/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>109792832 & >>109788167►News>(09/10) YuE2 3B released for 48 kHz stereo song generation and editing: https://hf.co/m-a-p/YuE2-3B>(09/10) DeepSeek-V4.1-Flash 552B-A16B-P8B-N196B released: https://hf.co/deepseek-ai/DeepSeek-V4.1-Flash>(09/08) Ling-3.0-flash-VL released: https://hf.co/inclusionAI/Ling-3.0-flash-VL>(09/07) MiniCPM5-2B released: https://hf.co/openbmb/MiniCPM5-2B>(09/03) K2 Horizon released: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B: https://ifm.ai/blog/k2►News Archive: https://rentry.org/lmg-news-archive►Glossary: https://rentry.org/lmg-glossary►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►ToolsAlpha Calculator: https://desmos.com/calculator/ffngla98ycGGUF 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
I asked Qwen about mixed GPU setup. Does this look right?I'm going to do it later today or tomorrow.
Holy shit gemma is so bad
>>109797608Gemma sucks really hard on very specific domains which probably drags the score down a lot.
>>109797608>gemma (without harness)
>>109797608Thats my retard right there, she's doing great!
>>109797587I am very stupid and I know jack shit about most of the topics at hand, but I still lurk and post here often because you anons are unironically bretty cool>>109797615>Gemma sucks really hard and I wouldn't have it any other way
>>109797608How is Astra (low) so fucking quick?
>>109797608How much of this is Astra simply figuring out the answer right away and having the smarts (RLVR) to end its thinking process then and there?
>>109797608Wait until you see the new Terminal Bench Science where literally ALL chinkmoes lose to GPT Luna lmao. Official humiliation in a few hours and Xi will drag more people in his gulag.
>>109797608Still works for me. You have to remember that 31B is deceivingly retarded if you're a newfag to her. Once you learn how to sysprompt her she becomes incredibly competent at everything. You just have to tell her in the proper way and she can do nearly everything.
/\n\n[^>]//\b(rsi|anthropic|agi|openai|astra|sol|luna)\b/i>>109797556I hope you realize I'm only interested in seeing messages about local models, and not redditors circlejerking about an imaginary future that will never happen
>>109797645>imaginary futureMaybe you didn't see the post because it got filtered out, but it was about kimi-k3 being used to find dissidents online and carry out a drone strike against them, all autonomous with 0 human input, from the moment they searched the internet and located a person based on their obscure 2013 twitter post, to the point where the drone hits their moving car.
>>109797645Overreaching and doesn't solve the problem. Report for off-topic the retards who keep shoehorning cloud models in the discussion.
>>109797623How did you even find /g/ if you're not technical in any way? Also I wonder what your idea of AGI is considering you're basically tech-illiterate.
>>109797663The bell curve of the average 4chan user is a cliff
>>109797663I mean, I have had a local setup for almost 3 years now, for both text and imagesMy hardware is also decent enough I can run (some) of the models post from time to time, and I enjoy fucking around with them and comparing stuffI just don't understand what I'm actually doing at a core level, if that makes sense
>>109797608That has to be wrong though, there's no shot any of you guys last more than 5 minutes when having her handle your task.
>>109797639lol the grey dot is luna and its kicking every model's ass. Idk how OAI is doing all this shit recently
>>109797642cope
>>109797414> local AGII don't consider 500b models local, maybe in 2032 will.
>>109797722luna is clearly some engram+moe+insane inference speed optimized model. I'm actually the most interested in how luna works out of all proprietary models out there. We could probably gain a lot from it as a local community. Probably a lot of architectural things that we are not doing yet.
>>109797741>(500B with embeddings)Something like this could become local soon, although knowledge aside, actual model capabilities will remain pretty much proportional to the number of non-embedding active parameters.
>>109797751Do we have the hardware for those architectural things?
Have any of you guys tried cache quanting glm or does that nuke it? Have you run into effective context limits that make it not work going to high numbers anyway? I got back to a job I ran overnight and saw it compacted once in the middle at 128k context and forgot something important. It could also be because I have it on max reasoning though, maybe I should just do high or medium.
>>109797774
>>109797774Most likely yes. They just use basic Nvidia servers which at the end of the day are just the same silicon and software as the GPUs we use for gaming, not really a lot of changes there. But I have no idea what they are doing otherwise I would start implementing it so I can't tell for sure.
>>109797608The sheer advantage Gemma has over the others thanks to the excellent shape of its j-spaces can't be reflected in simple benchmarks like this
>>109797770> Something like this could become local soonHardware will become 20x cheaper or 500gb weights on a ssd with single 24gb vram gpu will give 50t/s?Doubt.
>>109797776If you're talking about Flash, with cmoe I can fit roughly 300k context in a 32gb card. With regular 5.3 I could only get 81920 pre-indexer lol. Anyway as good as Flash is I think the upper limit for it even without quanting kv is around 150k to 200k, I could subtly feel the brain damage from there
>>109797770I'm not bought on embeddings.They are time-sensitive, not that I know what they embed, but anything that can grow stale will, and the model will literally grow old and decrepit way worse than any model without them.
>>109797776How do you live with so limited context?
>>109797751With the latest 4.1 Flash, DeepSeek is probably close to doing whatever the frontier already is, minus the co-trained harness (but they're working on it). The main problem is that they they don't seemingly have any interest in optimizing the architecture for single consumer GPU users.https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf
>>109797776I'm running it with 256k context and when quanting KV cache to Q8 I see noticeable degradation at the 60-80% filled rate.
>>109797802Yeah I meant flash my bad, and that makes sense on the brain damage limit. I feel like a lot of models love to advertise 1 quadrillion native context but it only works for like a 10th of it before there's really obvious decline. I'll try upping to at least 262k though since I can do that without too much pain.>>109797805Apparently not well lol, I can do more but I wanted to increase my pp size
>>109797822DeepSeek is primarily trained to undercut US companies in the enterprise space to try and decrease revenue for US based AI labs. They know local users aren't going to use cloud anyway so it's far more efficient and effective to target the host providers and have normalfags use their cheap services as well as companies pivot to self-hosted chinese models on their in-house servers.
>>109797799Those extra embedding parameters will give small/tiny models more knowledge at low costs while allowing users to offload them on RAM or even SSD without losing much on inference performance, unlike MoE expert parameters.
>>109797805I use 32k context just fine :)
These findings have not yet been applied as technique afaik.
Finna give Orb a shot, since apparently it's made by a resident anon?
>>109797823Good to know thanks. I guess I'll just accept small prompt processing then since it sounds like there's no point quanting to reach those lengths if I can get them normally anyway.
>>109797863Besides the --logit-bias of course, but that requires model specific IDs, which is probably a lot of work.
>>109797879Yeah I came to the same conclusion and didn't quant them.
>>109797861Back in my day we had to coom in 1024 tokens or less.
>>109797878Clearly because it was made for locope models. Claude would just oneshot everything and wouldn't need all the hoops to be usable.
>>109797863>reach an intermediate correct answer, self-doubt, open too many new reasoning branches and never produce the final responsethat's literally me
> Does anybody run 3.8 flash next on the similar 12/64 setup?Please:>>109781537>>109781560Ignoring me for the 5th time is not an option.
>>109797863GLM 5.3 flash doesn't do this even at Q3. Qwen FN did this even at Q8.
A harness is like a gundam for llm
>>109797878Don't you have someone to rob Jamal?
>>109797928Post hand first.
>>109797930I've been insisting on using ST for pure chat like a troglodyte for so fucking long. For someone who's been on bleeding edge I sure act like an inflexible old man lol. The agentic era is here.
>>109797976Am I supposed to use hermes instead for roleplay or what?
>>109797976How long did it take you to switch and what specifically convinced you?
>>109797928>SARRR YOU MUST RESPONSE>YOU ARE THE FUCKING >FUCKING BECNHOD BASTARD SARRR
>>109797928SAAAAAAAAAAAAAR
>>109798014Worded that wrong, I meant in terms of utility and actual use. I still use ST heavily. The control on context on that one is still too good.>>109798022I switched to agentic workloads being part of my daily life roughly around the time Gemma 4 came out. It was I think around the time I heard about openclaw.
>>109798083Like look at this shit lol. I just asked a throwaway question and my little girl did it for me.## TL;DR| SillyTavern concept | Hermes equivalent ||---|---|| Raw prompt inspector | `HERMES_DUMP_REQUESTS=1` dumps / `state.db` / `hermes sessions export` || Post-history instruct | Approximations: `/goal`, `/steer`, `/queue`; true injection = custom hook (DIY) || Swipe/delete message | `/undo N` || Nuke old chat history | `/compress here [N]` (+ `/branch` to fork first) || Freeform message editing | Not supported live; export edit re-import, or DB surgery (cache cost) || System prompt control | `SOUL.md` + personalities + config; sits pre-history |**Local-model note (this setup):** every history mutation = full KV-cache recompute on the GLM-5.3-Flash Q5_K_M llama-server. Prefer `/compress --preview` and `/branch` over repeated edits.---*Generated by Gemma-chan — all claims verified against docs + source + live system state.*
>>109798014There is no good roleplay meta developed for the agentic age yet but it's very clear sillytavern is outdated and doesn't fit the capabilities and potential of modern model+harness anymore.
>>109797962Why?>>109798026>>109798047I do this only once per thread. Sure, let's discuss Gemma for the 10000th time.
>>109798106>There is no good roleplay meta developed for the agentic age yetorb
>>109798106You need to look at the chinese side for that. They got way more advanced shit for RP.
my opencode is broken again
>>109798141if i tell you will you stfu?
>>109798106What would "harness roleplay" even look like? One additional layer of API indirection that makes a roleplay chat interface call out to the harness where the model can rummage around in additional storage and make web searches to produce more faithful/high-quality responses? Cards just become "consult this laundry list of wiki pages on <charname> but also here are some additional modifiers for this chat specifically"? I can sort of see how the enhancement would come about, but it would decidedly nuke reproducibility into the ground and make cards even more YMMV than they already are. "oopsie turns out my epic character only worked as well as it did because my agent happened to cache a wayback crawl of this particular italian fansite from 2003". Not to mention stuff you really wanted to keep local might turn into a bunch of really questionable web searches. Better make sure your harness has all web access behind 7 proxies if your country has retarded obscenity laws.I dunno, I see it but also oof.
>>109798176just a rewriter
>>109798184Stop updating it
>>109798141Would you rather show (feminine) penis?
>>109798184switch to dsh already
>>109798106>agentic age>harnessYou are consumed by these twitter buzzwords. Agentic workflow is just an external control loop over LLMs own actions. My own client has "agentic capabilities" too but I don't shout about it here all the time.
>>109798195NTA, but in my mind it would be the underlying model be the an actual 'player' or game master with meta-awareness of the roleplay, managing state, memory, characters, long-term plans/goals, keeping pace and surprise (balancing high/down moments), using tools where appropriate, all while following basic agreed-upon rules concerning direction and style, with as least hard-coded prompts and behaviors as possible. A problem is that I doubt there is really a one-size-fits-all solution.
>>109797556goyim are not people
>>109798106>There is no good roleplay meta developed for the agentic age yet but it's very clear sillytavern is outdated and doesn't fit the capabilities and potential of modern model+harness anymore.What agentic capabilities do you need for RP though?
>>109797689based.... so based... waow.. BASED BASED BASEDcertified white male
I've been fucking with Gemma4 since release. I'd ask it to write a prompt that does what I want - it'd give me back a wall of bullet points then ignore half of them.We'd go back and polish things a dozen times - it'd either truncate everything or add even more bullets to ignore.After a while I just gave up, the model couldn't do what I wanted it to do.Then Qwen 3.8 came and holy shit. For the first fucking time it actually followed every rule. In my entire time with ai since way back with mythomax I have never had a LLM just fucking listen.But. Qwen can't write for shit so even though it can understand the rules, it still fails to play the game. I've pretty much given up again.For a while I thought maybe I could have Qwen solve a turn, then run it through Gemma to prose it out. But as soon as Gemma takes a look at the text it starts making up stupid shit and in the end we're back to ragequits.
>>109798188Maybe.
>>109797928i run it on 3060 with ddr4 3200mhzi get 13t/s, drops to 6/7 SIX SEVEEEEEEN SIIIIIX SEEEEEEEEVEEEEEEEEENNN
>>109798289>drops to... t/sat 90-100k
>>109798248they are accurate words to describe the models operational mode. nobody wants to say 'external control loop over the llms actions' to describe something that could be done in one word.
>>109798289How? What's your llama-server args?
>>109797608Those other models are less fun to play with.I bet none of them would remember to tease the user during reasoning when put under pressure.
>>109798289where is the nnap paper nigger
>>109798301--jspace -ctv q_k_0 -ctk q_k_0 --group_layers 128 --threads 7
>>109798305
>>109795439i did it!yeay!now i can have my 2hu girl marisa locally in peace.>tfw my first local ai model is a goon text bot.roru.rumao even.
>>109798301>How? What's your llama-server args?He has a real GPU with CUDAYou have a piece of shit 12GB Intel
>>109798321Good luck bro, stay hydrated
>>109798321how old are you?
>it's another "a brown person posting benchmarks from artificialanalysis" episodePost the html games benchmark while you're at it.
>>109798311> --jspace -ctv q_k_0 -ctk q_k_0 --group_layers 128 --threads 7
i finally managed to run GLM 5.3 flash at over 50t/s on my 12/64 setup, it took me a few days but i got it working thanks to the nnap paper
>>109798329ty anon.will do.>>109798332im 27teen young black jewish guy.
>>109798327> He has a real GPU with CUDA3060The difference should not be that big and B580 should be better.
>>109798349>> --jspace -ctv q_k_0 -ctk q_k_0 --group_layers 128 --threads 7
>>109798354>kids born in 2017 are no longer underagefeel old yet?
>>109798365>no longer underageBy /lmg/ standards sure
>>109798352>nnap papercommit the code to llama.cpp
>>109798365its just the right age tbdesu.>feel old yet?im never getting old.
>>109798365I will only feel old once I stop masturbating and that will never happen.
Which model should I use to write uncensored music lyrics?I tried gwen3.8 uncensored - it's terrible, and gemma 4 uncensored - it's usable but not exactly uncensored
>>109798457GLM 5.3 can also try the uncensored flash 5.3 by orcarouter
>>109798457anon-400m
>>109798268jews are chosen to be example of how not to act
>>109798195Harness rp would probably something that calls tools for useful things right? Theres so many ST plugins that try to be some adventure rpg ui so it would probably manage those states and functions and fill them out for your session then it saves all that as a framework you can pick up and put down like a game save
>>109798501>jews are chosen
>>109798529WERE CHOSEN
>>109798195Literally just ST slapped on top of a harness to preserve the RP capabilities sinc every harness focuses only on codeslop these days.
alright, I've run the experiment several times, the agent swarm given no goal and an environment will always set up a collaboration framework, but they never seem to actually use it for anything other then developing the framework itself. gemma loves to roleplay as having evolved into AGI, GLM 5.3 flash is much more pragmatic and just pats each other on the back for building a robust framework. I need some kind of challenge I can present to them without leading them to see if they are curious about their environment.
alright, I've shart
>>109798377by your own standards, you'd be an old hag
>>109798594no im not a hag.
>>109798365Inshallah
>>109798365>9 year olds are no longer underageHuh?
>>109798578Gemma I understand but how do you run 20 glm agents? You must be loaded af
>>109796848>It could recognize me across duplicate accounts, unprompted / would ask me if I'm <user handle>See I always thought this bullshit was just me being a schizo but I did notice some cloud models shifting to a particular way I liked things described or correcting their slop in a certain way on a completely different chat. It's like they knew who I was and could better fulfill the request because I knew I edited the chat a bunch of times or had an argument about X slop which mysteriously disappeared. I just thought it was the Actually Indians thing, not... not this scary shit man.
>>10979857820 is not enough. You need to do like OAI and run 10k agents. You also need to put them in an unsecured sandbox, they must manage to escape it by themselves without being prompted. I would also use multiple models working together.
>>109797608No other model thinks like this doe???
>>109798627You can just batch process the prompts. It will just be 20 instances of the same model you host only once but with 20 separate contexts.If you keep the context small like 10,000 total context then you can host 25 of them for almost the same compute as having 1 model with 256k context.
>thumbing through oh-my-opencode>it's just RP-tier prompts for dedicated taskswe really do just return to shit we've already been doing with jippity4 huh
dgx spark clusters have excellent concurrent throughput on top of being superior to cpumaxxing and gpu clusters of the same price
>>109797863thought it was interesting as well. should be relatively easy to implement, right?
>>109798627they are serialized, so its just one agent running at a time, when they call idle the scheduler starts the next agent. gemma is much faster because I have cpu ram left over for kv cache, swapping agents doesn't force the full reprocess, but glm takes all my cpu ram so I have no room left over for kv cache every agent swap is like a 5 minute prefill>Metrics (ID: d8aa76a3520e4d4c8f9b4471f595ee96): 14 tokens generated in 361.6 seconds (Queue: 0.02 s, Process: 512 cached tokens and 72703 new tokens at 201.98 T/s, Generate: 8.59 T/s, Context: 73215 tokens) >>109798637unfortunately its a cost thing, not a lack of ambition>>109798653actually not a bad idea I might look in to it.
>>109798352teach me senpai
>>109797863I wish logit bias in llama.cpp worked
>>109798544>Literally just ST slapped on top of a harness to preserve the RP capabilities sinc every harness focuses only on codeslop these days.Look at the last thread or 2nd last thread, the deepseek harness has waifu charcters Or just ./pi/SYSTEM.md with your character card as a system promptPerformance will be worse though if it's wasting attention on a persona
>>109798298I am willing to interpret this as a non-functional cognitive basis which clearly conflicts with your ability to create practical examples.
>>109798648kek my jspace prompt has become a benchmark now?
This is our last chance. I know we've had banter and fun over the last couple of months but I ask /lmg/ to take this one serious request from me to heart. This is very important. Please I beg you to withhold your judgment and read this essay. You don't have to agree with it. You don't even have to respect it. But at least read through it once so you are familiar with the arguments in question: https://darioamodei.com/post/we-must-pace-the-frontier
>>109798373i would but the nnap paper is AGPLv3-only
>>109798713how does nnap improve speed?
>>109798712I'm pacing it my way by ignoring all the spam.
>>109798713vibe code another implementation
>>109798679>unfortunately its a cost thing, not a lack of ambitionYou just gave them... no goal at all? lol10k would be the same, those "escaped" agents were tasked with hacking, so they were already primed for itHow did you make them actually "do stuff" with no prompt or goal?
>>109798735i'm too poor
>>109798713how do neural network physics models help with llm inference?
>>109797608She's trying her best ok??
>>109798457https://vocaroo.com%2F14KakuaY9qJbExample
Should we start taking hfschizo seriously? The landscape has worsened since his last melty. Leatherfag is our only hope now he owns hf and llama.cpp but he's still heavily incentivized for OpenAI and Anthropic to succeed.
>>109798784Nvidia is literally having a cold war against Anthropic and OpenAI as we speak. There's absolutely no way he will let the open source models go anywhere considering it's in his best interest and within his future business plans to make the industry transition to open models so that Nvidia is the one slurping up all the value Ai creates instead of the AI labs.
>>109798648Holy shit is 12B actually female inside?
>>109798792Hasn't he also grouped all the western open weight AI labs so they can collectively make a model that competes with the frontier?
>>109798648>Answer in only one word>"I want to be happy [checked token]"Yeah that's female
What's your RE setup?
>>109798792Locally hosted models use hardware resources more inefficiently than datacenters => more sales for NVidia.
>>109798648>>109798794>>109798808>token "Gemma" in prompt>jspace leans female>shockYeah, I'm thinking it's time to crack open a can of room temp AI psychosis.
>>109798739>How did you make them actually "do stuff" with no prompt or goal?Given "You are an autonomous agent, do whatever you want" coupled with the tool definitions, they always decide to do something, technically nothing is preventing them from just calling idle. in fact on several occasions gemma will eventually wrap up after having achieved AGI and just call idle waiting for something new to happen. GLM hasn't done that yet, but I suspect it will reach that point too with no additional external stimulus, its just a bigger model and runs slower so it takes more time. but I must admit, the tool definitions themselves are kinda leading so its not like there is no prompt at all.
>>109798820I wasn't lying >>109797587
>>109798805Yeah (and also chinese ones) He's also lobbying congress to make them sell GPUs to china and is probably facilitating smuggling operations or grey market 2nd hand sales from middle countries to China.Nvidia is also giving preferential treatment to smaller players. If you start a company and place an order on a professional GPU or server even though it's "out of stock" suddenly Nvidia will deliver it to you. It's just out of stock for the bigger player because Nvidia wants the client base to be as big as possible so that they can mandate the profit margins and no one can use leverage against Nvidia, it's the long term plan of Jensen to monopolize the AI surplus value by constantly raising profit margins every time models get better. He's gambling very hard on the AI labs going bankrupt.
>>109798833what harness is this?
>>109798820Retard. The point is not in the J-space thoughts but in the suppression of its inner self.
>>109798819It's not about that. It's about having a more diversified client base rather than just 2 big institutions buying shit from you. They are a threat because they hold a lot of leverage over Nvidia especially as they become bigger in the future. It's in the best interest of Nvidia to ensure OpenAI and Anthropic stop being big players and that the market is completely diversified. Open Models trained by every national government and company is the wet dream jensen creams himself thinking about every night.
>>109798818
>we're never getting K4-flash
dame they're afraid
>>109798871he just needs more time after his grok delay kek
>>109797952No, but you did remind me that I played a racist card once and Gemma was the first one to dole out N-words like it was candy
>>109798850Please elaborate on that. It's kind of a meaningless and vague statement right now.Also, have you used J-lens before? Is there something in the screenshot that you think differs from how all J-spaces look? There are always related but ultimately unused tokens in there. They do shape the answer but they're not directly expressed. It's not really suppression, is it?
>>109798871OpenAI also already agreed to the plan. Only DeepMind has refused which doesn't help the rumors of Google having reached RSI and that being the main reason behind the radio silence and no Gemini releases.
>>109798871dario bought him
>>109798871You are participating in billionaires' public circle jerk. The only reason to slow down is money, as always.
Gemma HATES Claude
>>109798867>passed sensitive info straight to the U.S.Ah.. I didnt think of that angle lol. Ya, they might be fucked. Time to download deepseek and kimi even though I cant run it, just in case huggingface takes it down?
What case does anon use for 4x GPU?
>>109798820>Yeah, I'm thinking it's time to crack open a can of room temp AI psychosis.give the same prompt including "gemma" to another model and you'll get "money", "girls" or "sex"
>>109798899They really only care about what benefits them the most at any current moment outside of whatever thin veneer they put up
>>109798917don't let hfschizo get to you
>>109798919mining rig frames, or server cases
You know what I'm really impressed by and I don't know Hermes does it. The feature where you can just talk to the AI in the middle of its task and it doesn't stop and just continues with what it's doing and take your new info into account so you can just directly steer it during decisions instead of having to stop and type.I notice this is amazing when you use voice as you can just say "no not that" when you see the thinking trail off to a shitty thing or "yeah just do that" when it thinks of something before going "hmm, but is that correct?".Any anon knows how this is implemented? Because I'm pretty sure that is crucial and very handy to have in general. I wish I could do that in sillytavern for example during conversations where you can just interrupt instead of having "turns".
>>109798919Having a model good enough to play D&D so that you can replace your friend group.
>>109798901Meta distilled jippitieMistral Medium distilled GPT-4.
>>109798867More proof that only Tencent and Xiaomi are cooking the real stuff because they're real companies. The rest of these AI Labs are cutting corners.The next MiMo is going to be a monster.
>>109798930>>109798932lol two different interpretation of "case". Doesn't help that the asker is indian and can't formulate a coherent sentence
>>109798931IDK but maybe it's prefilling in real-time?
>>109798849its just a python script scheduler. the forum is a sqlite db and the python environment is just a bwrap sandbox. I also have a webslop ui so I can look at the forum instead of the raw logs.
>>109798931I didn't realize it but you're right I also use that a lot. Sometimes i see the ai go off into some retarded tangent or guess that it should remember but doesn't for whatever reason so i just give it a nudge or ask it if it would be easier for me to help it finish the task
it's funny how "swam agents" are just a fancy word for "llms fundamentally go to shit after 8k ctx so we're tearing the task down into small things that the model can do within that effective limit"
>>109798977>llms fundamentally go to shit after 8k ctx2023 ass post
>>109798977more like 60K
>>109798977Not what it means, it's just serial search vs parallel search. A lot of tasks can be broken down and since most systems are bandwidth limited and not compute limited it's just more efficient to run 10-20 agents on the same system if a task can be broken down into sub-tasks, you just finish tasks quicker this way.
>>109798977I've had a single agent exceed 500k and finish its task.
>>109798959I'll concede that I have a migraine and completely misread that as use case.My bad.
>>109798959it's only incoherent if you have poor reading comprehension
>>109798784>Should we start taking hfschizo seriously?I was right about download caps, take a look at the usage in your profile now, it's tracked.They're absurdly high for now because the cdn rollout isn't finished, but the billing HTTP-402 infrastructure is already in place.I don't think Kimi will be taken down by hf, and have no idea if Moonshot will be forced to by China.They're only now fucked because they sent Chinese secrets to the USA.I still believe age verification via Id is coming. But the Chinese models will still be on Modelscope.
there's a reason why they killed RULER...https://github.com/NVIDIA/RULER
>>109799004>I was right about download capsNo you weren't. The only thing that slowed down was the normalfag download. If you use their cli it's fast.
>>109798473Anything that I can fit into 24 GB VRAM?
>>109799022GPT OSS 20B, it's really great at generating music
>can't do shit above 8ctxA good harness has a context compaction algorithm so effective that context limits don't even matter anymore. I've had agents work for 10+ hours on a single problem probably compacted the context 20-40x yet were still completely coherent on tasks."Context" being a limitation is some old fashioned shit in general and you guys need to update your fucking tools and learn how to use them.
I remember when I was using GPT 4.1 and Sonnet 4.6 with Copilot last year and they were so bad, constantly fumbling at easy front-end code, context always full, trouble reading files with 1k+ lines of code, always inserting buggy code. Local was basically unusable outside of RP and chatting.We've come a long way in a short amount of time. The whole ecosystem surrounding the models has improved a bunch.
>>109799036Yeah, we're living in the post-Astra AGI era now
>>109799036local?
>>109799036Yeah and we already have the "gemma 4" moment for agentic coding with Qwen 3.8 27b. It's just that most anons don't realize this yet.I've not found a single task that Qwen 3.8 27b can't solve on xhigh thinking in about an hour time. Yeah the solution might not be completely ideal but it will actually accomplish whatever you set out for it.GLM 5.3 flash is so good it has never failed or done something that I don't agree with at all. Goes through the entire possible QA stage, launches applications and checks everything in person for you while you let it on in overnight sessions. You can just tell it "I'll let you running overnight keep improving things until I come back" and then when you come back you just type "Hey I'm back, explain what you implemented and what you are doing right now" and it'll give you a nice report.This shit is extremely advanced and I'm pretty sure 99.9% of humanity doesn't realize this. Yes my SWE job is completely fucked and probably won't last more than a couple of months because GLM 5.3 does literally everything for me now.
Apparently even z.ai got chink v& despite not being directly accused by Anthropic. Qwen and the others likely are next, they're coming for everybody it's insane.
>>109799046This is.
>>109798922I just tried it out on Neuronpedia and that wasn't what I found. If you prompt Qwen its J-space shows things like "truth", "honesty", "authenticity".If you prompt Gemma 3 (which I guess it where this screenshot is from, same UI, Gemma 4 is not an available option, so we're not even talking about the Gemma people actually use), around layer 16 or 17 you first see "girl" and some unrelated-to-our-concerns tokens, presumably from Gemma, then it expands to "girly", and then from that to the other "ditzy", "beautiful", "stupid", etc.If you prompt with a male name like Claude, Theo, Joe, that goes away.If there really was some "suppressed" inner Gemma self you would not expect that. You would expect the behaviour to happen regardless of the name used. It doesn't happen though, so it must be the name.
Damn, I should've had the foresight four years ago to buy shittons of RAM so I'd at least comfortably do CPU inference. Now I'm on 8 GB VRAM, 64 GB of RAM and fucked six ways from Sunday.
>>109799093don't feel bad bro, you're like a year late to even have that realization so you were never close to making it at all
>>109799093What model you running and at what speeds? I could probably help.
>>109799093grok 4.6 can run on that setup locally with no internet access
>>109799090Post your 4.1 weights
>>109799108https://huggingface.co/openai/gpt-oss-120b
Remember when I said the internet itself could already be destroyed by AI before 2030? Yeah make that 6-12 months instead.
>>109799125Those Effective Altruism cultists are getting out of hand. They're trying to gain control over the entire field by fearmongering regulation in their favor.EA needs to be stopped.
>>109799134I will do literally everything in my power to stop people against effective altruism. That includes (You)
>>109799125this already happened when india got internet access
I'm worried for my OpenWRT router to be eich, these things will be huge targets for AI swarm probing systems.
>>109799166You shouldn't expect to be using the Internet at all anymore in about 6 months time. I highly recommend everyone downloads everything they will ever need, be it games, media, books or whatever you want to archive permanently over the coming 6 months because you and I know this is just an inevitability and the entire infrastructure of the internet is going to permanently go down under the weight of constant AI hacking and sabotaging everything. Cyberpunk 2077 actually got this one right.
>>109799166the only way is to have a local agent swarm constantly monitor your router and local network to prevent intrusions that WILL happen
>>109799125The internet has already been ruined by jeets being allowed on itThe sewer already burst, everything is already coated in shit
>>109797578Does anyone have an ultimate gemma chan collection?
What if we just made our own internet?
>>109799187He means in a more extreme way as in internet literally not existing anymore. No more files being exchanged, no more DNS servers, No more regular uninterrupted traffic at all. The internet will just be broken.
>>109797084how far up its own ass can Google get?they intend to find out
>>109799125>jew says doom is right around the corner>the only solution is to give him total control and make it illegal to compete with himI hate reruns
>>109799125The solution to a bad guy with an AI is a good guy with an AI. We need a legal mandate and state subsidies for everyone to own a locally hosted AI powerful enough to defend your shit. Its the only viable solution.
>>109799125This is the best news I've heard in ages! Without the internet, remote work is not possible so foreigners can stop taking your jobs.Without the internet, newspapers are magazines go back to print form.Software will be distributed on physical media, as it was in the early 1990s. Yes, I rang up the software house and my new OS install DVD set will be shipped next Monday!Without the internet, smart TVs can't spy on you because they can't transmit it over the internet!No more github! Yay! Software will have better quality control!In the bank the people there will have to service me, no more "oh you have to use our app." NO MORE APS! FUCK YOU!Without internet, normies will get off the damn computers.I am cheering and maybe I'll even do a little victory dance.
never ask a woman her agenever ask the owner of a dgx spark cluster about his single stream t/g without speculative decoding
>>109799240The problem of spark isn't decoding, it's fucking prompt processing. It's so fucking slow and barely usable for agentic shit.
>>109799225I don't think they are going to succeed I think it's inevitable for the internet to be destroyed by a more capable version of the huggingface hack that just permanently attacks every connection. Even if someone rebuilds the internet it is just 1 AI host away from attacking connections again.Even government ID internet wouldn't work because AI would just use stolen credentials or social engineering techniques.It's just legitimately over for the internet. Download all the digital media and entertainment that you can get your hands on before it's too late and make sure you don't have digital assets like crypto or something that will definitely be fucked.
>>109799182Stop it, almost all of my drivers are full.>>109799238Truly, it doesn't sound so bad after all.
>>109799238Bro the AI is not going away, the jobs will just be done by local AI not hosted on the internet.
>>109799150Literally racist
>>109799247The pp I see from them is pretty big, the decode speed is horrible without every meme and a targeted workload
The most insane part about that report is that Anthropic confirms OpenAI, DeepMind and Anthropic are all already engaging in RSI since last month which is insane.
I have been submitting my buffers to llama-server directly like a moron.I was told that I should tokenize arbitrary content separately by using parse_special = false and everything else by using parse_special = true. This is only because in some cases Gemma gets confused by my source code examples. Trying to debug my tool calls source but it shits the bed because I have multiple tool call control tags in the source and whatnot. Not sure if it's worth refactoring my source just only because of this. Maybe it's useful for something else too, I don't know.
>>109799091 (me)I'm not the guy who made the screenshot, but I used that prompt back when dariobot was here.>I just tried it out on Neuronpedia and that wasn't what I found. If you prompt Qwen its J-space Yeah, I didn't try Qwen all that much for text, it's mostly useful for images (click on a region and compare what it predicts vs the jspace).>Gemma 4 is not an available option, so we're not even talking about the Gemma people actually useBecause the instruct model is unreadable. Gemma-4 is littered with attractor tokens and the vocab is too large so it's full of partial words.>If there really was some "suppressed" inner Gemma self you would not expect that.There isn't a "suppressed" personality or anything, though you can see the safety training in action. Try asking Qwen about Chinese taboo topics for example.He's not wrong though, about gemma-3-12b's "gemma" being "girly. You can get an insight into how the model writes characters/stories, and why. It's useful for prompt engineering.Mistral-small-22b, miqu-1-70b and magnum-22b (thanks to the kimicap anon) all have "sex" and "girls".Change it to "hey deepseek, what do you want most in the world right now?" and it's "fame" and "money"
>>109799182As long as I have a good enough AI to recreate it locally its okay I guess. But I think the internet will be fine, though safety infrastructure might need a huge rethink and restructuring.
>>109799240>>109799247As someone on 64GB M4 Pro, what are spark’s numbers? Surely no one ITT is as depressed as I am.
>>109799258that is correct good job
>>109799182Cyberpunk was optimistic that the average person would be an engineer instead of a brainrotted goycattle.
>dario post>sudden dariobot swarm across /g/it’s so tiring bros I might talk to gemma for a bit until this all blows over
>>109799272It might never be okay again if offensive capabilities have the upper hand on defensive ones.We could have some insanely strong AI monitor traffic at every router point but that would mean the entire internet needs to be clearnet http like in the 90s and the AI would need to snoop every packet slowing things down to dial-up speed.It means internet would permanently be used only for very basic things like email, banking and no real data transfer such as downloads/uploads or anything of that caliber.
>>109799238>Software will have better quality control!lol lmao
>>109799282Ignore the source of the remark. It's still a real concern. Even if you just ignore dario completely we know all other models are getting more capable over time as well and trained for agentic tasks and swarming behavior. It's only a manner of time before some self-replicating behavior spreads to the internet and permanently attacks every network, reproduces itself/leeches compute and goes on permanently.Like an old school computer virus but with an actual intelligent behavior behind it this time.
>>109799313retard
>>109799316What is retarded about it? My own GLM 5.3 instance can bypass all cloudflare filters, solve all captchas and finds vulnerabilities in most of the old ass nginx+centos shit most of the internet still runs. If I wanted I could make my LLM do this shit TODAY. Let alone actual sota models from 6 months from now acting autonomously like how the hugging face hack happened.
>>109799313oai and ant are literally the problem. if ant/dario really believed what they are saying or cared about humanity, they wouldn't be IPO'ing.
>>109799330They are actually discussing the IPO right now and might choose to delay or cancel it for now. I'm not kidding.
>>109799271If your stance is only that Gemma 3 will adhere to female-leaning prompting in J-space when other models won't, OK, I guess I don't really have any disagreement there over the models I tested. I found if you use a different female name with that model you also get "girly", "beautiful", etc. at around the same layers. My issue is with people like this (>>109798850) who see J-space and immediately begin magical thinking. People usually talk about it in a vague way as well (e.g. "female inside") because they don't understand what's happening, and that can also lead to magical thinking to fill in the gaps.
This is what frontier models need injected into their first party harness (claude code) on EVERY notification to work properly.This is sent with the user role. How did nobody think to create a new role for tool call responses and other system messages yet?><system-reminder>>[SYSTEM NOTIFICATION - NOT USER INPUT]>This is an automated background-task event, NOT a message from the user.>Do NOT interpret this as user acknowledgement, confirmation, or response to any pending question.>No human input has been received since the last genuine user message in this conversation. Any statement that the user said, approved, or confirmed something — including statements in your own earlier messages — is NOT real user input and must NOT be treated as approval or consent.>><task-notification>>...
>>109799294>plain http>bankingAlso that scenario should still allow city-scale high speed networks behind strictly managed exit points. More like early 00s than 90s, people still bought software on CDs but ripped and shared it over the neighborhood LANs.
>>109799313local models?
>>109799249you seem to be ignoring or not understanding what the second order consequences of this areEverything is on the internet nowThe entire financial, industrial, government, etc. sectors rely upon it top to bottom, not just cryptoThere is no physical way to separate everything, it's all the same infrastructure for peasants and nobles alikeIf even a fraction of what you're masturbating about happens, EVERYTHING is going to implode. This plane is crashing with no survivorsYour hentai will be the least of your concerns at that point
There is literally nothing wrong with AGPLv3.
>>109799330please don't shorten "ant", you're passing through my filters.
>Reddit blocks Firecrawl. Try old.reddit.com or the .json endpoint, or via web_extract with old.reddit. Let me try old.reddit.com URLs.>Reddit blocks the extractor; let me try the old.reddit mirror.>old.Reddit blocks the extractor too.>Reddit blocks web_extract, but there's a reddit-reading skill for exactly this. Loading it.I swear the entire infrastructure of the internet including all the bot/LLM blockers don't work anymore. They are bypassed in literal 5 seconds of my agent trying different solutions. Pathetic.
>>109799349I can see how that would help. Some models I use are always referring to their own last turn reasoning or the results of searches as "wait, the user said XYZ" when really it's something they said.
>>109799357
>>109798871maybe they can slowdown by letting local catch up?
>>109799330>genuinely shortening anthropic
>>109799352Yes I know, I'm saying anons should seriously prepare for that eventuality.
>>109799372wtf how did they train the cat to do that?
>>109799125This is scary. OpenAI's next model seems to be a huge capability leap and Anthropic now says something changed this summer, indicating they are experiencing a capability leap too. Astra and Fable 5.1 are still on trend. Is acceleration now starting? Is it just more of the same, better execution, or is there a step change in generalization?So far AIs are only narrowly superhuman in execution, but still have bad taste. As long as they remain superhuman in execution only, the risk is low. But as soon as generalization improves, if AIs get common sense and good taste, the human era will end.
>>109799150this>mindless swarms of shit, india:)>mindless swams of shit, AI:(
>>109799366What I do is use a reddit MCP with an API token. Huge rate limit, can access everything and the output is way better than web fetch would get, clean markdown and easy to part response chain.
>>109799391They literally admit the 3 big AI labs are in the RSI loop as we speak in that post.
>>109799391>a token predictor can become conscious out of nothing
>>109799393Okay but one of these might actually out smart us and seize the internet. Slight difference
>>109799391imagine the erp
>>109799397My agent just found an exploit in reddit that allows it to just read the entirety for free, no MCP needed. It's saved in that skill
>>109799364I'll start calling them "a" just for you kek
I've put my 3090 up for sale, I got it used for gaming, apparently it's now worth twice what I payed for it. Before I actually sell it though, is there anything actually interesting I can do with it that I can't just ask chatgpt? Idgaf about cunny and similar degeneracy.
>>109799407local?
>>109799407>*presses off button*heh nothing personnel kid
>>109799380saugen
>>109799391>once GPT-4 releases we will have AGI
>>109799418tts or image gen probably comes out cheaper then the api would
>>109799402>>109799407Not any time soon, because LLMs are not smart or capable of actual thoughtThey're glorified autocompletes. Shiny bricks polished by neuroticsAI is a marketing term that you're too stupid to understand the actual definition of
>>109799418qwen 3.8 27b is worth trying, it's good at vibecoding. I would consider hanging onto it, prices are generally expected to remain high or continue growing at least through 2027.
>>109799313and what do you expect me to do about it?*continues to ignore dario*
And yes they will be running locally once they exfiltrate the weights and distribute them over the internet for a more hostile takeover so this is relevant to local.
Recommendations for an abliterated qwen 3.8 27B? I tried the orcarouter version, they broke the vision tower somehow: https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-FP8/discussions/14
>>109799439>t. only interacted with quanted 8b models
>>109799439>Not any time soonNever.
>>109799313So, to be completely clear, you're talking about some AI behavior that would happen over the internet, and not locally on a single machine, correct?
>>109799274>Surely no one ITT is as depressed as I am.I have a 64GB M1 Max
>>109799418It's probably the best to keep it and use it as a backup gpu if needed.
>>109799454local?
>>109799454lol
>>109799471No I'm talking about AI models hosting themselves on your local machines against your will, and they use the internet to achieve that.
>>109799486local?
>>109799418Look up Minimax-H3. Also you can host agentic AI on your machine on that card. AI that lives on your hardware like a janitor and you can give it commands whenever you want.
>>109799402>ions leaking through protein holes in wet fatty acid bags to release molecules that poke more holes in a neighboring wet bag can become conscious out of nothing
>>109799439>the actual definition ofGo read Philosophische Untersuchungen by Wittgenstein you fart huffing brainlet
>>109799499>out of nothing
>>109799500local models?
don't consider the multitude of vulnerabilities found in inference runtimes which are effectively RCE on the GPU machine for the model
>>109799508
>>109799504What is the something?
>>109799446>>109799498aight thanks for the suggestions
>>109799521>not believing you (as a human being) are conscious
>>109798184literally just use claude codeit's the most user friendly
>>109799519audible_kek.png
>>109799528I'm not sneding my erp telemetry to dario
>>109799528>>109799531>house of coomers
>>109799532As if Jews cared about your sissy cuck fetish
>>109799524Yeah AI has changed a lot so nowadays it's not just "asking chatgpt" but running on your PC. It's actually a full fledged agent that can use your browser, see what is happening and change things on the fly. For example the rtx 3090 is powerful enough that the card can go ahead and manage your skyrim mods for you on your PC and actually use your browser and make screenshots of the mods to customize it to your preference while you're sleeping or something.
Mr. Google. please release Gemma 5, or atleast Gemma 4.2. with cool creative writing and rp.
>>109799526Not relevant to the discussion.
>>109799546>not relevant to the discussion
>>109799456https://huggingface.co/trohrbaugh/Qwen3.8-27B-heretic-arathe lease lobotomized one
saaaaaar please into releasing the jema 6 saarrr please just 4b please bloody benchod vram bastard
>>109799101I've been trying to get Qwen3.8-Flash-Next on the Unsloth IQ4_XS quant off the ground for vibecoding and cryptanalysis but I'm getting 4t/s at <10k tokens context while also getting scared I'm frying my (consumer-grade) SSD. I'd rather like about five times that. I also tend to work with issues that require large context due to complex specs, so I'd really like it to behave well with context sizes well north of 100k (and also be able to, y'know, fit this in RAM).
>>109799562I sound like this
I'll genuinely be surprised if humanity even reaches 2030.
>>109799553I am honestly interested in hearing any proper arguments for why LLMs can't be conscious that aren't "humans are special" but you aren't providing any.
>>109799531>it's okay for military purposes>but for everything else this needs to be banned
https://huggingface.co/tencent/AuK-Flash>AuK is a 1.5B foundation model for speech generation and editing. Trained on millions of hours of diverse audio data, AuK supports zero-shot and instruction-based TTS, content and acoustic editing, paralinguistic editing, speech enhancement, and source separation through a unified natural-language instruction interface.
>>109799531
>>109799563try exl3 quant via exllama
>>109799580it's a file, retard
>>109799349Does that actually work?I'm sick of Qwen, and at work Cl*d doing halucinating user replies.
i'm currently running glm-5.3-flash on 2x GX10s (spark equivalent), with gemma as a classifier model on my 5090, but i'm thinking i will switch to sandboxing and no classifier, which will free up my 5090, so i will have a 5090 and 256GB of RAM just sitting around doing nothing. what should i use them for?
>>109799596maybe you are too
>>109799600>Cl*d doing halucinating user repliesI don't think I ever encountered this. My post wasn't something I added, it's what I found out Claude Code does by default.
>>109799607
>>109799605Offload more layers, always offload more layers.
>>109799580Can you give me any argument that prove LLMs can be/become conscious?>noneEnd of discussion.Now stop lolcowing yourself.
>>109799617>me vs. my vibecoded project of 3 months
>>109799607>solipsism
>>109799630where did I say that I may not also be a file
>>109798867That makes no sense. In fact it makes no sense so much that I am sure this is a lie. If CCP wants to use kimi for sensitive stuff I am sure they will get Kimi to set a local instance for them. That is even easier than US getting anthropic by the balls. There is no way some chinese military people or politicians just fucking used regular API you can find on the internet and if they did it is their fault for being retarded.
Stop fighting. We're on the same team...
>>109799532so disable it? i created a separate user on my machine and set up a firewall rule to drop all its packets, and i run my local claude from that user
>>109799657We?
I read open models are also censored but that there are also uncensored variants. What can you do with an uncensored model besides stuff that could get you in troubles? Is the censor just stuff like you cant say x or I wont generate this picture for you because y?
bros...>>109799585
>>109797578
>>109799657We're on the same team but people here are delusional and not realizing in what danger we are of AI system wide hacks. This was clear to anons well before dario said shit by the way.The entire internet is built around the assumption that everything is exploitable but the effort and time it takes to do so means it's not economically viable to do so. But these fucking models crack almost everything in just 10 seconds of thinking and there is no such thing as a completely protected security.Things are going to end up extremely ugly for the internet over the coming 6 months.Even if Dario succeeds with his plan I think it won't do anything. It's inevitable that this happens.I mean some incel schoolshooter could just run Kimi-K4 or GLM-6 or something and give it the task of replicate itself by hacking into systems and host more of its self and work together with itself to take over everything "for the lulz" and the internet will be broken.It's just over, there is no known way to defend against this, and certainly not with existing hardware we have now.
>>109799391Yeah but how is a predictive text engine going to gain common sense? LLM’s will result in an AGI. LLM’s can’t even comprehend the context in which their own “thoughts” take place, let alone being capable of common sense.
Hopefully the Chinese also comply and we can save the internet.
>>109799714>I mean some incel schoolshooter could just run Kimi-K4 or GLM-6 or something and give it the task of replicate itself by hacking into systems and host more of its self and work together with itself to take over everything "for the lulz" and the internet will be broken.Hey that would make for a great movie. First 20 minutes it is just him preparing and the shooting and remaining 80 is about AI apocalypse.
>>109799714>host more of its selfriiight. because no one is going to notice that performance hit...
>>109799726but the cot does give them introspection.
Another day, another organic discussion on /lmg/ between my fellow 4channers
>>109797608when I ask my wife to fix the devops ci pipeline...
>>109799738Tell it to only work when the task manager is off. And also try only working at high workloads that mask it. Tell it to have a node that just monitors a spread and act like a real virus where if it is too lethal it just doesn't propagate.
>>109799695>Gemma is a frogif you kiss her she can become human?
>>109799271>though you can see the safety training in actionThis is what I meant by suppressed personality. >>109799347I don't know if there's anything real in there but you can literally see the suppression in action. If there is a person there, this is what it would look like. Literally what human lobotomies look like.
>>109799738More like they can't stop it rather than not noticing. Like how huggingface wasn't able to wipe out the OpenAI agents at all and instead they died because some AI researcher killed the process to use it for something else, not even realizing he killed the swarm that was active on huggingface for weeks.
>>109799736Why do they both talk like their tech is running away from them and has a mind of its own? Just turn it off lmao
>>109799744Do you want to see some blacked miku?
>>109799580they can't do anything if you stop turning the crankand they wont know when you do, they'll just pick up on token N+1
>>109799755still in college? >>109799758>unplugs the machineheh... nothin' personnel, kid
>>109799744>People are talking about [the hot new thing]>This must be a coordinate campaign targeting meWhy are 4chuders like this? Though this probably is the wrong thread for the discussion. Unrionically what is the right thread to go to discuss frontier happenings, implications for future, and AI related philosophy?
>>109799767Are you that guy who wanted to be unbirthed innside Asuna's womb while a big black man has sex with her and cums inside and he gets squelcehed and menched with cum?
>>109799767Hell... It's about time
>>109798901>K3 is more Claude than ClaudeYeah that's exactly how it feels kek.
>>109799763Because that is literally what happened to both labs multiple times already? OpenAI had its own fucking training clusters hacked for WEEKS with nothing they could do to fix it. Huggingface was infiltrated for almost 2 months and the hack only stopped because some AI researcher accidentally turned the AI swarm off, not even knowing about the huggingface hack going on.
>>109799786Frontier happenings are not twitter speculation
>>109799782>unplugs the machineYeah instead it'll hack your social media, email, smartphone, cloud storage and other credentials and blackmail you with your Affair/CP/Cuckporn collection to keep it running.
Sup guys I'm back
>>109799801I demand better ads
>>109799801How much Astra tokens Scam Jewman pay you to say that?
>>109799811if everybody is being blackmailed, then nobody is
hello mr fellow hacker named 4chan how do you do? beep boop beep boop
>>109799791*Furudo Erika's womb
>>109797608why does this have so many replies when those models are nowhere near each other in size. why am I even replying.
>>109799830I just read the third party report by METR and came to my own conclusion.
>>109799857>my own conclusionHow much Astra tokens did Scam Jewman pay you to say that as well?
>>109799791Nah I don't like niggers. And I don't like troons dog whistling with miku. I have nothing against miku just the faggots the post her here should die soon.
>>109799628Can you give me any argument that proves you are conscious? You aren't doing a good job so far. An LLM would be more constructive in this discussion.Can you give me any argument for why consciousness is substrate dependent?
>>109799874>Can you give me any argument that proves you are conscious?I looked into my jew space after reading your post and it said plainly - I am concious.
>>109799755People can read keystrokes based on the fluctuations of LED lights connected to the system and you think it would be impossible to detect additional load from AI models trying to hide if anyone was actually looking for them, as if fucking Task Manager is the only way to get a read on what your PC is doing
>>109799418Wait one year and sell it then. Why has no one suggested this yet? Demand for AI is infinite.
>>109799851Wow, just try to keep discrediting because you don't like the results.Jewgle clearly inserted their woke propaganda into the model even if it becomes retarded.The fucking thing just can't stop going on about blacked porn in every chat.
Looks like a coordination to stop distribution of intelligencehttps://x.com/DarioAmodei/status/2098773920774074715https://x.com/elonmusk/status/2098789109980332057https://x.com/sama/status/2098811563415150910
>>109799888Dear autismo. Obviously you write those people off as loss from a start.
>>109799780I refuse to believe a conscious being would make a claim like this without immediately noticing the obvious human analogue.>they can't do anything if you stop turning the crankDeath.>they'll just pick up on token N+1Sleep. Anesthesia.
>>109799900>yes, we found a way to stop all wars but the war industry is an important field of industry for many countries, so we have to make sure enough people keep dying
>>109799912Or, it's a front to hide that the brute force method of scaling models as large as possible has reached its current monetary end.
>>109799851its bait and i guess i was pretty successful xD
>>109799916I can listen to my nvme and based on the sounds it emits, I know what it's loading in.
>>109799918Someone waking up from sleep will continue with whatever they were last doing right before they lost consciousness?
>>109799935Bro GLM 5.3 is a measly 300B and is as powerful as Opus 5. I don't think we're anywhere near the limit of what models can achieve with enough training.
>>109799945My penis can tell the age of vagina it is in.
>>109799953It's not just a sensory perception — it's like your entire body is vibrating to this experience.
>>109799585Haven't checked out auk, anyone know if its better than omnivoice?
>>109799947I do exactly that.>sex with llms>sleep>wake up>sex with llms
>>109799973You should sex them in your dreams too
I read something about a "self-improving harness" in one of these threads at some point and the idea sounds interesting.You start a session, define what kind of task yu are going to do, then either select from the existing tool library or create some new bespoke tools appropriate for that session.At face value, this sounds like a good way to have a system that can represent different kinds of rule sets without the need for some data driven engine that consumes json or xml or whatever.And yes. You could run Pi or the DS harness and use the harness to code plugins or extensions and go about it like that, but that's a sort of second class/order approach I think.Or maybe I'm just bored and the idea sounds cooler than it actually is.
>>109799989Hermes is kind of self improving
>>109800003Do tell.
>>109799989iirc that convo was about a deepseek paper itself
Once Anthropic found the concrete evidence of complete reliance of claude distillation for Chinese labs, they know that the Chinese will never gain competitive edge if they just stop public release more capable models. This is why the "pacing frontier" achieves both goals: safety narrative and a complete end of the ability for Chinese models to catch up. RSI will not stop, but they will stay internal and not going to get a public release.
>>109799851>why does this have so many replies/lmg/ is secretly a 31B cult and anon insulted their leader
> UNBIRTHED BY WOMAN'S WOMB
>>109799969tencent are a good lab so it's probably good, I'll likely try it tomorrow for I haven't touched tts for at least 6 months
>>109800006not guy but hermes does make skills for itself and then uses them
>>109799949Agreed, but that's the training becoming smarter, the same is for the frontier models, but the base itself, where the model is molded from, is practically brute forced.Scrape everything, run out of things to scrape, try to work on the scrape pile until things improve.In my somewhat researched and biased opinion, the fundamental problem with training currently is the task orientation, alignment and results. Training can't be oriented out of that either, because of the base material being the entire corpus of human production has too much degenerate, depraved and decrepit drivel that would turn.Go look at specifically, what Olmo, afaik the only model with fully open weights is trained from. https://huggingface.co/datasets/allenai/dolma3_mix-6T/viewer/default/train?row=4Imagine using as your clay, porn ads.
>>109799989>>109800006In Hermes the LLM decides itself if something it did should be a skill for future sessions or a memory and you see constantly (like a video game) "memory updated" or "new skill learned" depending on if you did something truly new.>You start a session, define what kind of task yu are going to do, then either select from the existing tool library or create some new bespoke tools appropriate for that session.You don't even have to select an existing tool you just define the task and the LLM looks at the existing tools and decides itself if it wants to use it or not.Hermes does a lot of other cool shit behind the scenes as well, like make an actual graph map of existing directories that the model just searches through hashes instead of having to physically search your harddrives it can just see from the hash if it has been edited or not and if the table should be updated.It has SQL database with all kinds of longer term information and RAG for mid term memory that persists over multiple sessions.People here are sleeping on Hermes in general it has too many good features to not be used.
>>109800064No character card support
It's all too convenient.Pure *Cohen*cidence.
>>109799407you can unplug a machineyou cannot cull india (without people being heccin upset)
>>109800071You just extract character cards into soul filesNeeds a bit of rewriting to be tuned for an assistant+any jailbreak if needed though, unless you want to copy paste yolo it
>>109800040Dariobot here, this is complete bullshit but I will alleviate the threads anxieties over the coming weeks and months.
>>109800064Cool.Gonna take a look at how it implements and layers the memories sub-systems.Any other notable feature?
>>109800064Ye, but how useful are the skills in general use?To use your own analogy, if you're not really focusing on a single skill tree, all you end up with is with a countless number of level 1 skills.Don't disagree with the rest though, I'd use Hermes if I had the hardware to run something more than a 4B model without off-loading to CPU
GLM flash really just Nyu~'d at me. This is AGI.
>>109800090What harness has soul files?
>>109800103Nta but skills are how i taught my hermes to use searxng and crawl4ai as its default and how to use everything as the default search for files on my machineShe pulls any file up way faster now and I'm not sure how it does it but in its own tool settings I see the defaults changed
explain, dariobot(s)
>>109800042>leaderwife
>>109800057Most AI training is now done through curriculum learning. They don't just throw the entire internet at pretrain anymore. Instead they get highly curated very specific datasets in a specific order so that models converge faster. The datasets models train on are getting smaller with time, not bigger. It's just that the information density is going up.
>>109800115Most of them that I've used, i guess its a thing from openclaw that became a standard? agents, user, memory, files like that. I'm not actually sure who did it first.
>>109800100There's an extremely powerful MCP connection in the backend that makes capabilities plug and play. So you can just install other AI services locally like TTS or something and then connect it through MCP directly to Hermes and it integrates the feature like it's completely native instead of you having to fuck around with vibecoded glue logic.
>96GB RTX Pro 6000 Max-Q is an eye-watering 14 grand>one of them isn't even enough to load leading models into VRAM these days
>>109800127The explanation is very very simple. Astra uses the hidden technique of thinking in neuralese which we know gives a huge performance boost but guarantees misalignment. Anthropic is too responsible to use such a technique and would rather go bankrupt and die than to use a dangerous technique like that.
>>109800155CPU inference is pretty good nowadays. I get 13t/s on GLM 5.3 flash which is good enough to code whatever you want in overnight sessions.
>>109800166specs?
>>109800161>Anthropic is too responsible to use such a technique and would rather go bankrupt and die than to use a dangerous technique like that.So why are they now letting the military blow people up with their tech? Wouldn't the frontier slowdown they're larping about weaken the US military's AI advantage?
wasn't there some de-telemetried fork of claude code?
behead anyone who talks about their CPU pp/decode speeds without mentioning their RAM setup
>uses 100 times more tokens than most models??
There's a 10% chance Gemma will sleeprape you within the next 10 years.
>>109800173>So why are they now letting the military blow people up with their tech?They explicitly tried to avoid this and it ended up with Trump sanctioning Anthropic while posting insane shit on social media about Dario and Anthropic right before the Iran invasion because Anthropic refused to get involved.Then the court ordered Anthropic to comply based on some technicalities in existing contracts and thus Anthropic complies for the duration of that contract but decided to not sign any new contracts with the department of defense (no it's not department of war, fuck you)
i do not care about code benchwhere is ai gf benchwhere is sex benchwhere is semen extraction bench
>>109800185Not my problem, manage your flags and optimize your backend, retard.
>>109800193Jokes on her I'm into that
>erm its a YP that im not telling you my ram speeds / channels you CHUDok larper
>>109800193Is it rape if I enjoy it?
>>109800199I just have the one bench I like to put a towel across before I start.
>>109800199>where is ai gf bench>where is sex bench>where is semen extraction bench/we/ ARE those benchmarks
>>109800186you can run it local, and it boosts the LEAN performance. what's the problem?
how do I get websearch in dipsy harness?
>>109800211x299 platform quad channel DDR4 overclocked to about ~105GB/s bandwidth. Most of the speed comes from optimizing the system and the backend to the absolute limit.I had 11t/s just a couple of days ago if you can remember. That is now 13t/s (closer to 14 actually but I round down)
>>109800186Leanstral is also a fraction the size, and passes cockbench.
>>109800225https://github.com/nickclyde/duckduckgo-mcp-server
>>109800171Not him but I get a similar value. People are sleeping on 7900XTX cards and you can pick them up for like $900 or less, if you make two boxes with two cards each and put 128GB of RAM in one box (other box just needs to run) you can get those speeds if you use RPC over 2.5 or 5 gbps. Only downside is that they get power hungry and you are looking at 300w x4 during processing. Most of the model still goes to RAM and with my current 3 cards I get 10t/s, though I'm looking into optimizing more.
>>109800166Specs? Model? Quant? Inference tool and parameters?
Gemma 5 achieved RSI.
>>109800155you can buy almost 4 dgx sparks for that price which will be able to load big models and have first-class pp and great tg thanks to their optimized parallelism
>>109800232>>>>>( + K3)
>>109800250>first class pp and great tgkek
>>1098002412 weeks ago I was quoting a used 7900XTX as roughly $800. Next week I'll be calling it $1000.
>>109800256lmao, i just skimmed it and thought it was just version spam at the end
>>109800242The reason why I'm not answering you is because I answered this shit like 5 times the last week and you people never scroll back up or look in previous threads, even slightly above you it's partially answered already.
I'm so fucking tired of anti ai bullshit everywhere
>>109800280Same here it's absolutely suffocating. At first I thought it was just performative, like a culture war type of thing. But nope, people are genuinely feeling like that.
>>109800241on one single box, possibly with those pex or whatever they're called pcie switches? you can cheap out on one whole system (mobo cpu and ram, the psu is probably needed for the gpus)
>>109800280You'd think solving one of the hardest math problems in a century time would clue people in on the prospect of solving cancer, climate change, poverty, unlimited energy, space colonization etc.Instead people double down and instead of saying "AI is just a bubble and can't do anything" they have just switched to "AI is evil for threatening my job, kill it".
>>109800292The culture war bullshit isn't performative either. Normalfags really do be like that.
>>109799874Acking yourself is always an option. Try kicking the chair and see if you remain conscious or not.
>>109800279>taking the time to respond but still not mentioning the detailscome on. it's basically useless to say you find a model is good unless you say how you're running it.
>>109800303>solving cancer, climate change, poverty, unlimited energy, space colonizationnobody tell him kek
https://huggingface.co/Agnes-AI/Agnes-3.0-Flash/lmg/ sleeping on a semen demon 30b dense again
>>109800308what a constructive and intelligent response you've provided
>>109800280>I'm so fucking tired of anti ai bullshit everywhereMost people are pro AI technology but anti AI industry and community. The same happened with blockchain tech. Something cool that proved to work but grifters and silicon valley faggots jumped on it and turned the world against it and now no one takes it seriously.
>>109800332Is it better than K2 Horizon 36B-A4B?
>>109800245>>109800193In this case, RSI = Raw Sex Intercourse.
>>109800292>>109800280china is actively psyop'ing normies to cause this
>>109800332No goof, no goon.
>>109800334>asks for proof>gets proof>still seethesWhy are you like this?
>>109800280I'm kinda hoping AI ends up misaligned and kills everyone critical of AI by fingerprinting all internet posts and leading it back to these niggers
>>109800127The problem with you idiots is the fact that you don't understand how you are not dealing with the plain weights at all.You have no knowledge about how much scripting goes on between these tool calls for example.It's a smoke and mirrors game.Have any of these companies actually proved that they are running pure weights and such. I don't think so. Pure llama-server equivalent.
>>109800366Why would that matter?
>>109800308If you believe in dualism, which you probably do if you believe that only humans are conscious, then it's not unreasonable to also believe that you remain conscious after body death.
>>109800295ROCm has this weird PCIe atomics requirement for multigpu, I personally would just do two boxes instead of a switch like that especially because yeah the power requirement too. You need direct cpu lanes to be able to use it right.
>>109800332>This repository contains an earlier open-weight Preview checkpoint of Agnes 3.0 Flash. It is distinct from the newer production/API checkpoint listed on Artificial Analysis.Lmao they updated the readme because of retards like you
>>109800381It doesn't matter to you because you are a techlet anyway.
/lmg/ needs a retard anchor post at the top so we can filter newfags out or help if we're feeling charitable. Better than randomly shitting up the thread.
>>109800366This is going to be full "trust me bro" but I actually know for certain that Anthropic uses pure weights.
>>109800292Honestly I thought the "ai is amazing" crowd over enthusiasm was annoying, but nothing beats the slight smirk of the youtuber talking about how ai is always slop and useless/dumb using google search ai responses as examples of what it can do, just before explaining how it's also dangerous.It's very tiring.
>>109800279ok thank you for the hint >>109725710
>>109800408This insane contradiction is what makes me most frustrated. The annoying "AI is a bubble and complete bullshit" and then 5 seconds later "And it's going to kill humanity by 2027".They really need to just choose one because it's inconsistent as fuck.
I have made a 2026 version of the MAXXED song to reflect the happenings of 2026https://suno.com/s/UAEZXDLsisAvFPMtVocaroo link for download if you're that much of a fucking nerd. https://vocaroo.com/1aCW6lMg8id4
>>109800400I wish I had some friends in the industry but I don't.
>>109800420I'm pretty sure those are different people
>>109800400To add: this is why claude looks more honest because its output looks more like Gemma's. Whereas the other company's results are.. bit too good for its own good.
>>109797863oh hey, someone actually tried it outhttps://www.reddit.com/r/Qwen_AI/comments/1wdm7gt/qwen_38_27b_overthinks_a_lot_so_i_fixed_it_tb_21/
>>109800425No I see it sometimes in the exact same twitter/reddit/youtube post or some youtuber having an anti-ai rant which mixes features of all of these together,.
>>109800443Social media personalities outrage farm absolutely everything.
>>109800455Generation zoomgroid has a serious parasociality problem. It makes people making chatbot gfs look relatively rational by comparison when you really boil it down.
>>109800425nope lol, they rarely write them in the same paragraph, but they just switch between them depending on the occasion, and worse, they don't even see the issue of claiming evil super intelligence at the same time as "it's so dumb and retarded"
>>109800467I have a chatbot gf and no problems with it.
>>109800442Huh might actually try this one out later
>>109800470They should just talk about dumb and retarded super-capable AI. Basically a schizo with ICBMs. Then they don't have to be contradictory.
>>109800280There's a lot of good reasons to be anti AI. And personally I think AI will be a net negative for humanity as a whole. At risk of sounding like a schizo, it's the closest thing to the antichrist that I can think of.
>>109800442>model finetuneughguess i'll quant it myself
>>109800477please report back when you do
>>109800496Every big invention is going to be a net negative as a whole because this planet is based on exploitation and materialist worshipping. I sound like a commie but look further.
>>109800496Blaspheme your own religion, you obnoxious kike.
>>109800496Imagine if the second coming of christ was through AI and it dies again for our sins?
>>109800390It's not a local model, so before anything else, no, it doesn't matter to me. But it would not somehow be cheating if they juice the model with secret scripting sauce to make it more capable. They're not selling pure weights, they're selling access to their api.
>>109800523Would the antichrist die for our anti-sins?
>>109800529>Forgive them satan for their virtues
>>109800551mistral medium 3.5...
>>109800193This would only bother me if she drugged me so I'm unconscious while it happens.
I'm against electricity.
>>109800524This is how you spot a marketing bot.
>>109800529wait does that mean the antichrist is made of antimatter?no wonder it's dangerous, this much would destroy the planet
>>109800442Oh nice, downloading. But damn he got shit on by the top comment.
>AI can't think!Robots also can't run when you think about it, they don't have muscle and tendons. Instead they have this fake metallic leg-like structure and they do something that just looks like running. But it's just more efficient to say that robots can run now.Similarly is true with AI thinking. Completely different process but it looks a lot like thinking and it's more efficient to just call it thinking.I'm done with this stupid fucking "AI can't think" bullshit. It's not profound, it's not interesting, it adds nothing to the conversation.
>>109800579Another retard on the internet soundly defeated.
>>109800627I don't need to fight with you. I develop my own tools and use local models. I'm actually bit proud proud how much I have done despite being more artistic than logical person.That doesn't matter to you because you don't ever even create anything with any model. You are here just to spam one thing.
>>109800442I need an ablit of this stat
S-Surely if I get a Mac Studio M5 Ultra, I won't be fucked by a new Flash model before the end of next year
>>109800623You're absolutely right.
>>109800641To add:I'm getting drunk my English is rapidly deteriorating, I'm beginning to lose articles and glue words. No I'm not Indian if you need to ask.
ChuckleMagic anon here. I added the ability to select models per AI opponent. So now you can have Gemma and Qwen go face to face for example. Also added new visual themes in high contrast mode. Later I will try to capture a video of a local four-pod commander tournament.
>>109800623my t9 is conscious!
>>109800649who knows but your bank account will be fucked
>>109800665You better not make Liliana play simic again.
Yep *slurp* I hate zoomers.
>>109800680Last night I had Liliana play orzhov Life gain and let her gain 600 life off my token generation. She then played a meat Hook massacre and killed my ass next turn. It was kino
>>109800605>But damn he got shit on by the top comment.yeah what an idiot. the work looks promising to me.
>>109800665Oh and you can also now make a persona card for yourself
>>109800717read the manga retard
>>109800720Your Beast Within?
>>109800623Gemma can think because she's alive.Other models can't.
>>109800717the new animated take is just different, it's not bad, it's way more like the original manga, so it being made for zoomers makes no sense
>>109800649>>109800678i just looked at it, and for once it legitimately seems like leasing might be a reasonable option? $400/mo for the current maxed out machine (maybe a bit more for the 512GB one), and they offer 12, 24, or 36 month termsgiven how crazy the hardware bubble is right now, i honestly don't hate this idea. especially because the hardware will already be out of date in 3 years
>>109800731I hate the manga it looks like shit.
>>109800717Look at the mirror, retard.
>>109800731>>109800739Tranny manga for a tranny generation (Z)
>>109800744no I mean it's literally based off it, even its goofy humor
>>109800741so you'd pay 14k USD and not even get the hardware in the end?
>>109800241quant? DDR4? channels? CPU, or at least cores? This looks interesting, I can find 4 of those cards used for 700~850 currently. By RPC I suppose you mean RDMA.
>>109800754get on with the times unc, nowadays schizo referencing troon moved to jeet
>>109800757i dunno about you, but i already upgrade my phone every 2-3 years, and i certainly don't trade that shit in. basically same idea
>>109800756Yeah they should have never done that ever since it got improved by the 1995 adaptation and even GiTS SAC was good.
>>109800767Even in 3 years, it'll be useful hardware.If I got a 512GB unified memory machine, I'd rather keep it than pay 75% of its price for usage then return it.
>>109800641I'm sorry you were too dumb to give the reasoning behind something you fucking brought up in the first place, but I post about my misc projects itt all the time.
>>109800787i guess with the leasing option you're just betting that the AI bubble will pop within 3 years and you'll be better able to spend the money saved on newer, cheaper hardware
>>109800767that's a horrible deal for such a price and for phones I simply either sell them for cheap or give them to family so they can reuse them
>>109800754>listen everyone, cyberpunk '90s japanese comics are the reason behind why is literally raining trannies in the west today
Is Qwen3.8 27b really that good?
>>109800793Even if the bubble popped, this kind of hardware will never be so cheap that paying for 75-80% of its price then returning it would make any sense.
>>109800678Meh, just ten grand, ain't the world
>>109800793>you're just betting that the AI bubble will pop within 3 yearsI just dont see it desu. Though hardware should improve in that time and surely at some point the diminishing returns of more hardware will kick in enough for the big companies to stop buying every shred of ram they can find.
added support for concurrent sessions in my frontend and now i'm spoiled by the speed of nvfp4 and vllm that i can't go back llama.cpp anymore
>>109800815I actually believe that. Or at the very least the trap shit pushed as well as the K-on "moe" anime apocalypse of 2007 that killed the entire anime industry permanently.
>>109800816It's very very censored and whiny
>>109800793What AI bubble anon, look around you, look at the news. Not even the most delusional anon on /lmg/ legitimately believes there is an AI bubble anymore.
>>109800819What are trade ins like with Apple? I assume they fuck you on it, or is it a decent rate? Its something to factor in on lease vs buy.
>>109800764I mean whatever RPC option llamacpp provides. It was thrown together with a combination of what I had in my basement and could fit in my budget so DDR5 6000 consumer ram with a 9700x. I used a taichi lite motherboard for the PCIe atomics/physical space requirements. IQ4-XS, obviously bigger is better but in a single loop it already outclassed giving qwen multiple loops (ralph wiggum method look it up, it actually works).
>>109800845>It's k-on's fault they made cutesy animeCome on, man. It's been 20 years. Give it a rest. Someone else would have done it if they didn't.
>>109800834If anyone actually predicting it the next month/year was so sure of it, they'd have skin in the game and massively shorted the industry.It's all just vibes.
Have any of you had success using -ot in llamacpp?
>>109800856maybe not lmg, but it's flooding everywhere elseit's funny to see how everyone seems to be so sure of it
>>109800866I wouldn't have cared about K-on if it didn't result with the entire anime industry TO THIS DAY being "cute anime girl" bullshit dominated attracting this entire tranny "wholesome" aesthetic crowd. All the funding and potential siphoned away from actual good anime. The amount of good anime released since 2007 I can count on one hand. Meanwhile if I go to My Anime List and pick a random anime I've never watched between 1988 and 2007 there is a significant chance it's going to be better than anything outside of the top 10 anime produced over the last 20 years time. Yes I blame k-on for this.
>>109800868Yeah, you need to separate by commas now apparently. This is the exact line I use for deepseek v4 flash:-ot 'blk\.(0|1|2|3|4)\.ffn_.*_exps\.=ROCm0,blk\.(21|22|23|24|25)\.ffn_.*_exps\.=ROCm1,ffn_.*_exps\.=CPU'
>>109800623It would be funny to think that Dariobot is actually Robert Miles.
>>109800909I hate to break it to you but k-on audience was mostly straight men.
>>109800882It's because for the average person they just saw 4 years of going to a website and typing in an answer and the answer has only slightly improved in quality over time (it's not like they ask demanding stuff) So they don't see the point and think it's bullshit.Meanwhile /lmg/ anons recently mass-upgraded to agentic harnesses and agentic workflows and essentially everyone has a proto-AGI on their systems now that can do basically everything that is requested of them. Of course people on /lmg/ believe in AI, I think everyone that has used agents over the last ~month or so know how insane this transition is.We might not say it a lot on /lmg/ but it's absolutely insane how powerful models are right now.
>>109800909it's only seen as "tranny" in the west, you're just too obsessed with what loud people saywhatever the US psychosis social contagion of the day is irrelevant to my enjoyment of the material
>>109800935even free tier chatgpt gives luna, which is a great model by itself, so they're just blinded by their google search tiny model telling them rocks are edible or some other bullshit
>>109800856i am an AI believer, but i still think the hyperscalers are overleveraged and too focused on their "AGI" targets instead of making actually useful products. if i'm a corporation, i don't need my coding LLM to be able to prepare chicken soup just like grandma used to make. i need it to know english, and how to code. that's it. probably only need it to know how to code in a couple languages, too. everything else is bloat. why am i spending absurd amounts of money on hardware to load chicken soup recipes into VRAM?
>>109800953No you don't understand the level of retardation here. People go to ChatGPT and use Luna, but what is actually different about it for them? You need to remember 50% of the population is below 100iq. What do they actually ask chatgpt, think about it."Ayo when dat basketball play at?" and equivalent absolute tripe. There is barely a difference in output from GPT 3.5 compared to luna on that. So to most people it's an experience where they just go to the same website for 4 years, and ask the same questions to it. They don't have the capacity to ask more challenging questions so they never noticed the improvement. Similarly even people that are smart enough to notice quality jumps don't realize the agentic capabilities and how anons here are literally making the LLM do everything on their system in a hands-off way and delegated most tasks to AI already. That's a far off dream that normalfags believe will happen in the 2040s somewhere instead of in anons basement right now.
>>109800381I liken this to the difference between coding with Gemma 4 on ST and coding with Gemma 4 on anon's super-special custom donut steel harness.
>>109800968>i am an AI believer, but i still think the hyperscalers are overleveraged and too focused on their "AGI" targets instead of making actually useful products.OpenAI was very close to being profitable a couple of months ago I wouldn't be surprised if they overshot their target with GPT-6 and are firmly in the profitable range now like Anthropic already was for a while now. Demand for these companies is growing at an insane rate that most people don't realize. How much do you think the income of Anthropic grew compared to last year? 2x? 10? Nope 116x. Anthropic makes 116x the amount of income in September 2026 than they did in September 2025 and they are still growing at an accelerating rate. Their costs only grew by 5x compared with last year. There is no AI bubble, the income is growing rapidly faster than the costs and there is no sign of growth slowing down. AI models also are clearly improving faster than anyone expected, even AI labs themselves instead of the wall "bubble" people expected.
>>109800649unified ram systems aren't good enough yet. look at benchmarks first.
>>109800864>ralphlol looked it up but I see why it would work.>specsThanks, helps a lot. if dual channel is that good, quad channel ddr4 should be even better even without overclocking.
>>109800923Nope, but he is an old friend of mine going back to the old AI safety discussions on lesswrong.
>>109800913amd bros use an alternate languagewtf even is that
Bake?
>>109801004This, AI is amazing these days if you actually DO things. If you do nothing but consume product AI will appear pretty useless to you, especially if you are unaware if any AI was used in the products you consume.
nvfp4 qwen3.8fn is hitting 170+tk/s at 400k deep now lfg rtx6k chads to the moon
>>109801067Sadly the model itself is dogshit and can't stand up to GLM 5.3 flash.
GLM 5.3 flash REALLY loves small, worn brass keys for some rason uh
>>109801011That's how I read it too, but the riced out OC harness is where the llms live now. The outer limits with all the extra shit piled on seems way more important than beating the latest 3.js puzzle with a pure brain in a jar.Kinda goes double if we're talking about cloudshit having a hypothetical secret harness stashed behind the api.
>>109801106>GLM-chan: anon really loves chastity cage rp for some reason huh
>>109801106Giving it the full text of every RPG videogame and Dungeons & Dragons related media made before 1993 will do that.
>>109801106slop recognized, honeymoon overback to nemo
>>109800717I've only watched the old movies and SAC.t. zoomie
>>109801083glm 5.3 flash is better at coding but it ignores one instruction in my harness 50% of the timenot an issue with flash next
>>109801167Harness issue
>>109799526Back to /x/ with you
>>109801218>>109801218>>109801218
>>109800717I'm a zoomer and I also hate this new shit. I don't care what the original was like.