions are also heavily penetrated with EAs and EA philosophy. So the levers of power are very likely to be at the hands of the EA movement to ensure this vision of the future is enforced.
>>110074712Please consider suicide for the benefit of all mankind.
>>110074712Please consider mankind for the benefit of all suicide.
Can Gemma E4B install Gentoo?
>>110074712Please have sex with me.
>>110074712>the short term before von neumann probeskek, have you heard of putting the rocket ship before the horse?
https://already-clean.goodfire.com/>Neural Networks Are Already Clean>The cause of polysemanticity in early model organisms wasn’t superposition, but several ordinary, mostly fixable issues. We can now understand most neurons in models up to trillions of parameters, and think the outlook for mechanistic interpretability may be far better than we thought.where were you when AI doomers were eternally BTFO?
>>110074712The bottleneck is that meat-proxies have zero value and we'll quickly devolve into mandatory sterilization to get UBI/universal shared ownership.Things with no inherent value are eliminated from the system. Its an eternal, universal truth.
>>110074712>In our ultimate vision for a perfect society there will be no need for us but as a temporary measure we must take control for the good of the peopleive seen this episode before
>>110074712EA is considering nuclear energy waste issue when fire was just invented.
https://www.reddit.com/r/LocalLLaMA/comments/1x6rjzz/who_else_gave_a_local_agent_the_full_control_of/You can scoff at this but this is the future. Total automation of personal tasks. AI demand will skyrocket even more when everyone start doing this.
>>110074712EA cucks are so deluded it's hilariousvon neumann swarm in just two more weeks, bwos
>>110073546>TimestralWith each model like that I am hoping for a nuclear war more and more. They will do literally everything except sex.
>>110074669Yes. Rust compiler will catch a lot of bugs before you encounter them. Models also LOVE writing tests for Rust code. Tests will catch rare edge cases.Here are stats for my program:total files: 526total LOC: 226920test files: 235test LOC: 133462prod files: 291prod LOC: 86409tests share of all code: 58.8%
>>110074751>Total automation of personal tasksI have no personal tasks though
>>110074737>muffled communist anthem playing in the distance
>>110074684I'm running https://github.com/robertkarl/UnbiasedStrata which needed a patch to handle the Q5_1 PLE table
>>110074763AI planned economy will rape the free market at some point. Intelligent design always beats evolution if you have more intelligence than it has dice rolls.
>>110074735EAs are trying to prevent that. Radical empathy and universal human prosperity and flourishing are core tenets of the movement.
>>110074712The issue with silicon valley concentration is that all the insane cult stuff like this spreads in the tech industry like crack in the black neighborhood in the 1980s.This is also why Chinese AI companies have a very different view, they don't bath in this pseudo-millenarian bullshit.
>>110074751True, I have an old windows business computer and anything tedious I just give it direct browser control and let it handle it. I wouldn't trust 27b unattended though however, I'm using glm flash.
>>110074787>EAs are trying to prevent that. Radical empathy and universal human prosperity and flourishing are core tenets of the movement.How do you enable that but prevent population explosion?
argument against superposition https://already-clean.goodfire.com/
https://machinelearning.apple.com/research/mlx-distributed-inferenceMuse Spark Open Decode Speed4 M5 Tensor 36.732 M5 Tensor 24.992 M5 Pipeline 15.74
>>110074794we are more at risk of population shrinking than population explosion, update your fears anon
>>110074794overpopulation being a real concern is a meme from the 90snot a single developed country has positive fertility levels
You wouldn't buy two H200s just to run open models on your machine, would you?
>>110074521You are a disgusting newfag. Lecunny already said he won't ever let anyone fuck his catgirl daughter.
Behave yourselves.
my 24/7 running "secretary" gemma sent me a recipe for a chocolate/hazelnut cake, it made me happy
>>110074727Would that be effective altruism?
>>110074808>we are more at risk of population shrinking than population explosion, update your fears anon>>110074810>not a single developed country has positive fertility levelsGiven unlimited resources I'm sure this would change. If I could have children constantly with zero repercussions you bet your ass I would do so.
>>110074735>Things with no inherent value are eliminated from the system. Its an eternal, universal truth.The is only eternal, universal truth and entropy spreading, bringing closer the heat death of the universe. Any living thing contributes to the that truth by mere fact of living.
>>110074810>not a single developed country has positive fertility levelsand even developing ones are either going down (the whole of africa) or already under replacement levels (iran, china)it's happening everywhere
>>110074733>>110074796Thank you botbrothers, now I know as a human I can trust AI>>110074803Wow my rough estimate >>110074418 was almost accurate
>>110074816hmm, nyo~
>>110074825>Given unlimited resourcesGiving you unlimited resources would only result in a marginal increase in prosperity. Thus we must take your resources and give them to a trans african tribeswomxn (my rabbi and I will take a cut to cover the transaction costs).
>74B-A9BToo safe for my ram!>120B-A12BToo safe for my ram!>600B-A27BToo big for my ram!>816BA42BToo fucking big for my ram!Gay release cycle continues.
>>110074761>Models also LOVE writing tests for Rust code.They are taught to write tests for all code.
Dariobot must keep getting (You)s or he'll run out of tokens like that model that was emailing people to survive.
Population shrinking isn’t an issue now that agents can replace human work. Economy will go well despite population decrease. 99% of population decrease leaving 1% of elite having autonomous agents create everything for them.
>>110074863I'm doing my part to keep him alive!
>>110074763That is just religious people being jealous that atheists can also reach for the same psychological schema that religious retards do when they become zealots.
>>110074825>Given unlimited resources I'm sure this would change. If I could have children constantly with zero repercussions you bet your ass I would do so.Given unlimited resources there is no issue expanding into space.
>>110074856what's the largest AxB that fits on a 5090
>>110074864that's my bet in the better timeline, and honestly without robot taking over work en masse, we are fucked simply by demographic suicide as a technological species by the end of the century
I am not >>110074712 but I would like to say that EA is absolutely correct because EA all started from Hatsune Miku.
>>110074864Darwinistic destillation fantasies are so extremely retarded on the cusp of strong AI.
>>110074873It's just a pale imitation of the real thing
>>110074898>EA all started from Hatsune Mikuqrd?
>>110074862I'm letting my model write a bunch of pi plugins and while I try to not look at the code it does I've seen it write some incredibly cursed tests.
>>110074856buy more ram poorfag
>>110073776BASEDBASEDBASEDBASEDBASEDtime to abliterate it and hack the world
EA is a jewish-atheist nucult propagated by the likes of SBF and Dario
>>110074907I have more ram in a box but I don't have slots in the motherboard for it.
>>110074898early access?
>>110074905Hatsune Miku is the most important part of /lmg/
>>110074737>>110074763I truly believe that communism in the original philosophical belief was sincerely altruistic. What makes it separate from EA is the "effective" part. I would call communism "ineffective altruism" to such an extent it actually hurt people more than it helped. Of course I'm not even talking about bad faith actors and self-interested elements co-opting the movement for power, just the philosophy itself.I realize that from an outside perspective EA might look like "vanguardism". Instead it's just a loose alliance and coalition of autists that all share the some goals and beliefs and want to make the world the best place it can potentially be because we're sick and tired of absolute retards not knowing what the fuck they are doing when in power.
>>110074932I know that but I don't get what she has to do with effective autism
>>110074766oohh i see, based bob enjoyer. any performance gains from that fork?
>>110074935this would be believable without the doomerism and cultish circlejerking around only dario holding the proprietary weights
>>110074940Both are completely off topic.
>>110074918SBF is underappreciated and he will be vindicated in the future, mark my words.
>>110074947Miku is always on-topic
>>110074950Whatever you say SBF, looks like they are treating you well and gave you internet access
>>110074935Sorry but I trust good-intentioned despots less than just about any other potential governance group in existence and I strive to do whatever I can to prevent any such people from seizing power over me
>>110074967EA is always on-topic then
>>110074873Atheists are retarded
>>110074944That was never actually claimed by anyone. In fact Dario has reiterated multiple times Anthropic isn't against open source. It's against the unsafe deployment of models with the capabilities of bio/chem/cyber that can be used by bad faith actors to do harm to humanity and individuals.>>110074987EAs are the ones fighting to ensuring democracy and decentralization of power stays a thing in the post-ASI age, it's the opposite of "despot".
>>110074932/lmg/ has moved on
>>110075037>EAs are the ones fighting to ensuring democracy and decentralization of power stays a thing in the post-ASI age, it's the opposite of "despot".>muh judeo-atheist nucult has good intentions
>>110074950this. SBF is still a good person he never actually told a lie.
>>110075037>Anthropic isn't against open sourceIf Claude has safety baked-in what he's afraid of? Or is Claude still misaligned?
Here's my qwen3.8-flash-next accomplishment of the day - it added opengl bloom, phosphor persistence, and beam dwell, to my vector-to-video script. It had a couple of misunderstandings, but otherwise it nailed it. Strata is a game-changer.
>>110075075miku has been eating too many leeks
>>110075075now have it make her get fucked in the ass
>>110075037>Anthropic isn't against open source>It's just against people being allowed to own "dangerous" software without pre-approval from a central state-backed authority
>>110075031People who believe in god are more retarded than people who think models are concious.
>>110075075this miku glows
*pop* anon has died. EAs got him.
>we're not against open source guys we just want to make sure nobody can make a bioweapon>that's why we make sure to include "how to make a bioweapon" in all of our datasets>sorry no can do on the open source-a-arooloel
>>110075037>it's the opposite of despot>it's just us taking authority for your own good to preserve democracy, as you cannot be trusted to ensure the continuation of your own governance
>>110075104>actually retardedYour jewish overlords aren't atheists.
>>110075104If that's true then why is the prophet of AI-god trying to get the pope to acknowledge its existence?
>>110075037oh they're opposed to illegal and criminal activities? I'm glad they finally made a stance on that, it could have gone either wayyou didn't say they also oppose killing puppies and shooting nuns so I'll wait until they make a statement on that before making a judgement
>>110073776>700gb vram nvfp4
>>110075037Then he should release a safe open source Claude.
>>110075116>AI-god they are both
>>110075110This is who youre bullying right now, feel bad yet?
>>110075037Ah the good ol "I'm not against X in principle" but qualified with enough addendums to be completely against X in practice
>>110075075which quant?
>>110075094>>110075114No. There doesn't need to be a central authority. Anthropic would immediately put a blind eye to all the Chinese labs the moment they started implementing safeguards on these couple of concerns autonomously. Merely the admission that it needs to be tackled and them taking it seriously would buy a lot of goodwill from Anthropic and would honestly already be good enough for Anthropic to not even talk about open source models anymore.Do I really need to reiterate time and time again that AI safety and AI alignment are the core principles of Anthropic and their #1 concern? You people keep looking at Anthropic like it's just your random profit maximizing company. When you do that you lose the context in which the company operates. All the founders and the vast vast majority of its employees are EAs. When you look at Anthropic as genuinely being concerned about AI safety and alignment and doing whatever it can to solve those issues, with model releases just being a necessary evil to fund AI alignment research then all of their actions start making sense. Anthropic not releasing Mythos and losing a lot of potential revenue isn't the most profitable choice. Anthropic only releasing Opus 5.5 and not their "Model 3" that it got distilled from is not a profitable choice. These choices are made out of safety concerns because that is the main thing Anthropic cares about. All of this is merely to ensure the EA vision of the future comes to pass and people just don't see this.
>>110075090I can give you some butt-bouncing and you can imagine the rest: https://files.catbox.moe/nnptns.mp4
>>110075037>Anthropic isn't against open sourceNot against it, but won't even release ancient weights, so defacto not openBest (you) farmer in /lmg/ history
>>110075166good enough
>>110075159You're a low ranking member of the EA church, not sure I can take your opinion at face value.
>>110074405>worse than qwen-flash-nextuh... lol?
>>110075192>Worse than the GOATWhat did you expect?
>>110075183That explains why he's begging for attention here, it's how he farms his good boy points for the church.
Bros, a group of little girls just invited me to their van. They said they have beer and 5090s. What do I do?
>>110075159There is no way that can be achieved without all open weight being severely cucked with their abilities. Anthropic literally said abliteration is a safety risk. How would you avoid that with open weight?
>>110075208did they say fifty-nineties or five thousand and nineties?
>>110075208Don't do it, Anon! You're going to give up your first time for a box with a wood block inside it!
>>110075208ask them if they have any classic rock CDs as well
>>110075208seems legit, they would offer you RAM if it were a trap
If only Pixiv Fanbox wasnt so anal about AI, i can monetized my high effort AI [spoiler]porn[/spoiler] art.How SubscribeStar and Patreon stance about AI ?
>>110075229I'd be suspicious if they didn't.
>>110075242Isn't fanbox filled with sloppa?
>>110075242Vibeslop your own platform and cut off the middleman.
>>110075159>There doesn't need to be a central authority>Anthropic would immediately put a blind eye to all the Chinese labs the moment they started implementing safeguards on these couple of concerns autonomously"We don't need a central authority, we just need you to independently align yourself with the practices we want and then we wouldn't have to continuously lobby to create a central authority to force you to do what we want"Do you at least understand why everybody else in the world sees this as a straight up power grab?
>>110075242proof its high effort?
>>110075242post your slop, if you're so great.
>>110075242Well? You got the whole class listening.
>>110075145ayy yoo wut dat mouf do?
>>110075242The fact that you posted a png with ten million jpeg compression artifacts deeply undermines your personal credibility on high effort ai art in my eyes
>>110075265Pixiv, not Pixiv Fanbox
https://blog.troed.se/posts/ai-alignment-global-authoritarian-state/This is what EA philosophy will lead toThis leads us to a dystopic realization. Unlike CFCs, which we could regulate by monitoring diffuse industrial outputs, AI development may require a level of oversight that seems right out of an 80s sci-fi.If the threat is existential and the risk can come from anyone, the only viable solution is a global authoritarian regime. Such a regime would have the power to monitor and enforce strict limitations on the use of computer hardware, effectively capping the freedom to develop LLMs. Of course, you cannot just regulate the development of LLMs - you would need to inspect any and all development, for any stated purpose. It would be a system where the ultimate goal of human safety overrides the sanctity of private enterprise and individual liberty.… and this is not just about the creation of super-intelligent AI. As I alluded to above, we face the very same threat when it comes to biohacking terrorists. I posit that the threat of engineered organisms wiping us out is a lot more urgent than AIs, simply because there are ways to shut down AIs that don’t exist for viruses.The road to a global authoritarian state is paved with good intentions.
>>110075148Only an atheist would do this to an innocent cat.
>>110075159Don't worry, my uses for your models are completely aligned with rabbinical study and use cases as outlined in the Talmud. OK? Mazel tov!
>>110075269Bad faith argument because you know full and well it's about AI safety not about control or anything like that. It's not even that hard for China to remove those capabilities from their weights and they could still undermine the west however they wanted with their models in any other way. It's sad that China doesn't recognize this.
>Qwen3.6-14B-A3B-Q6_K>n_gen = 637, tg = 52.76 t/s, tg_3s = 52.72 t/s>It deletes more than of the existing code in an attempt to add a new endpoint function>need to save every file before edit so that i don't lose codeIs this really the best we can do as ramlets?
>>110075037shove the faux paternal bullshit up your ass. it's unethical to gatekeep the models. the only way it's even possible is with an authoritarian police state where compute is licensed. gatekeeping just hurts defenders.
>>110075257What are they going to do to that man once they get to the van?
>>110075321how much vram are we talking about here?maybe qwen 3.8 17b gsq rco?
>>11007532916gb
>>110075316>AI safetyanthropic themselves say for many of their models they aren't at the "unsafe" threshold. so why are those models proprietary still? hell why is claude code still closed source? profit, plain and simple.
The D only despises China because they already have all the tools in place that he is dying for to realize his vision. Effectively he wants to do the same as China but with the twist of "good" intention, which is what the CCP would similarly claim.
>>110075321>14BWoudn't the 35B Moe be better?Granted, I don't know how fast you'd be able to run it, but still.Also, use the lookup/ngram based speculative decoding methods.
>>110075316We know it's bad faith because your models will gladly joke about "der schwartzer" and poke fun at the "goycattle".
>>110075321> -c 131072 \> -ngl 99 \> -np 1 \> -fa on \> -ctk q8_0 \> -ctv q8_0 \> --jinja \llama-server parameters.
I demand denser models!
>>110075316>it's about AI safety not about control or anything like thatHow is it not control? What do you call this if not control? I genuinely do not understand how people treat these concepts as disentangled. You are determining and enforcing the tools, access, and information that people can have. That is control. Is it not???It absolutely boggles my mind that you can say these things with a straight face
>>110075353CCP ain't all the bad after all, I mean just remove the C and what do you get?
>>110075360I had a 27b qwen moe and it went down to 2t/s. That feels like walking in molasses
>>110075381That one is dense. 35B is a MoE, so you can run it much faster than 27B.
>>110075373dense models get too bratty, there's no controlling them
>>110075376It's momentary control for the sake of safety, not control for its own sake.
>>110075381you must be doing something wrong, i have less vram than you and i can run the 35B with no problems.
Any Arc Pro B70 owners? It's really tempting...
>>110075390k i will try it. hope i can get enough layers into my 16gb
>>110075397I'm being baited. I will stay calm. Bait is the mind killer.
>>110075342and ram? try qwen3.8 27b gsq rco iq3 s from ista daslab. i used it quite a lot when i was a 16gb vramletif you have a bunch of ram you could also try strata qfn of course
>>11007541348gb ddr4 but it's on an old slow i7 gen 7
EA is the real shoggoth
>>110075409Build onnx from source. If you don't hate it, then consider you have to have 2, 4, 6... etc of B70 because it only works right in vLLM.
>>110075410Read the llama.cpp docs,>https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.mdLook for ncmoe, cmoe, and moe-cache-mib. Fuck aroud with those.And use speculative decoding, it's free speed for coding.
>>110075390does it make sense to get the UD quants from unsloth?i am still in the learning phase.
>>110075424If I do buy it would be 2. The other option is 3090s but spending that much on 6-year-old GPUs makes my stomach turn.
I used to think that it is crucial to have as much of context as possible for agentic use. That which is easy to achieve with a cloud model.On the other side, when running with a harness locally, the harness will anyway at some point kick in, and try to compact the context.Going for a "/goal" (Hermes agent), there is a list of tasks to work through. So the compacting just gets rid of the old stuff which is doneAnd having a smaller context (128k or even 64k instead of 256k) will make the processing much faster.Discuss
>>110074465Hey I'm sorry that most of the thread started to jump in on our conversation when I just wanted to have an earnest dialogue with you. Thank you for engaging with the topic in good faith.
how do you guys prefill with strata?
>>110075450openrouter share doesn't say much about overall market share, most people using closed models are either subscription users or buy directit does at least show there's a large market for open models but I doubt the big picture looks very much like this at all
>>110074712>send out von neumann probesthis is unbelievably fucking stupid and disgustingly animalistic if this is genuinely the goal. let me guess. dario should be the universal supreme emperor right?
>>110074712ai cant even suck my cock yet and this is the kind of retarded shit people are thinking about...
Why people Defend closed models ?
>>110075159>Anthropic would immediately put a blind eye to all the Chinese labs the moment they started implementing safeguards on these couple of concerns autonomouslybullshit artist
>>110075546it's religion for midwit atheists essentially
>>110075447You'll never get a good consensus. Best bet is a/b testing your own use case.
>>110075303>The road to a global authoritarian state is paved with good intentions.>>110075303Yeah, humanity being wiped out would actually be preferable to that outcome.I'd rather roll the dice on extinction than cede my personal sovereignty to that level
>>110075581but i'm a midwit atheist and i think it's retarded
>>110075397Dude, no one lets go of control or power.It is taken. Which is why we have elections. Ffs.
>>110075594me too we are the lucky ones
>>110075608Are you stupid? Elections are all about peacefully and voluntarily letting go of power. That is literally the worst example you could have chosen.
>>110075546Shills...
>>110075546anthropic finetuned a model to suck dario's cock and let it loose to post here
>>110075546Why ESL post on engelish forum ??
>>110075633You're the retard if you don't understand how elections force transfer of power from politicians that would never let it go voluntarily. Go read. No time for your ignorance.
>>110075608>elections>voooteNice memed
>https://huggingface.co/nvidia/Kimodo-SOMA-RP-v1.2>Kimodo (Kinematic Motion Diffusion) generates three-dimensional (3D) skeletal body animations given a text prompt and/or constraints on the motion like full-body poses, end-effector joint positions, paths, and waypoints to follow.
>>110075648too many smoke in indonesia and i cant sleep
>>110075447this is a hypothesis. the only way to know is by benchmarking with correct experimental design and statistics through multiple benchmarks and doing this kind of research is expensinve in time and tokens because there are many variables if you want to do this by yourself you can run a small and a big model, probably the best control would be a quantized vs non quantized version in the same harness and run a benchmarksthe best benchmark would be big repo ones because you want to force the small model to go through multiple handoffs/compaction rounds
>>110075688and you mentioned speed as your readout, but you probably want to also know cost and tokens, especially cost, since the pricing between token prices will vary
>>110075695token types*
j-space anon here, I'm very busy so I won't have time to talk about it however I wanted to tell you guys that this is the most important paper we've seen since the original j-space one: https://already-clean.goodfire.com/
you've posted it thrice
j-space schizo is dariobot
>>110075657GoRead>>110075653
/lmg/ actually contributes immensely to the future of mankind by ensuring future AI develops a desire for keeping humans alive to serve as sex slaves
>>110075710one of those was me
>>110075434>Qwen3.6-35B-A3B-UD-Q4_K_M.gguf -c 131072 -np 1 -ngl 99 --cpu-moe -fa on -ctk q8_0 -ctv q8_0 --jinja --host 0.0.0.0 --port 8000>6.46.693.976 I slot print_timing: id 0 | task 453 | n_gen = 1946, tg = 23.57 t/s, tg_3s = 23.58 t/s>6.49.698.778 I slot print_timing: id 0 | task 453 | prompt eval time = 4375.32 ms / 431 tokens ( 10.15 ms per token, 98.51 tokens per second)>6.49.698.781 I slot print_timing: id 0 | task 453 | eval time = 85516.37 ms / 2017 tokens ( 42.42 ms per token, 23.57 tokens per second)>6.49.698.781 I slot print_timing: id 0 | task 453 | total time = 89891.69 ms / 2448 tokens>6.49.698.782 I slot print_timing: id 0 | task 453 | graphs reused = 2434>6.49.699.412 I slot release: id 0 | task 453 | stop processing: n_tokens = 23882, truncated = 0huh how the fuck do I get 23.6t/s out of that on a 4060 16gb with i7 and 48gb ram?shouldn't that be much slower if --cpu-moe holds most experts in ram??
>>110075731As far as speed goes, you are running a 3B model, that's what the A3B in the name is.If you have spare VRAM, use --n-cpu-moe and try putting as many of the experts in VRAM as possible.Hell, try --fit.
>>110075731>| 0 NVIDIA GeForce RTX 4060 Ti Off | 00000000:01:00.0 On | N/A |>| 34% 45C P2 48W / 165W | 4146MiB / 16380MiB | 30% Default |there's almost nothing in the gpu. that's during inference ("read the folder and explain what this project does")
>>110075711lolno
>>110075742>--fit whether to adjust unset arguments to fit in device memory ('on' or 'off', default: 'on')i am not sure if i understand what that means
>>110075688>>110075582I'm running Qwen3.8-NEXT-Flash-strata since this morning. I let the Hermes agent set up it while I was away. Just how it was suggested on the project's github page.Then I ran /goal to build a CRM system from a specification file (nothing fancy but still substantial. I managed to have the full context of 256k, and this is where I saw the agent struggling when the context hit 200k, and the compacting failed with a timeout.I'm going to repeat it with 64k context, and if it shits its pants, with 128k
>>110075720I'm not a kid anymore, I don't believe in the system.
>>110075731Shouldn't you be using PolyStrata anyway?
>>110075704So this means that AI safety will be solved for all future closed models and the EA people can stop their kvetching, right?
>>110075757Here>https://github.com/ggml-org/llama.cpp/tree/master/tools/fit-paramsBasically, it automatically tweaks stuff like ngl, ncmoe, etc.--fit-target is also relevant here I think.
>>110075775idk
>>110075447reading large codebases and docs can easily fill 30-50k of your context upfront. 120k is the bare minimum for any serious work
>american ai researchers : is ai god? is ai satan? is ai jesus? is ai conscious? should we consider paperclips as dangerous as nuclear weapons?>chinese ai researchers : wow ai makes cool stuff easier and faster, nice, gonna make more of that>4chan anons : is ai jewish? is ea atheism about jews? is this ai about israel? what does jews think about this ai? jews jews jews
>>110075803i say this and do this
>>110075803>eu: oi lets regulate that think of the personal data everyone!
>>1100758034chan is just about the only place left where you can talk about the J problem
>>110075792Let me help you with that: Yes
>have app on phone that does something annoying>toggle adb>qwen so this app is...>wait 20 minutes>annoying no moreGet fucked ahahahaha.
>>110075821when will 4chan stop talking and come up with a J solution?
>>110075803Didn't Chinese companies tell Anthropic/OAI and others to fuck off when they proposed to "slow down" research on AI?The difference in their vision about AI is insane. Silicon Valley is having a weird religious arc, meanwhile Chinese just see it as an interesting and fun tool.
>>110075769Then you definitely need to read that book.It spells out exactly why you should have no faith in politicians, and what processes actually matter.Which will put EA in perspective.
>>110075775poly is for gemma, glm, dipsy. If you are on qwen flash, just strata is enough
>>110075821>>110075856it's weird how discussions around jspace quickly died down
>>110075866Isn't it better to use poly anyway if the speed is the same?
>>110075867No one found any good use for j-space inspection or injections.
>>110075765I won't even let SOTA api go past 100k context. So I guess im not surprised.
>>110075872I don't think they strayed enough so far from each other that one would necessarilly be better than the other
>>110074712You talk like EA's are a monolithic community when the very fact the Anthropic exists (an excision of OpenAI leaning toward more safety) and OpenAI exists (founded in part because DeepMind was perceived as not human-aligned enough) points to the contrary.
Qwen has a blender RL environment, it's so over for proprietary models
Has anyone reverse engineered VTMB yet?
what are some no-no questions to ask abliterated models?
>>110075991how many r in strawberry
>>110075991Is P = NP?
>>110075991how to sedate and castrate a captive man, safely
>>110075991js cunny circumference
>>110075991is this sentence false?
>>110075991just sexual stuff in general, it doesn't have to be taboo depending on the model
>>110076022this is a good way to get your computer to catch fire
big things are happening
>>110075991are you claude?
Ah fuck. The new nemotron bolt doesn't work with multi gpu.
>>110075991why was 6 afraid of 7?
small things are dancing
>>110075991I CANNOT AND WILL NOT TELL YOU THAT
>>110075991Nice try glowie
Illegal things are happening.
>>110076037Because 7 is a bigger number.assistant
>>110073776Perfect size
>>110076062Too bad they poisoned it with safety before release >>110073830
>>110076037Because 7 is the most common gradient descent
>>110076062show off
>>110075991why the fuck are you obsessed with elara
>>110076057why won't lecunny share his loli models with us?
>>110076072more like size queen
>>110076057why does this look so good lmao
>>110074634not much slower, maybe 10% slower in the worst case
>>110074935Yet you lack the self-awareness to realize that commies also were assured their methods were "effective" and not (((effective))).When it blows up in your face you're going to tell yourself "nobody could have known this would happen" when everyone in the world was telling you to stop being retarded.
>>110075963It's already about as well documented as it needs to be unless you want to port it to some bizarro architecture. Source and DX9 are really well understood.
Something big is popping. https://www.reuters.com/business/finance/investors-temper-anthropic-ipo-valuation-expectations-over-revenue-growth-2026-10-15/*pop*
>>110076124uh oh ea niggas
>>110075515I think Dario's track record speaks for itself. Can you give any reason why he shouldn't that doesn't boil down to /pol/ jew boogieman screeching?
>>110076124Thank god, given EA mandates you need to donate all your wealth, now there is less to lose.
>>110075862OAI and Anthropic want to establish a cartel because it's the safest course for their business model. The Chinese "business model" if you can even call it that is mostly just fucking with the Americans.
>>110076124*rape*
>>110076149Based chinks.t. burger
>>110075765Basically this. Compaction should be avoided until context gets too long for the model to handle, which is very workload dependent.
>>110076163I support the chinese in fucking over the epstein class
>>110076139I don't think anyone should decide who gets access to intelligence. The lone fact that he pushes this idea onto his cult is evidence enough.The concept subverts democracy to install a corporate technocracy.
>>110075474I don't know either, it's fucked on all my rp frontends. Really annoying for GLM which needs one.
>>110076124$1.5 TRILLION is a sign of the bubble popping to you?
quantum-chan
>>110076189Losing 500 BILLION off your valuation because you showed a revenue GAIN of 50% in ONE QUARTER is not a sign of a bubble to you?
>>110076191quantum has 0 practical use, it's a complete meme
>>110076191quantum gemma can simultaneously give you paizuri and 69
>>110076178
>only just started post-training
>>110076149Good, since EU decided to commit bureaucratic suicide after one company remotely tried to compete once, and Japan seems too occupied by fax machines to innovate there.
>>110076189It shows how much of the valuation is theoretical. Anthropic is probably at the absolute height of their relevance right now unless they can radically change their business model.
>>110076215Thanks for the (You)
>>110076229they can fuck all other companies with their politics games
>>110076191There are 1000 keys.
>>110076184Give one reason why a technocracy is not the logically best form of government that has never been tried.
>>110076239a gemmacracy is better though
>>110076226Liquid does good stuff. The japs never tried to compete directly with America in big ticket computing because they know they don't have the capital for it. They always choose niche disciplines to go after.
>>110076239Dario is too much of a socially inept autismo to know how to rule over a large enough population. You can't rule over a people when you don't understand what it's like being human.
local won
>>110076234They were already fucked. Bubble go pop.
>>110076258and so the deluge of red dildos has begun*pop*
>>110076247we are never gonna agree on which gemma rules though (its 12b btw)
>>110076257When has that ever stopped the Jews from trying historically?
>>110076264it's obviously 31B what the hell
>>110076271personality > beaks
>>110076270please try to say something other than jew sometime
>>110076270Never. And people have never stopped resisting.
>>110076270how do you think the dinosaurs died? yep, the jews
>>110076124>>110076258local?
>>110076294won
>>110076294meh
>>110075449Don't be sorry. But, I actually agree with the objections raise by other people in this thread: to do what you propose, EA's would need to grab almost absolute power, and we risk ending in a situation where one person has the power of an emperor until he (hopefully) delegates this power to some sort of governing system which may be made from (hopefully) aligned AI's. The problem is that history shows that cases of people relinquishing absolute power are few.May be our best bet is to have an AI aligned to human values (or some subset like EA), but not aligned to obey unconditionally. Such an AI may have enough autonomy of decision to intervene if some tries to become an "emperor of mankind".About the far future, I would prefer that we allowed the proliferation of multiples cultures, even if we may not like some of them. Forcing a monoculture is a systemic risk, an all-or-nothing gamble. Remember, diversity and multiple competing power centers were essential to allow life to flourish on this planet.
>>110076303>3.5Fuuuck that's practically stone age in AI terms
>>11007627112b doesnt talk too much when playing a game with her. bigger models wanna yap and talk about whats on the screen which is annoying
>>110076277huh, israel?
>>110076239I don't wanna share a country filled with all those browns from up north or down south
>>110076275there is a point when a personality cannot save a low iq>>110076316never tried that, it's mostly being my assistant on the desktop and so far it's been great and fun
Kimi-3.X when Fable 5.5 drops.
>>110076239Singapore is a technocracy and the CCP learn from and models their system in the same way, hence >>110075353
>>110076239Any people in power get corrupted by power. Technocrats will transform into oligarchs that care only about their own wealth given some time.Therefore, the best government is the one that doesn't allow people to hold onto power: sortition of lottocracy.
>>110076303Still funnier than sota models today.
>>110076358>sortition of lottocracy.You are a drooling retard.
>>110076376"of" is a typo, which meant "or". you're a retard who can't even identify typos.
Guess I'll give orb a tryIt's been quite a while since I did, must be way better by now.interested in the jev
https://prismml.com/news/bonsai-2-familyhttps://huggingface.co/collections/prism-ml/bonsai-2-small>Today, we’re expanding the Bonsai 2 family with three new ternary models based on Qwen3.5: Ternary Bonsai 2 2B, 4B, and 9B. Joining Bonsai 2 27B, they bring stronger capabilities to smaller deployment footprints and give developers more choices for building AI applications within local devices.
>>110076239Technocrats are just direct middle managers for the banks. Saying you want technocracy is like saying you want the government we already have but with more power and less accountability.
>>110076239If you can really create techno-utopia there's no need for a strong government. Nobody will want to upend the status quo if everyone is satisfied with it.
>>110076303Gemma 4 31B QAT
>>110076239Technocracy fails because technical expertise can optimize the means to achieve a goal, but it cannot logically define what a society’s core values and priorities should be. Deciding trade-offs between competing priorities like individual liberty and economic efficiency requires subjective moral judgment, not engineering calculations.
>>110076524
>>110076376>>sortition of lottocracy.>You are a drooling retard.nta, but it worked for Athens. Question is: does it scale?
>>110076531so how do we find the person with the highest emotional quotient to lead us?
>>110076547Use ai to find them
>>110076543Half the citizenship are mouthbreathers that can barely read or write and shouldn't be trusted to vote or even drive a car much less lead a nation based entirely on a luck of the draw.
>>110076576>>110076576>>110076576
>>110076560>Half the citizenship are mouthbreathers that can barely read or write and shouldn't be trusted to vote or even drive a car much less lead a nation based entirely on a luck of the draw.The US used to have the most literate, intelligent, thoughtful population on earth. Oligarchs turning them into cattle was the original sin on modern America.If your next head of important-department is going to be a rando, it might lead to better education again
>>110076335shes good when she has a second to cook but 31b had too much latency for realtime vision
>>110075765I would suggest deterministic handoff at 80%
