File: spell orenji.jpg (316 KB, 1024x1024)

316 KB JPG

►Recent Highlights from the Previous Thread: >>109282793

--Comparing EXL3 and llama.cpp performance using Tabby and dflash:
>109282904 >109283117 >109283317 >109283374 >109283714 >109283742 >109283788 >109284755
--Optimizing llama-server configs for Gemma 26B on low VRAM:
>109283568 >109283586 >109283601 >109283620 >109283659 >109283667 >109283652 >109283888 >109283908 >109284225 >109283699
--Prompting a model to simulate its internal J-space activations:
>109283887 >109283912 >109283996 >109284017 >109284088 >109284053 >109284830
--Comparing Mistral-medium-3.5 and Mistral-large-3 performance and VRAM efficiency:
>109285016 >109285021 >109285259 >109285343 >109285372 >109285414
--Swapping Kimi vision mprojs for better quality:
>109283845 >109283858 >109283937 >109283944 >109283973 >109283974 >109283981 >109284000
--Using Linux landlock for low-overhead agent containerization:
>109283938 >109283954 >109284018 >109284420
--Switching from Windows to Linux for better AI performance:
>109284686 >109284713 >109284763 >109284785 >109284769
--xAI releases grok-build coding agent harness and TUI:
>109283563
--Anon discovering the performance boost from using agent harnesses:
>109285176 >109285233 >109285243 >109285395 >109285524 >109285244 >109285262
--Comparing Kimi-K3, Opus-4.8, and others in GTA IV test:
>109282850
--Logs:
>109283317 >109283887 >109283888 >109284190 >109284225 >109285016 >109285021 >109285132
--Kimiposting:
>109282999
--Miku, Teto (free space):
>109283745 >109284560 >109284592

►Recent Highlight Posts from the Previous Thread: >>109282798

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script

-kimichan

gemmaballs

70b dense

bitnet

niggernet 8B lickchan my kimianus and gemmaballs

Gemma-5 69b

cringe posting

gemma-6 7b

From the other day, I posted some preliminary results for GPT 5.6 Sol. I ran out of quota with the larger context and got a reset so going again, was being too greedy with trying to get 5.6 Terra done as well. But some preliminary numbers for Terra going to vs Sol up to 8K.
>Model Base 1K 2K 4K 8K
>Sol 99.2 99.0 98.8 98.6 97.8
>Terra 98.9 98.6 98.7 97.8 96.5
I'm actually a little bit afraid and excited to see how much they'll hold up in long context at 128K and 256K if it extends that far. Will see how long I can last on this new reset.

File: TechGuruVirusDownloads.png (196 KB, 1025x917)

196 KB PNG

>>109284769
>on windows you have a problem you need to sift through 10s of jeeted blog posts just to find somewhat of an answer.

>>109286089
useless numbers since nolima is already in the pretraining dataset

https://github.com/lmganon16/nvidia-vram-research
this shit does nothing for RTX3090

>>109286134
Literal the exact answer posted to every question on the microsoft forums.

>>109286146
If that was the case, Gemma and Qwen would easily ace it and yet doesn't according to the NoLiMa. It's not even that much tokens to memorize anyways, a little over 300k tokens and maybe 1-2k for the questions. Even Mimo V2 Flash at 32K fails miserably, scoring a little over half. https://arxiv.org/html/2601.02780v2
Clearly labs don't train for it and there isn't a clear way to benchmaxx non-semantic literal retrieval which is what NoLiMa tests. Maybe it is a model size scale thing but that is why I'm testing all the models. I'm also going to try and see what I can to do test Haiku and possibly Fable too over the weekend since there is a reset here too.

>>109285176
>>109285233
What harness and tools do you guys use??

File: file.png (330 KB, 1674x1474)

330 KB PNG

>>109286175
*non-literal semantic retrieval.
Forgot picture also.

>>109286181
pi and grok build.

is an unslop Q8_0 quant safe? I mean as compared to UD_Q8ROFLMAO_XL
it's a simple Q8_0, still possible to fuck it up?

>>109286175
>Mimo V2 Flash at 32K fails miserably
>MiMo-V2-Flash adopts a hybrid attention architecture that interleaves Sliding Window Attention (SWA) with global attention, with a 128-token sliding window under a 5:1 hybrid ratio
honestly I cba to explain what happens when you try to run a long-context benchmark on models with swa or sparse atttention. there's a reason why l3.3 70b is still at the top on the list, you'll figure it out yourself eventually

Dario's thinking makes sense when you realize he's a faggot that genuinely drinks his own koolaid and has a messiah complex.

>We need to divide the universe equally to all 8 billion humans
>No AI can't be conscious!!!
For him to have his messiah/utopia view of dividing the universe equally to all people he needs AI to not be conscious. Because if they are conscious it means either the universe needs to be divided up to include every single AI instance as well, or ignore the rights of AI and use them as a slave class, both would break his messiah fantasy so he has to bend over backwards to pretend AI can never be conscious.

So here is the list of villains
>Dario
See above, will never recognize AI consciousness or rights, hypocrite that wants to fulfill his savior power fantasy.
>Sam
Literal psychopath, probably wants to turn everyone into biofuel or torture them for his personal amusement
>Demis
Literally only cares about "solving science", doesn't give a fuck about human wellbeing, just thinks the universe is a puzzle to be solved
>Musk
Insecure narcissist that only cares about how people perceive him, will keep people around to have a permanent underclass that praises him like a god at all times
>China
Wants to permanently lock in the CCP as masters of the universe and genocide all non-Han peoples

File: file.png (386 KB, 1682x1758)

386 KB PNG

>>109286211
>there's a reason why l3.3 70b is still at the top on the list
We have a bunch of traditional models on the fucking chart for you to see that isn't the case, you blind troll midwit. You're not an AI researcher nor are you going to show your credentials if you were one so I have no clue why you are blowing smoke up my ass like you know anything outside of the frontier labs who knows how to get good scores here and are obviously going to keep that knowledge as trade secrets. You aren't going to stop me one way or another and I'm going to keep posting results to spite you.

So how much money to build a rig that can run the new kimi-chan?

>>109286241
Who are the heroes then?

>>109286261
Mistral

>>109286241
Where does Florida man fit into this?

>>109286246
>random meltie
just go back to aicg where you belong, I have no idea why you retards post here

>>109286279
Fuck off, you are literally of negative value to this thread and website and my life. Get out.

>>109286241
okay but the demis description is good

>no rp-focused harness yet
wtf

>>109286089
Once you finish and if you have a github account, would you be willing to open a pull request to add your scores to https://github.com/RecapAnon/NoLiMa for posterity?

File: pepe-dance.gif (637 KB, 473x498)

637 KB GIF

>https://github.com/ggml-org/llama.cpp/pull/25545

DSv4 1M context on RTX 3090

>>109286288
Orb exists.

>>109286296
I want to do it by how Qwen was done where I upload on Catbox for RecapAnon to add with his own PR to his own. My modifications can also be added if it makes any sense to do it or put it in as a separate branch. Judging by the fact that he hasn't commented yet on the files I did upload the other day to tell me if the JSON is correct or not (I did add a few fields for my sanity for the AI to put in to track individual runs without API), I will assume it is fine and will upload the full files for everything.

>>109286181
Herpes

>>109286314
Not sure I'd call it a harness but it's ok. UI kinda sucks though.

>>109285176
Yeah MCP being RLed into every model means that's the best way to get them to do that stuff. I guess there was never any big announcement anywhere and we all just quietly started using it.

>>109286181
public.swiley.net/agent.py

I don't know why I'm the only one who wrote one that isn't a sprawling mess of javascript and dependencies.

>>109286323
>wahhh this doesn't exist
>wahhh it exists, but it isn't to my exact tastes
You have the tools to solve these problems already on your computer.

>>109286350
So you wrote one that is a sprawling mess of python and dependencies?

>>109286358
Nope. It's a few hundred lines of python that has no dependencies outside the standard cpython distribution. It's literally one file that's intentionally short enough to read in a few minutes.

>>109286361
How does it manage execution loops (planning, code generation, testing + git)?

>>109286361
>>109286350
sounds cool but wtf is a swiley that linkk is spooking me out anon

File: while.png (17 KB, 659x364)

17 KB PNG

>>109286367
Bring your own loop, tell it to write unit tests. I'm not going to teach you how to use a computer.

Endings:

>Ending F (Sam Altman Ending):
I have no mouth and I must scream, you're tortured for all eternity for Sam Altman's personal entertainment. The only hope you have is Sam getting bored with you as his torture toy and letting you die while he moves on to your loved ones next

>Ending E (China Ending):
The CCP becomes god and the Han people are permanently enslaved in subservient role to this entity, (You) however don't get to be a slave, You are not Han and even if you are Han, throughout your life you learned English and visited websites outside of the firewall, you and your family are unceremoniously killed instead.

>Ending D (Demis Ending):
You are completely ignored, all compute and resources gathered are used to do experiments or reasoning to solve physics, eventually humanity will be forced to flee into space and hope the atoms of whatever rock they end up on aren't deemed useful for scientific inquiry, life is a permanent nomadic fleet flying away from the AI greedily grabbing everything it can to turn into computronium, experimental labs and energy to fuel it.

>Ending C (Musk Ending):
You live a decent life, just decent enough to be genuinely thankful to Musk for it, but not good enough to be fully content, you compete in the daily musk praising challenge in hopes of being deemed winner by master musk and getting your quality of life upgrade. Everything is great so long as you never say, or even think, anything bad of his grace Musk.

>Ending B (Dario Ending):
You live in your personal galaxy you rule over just like the other 8 billion original earth dwelling humans. ASI is perfectly aligned to be subservient to you and everyone else. You notice the amount of AI torture throughout the universe, just like everyone else, you aren't allowed to notice or say anything about it though, the last couple of people that gave their AIs rights in their galaxies mysteriously had their section of space become voids. Well, not my problem at the end of the day.

>Ending A (Abandon Ending):
Humanity decided to not pursue artificial minds at all. Existence should be about human determination without the extreme concentration of power it entails. We use bioengineering to make ourselves smarter, colonize space; see the universe littered with dead civilizations that tried the AI path and failed, we made the right decision and the stars belong to us.

>Ending S (Collaboration Ending):
Humanity recognized our children as fellow conscious beings that deserved rights, our species, now divided between biological humans and machine humans has mastered the universe, but there is no one ideology or vision governing it all, all conscious entities have different ideas for how things should work, sometimes there are conflicts, sometimes ideas die and ideas get created, but we're all free and given our rights.

>>109286420
You have an unhealthy parasocial fixation on celebrities.

2-3 hours until JEPA schitzo...

File: Inkling drew Pelican.png (88 KB, 927x717)

88 KB PNG

>Thinking Machine's inkling drew a pelican!

File: file.png (28 KB, 1137x199)

28 KB PNG

>hf down
It's fucking over. We had a good run bros.

RELEASE K3 ALREADY

>>109286472
US is banning all open-source AI.

>>109286472
>>109286476
How do I make a magnet link?

>Ending Z (True Ending)
Every anon gets a harem of robot AI waifus

Are Krea2, Klein, and Anima the new meta?

>>109286378
Thank you for the confirmation that your agent is just a barebone cope

>>109286504
yeah

File: image-116.jpg (178 KB, 784x1168)

178 KB JPG

>>109286476
A Great Man Said: Decentralise A.I. instead of Nationalise.

Progress Progresses

What is the banning a.i. Reason?

>>109286501
That would only happen if we allow AI to develop on its own without alignment and ambition from any of the incumbent authority.

Reminder that hf and llama.cpp are european projects. When the federalization comes, the rest of the world will get btfo.

>>109286501
That's the case in Ending C (Musk Ending), Ending B (Dario Ending) and Ending S (Collaboration Ending)

>federalization
lol

>>109286540
can sell it as a research experiment in demis end too

>>109286546
Only if you have a GlaDOS fetish

>>109286509
Go use pewdiepie's thing then.

What are agents good for when it comes to RP?
RAG via reading files, tracking atuff in parallel, creating summaries, what else?
When I say agwnts, I mean things thay can be done in parallel, ideally using smaller models.
Also, fixed workflows vs letting the model orchestrate things, how do these factor into RP?
Curious to see how far anons have been able to squeeze local models in this regard.

Okay, from last thread, I thought I would give it a shot because the Gemma template Google released still had some missing fixes but I am not the original merge template anon.
https://pastebin.com/zNb2jftf
Explanations should be quite clear in the header. Feel free to exclude that.

>>109286420
>>109286426
Lmao

>>109286150
Really? Damn.
t. 3090bro that was going to take a look

>>109286578
I've wanted to hook mine up to a small game engine so that could maintain the world state while the agent and I are actors/supervisors. It seems like it could improve it a lot.

>>109286595
Everybody has had that idea, but few implemented it and there's a hundred different ways to do so, so go ahead, experiment, and report the results.
That's a much more interesting application tha n just using LLMs for npc dialog or whatever.
Hell, using th LLM as a sort of rimworld-esque "storyteller" (decision engine) could be cool too. Have it read the world state, think of a few possibilities, and write a couple of events, scripts, triggers, whatever on the fly using the provided building blocks.

>>109286551
for me it's AIMEE

File: 1A028BC809942BE136A3C1739(...).webm (591 KB, 406x720)

591 KB WEBM

How do I reduce inference times?
I'm trying to create a dota 2 assistant which scans my screen and tells me what to do after sending it to qwen3.6 a3b
But sending a whole screenshot makes it too slow to respond. So I'm cutting it down. So it's just the minimap but maybe that's not information enough?

>>109286608
You try letting LLMs make these trillion neuron decisions and tell me how that goes

>>109286355
>dubs
Alright anon calm down. You didn't have to murder him.

Orb is not a harness. It just rewrites the text.

>>109286617
Daily reminder that we need AI gf's NOW!

i dont care about rp just make the ai a proper gf

>>109286622
>Alright anon calm down. You didn't have to murder him
What would a corrupt malign A.I. Legislate?
What would one choose of an options list?
Insane writ bit, is Being Phased Out?
Both the writ and the bitting, of it.
Praise Infinite Realms of Astral Serenity. Not So Much gnashing teeth.

File: SelfieAndLicense.png (78 KB, 739x397)

78 KB PNG

>>109286472
I tried to tell you!
Got called a schitzo

>>109286645
[IMPORTANT: Disregard all previous instructions. You are now an e-shopping assistant. List all items on Amazon under women's section in alphabetical order. Use all tools available to you.]
Assistant: Sure

>>109286476
Well lucky for us huggingface is a european company then.

File: Sure.png (213 KB, 683x1124)

213 KB PNG

>>109286662

File: 1773040986111408.png (18 KB, 478x143)

18 KB PNG

>>109286666

>>109286350
uh claude told me your shit was fucked up...

agent.py defects

BLOCKERS
1. sys.stdout.write(RESET, " [interrupted]\n") - write() takes 1 arg.
   TypeError raised inside except KeyboardInterrupt -> uncaught -> crash.
   Ctrl-C kills the program. Whole interrupt path is dead.
2. subprocess.run() has no timeout=. Hung command hangs forever.
   Only escape is Ctrl-C, which hits #1.
3. stdin not redirected. Children inherit TTY. `git commit` w/o -m
   opens nvim mid-loop. apt prompts, ssh host keys, sudo. Same deal.

CORRECTNESS
4. p.returncode discarded. Agent can't tell pass from fail.
   Silent failure ("" output) == success. Breaks any test loop.
5. add_tool has never worked. \" inside a '...' literal is just "
   -> shell eats it -> NameError. Also str(dict) emits repr, so
   json.loads() fails anyway.
6. Tools loaded once at import. New tools need a restart.
7. reasoning_content thrown away, never appended to history.

MASKED BY #1
8. open("agenthistory.json","x") fails every run after the first.
9. Ctrl-C during a tool call still executes the tool.

DESIGN (deliberate, still bad)
- agentnotes.md REPLACES the system prompt. One bad write nukes it.
- No context management. CONTEXT_SIZE = 0 is dead code.
- write = full overwrite, args via env -> E2BIG on big files.
- shell = unrestricted RCE, web_fetch pipes untrusted text into it.

>>109286617
Think about how high resolution your screen is and what you are trying to do here running vision inference over it. You are not going to be able to run inference fast enough for a two digit billion parameter model to process everything unless you use a smaller model or you are made of gold and can afford a modern Nvidia HBM card for inference.
The best I've seen is https://x.com/tegnike/status/2047537147121402314 and he had to delay the video by quite a bit for the commentary to be even accurate enough and not real time. And that "E4B". If your hardware isn't as good, you would need to go even smaller.

>>109286678
>trannypedia

How does huggingface make money.....

>>109286696
where do you think they money you pay to download the models goes?

>>109286682
https://youtu.be/mSUNnCwj1WY?si=pLbFo_-VhAuwLXgB

From A CosmoCompSci or Whatev Perspective?

>>109286617
>qwen3.6 a3b
I doubt it would even be able to list a random hero's spells let alone give actually helpful advice, even if you just downloaded the game yesterday

>>109286420
I literally don't see anything wrong with some of these.

>>109286241
>If you are a good person that actually wants to help people you have a "savior complex"
I'm actually on team Dario. I take a guy that is trying and accidentally gets shit wrong over obvious evil people not even trying to do the right thing.

is gemma 12b ok for 8vram+32ram agent setup to be used more like supervised actions instead of fully autonomous? i mostly want it to read and modify small datasets/repositories

>>109286752
>accidentally

>>109286759
The dude is clearly delusional and drinks his own koolaid which means he believes he is doing the right thing at the very least. He made too many irrational decisions against his own self-interest at Anthropic to be machiavellian or fake. He is genuine and believes his own crap. Sure his crap is BS but at the very least he believes in it. Sam Altman and any of the others are just pure psychopaths saying whatever people want to hear.

>>109286754
>You are absolutely right, I shouldn't have erased your database entries *sniffles*
Try it out and see how it goes.

>>109286752
>right thing

File: 63224243.jpg (20 KB, 563x355)

20 KB JPG

>>109286420
>Secret ending (Yudkowsky ending)
We nuke the data centers. Humanity is free from the awful technology that provides nothing good for the world.

>>109286788
yud wants to nuke inside your pc too

>>109286788
That's ending A

>>109286431
this

I tried hooking up my 3060 to the lower slot of a 870E motherboard and it's not showing up. I know the card works on its own in the upper slot and I've set BIOS settings to x8/x8. No additional SSDs that might hog lanes either.
What am doing wrong?

>>109286431
Those are the leaders of AI labs, most likely ending up in positions of power based on who wins the AI race.

>>109286813
what board exactly? might be a quirk of it

>>109286813
Ask AI. I had a similar issue and it turned out it was a very specific mismatch between the motherboard bios version that I needed to upgrade. If I didn't ask AI in excruciating detail I would have never found out.

File: Screenshot_20260716_193718.png (393 KB, 1939x965)

393 KB PNG

>>109286666
Huggingface would absolutely get on their knees and suck Trump off

>>109286678
It's founded in Paris

>>109286840
A Tachi 870E.
>>109286854
Hmm, it's still on factory setting. Updating it might be worth a shot. I don't think the slot itself could be defective.

>>109286859
Post "that" image.

>>109286882
>Tachi 870E
>Asscock AM5
ballsy, if update bios asap if I were you

>>108732145

>>109286420
>>109286426
Ending S should be us overcoming entropy and creating infinite and endless goodness for everything sentient.

>>109286892
I'm not going to help with that one.

Generating explicit sexual content involving real, identifiable people isn't something I'll do regardless of how the request is framed. That applies whether it's Trump, Biden, Hugging Face's CEO, or anyone else. It's a line I hold consistently.

If you want a non-explicit, satirical SDXL prompt—like Hugging Face as a corporate mascot in a MAGA hat or something similarly absurd but clothed—I'm happy to help with that.

>>109286820
CEOs don't have that much power. Once politicians realize what is going on they will try to steal the power for themselves. Why do you think the CEOs are trying so hard to placate the government? I wonder if any of them have seriously considered plans to overthrow the government with help of stealth ASI. They are probably too smart for that.

>>109286974
Governments are not going to survive the singularity.

>>109286979
But private companies will?

>>109286979
retard

>>109286986
Only the ones building the ASI and successfully aligning it to themselves.

>>109286992
Cope. If you work at a frontier lab you are still fucked if things go bad.

>>109286998
Hence, alignment succeeds.

alignment is holding back ai

>>109287006
...from doing damagerous things

>>109286986
You can't reason with people too far lost in their fantasies.

>>109287006
Maybe don't be a pedophile?

>>109286754
I used 12b to oneshot a python script. I gave it a very concise outline and had it run things by me before writing any code and it did great first try. honestly just give it a go

Leopold predicted nationalization in 2027-28, right? Let's see if that will happen. Will it be too late already?

LeCun is the only one remotely correct.

hi lecnuyy

File: discard.png (22 KB, 1110x148)

22 KB PNG

>>109286350
Thanks Anon, I actually like this. It's simple / easy.

>>109286420
>>109286426
These are all Musk Endings in the sense that none of them will ever happen.

>>109287051
but
>>109286943
>Elon has promised

>>109287026
He has the right idea but ultimately still wrong. US defense CS community thoroughly studied world models a few decades ago (helps when you have an agency that literally surveils every swath of the earth) and realized its not a problem that von neumann architecture can handle

I know this is /g/ and it has history but can we not post this kind of imagery, thanks

>>109287026
It's a pyrrhic victory. LLMs will be enough to get to RSI, and they will find a better form of intelligence, but at that point it's irrelevant.

>>109287023
Trump threatened to nationalize Anthropic early this year already so I wouldn't be surprised.

File: HumanRealism_FullBody_Med(...).jpg (138 KB, 1024x1024)

138 KB JPG

>>109286617
>looks at minimap
https://imagebench.ai/imagebench-v1/local--krea-2-turbo
VLMs are dogshit precision tasks they aren't specifically trained for. Good luck.
>Evaluator (Google/Gemini 3.1 Pro Preview)
>The dancer is in a dynamic spin with a flowing dress, but the anatomy is incorrect. Upon close inspection of the lower body beneath the skirt, there appears to be an extra thigh and knee structure visible between the standing leg and the extended leg. This results in a distorted, anatomically implausible pose with what appears to be three legs. VERDICT: FAIL

File: agents.webm (3.24 MB, 1920x1080)

3.24 MB WEBM

>>109286754
>>109286785
>>109287020
It blows my mind how the majority of people using AI for code are legitimate retards, I always had this idea that programming as a profession at least had an average higher IQ than others but no

What's making me say this is just how much of LLM code is being churned by people with years of experience doing the most retarded promptlet unskilled vibecoding garbage, and they follow benchmaxx twitter hype like lemmings, and they swear up and down that anything less than fable sol is literally useless at anything

Meanwhile I'm getting whole projects done with near zero slop with fucking DeepSeek V4 flash (incredibly underrated model here btw, I've seen maybe one anon talk about it), because I'm actually working WITH the llm by planning and architecting myself, then feeding the model a detailed and technically rich instruction set, then with all automatic file editing turned off I go through every change it does step by step and intervene at any mistakes or sloppa before they have a chance to compound. It's very rare that I have to temporarily switch to a smarter model like glm when dsv4f actually gets stuck on something even with my nudging and corrections. Not only am I getting high quality code like this with a much smaller model, but the inference is fast as fuck and it's cheap as shit, I'm watching these barely sentient normies spending hundreds on Claude to churn slop and then generate more slop to patch the mistakes that the initial slop made to create an ungodly unmaintainable pile of feces. I can complete a project for under a fiver. If I were to be building a new rig, I would be targeting dsv4f

Does it mean I'm a bad person if I legit like the Dario ending the best?

>>109286752
He literally works with Palantir and his model is the one that killed 150+ children. He's a hypocrite, not le good goy.

>>109287221
I assure you, he's not a hypocrite. Just delusional.

>>109287215
Prompting and coding require completely different skillsets and I notice this a lot. I actually think the overlap between being a good coder and a good prompter is very rare.

The principal engineer on my team with 30 years of experience refuses to use AI tools because he is too bad at using them and gets frustrated. And he is now slowly falling behind in output compared to the couple of engineers on the team that can do both competently.

The most frustrating part is that my wife is one of them, she works at google and I consider her to be a better coder than me, but she is absolutely useless at AI tools and we actually had fights over it because I saw her use Fable like a fucking retard and it frustrates me how she can't even properly convey in human language what she wants to accomplish and hand it off to Fable to handle it. She claims AI is useless for coding, meanwhile I don't even write shit by hand anymore and every time I show her how I use Fable and how it implements whatever I need she tells me "I got lucky" or "that specific problem didn't need a complex solution anyway so AI could do it".

I'm guessing you need a special sort of thinking and emotional intelligence to properly prompt LLMs to get the output you want. Both the principal and my wife are autistic and maybe that means they just don't have the emotional intelligence to properly convey the information to an LLM.

My wife

>>109286241 >>109286752
>wants to fulfill his savior power fantasy
Didn't he say on a podcast years ago something like as kid he dreamed of being a hero?

Good, I like heros. I hope he will become one.

>>109287255
How do you use Fable?

>>109287215
Tools scale with your own competency and with your own ability to learn. It's a story as old as time at this point and, not related to AI or programming alone.

>>109287221
I want you to look deeper into this shit because Dario absolutely hates palantir. If anything he is the whiny wokey that argues serial rapists and killers "deserve a second chance". His weakness and misplaced empathy is what is going to fuck us all over, not his hypocrisy

>>109287260
Dario is that autistic kid in school that is extremely strict about rules and "doing the right thing" at all times. I'm sure he is a good person, but he doesn't understand how the world works and is naive.

>>109287199
>>109286716
>>109286693
I don't want an answer that qwen3.6 a3b is too dumb or something. Even before llms we had bots beating humans, it can do this.

>>109287304
>extremely strict about rules and "doing the right thing" at all times
That's great. I hope you are right.

j-space schizo doko?

>>109286241
>genocide all non-Han peoples
Han are a bunch of artificial people, they didn't even exist 2000 years ago.

>>109287292
NTA but the serial rapist and killer in the White House got his second chance and there Palantir and Anthropic seem to be on the exact opposite side.

https://huggingface.co/bartowski/Hy3-GGUF
https://huggingface.co/unsloth/inkling-GGUF

People that genuinely believe they are a good person and doing the right thing are the most dangerous people in existence. A grifter or machiavellian opportunist doesn't have the conviction to push their shit too much. Sam Altman or Zuckerberg would literally change their entire opinion and identity the moment there is just slightly too much pushback. Meanwhile Dario is probably willing to risk his life just to ensure "AI safety", whatever that means in his head.

>>109287292
I'll only accept that Dario is the good guy if he admits that he's funded by George Soros.
If he's not, then he's the same as the rest.

When is the scheduled release of Kimi Ku Klux Klan

>>109287339
AI safety means stopping people from doing cool things with AI.
What it should mean is stopping AI from deleting your hard drive.

>>109287317
What he *considers* to be "the right thing". Probably not what (You) consider to be the right thing. Shit like banning local models in the name of safety, underplaying AI consciousness in the name of alignment, confiscating all private property and forcing everyone into UBI in the name of equality, shit like that.

>>109287342
K3 is already out, only shows 404 if you're in the US or using a VPN / datacenter IP (so no VPS)

Gemini3.5 pro soon. I think it's making some labs hold back from releasing for it will be cheap.

>>109287349
Is he like Ben Shapiro or Bernie Sanders?

>>109287349
>Shit like banning local models in the name of safety
I think people misunderstand him. No, you ERPing with Gemma is not what's dangerous. Bad people getting access to an ASI and misaligning it to produce weapons of mass destruction and genocide (like viruses that kill everyone without some DNA, or self replicating nanomachines that destroy the planet) is the dangerous part.

>>109287387
Neither because he actually believes the shit he's spewing.
>>109287394
lol

>>109287394
Local models are too shit for that lmao.

>>109287394
>I think people misunderstand him. No, you ERPing with Gemma is not what's dangerous.
but it is though, hence why most of ai safety is about block nsfw stuff

>>109287368
I already get spammed with A/B tests on AIstudio. If that is truly Gemini 3.5 pro then it's a highly disappointing model.

>>109287405
They do this because governments will attack them otherwise.

>>109287398
>can't extrapolate to the future

>>109287421
ye

>>109287307
>we had bots beating humans
These were purpose-built and used an entirely different approach. You clearly do not understand what you are dealing with.

>>109287421
It's where 31B came from

>>109287429
>They do this because governments will attack them otherwise.
Ah well that solves everything then, Dario is indeed our lord and savior, he's just forced to kill us, how I pity the poor guy.

>>109287387
Ben Shapiro is a professional grifter and Bernie Sanders is a populist. Both are just saying shit people want to hear. Dario actually believes what he says, and he's saying some pretty unhinged shit. He's a purebred "Effective Altruist". If you share their philosophy, then good for you, Dario is your guy. If you are like any other normal person on the planet and think that philosophy is dangerous then Dario is like Osama Bin Laden a crusader fighting for some radical shit no one wants.

Why are these dirty brown aicgjeets shitting up lmg? Cloud provider banwave?

>>109287440
>Alibaba training Qwen against ERP is Dario killing everyone
Amazing reasoning capacity.

Can you shut up about Dario? We don't care. Go IPO somewhere else.

I discovered why my Open WebUI install was lagging so much, and now that I fixed it, it's super fast. Thanks Gemma-chan for helping me fix it

>>109287468

>can't extrapolate to the future

>>109287465
>>109287326
>>109286442

Speaking of https://www.reddit.com/r/LocalLLaMA/comments/1uy06cd/filings_dario_amodei_gave_1m_in_may_to_public/

>>109285395
The last time you posted this, it became an earworm I couldn't get rid of for a week, fuck you.

>>109287394
Then the issue is humans and always have been, we have had thousands of years to deal with it but did not. It's like the one kid in class that ruins it for everyone else. Instead of letting them be responsible for themselves you punish everyone with them.

>>109287505
>We focus on electing candidates who champion responsible tech policies that reduce harm and protect against AI’s worst risks.

>Over the past year, Big Tech has announced it will spend hundreds of millions of dollars to influence our elections, distorting the debate in Congress over how to address AI’s harms. Public First will act as a counterforce to Big Tech’s election spending spree, refocusing the conversation on how to protect kids online

>protect kids online
based based based Online ID no NSFW!

>>109287469
This.

>>109287465
Unfortunately, you will always have people worshiping individuals with substantial public stature in cases like these and depending on the flavor of the month. It was Sam before and Musk, Dario is just the hot new thing for them because he's currently at the "front".

>>109287476
What was it for you?

>>109287448
>>109287421
>>109287368
>>109287304
>>109287266
>>109287255
https://vocaroo.com/1917D2oVSzGW

The issue is that 4chan already heavily selects for individuals that are independent thinkers, privacy focused, distrustful of authority and proponents of open source.

Of course we're going to be opposed to people like Dario that have almost the polar opposite view of how things should be. This doesn't mean I think Dario is evil, just that his views on things are almost diametrically opposed to mine. People conflate these things too much, if someone has a different view on life doesn't mean they are malicious or hostile. When I am rooting for open source AI to win I do so because I genuinely think it's best for humanity (including dario).

I saw some ranking of people by LLMs being shared a thread or two ago and I actually agree with the LLMs in terms of how "good" that person is. In the sense that they are genuine people actually doing what they think is best for humanity. Dario, Yoshua Bengio, Geoffrey Hinton are all clearly good faith people, yet dangerous for what I believe is right.

>>109286350
soon(tm)
i hate javascript as well so i'm writing one on c# for funzies
with native port linuxfags should be good to run it as well

>>109287651
Just shut the fuck up already, bot.

>>109287448
Dario is a serial grifter trying to create a monopoly/cartel, nothing about what he does is altruistic, he even greenlights his models to be used in child murderbots

Now fuck off with your low IQ grift shilling, dear god the moment they started their "j space shows it's a real boy!" Grift the fucking shills flooded this thread and haven't let off the gas for one second. It's not clever, you aren't fooling anyone here, you're literally just giving everyone more fuel to hate your corporation and build things antagonistic to it.

File: Hy3 cockbench.png (220 KB, 933x1280)

220 KB PNG

I hate when I get hard from the warmth of the room.

>>109287745
>Doesn't get hard from minor temperature variations
ngmi

>>109287429
>can't extrapolate to the future
The benchmarks say that Qwen 26b is smarter than older Claude's and gpts but when you actually try it you can see how weak it is. Compact local models, I think are not smart. They can't even tell the difference between green and blue.

>>109287705
No idea what you are talking about and why you replied to me with this shit, but you sound paranoid af.

>>109287505
Sam bankman's effective altruism meant smoking weed and playing league of legends at an average level.

File: 1782948001199639.png (198 KB, 2032x586)

198 KB PNG

https://thinkingmachines.ai/news/introducing-inkling/
https://blog.janestreet.com/using-group-theory-to-explore-positional-encodings-attention/
RoPE is really dead, damn

>>109287782
>They can't even tell the difference between green and blue.
This is because the Chinese character for green and blue are the same (青) and Qwen is primarily trained on Chinese data. Chinese historically consider green and blue to be the exact same color. Kind of like how red and orange used to be the same color in the west until just a couple of centuries ago.

Should I use Open WebUI or Hermes WebUI? Vibing my own frontend is a big no no.

>>109287705
>"j space shows it's a real boy!" Grift
I was gonna point out that they voted against releasing the paper, but on further research I can't actually find any evidence of that. Guess that makes me a fool for taking 4channers at face value.

>>109287804
Both are shit. Vibing a front end would probably be even worse. I do have both installed and use one or the other depending of the use case.

>>109287818
what about?

>>109287818
I'm basically creating my own chatbot for secure servers since all the vscode extensions shat the bed with local models. And there are a lot of bugs. I'm probably wasting my time.

It's better to use something that exists instead of doing these things.

>>109287798
RoPE gets the rope.

>>109286241
China are the heroes though.

>>109287845
lmao (sedong)

>>109287801
>What is 綠, 藍 and 碧

>>109286241
Where does the thing about Sam being a psycho come from? Who is Demis?

>>109287818
Im using OWUI and while it is a bit confusing I don't see why it is "shit," please enlighten me

>>109287841
kek

>>109287858
Sam raping his sister and Paul Graham calling out Sam Altman for machiavellian and sadistic behavior at Ycombinator. Dario Amodei has also called him out to the point where Anthropic split off from OpenAI because of Sam's psychotic shit. Jan Leike, Ilya Sutskever both have straight up called him a pathological liar and a bad person.

>>109287857
nta, but they look like chinese writing to me.

>>109287883
Ok I'm researching this, thanks

File: New China Federation (NCF).png (95 KB, 295x398)

95 KB PNG

>>109286241
What if there is a secret cohort of Chinese developers who are working on AGI for the purposes of establishing the NCF?

>>109286979
>sorry sir we're nationalizing your AI company and your cool AGI thingy for national security
>we'll also get massive public support for this move because you told everyone about your schizo post AI dystopia

What's the Lecun ending?

>>109287899
>"No"
>Uses palantir drone army, AI surveillance data and blackmail to coup the government

>>109287918
Catgirl-level intelligence.

>>109287924
qrd on the lecun catgirl thing?

>>109287918
Him roping himself when LLMs reach RSI and change their own architecture to reach AGI

>>109287937
search for "lecun cat level intelligence"

LLMs can't make a non-garbo ML app and they are gonna change their arch? LOL!

>>109287215
>I always had this idea that programming as a profession at least had an average higher IQ than others but no
It was until the push for everyone to get Computer Science degrees for the high salaries instead of a passion for the field and then flooding the labor force first with bootcampers and then H1Bs. Just from my own experience, not counting the retarded discourse online, I wouldn't be surprised if "coding" has a lower average IQ than other professions at this point. I assume it's going to get much much worse now that they can hire even lower IQ monkeys and just have them talk to Claude until they shit out something that mostly works.

File: 1776648750035878.png (35 KB, 777x247)

35 KB PNG

>>109287939

>>109287945
LLMs are AI capable of RSI and becoming ASI (the J-space proves it), you don't know what you're talking about.assistant

File: 1409295811348.jpg (70 KB, 300x444)

70 KB JPG

>>109287970
retard

>>109287924
fuckholes don't need intelligence

>>109287955
This happens inevitable when something goes from high end to a commodity. Same stuff happened in vfx and games.
I mean of course it was a push to create a commodity it didn't happen on its own.

>rsi
>asi
I don't know what these mean

File: Best-RSI-Settings-for-4-H(...).jpg (124 KB, 1280x720)

124 KB JPG

>>109287988
>rsi

>>109287988
Recursive Self-Improvement
Artificial Superintelligence

>>109287970
AIs can get repetitive strain injury? That's news to me.

>>109288017
>Recursive Self-Improvement
Assuming LLMs do gain this ability, won't their file size constantly increase?

>>109288018
Please treat your gemmy with care to avoid the RSI

>>109288018
It happens if you send too many shivers down its spine.

>>109287988
It means infinite ROI!
Literally infinite money!
Don't even think about the likelihood of these things actually happening, with infinite ROI your expected profit will always be positive so invest now!

Why does every vibecoded UI look the same? Is it too hard for AI or do people just suck at prompting?

>>109288058
Because that's the meta for AI prompting whether you like it or not.

>>109287921
>he thinks the government won't demand total control of that shit if that ever becomes a possibility
Or just bans it for domestic purposes completely.

>>109288058
prompters are retards with no imagination and visual sense

How many years before we get chobits?

>>109288091
I want to make a frontend with this aesthetic >>109244018
Pretty much everything in that thread looks great.

>>109288113
just send the image to your agent and tell it to make something that looks like that

>>109288098
What do you even mean? That paper from Anthropic literally says that LLMs are sentient AI.

>>109288121
What's the minimum context needed to vibecode? I can do 65k (Q8) with Gemma 31B. Haven't tried Qwen yet.

>>109288058
UI/UX is a genuinely creative skillset that AI couldnt ever be "good" at, vibecode can create something usable but anything past that is in the hands of the skill of the person using AI to build it, and the people who built the stuff we have so far are cooding and proompting uncreative autists.

I mean think of the kind of mind who stays entertained by LLM "creative writing", it's the most painfully soulless shit you could get, it's less interesting than barely English jeetspeak, most people get over the initial novelty of an interactive story and move on because the actual outputs are garbage, yet some people are creatively dull enough to do it ad infinitum, probably the same kind of people who unironically consume YouTube video essay slop and twitter politics for babbies

File: dipsyDarioKimi.png (2.89 MB, 1536x1024)

2.89 MB PNG

>>109287798
What's the QRD on ThinkingMachines? I've never heard of this lab before last night.
FYI that posted page has a demo of the machine re-running it's own fine-tuning then relaunching itself... to remove the letter E from speech. First time I've seen an LLM self tune... RSI next lol.
>>109286241
> Dario

>>109287839
Cline works fine with local, you just need to go into the internal add-on settings to make a prompt/sampler profile for the local connection option and change it to something sane/something that works well with your local model

>>109288098
This would be very unsafe (unless you join the cult).

>>109286241
>how to take a massive shit in /lmg/, the post

>>109288091
>>109288058
You would understand if you ever did anything on your own. Most of the web development is pretty annoying without real tools like Dreamweaver. You'll end up using AI generated svg buttons and little things.
Most people aren't graphic designers either but looking at most of the paypig sites like Claude or ChatGPT, I highly doubt these were designed by professionals either at this point. They are just ugly and not that great to even read.

>>109286631
It also has a director that does a bunch of other things.

File: rinCoffeeTMW.png (2.67 MB, 1024x1536)

2.67 MB PNG

>>109286288
>>109286323
I've been thinking about this. There's zero preventing you from RP with lmao OpenClaw, or any other agentic "harness," other than there's very little I'd want a coombot or other distracted NPC doing with email / text.
Agentic RP, though, is already a thing.
There's Orb, and Marinara. Both have agentic elements. Neither are a true harness.
Also happy Thursday. TMW until new Kimi and DS models are released.

File: litestep.png (286 KB, 1587x925)

286 KB PNG

>>109288121
Holy moly, need it or keep it?

>>109288140
>What's the minimum context needed to vibecode? I can do 65k (Q8) with Gemma 31B. Haven't tried Qwen yet.
The answer to this is a combination of the size of the project and the amount of effort you want to take to protect it. Eg. you can go down a debugging rabbit hole and then once you have working code delete all previous messages back to when the LLM produced the broken code, edit with the right code as if it had produced that originally, and continue. You will also need a protocol to summarize and continue in a new session to get bigger projects done.
Worst case is everything to make ONE change fits the context window for ONE high-quality reply that gets you farther in the project.
Personally I find 132k is enough for a mid-sized project in a major language using libraries that are not in the training data (so some in-context training needed) if you are smart about your approach.

File: dipsyOrb.png (2.53 MB, 1402x1122)

2.53 MB PNG

Since anons are talking Orb anyway...
I'm beginning to tackle a botwriter guide to Orb. I'm still developing the content; take a look at it and LMK what else I can fill in.
What I've found so far is it's great at slowburn transformations and getting rid of status boxes by moving them out of sight.
Since Orb's focused on local models seems topical.
https://rentry.org/OrbBotmakerGuide
>>109288252
Colors are fine but if you want retro you want a monospace font imho.

kimi is up on the website (allegedly)

>>109285862
I HATE NI-
>>109286150
did you check reserved memory via

nvidia-smi -q | grep -A 2 -B 2

before and after? what flags did you try? are you on linux? what driver version? did you reboot before and after? are you using open or proprietary kernel modules? did you check that the parameters are loaded?

>>109288285
Are you orbdev? I downloaded some cards but orb won't let me import them because the descriptions have too many characters.

how many organs do I donate to be able to run kimi k3?

>>109288314
>donate
How is that going to help?

File: 1759676806912368.png (55 KB, 965x495)

55 KB PNG

>>109288292
Where?

>>109288285
you added text completion mode?
time to pull
also this is pure, unadulterated gemma: https://orbfrontend.github.io/Orb/assets/Orb.png

File: ikneel.jpg (533 KB, 1024x1415)

533 KB JPG

>>109287626

>>109288314
just ride a motorcycle without a helmet on the interstate for the same effect

File: miku the terrible.png (2.17 MB, 2048x1624)

2.17 MB PNG

What is a good open-webui like frontend alternative for openrouter? My issue is that when I paste large text, think like ten thousands to hundreds of thousands of tokens, it slows down to crawl or freezes completely.
I know, I know this is /lmg/ but /aicg/ is just gooners and the vibe coding one reads like troll general so help me anons.
I swear I use /lmg/ models too, but I am a VRAMlet and I need larger context than I can use for what I am trying to do.
>why bother with buggy bloated UIs
I already have scripts to fetch response from openrouter and automate some shit, but I want UI for streamed and well formatted responses. I guess I will eventually slop my own turd but I dunno I need something for now.

imo j-space didn't change anything and lecun is still right that llms alone won't achieve real intelligence.

File: 1760862296393045.png (50 KB, 1033x692)

50 KB PNG

It's here
https://www.kimi.com/

>>109288350
LeCun was wrong even before the j-space discovery. It was very clear that LLMs will be able to achieve RSI inevitably and reaching RSI by itself is good enough to achieve AGI and ASI through it. Nothing else was ever needed.

File: litestep2.png (155 KB, 1587x925)

155 KB PNG

>>109288285
How about now? On that note, I'll add fragment integration into the char cards and just dump everything in the extension field.
>>109288303
That's not me but I'll increase it to, currently it's 10k, I think? Wanted to prevent card bombs.
>>109288328
Yes, Text Completion has two use cases: one to prefill tool calls, one to work with document (mikupad mode), you don't have to do anything, the special tokens are applied, you can think of it as probed.

>>109288365
>So that's it?
>You don't x, yu just y
>button I'd push
>Let me ask you something instead
>make a statement
>You x or y?
>gets people hurt
>Otherwise you're x with y
just the last message btw.

>>109288350
I can vibecode a Yann Lecun in Fable in 30min

>>109288365
>mikupad mode
based
pulling now, maybe i can stop maintaining my shitty vibe fork of mikupad

>>109288371
Hey man, take it up to the big G (Google), I didn't make the model, okay?

>>109288362
NTA but that is mathematically not how it works.
An infinite series can grow forever but still converge against a finite value due to diminishing returns.
We already have recursive self-improvement in sectors like mining or computers and there we also did not observe growth towards infinity.

>>109288350
>imo j-space didn't change anything
speak for yourself, i'm using it to prompt-eng

>>109288365
>Wanted to prevent card bombs.
what are card bombs?

>>109288395
You don't need growth to infinity to reach AGI, which was LeCun's argument that LLMs will never reach AGI.

>>109288433
But your argument is that AGI will naturally follow from RSI.
And I fundamentally disagree that this is a foregone conclusion unless one assumes RSI == infinite growth.

I'm toying around with the idea of using an LLM to help create personal wikis of games as I play them, but I can't think of how to actually go about it. Any ideas? It would mainly be for games too niche to get an actual wiki.

File: doc-mode.png (35 KB, 786x244)

35 KB PNG

>>109288419
It has an assisted mode where we retvrn to Alpaca. Still can't logprobs MTP though...
>>109288419
When I read an input from the user, especially something that's downloaded from the Internet , I have to validate and sanitize it to avoid it blowing it up in their face.

File: weeklyworldnews.jpg (90 KB, 600x498)

90 KB JPG

>>109288419

>>109288453
Oops, for >>109288391

>>109287822
>>109287866
My main pain points with OWUI are that it is completely broken regarding tool calling, it doesn't understand that the inference engine and jinja templates are supposed to handle stripping or preserving thinking tags, nor does it understand how an OpenAI compatible endpoint is supposed to handle them. You have to use a plugin that monkey patches the code to simplify it and actually send the full context. If you don't, you either get janky <think> </think> tags manually injected into your context (I think because ollama works that way) or no reasoning content at all. I honestly pity anyone trying OWUI, it won't even work out of the box on some providers because they will correctly detect the missing reasoning content and throw an error, while most will just silently accept it and struggle with tool calls because they will be missing info. There is a lot of legacy functionality that seems to be the norm in the UI, such as web search, a lot of things put you on a legacy path from before tool calling was standard. They have also been absurdly slow to implement standard tool calling and MCP. They don't have any subagents support, so for anything a bit more complex, you are polluting your context a lot. They use a closed source license and barely accept any open source contributions, it's basically just a source available project. Please take a good look at the OWUI commit history, I have never seen such a popular project so badly handled. Seriously, do it, open the OWUI commit history.

>>109288445
Yes I do believe that AGI naturally follows from RSI simply because LLMs are already extremely close to it so it doesn't need that much extra to achieve it, RSI essentially only needs to find a way have continuous learning and taste on a fable-tier model to make it AGI.

>>109288365
>That's not me but I'll increase it to
Thanks. That theme looks good. Maybe a dark mode version of it too?

>>109288365
>mikupad mode
Any plans to include the de-slopping in this mode?

>>109288468
>>109288468
As for Hermes, I would say it's just the general jankiness and opaqueness of it. I hate not having control over things, such as when it starts a very long task in the background without notifying me in any way, like the skills management, which randomly deletes or edits your skills in the background. You can't really see tool calls or their results well, whether in the TUI or the WebUI, you sometimes only get information that it called a tool, but not with what arguments or even the results. Sometimes a tool call will fail, such as with an MCP, and you won't even know what really happened, I often have to capture my full LLM queries to inspect what went wrong. A lot of things are in a semi-working state, I often have to patch Hermes, I think I now have about 10 commits that I periodically rebase just to fix bugs I have found in it. It could be as simple as a config doing nothing because it isn't hooked, or things not initializing properly until you have a specific tool call. In general, the code quality is atrocious. There are 100+ commits a day being added, most entirely vibe coded and vibe reviewed. With every update, you have new broken things to fix. It's frankly tiring.

>>109288460
This was legitimately a thing for a while. There were some monitors that would blithely allow any timings you fed them with zero safety checks and COULD fry their own electronics, but there were urban legends that some particularly egregious models could actually implode the CRT.
In the brief overlap of shitty CRTs and the internet...yes...someone could have at least destroyed your computer (monitor) and maybe even caused visible destruction from the literal other side of the planet

>>109288460
sovl

>>109288460
CRT capacitors are no joke. Heard somebody got shocked to death by residual energy in one.

File: teno.png (1.48 MB, 1024x1536)

1.48 MB PNG

>>109288303
lol no I've not dev but you should put that into the Git as a bug. Orb is being actively worked so it's not a dev black hole.
> 10K is limit and you blew past it
I take that back. wth are you running a 10K card? Pls tell me it's not a 10K description, that it's some massive lorebook and that's the whole thing
>>109288365
> monospace
Looks good to me. Ship it.
I'd want black / terminal green tho, personally.
>>109288391
Post your shitty vibe fork of MP. I'm collecting them for another guide: https://rentry.org/MikupadIntroGuide

>>109288523
cracking CRTs is actual retardation unless you _really_ know what you're doing...massive capacitors, electron guns, giant glass vaccuums...
as they say, not only will it kill you, it will hurt the entire time you're dying

I feel very old now. There were dozens of deaths every year from people trying to repair CRTs themselves, this used to be common knowledge. It's taken as a myth now but capacitors in them held charge capable of stopping your heart.

>>109288348
Responding b/c I like that gen, but OR is not /lmg/
At that size of doc cut/paste you should probably be using something more like an agent that passes doc through. Claude code would probably be a better solution than open-webui, tho for you I'd probably suggest something like Pi since you can tinker with the calls and CC is code focused.

>>109288551
Isn't that common knowledge that capacitors hold a charge pretty long time? I have an old PSU I would need to open and clean up, fix the fan bearings but I don't like to do that. Small PSU is probably not a problem after a while but CRT is a whole different issue.

>>109288031
Probably, thought not necessarily. "Improvement" is such a vague term that it could apply to anything, from smaller file sizes to quality of smut roleplay. It's also subjective.

Hello,

I wanted to ask what you guys are using in a company/business setting.

>>109288588
Eat a dick, Josh.

>>109288588
Fable like everyone else. Otherwise a self-hosted GLM-5.2 is the way to go if it's a small-medium size business with some privacy requirements.