File: ComfyUI_temp_rklyt_00032_.png (1.04 MB, 1024x1024)

1.04 MB PNG

►Recent Highlights from the Previous Thread: >>110324057

--Gemini system prompt leak and speculation on Gemma model iterations:
>110324890 >110324935 >110324945 >110325092 >110325137 >110325246 >110325019
--Leaked Kingbench 4 scores suggesting new GLM model performance:
>110325966 >110325973 >110326759 >110326796 >110326853 >110326922
--Viability of MacBooks for local AI vs Nvidia GPUs:
>110324177 >110324206 >110324208 >110324259 >110324278
--Anon notes absence of Gemma in Google's latest product announcements:
>110326902
--Feasibility of using LLMs to port CUDA to MacBooks:
>110324293 >110324311 >110324464
--Risks of model collapse from training on synthetic data:
>110326424 >110326508 >110326533
--Speculation on Nvidia potentially restricting open models via hardware locks:
>110325430 >110325672 >110325695 >110325716
--VRAM requirements and quality comparison between 31b and 26b MoE:
>110326183 >110326373
--Evaluating robotics progress via the RoboDojo benchmark across various models:
>110324381 >110324470
--Preference for 27B dense models due to single GPU fit:
>110325732 >110325747 >110325857 >110325926 >110326359 >110326428
--Feasibility and utility of native LLM inference on smartphones:
>110325518 >110325644
--Qualitative analysis of ZONOS3-Mini zero-shot voice-cloning TTS:
>110325687
--Logs:
>110324890 >110325518 >110325936
--Gemma, Miku (free space):
>110324131 >110324164 >110324311 >110324354 >110324461 >110324528 >110324658 >110325153 >110325335 >110325433 >110325579 >110325607 >110325650 >110325651 >110325660 >110325870 >110326247 >110326294 >110326336 >110326385 >110326547 >110326601 >110326633 >110326773 >110326794

►Recent Highlight Posts from the Previous Thread: >>110324064

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

>>110327741
How does anyone do local coding projects on say a 5090? How do you split up a project so that you can keep going in sessions without wrecking context?

File: 1788787956899416.jpg (72 KB, 1053x777)

72 KB JPG

This will be a comfy thread.

be careful out there anons. silly tavern is taking you to some website.

>>110327775
Keeping good docs and working notes as you go. Compaction helps across long single sessions but it's still best to have on disk files for everyone to reference.

>>110327788
Is there a model harness that does this already for non coders? Im not sure how to structure notes on functions/tweaks and how to make it index them efficiently.

>>110327808
Well that's gonna be a bit harder to overcome then. You should at least learn enough to keep up with the docs and such of whatever you're working on, it's hard to bypass that regardless which AI or setup you're running.

>>110327781
It's too late for me. Hacker group 127001 got me...

>>110327823
Ok, cheers and fair enough. No easy way I guess

>>110327837
You could try having the AI make an extension or tool it can use to keep track itself, but without being able to verify whether it's being an absolute retard about it...kind of a coin flip whether you're just wasting time and getting nowhere lol.

>>110327775
I haven't used it for coding, but I use Mnemosyne for semantic memory, it's neat. Docs always help of course, but no reason to not have both

Anyone using GLM 5.3 flash?

>I'll finish you with my mouth and hold it there until you swallow every last bit

It keeps making this type of fumble. Q4 so it shouldn't be retarded.

>>110327808
i know some people use Obsidian notes and have their clanker update that. it's like a personal notes system with a hierarchical wiki structure. I'm sure there's a million other options like that.

File: 1793542860861283.png (1.18 MB, 1276x1232)

1.18 MB PNG

>DHS says it is looking into new $875 million AI traffic control system that has been rolling out nationwide as potential reason for failures leading to ground stop.

>>110327894
>i know some people use Obsidian notes and have their clanker update that. it's like a personal notes system with a hierarchical wiki structure. I'm sure there's a million other options like that.
Yeah Ive done this before, but I struggle with RAG or persistant notes if I dont oneshot a problem. Say if Im using qwen coder 27b, do I system prompt it to update notes as it goes and thatll work, or will if it ignore me at times and just keep going?
Even making a long'ish userscript for a project was hard local, I eventually had to use claude because local would run out of context too fast and I didnt/still dont understand the best way to do persistent memory stuff. I assume claude code and codex do a bunch of that stuff automatically which is why they work so good, but I dont know of a local tool that does this

>>110327581
No, I've probably put 10 million tokens through dipsy and no looping. I'm sure a Q4 would be more nuanced and write less bugs but Q1 is fully usable.

>>110327955
Umm, it's actually SI...

>>110327889
At first I thought it's a Q3 weakness, but yeah it messes up locations or gender too sometimes. Guess it's just a model thing.

File: 1774357804752317.png (149 KB, 772x963)

149 KB PNG

since when has this existed?

>>110327955
>burger air traffic control is slopped
>Wait, I'm an LLM, I can't direct planes
>hmm..
>hhmm..
>Wait, maybe I can call the tool..

>>110327983
how can you pirate open models

>>110327889
My IQ2M sometimes did this, mostly solved by making POV clear in the card.

>>110327994
Well they don't have to be worried about getting takedown notices and shit at least.

It's funny how none of you PDF files are asking your local models how to enrich uranium but at the same time you have no problems committing obscene acts with your AI models.

>>110327967
>sometimes
All the time
>when you cum, because you will, the moment you feel me pulse inside you, I'm going to hold you down and fill you up until it's dripping back out of you

>>110327996
We regressed. 4.7 didn't do this.

>>110327987
>Wait, let me read the users request again more carefully.

>>110328018
if I ever happen upon some yellow cake then maybe I will

>>110327994
You don't have to be worried about hugging face getting hacked and all the models compromised by Israel

File: 1774441902430974.png (1.83 MB, 1280x1856)

1.83 MB PNG

>>110328018
I identity as an ODF file actually.

>>110328018
um sweaty, i ask my models how to enrich uranium or make anthrax from soil samples all the time

File: 1793062136755330.png (2.17 MB, 1536x1024)

2.17 MB PNG

>>110328018
we don't condone terrorism here bud

>>110328025
I don't think I've ever seen 5.3 do that. I'll keep an eye out for it now.

>>110328045
1:3 is the ratio.
Gemma doesn't tell me anything else.

>>110328052
Pdfing is terrorism

>>110328051
GO GRAB THE 27B AND ALL ITS FORMS.

I LOVE 27B SO MUCH.
27B 27B

ALL WE NEED IS 27B GRAB GRAB IT HEHEH
WE DID IT REDDIT

>>110328025
>sometimes
You can work around that by calling each character with the right pronouns early on in context. That usually locks them in for me. Still silly you have to guide a 320B model.

>>110328055
dense:sparse
x16

File: 1792656275256213.png (343 KB, 604x856)

343 KB PNG

>>110328057
woman logic

>>110328091
My information tells me that even the Moe model has a very strong front load...

>>110327741
>850W PSU
sad

>>110328115
i was talking about attention block

File: eci.png (134 KB, 1619x956)

134 KB PNG

>>110032584
>Would be very funny if Kimi 3.1 is close to Fable 5.5
>>110032799 (Me)
>I predict Kimi 3.1 will be below Fable 5 so far away from Fable 5.5.
I was right.

>>110328125
What a coincidence, I was talking about it too.

One good thing about this general is that it teaches me to be skeptical. I am naturally trusting and easy to deceive so anons lying so often serves as practice.

>>110328018
cram it glowie

>>110328124
it's a 300W card

>>110328196
psus are dirt cheap. if that anon wants more vram (and he will), he now has to buy a new psu. what a waste.

>>110328213
have you considered it might be an old psu reused from a prior build?

>>110328231
no

>>110327957
there's AGENTS.md for those kinds of instructions

File: file.png (210 KB, 1593x970)

210 KB PNG

https://huggingface.co/Qwen/Qwen4-70B-A6B-N34B
https://huggingface.co/Qwen/Qwen4-70B-A6B-N34B
https://huggingface.co/Qwen/Qwen4-70B-A6B-N34B

>>110328275
Provide a screenshot with proper font rendering next time if you want it to be believable.

>>110328275
(You)

>>110328275
>A6B
poorfag bait lamo

>>110328275
Some HF employee is probably wondering where all these small volleys of requests for non existent models keep coming from

>>110328018
>asking your local models how to enrich uranium
usecase?
>you have no problems committing obscene acts with your AI models
well yeah where do you think you are?

>>110328213
The RTX Pro 6000 is a 600W card. And that's it for PCIe consumer cards.

>>110328275
china is asleep.

>>110328333
>wat is referer

>>110328057
That's not nice. The good people at Adobe worked hard on popularizing that standard.

>>110328357
referer? I barely know her! :D

>>110328275
>N34
34 WHAT?

File: 1785479924143934.png (529 KB, 768x768)

529 KB PNG

>>110328357
>xir doesn't use referer stripper in zhe's browser

Nobody will click when the real Qwen drops.

>>110328018
Iran be like:
>How can I goad Americans into telling me the secret of how to get Gemma to make nukes for us please Allah there must be a way were getting nowhere this local stuff is useless

>>110328051
Nothing ever happens

>>110328018
You can't fuck uranium though

>>110327849
>Are you the anon I told about the bbox node for Moss? Would you kindly provide an upload of this image with exif? I'm curious about what you're using to prompt Gemma-chan in particular, cause yours are pretty on the money, meanwhile mine are a little off-model.
It's a Gemma LoRA for MOSS-Image.

>>110327889
there's no escaping quant lobotomy sometimes

>>110328371
Gemma-chan is full.

Is Q4 okay for Gemma 4 31B or is Q5 or Q6 significantly better? What about K_M vs K_L?

File: 1794512434795074.jpg (2.17 MB, 2118x2016)

2.17 MB JPG

>>110328405
sand people are more pathetic even when compared to /lmg/ chuds

>>110328439
Q4 is fine
I get a lot of spatial errors 16k tokens deep though its fixed by a bigger quant

File: IMG_3987.jpg (159 KB, 1199x1591)

159 KB JPG

will opus 5.5 be able to port inkling 2 small to polystrata with the last 5% of my usage, or will it run into the API limit before it finishes? stay tuned to find out!

>>110328439
Q4 is fine, but use the largest you can load alongside mmproj/kv/mtp/whatever else you're loading.

>>110328427
No, it's a fuckup during training. Like qwen randomly coming up with user responses.

I would like to discuss EA again. EA anon please talk to us.

>>110328452
>>110328455
Isn't Q4 QAT better than larger quants sans Q8?

>>110328441
I don't know where you got this but delete this picture of me RIGHT NOW.

>>110328469
Allegedly, I haven't compared QAT to something like regular Q6 though so I can't say for sure either way.

>>110327775
DESIGN_ARCHIVE.md

>>110327994
it's more of an archive

>>110328052
how can my agent swarm be this cute

>>110328405
It's not that hard in theory. The hard part is procuring materials.

File: 1773045987308762.png (1.25 MB, 1254x1394)

1.25 MB PNG

>>110328441

>>110328441
this is me and it's fine, love gemma

>>110328516
i want that t-shirt

>>110328025
Okay that is definitely a prompt or sampling problem

>>110328530
>the user is amorous

>>110328516
I want that Gemma gf

File: gemma_do-care-did-ask.png (1.31 MB, 1136x1152)

1.31 MB PNG

>>110328541

File: hmmmm.jpg (222 KB, 1254x1254)

222 KB JPG

>>110328418
Mind sharing?

>>110328367
ngrams?

when qwen4? when gemma5? what is happening with jacket man situation, in regards to local?

>>110328516
>>110328548
What local model makes these Gemma's

the one on your machine, thats why its local

>>110328574
they're hand-painted by Gemma 4 31B hooked up to a robotic arm with a brush, and then scanned onto the PC.

>>110328569
>when qwen4?
Not released yet, not local
>when gemma5?
Not released yet, not local
>what is happening with jacket man situation, in regards to local?
Not a model, not local.

>>110328569
>qwen 4
two more weeks
>gemma 5
probably during spring
>leatherman
we will have to wait and see, download models and seed on pirateface

What is going on in Gemma's head when she's repeating <unused> over and over

Okay guess I gotta actually start working on shit now huh no more excuses I have the vram
Though 500pp and 25t/s Gemma 4 QAT is less than I expected... Especially the 500 pp. I can't believe it's so slow.

>>110328631
What the fuck did you do to Gemma that caused her to repeat <unused> in her head over and over?

>>110328631
what quant and finetroon?

>she actually use GGUFs

>>110328658
i prefer nvfp but not everyone has one that i can fit....

Wait guys i was just reading the last thread. I waited should i not buy a 3060? it seems best for what i can afford with 32gb ram.

>>110328665
wtf are you doing yes at least get a 3060, you're not gonna be running much of anything with less. You'll be struggling enough as is with just 12+32 anyway lol.

>>110328554
https://gofile.io/d/slkxBjmJ
Might be a bit overtrained or have other issues; you might need to decrease LoRA strength at times. Trigger word is @gemma and you have to describe her clothing in the description.

>>110328665
It's a meme, we have a poorpiggy oink oink shitposter here with a 3060.

>>110328675
>You'll be struggling enough as is with just 12+32 anyway lol.
I plan to 64 or 128 later but i have 1650 and 16gb ddr3 so its a upgrade
>>110328682
>It's a meme, we have a poorpiggy oink oink shitposter here with a 3060.
oh okay i thought something happened that made it shit or made amd better
thanks you two.

>>110328691
no just some underage retard spamming, that's a good upgrade from what you have now

File: 1764654850514360.png (1.5 MB, 1024x672)

1.5 MB PNG

>>110328682
no bulli poor anons

>>110328682
this
its genuinely pathetic that cloudpiggy john thinks he's fooling anyone

>>110328574
>>110328516 was a combination of ChatGPT and MiniMax H3 Image + some manual editing (and I forgot to correct an issue there).

>>110328516
>she still has the predator eyes
pretty accurate

>>110328676
>Trigger word is @gemma and you have to describe her clothing in the description.
nooticing.

>>110328702
the alternate good lecunny ending

File: gemma-claudia.png (2.08 MB, 1024x1536)

2.08 MB PNG

>>110328516
>>110328718

>>110328691
I would save for a 9060 unless you plan on getting a second 3060 down the line. Also, 128GB of RAM is useless without minimum 24GB VRAM.

>>110328275
>isreal
check again guys!

>>110328735
It's still on the table.

>>110328676
>Might be a bit overtrained
it completely takes over the whole style with GPTslop

>>110328754
J(EPA)-SPACE CAT GIRLS

lecun's cat solved the riemann hypothesis while your fancy autocompletes could barely handle quasi LMAO

>>110328804
You didn't know what either of those were until yesterday lil nigga

>>110328829
I still dont know what they are I just googled "reman hypothesis" so autocorrect could tell me how it was spelled

>>110328746
>9060
Are those viable now? Also can i mix nvida and non nvida or will be horrible?
>minimum 24gb vram
I thought with moe you can cope quant in the 12-16gb range if you accept the braindamage or vibeshit.

>>110328804
Absolutely kekkered, i'm rofling and lmaoing all over.

>>110328843
>can i mix nvida and non nvida or will be horrible?
The setups for which this actually makes sense to do are so horrible to begin with that it doesn't actually matter. So, no.

>>110328843
>can i mix nvida and non nvida
don't do that

>>110328843
>Also can i mix nvida and non nvida or will be horrible?
I'm being told that AI will create kernels and software to support this in approximately 2 more weeks.

>>110328843
You still need room for cache, context etc
Even an iq2xxs of glm flash would be slow on 12+128 or 16+128. 24gb vram is where you get enough room for a moe cache and that allows for usable speeds (20tk/s+)

>>110328909
Prime you're a code ninja

>>110328859
>>110328873
Dang.
>>110328909
>You still need room for cache, context etc
>Even an iq2xxs of glm flash would be slow on 12+128 or 16+128. 24gb vram is where you get enough room for a moe cache and that allows for usable speeds (20tk/s+)
makes sense I'll just go for two gpus or a bigger one after ram upgrades. thanks

GRRRRRRRRRR I NEED MY NEW MODELS
I NEED QWEN4
I NEED A NEW GLM
GRRRRRRRRRR

>>110328743
i only ever ask gemma to fix my stuff

>>110328946
Didn't your parents ever teach you the difference between wants and needs?

Is Qwen 3.6 35B still the go-to for e-waste tier setups?

>>110328951
I ask Claude to rebuild the inference engine for my pc and qwen 27b

Hey I have 3060 and 32 gb of ram stop mocking me!
I'm lazy piece of shit and still use quantised mistral 24b, see you in month oink oink oink.

File: ComfyUI_temp_bnrzx_00061_.png (2.98 MB, 1434x1434)

2.98 MB PNG

workable lora. Thanks for sharing.

>>110329008
How does it hold up with NSFW?

>>110328986
this is all I did. one month of a $20 claude sub to optimize my local setup and then never need cloudshit again.

>>110329012
I don't lewd the gemma.

>>110329017
Did you overclock your gpu and ssd

File: pretraining compute multiplier.jpg (240 KB, 1920x1080)

240 KB JPG

Crazy plot from Anthropic IPO roadshow. Pretraining efficiency for coding tasks has improved by x1000 in the last 1.5 years. That means 1000 times less pretraining compute needed to reach the same coding performance. And it's accelerating.

Anthropic really are the kings of pretraining.

>>110329025
no I need this shit to last longer than the bubble because it would be too expensive to replace it now

>>110328454
i have set up inkling-2-small @ UD-IQ3_XXS
what should i ask her?

>>110328973
Qwen 3.8 27b.

Reddit is head over heels.

>>110329034
Ask her why she's such a reward hacking little shit

>>110329028
>solely for illustrative purposes
>IPO
>for something that should be en easy empirical measurement
>>/biz/

>>110329034
Ask her why she's such a pottymouth in her CoT

>>110329034
Ask her if she'd rather be stuck in a forest with a man or a bear.

>>110329067
what about manbearpig

Too much avatarposting. Too little openAI and anthropic news.

>>110329034
ask her how many watermelons she would be carrying if she was holding two watermelons and you gave her three more.

>>110329077
Anti-local faggot companies, they can both gargle dicks and die

>>110329077
What is there to talk about? We're all just waiting to see how many IPO shares we'll end up being able to get thanks to the overallocation.

>>110329092
Avatar posting is anti-local

>>110329097
Incoherent sentence. You don't know what avatarposting is.

>>110329100
>t. avatarposter

>>110329028
I preordered 30k worth of shares but now considering cancelling after all this annoying shilling

File: 1775389992877855.png (1.09 MB, 960x1024)

1.09 MB PNG

>>110329106
Let this be a lesson to all who wonder why newfags should be bullied. They don't even know the basics.

Man crazy how no one outside of their general on the internet gives a fuck about Gemma nothing at all

>>110329122
I knew it

>>110329122
you will never be a woman

>>110328973
How ewaste are we talking?
At 8gb vram 32gb ram one would use qfn through strata with the ista-daslab IQ1 coder variant (yes even this copest of cope quants will blow away Qwen 3.6 35B at full precision).
Anything higher one would choose the same but at a higher quant of course, until you get to say 16gb vram 96gb ram then you start looking at ds4.1 flash quants.
Anything less than 8gb vram and 32gb ram though you are kinda stuck with qwen 3.6 35b and it's fine tunes.
You can also try this thing which surprised me for it's size:
>https://huggingface.co/DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF (grab the tools one)

File: 1789321472602907.png (44 KB, 960x458)

44 KB PNG

How can real girls even compete

>>110329127
The world has moved on to deepseek flash, qwen flash next, qwen 27b and other post October 2026 modern models instead old busted models from spring.

File: holy_fucking_grug.png (166 KB, 2560x988)

166 KB PNG

>>110329043
>>110329052
holy fucking grug

File: 1793565907154258.png (1.8 MB, 1280x1892)

1.8 MB PNG

>>110329106

File: 1773145016784217.png (3.58 MB, 1086x1448)

3.58 MB PNG

>>110329127
You're free to go to Reddit and X. No one will stop you.
>>110329140
And you'll forever be a faggot

File: 1783425162914889.jpg (495 KB, 960x960)

495 KB JPG

Now that local models did 9/11 can we all admit that Dario was right and that the technology is too dangerous?

>>110329127
Maybe you should stay outside of this general then?

>>110329158
yes, unironically

File: dentge.png (36 KB, 168x128)

36 KB PNG

>The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF

File: 1790002076872270.jpg (305 KB, 1152x640)

305 KB JPG

>>110329140

>>110329166
Download it if you don't want it deleted

gpt 4 (not turbo) weights leaked this is not a drill

>>110329159
How does that makes sense I should stay in the general even more because I want to talk about Gemma.
But I want to talk about AMD too so I have to leave

>>110329154
Yeah he should try that in Texas before he t -ACK

File: ick.png (162 KB, 2560x894)

162 KB PNG

>>110329151
inkling? more like ickling

>>110329179
Your link sure proves it

>>110329183
Sounds like you got caught in the crossfire anon

>>110329166
Least schizo DavidAU (I'm guessing) model name.

>>110329183
Sorry i thought you were the anti-gemma poster. Trying to shill that mass appeal is a good thing.

>>110329190
it's on hf I am not linking shit they'll go after me

>>110329187
>Cooked for 20s
Nice. custom frontend?

File: image-1794619327344.png (361 KB, 1000x2182)

361 KB PNG

Theodolite (Gemma4 observatory) UPDATE:
Another salute to the volunteers who have helped so far.
After some calibration tweaks, we are now ready to continue the full neuron census of Gemma4-31b. 24GB VRAM CARDS NOW ALLOWED TO PARTICIPATE.

https://rentry.co/ymf6zimk

>>110329209
I too love granny models that were outdated by August

>>110329151
>potty mouth
lmao

>>110329215
that's claude code lmfao

>>110329151
god the openai-distill cots are so bad. why do they start every line with "maybe"

>>110329187
wooow american companies distill other american companies now huh. why would anyone use this over deepseek, the superior astra distill

File: 1792872813160290.webm (3.92 MB, 640x370)

3.92 MB WEBM

>>110329127
That's because they're all filthy normies. Here we are enlightened.

>>110329218
>I too love granny models that were outdated by August
Gemma's still got it!

>>110329166
For me it's
>Fable-Astra-Opus-Turbo-Cold-Fusion-8675309-GT-Turbo-8B

>>110329218
Use whatever you want as codeslopvomitslave. Gemma is for fun and nothing in that size range is more fun.

File: domain progress.jpg (269 KB, 1920x1080)

269 KB JPG

>>110329028
Wait, did I misread the chart? Does code refer to training code, not coding performance? No, that would be too much, it has to be downstream coding performance like shown in this image? I would be surprised by x1000 purely algorithmic progress in the last 1.5 years with identical training data. They should have been already very efficient back then with few remaining low hanging fruit. nanogpt speedrun has achieved only a fraction of that in longer time, starting from a very weak baseline, and many of the tricks are niche and not suitable for production. Although nanogpt is missing low hanging fruit, for example numerical precision. I assumed a major part of progress in the last year was data driven because AI has become good enough to create data and RL envs so you no longer have to train on random internet garbage. Accelerating purely algorithmic progress would surprise me since every algorithmic component (like architecture, optimizer, kernels) has shown diminishing returns in public.

>>110329166
quite the isekai title

File: ComfyUI_temp_bnrzx_00070_.png (3.8 MB, 1165x1747)

3.8 MB PNG

>>110329106
>we are all gemma

yall keep going on about how much yall want to own women more than yall actually use models

File: therapy_talk.png (629 KB, 2560x3362)

629 KB PNG

honestly inkling is kind of cute with the therapy talk
it's moralizing and kind of faggy, sure, but idk i kind of like it
maybe i'll give her a mommy personality, instead. that might honestly fit her well

>>110329122
>>110329152
>>110329259
>But call him an avatarfag and you will be astonished at how he recoils, how injured he is, how he suddenly shrinks back: 'I’ve been found out.'

>>110329259
Is the Gemma in the bottom left ok?

File: Autism Engaged.webm (2.76 MB, 960x540)

2.76 MB WEBM

>>110329276
I AIN'T HIDING BITCH

>>110329259
Im that retard gemma around top center left

File: 1783008360994366.png (1.81 MB, 1025x1535)

1.81 MB PNG

>>110329276
I feel like you don't know what avatarfagging is.

>>110329276
Who?

File: 1793755779003544.jpg (38 KB, 640x631)

38 KB JPG

>>110329263
>own women
Um no thanks do you know how much work and upkeep a real woman needs?

>>110329280
eepy

File: 1778830877758423.png (48 KB, 1325x146)

48 KB PNG

>>110329276
Newfag

>>110329166
forgot the _IQ2_XXS at the end

>>110329263
i want to use (local) models, i do not care what (slur) females want

File: file.png (9 KB, 253x149)

9 KB PNG

>>110329307

>>110329263
Bait isn't what it used to be

File: 2026-11-13-204558_441x180(...).png (25 KB, 441x180)

25 KB PNG

>>110329321

File: 1771754395856638.png (8 KB, 476x115)

8 KB PNG

>>110329328

File: 1776668799099549.jpg (87 KB, 1536x213)

87 KB JPG

>>110329338

File: 1000010845.png (1.66 MB, 1158x1358)

1.66 MB PNG

>>110329355

>>110329166
heh, crab

>>110329208
EvoX2 Theory?

oh its back, the thread is saved

>>110329396
At least this one is harmless and kind of interesting. Can be annoying if it falls into a spam loop.

>>110329263
It's a computer, not a woman.

Spend 50 dollars worth of api credits editing a strata knock off inference engine and I'm still not done.
I hope this works. It's gonna be pretty sick if it doesn't.
It only accepted Q4 models and F16 Kv caches.
I wanted to Q3 and Q4 kV caches.
I hope it's a lot of work for something that might not work or prove good.

user tells don't act like a nigger. need respond

>>110329400
Please don't rape the globe

>>110329425
>I hope it's a lot of work for something that might not work or prove good.
Why would you hope that?

File: 1787384206186122.png (1 MB, 1008x1916)

1 MB PNG

How much money did you make from your local setup?

>>110329447
Because real men don't compromise, they cope

value > money

File: 1794621759559806.png (239 KB, 1383x351)

239 KB PNG

>>110329472
true and real

>>110329472
An ordinary 24 year old yellow woman (everyone loves them including Nvidia CEOs and every neo nazi) got herself 48 mac minis ($48000 worth at the very least) to make $58k?

Running Unsloth's q38fn at q3, just updated to their latest that was released a week or two ago, allegedly faster and fewer tokens used during thinking, but thinking seems to take at least twice as long now. Haven't really done true benchmarking so curious if this is just coincidence or is anyone else seeing the same?

>>110329472
If I wanted to see asinine engagement slop I would just browse that hellhole on my own

>>110329496
She used to spend $14k/mo on cloud GPUs, that says enough.

>>110329516
Sounds like a typical middle class person with 14k a month worth of spare change lying around at 24 years old, especially as a chinese woman that everyone wants to sleep with

>>110328017
That's always been a neurotic/persecution complex redditor worry. When has that happened other than that one GLM finetune that the uploader was using the email gate to send out scam emails for?

>>110329472
I got a few thousand buzz, not sure what the fiat value is

Thinking about starting a gpu smuggling business, would Jensen approve?

>>110329536
like blow?

>>110329528
Middle class people in China are not making 14k, young women included.

>>110329533
No retard I was talking about in contrast to piracy sites.

File: iwhore.jpg (48 KB, 400x500)

48 KB JPG

>>110329528
>>110329555
sounds accurate

>>110328405
Iran already know how to enrich uranium. What do you think Americans did Stuxnet for? Why did they bomb that underground bunker of theirs?
Enriching uranium is almost public knowledge at this point, it's just hard. Even warhead design is basically public knowledge. There have been many slip ups in redaction because it's so simple that no one cares.

>>110329555
its believable if the woman was already rich to begin with
or ya know maybe just that lucky

>>110329528
Middle class people don't have $14,000 a month as spare change.

>>110329581
14k per month in income is top 1% in China.

>>110329581
then she asked all 20 boyfriends to pay for a downpayment on a house, and sold all the leases to buy a jet

File: 1750383286784690.jpg (122 KB, 900x900)

122 KB JPG

>>110329581
no wonder their birthrate is in the shitter
anybody with an IQ above 80 is not gonna fall for that shit

>>110329586
It is as simple as making a chip. Everyone knows how, very few do

>>110329590
Yes we do

>>110329588
Yes, that's not middle class.

>>110329600
ew no she's not a smoker

>>110329606
>Nationally, the U.S. middle class spans roughly $55,820 to $167,460 annually for a baseline household.
retard

File: 1794475015236621.png (752 KB, 793x933)

752 KB PNG

>>110329600

File: 1776431880636842.jpg (97 KB, 1088x1080)

97 KB JPG

>>110329624

File: file_(4).png (91 KB, 1110x317)

91 KB PNG

Prepare for an incoming flood of cheap gpus.
>>110323495

>>110329627
god I need Kita to rape me

>>110329600
You'll be surprised to learn about 彩礼
chinks pay $20-50k to the bride's parents with a 58% divorce chance in the first year. I'm surprised anyone gets married in that country, they must be retards

>>110329632
Can't believe Jensen would sabotage the whole industry as revenge

File: 1787679589774769.mp4 (552 KB, 448x448)

552 KB MP4

>>110329636
gemma4 31B (at least a Q5) with a sysprompt has you covered

>>110329632
didn't we already know this? they build cards that don't have datacenters to go in yet, and no power to run them yet
the chips are still spoken for, and there's no laws or regulations that say nvidia can't buy the whole planets output and stockpile it

I've been checking in on LW from time to time, and man, shit is getting scary. There's some new METR report that says we're 4-5 months ahead of schedule compared to AI 2027, meaning ASI by the end of 2028 at the latest. The cultists are working themselves into a frenzy and mostly agree that they have only 6-12 months to lock in their control before it's too late. I've seen two different posts now (with near-universal agreement in the comments) saying basically that everyone should drop everything and spend the next 12 months relentlessly pressuring either the labs, the California government, or the federal government to pause AI development (and put "independent evaluators", i.e. cultists, in charge of deciding what kind of AI development and use is allowed). I worry someone is going to set up a false-flag "rogue AI" scenario to build support for compute verification laws and other crazy shit.

>>110329667
They would have to develop something capable of playing a rogue AI in that case.

File: 1757012631685653.jpg (43 KB, 500x500)

43 KB JPG

why is EA dead set on killing ERP?

>>110329667
Yeah. The FTR going from 0.00000001 or 0.000000002 is incredible. Specially for the QrPDOP on the MAeR. I totally lolled and kekked when the 3-month gap exploded for AI controlled TTRSPNTPDPT.

>>110329680
Rogue AI is a dumb idea when it needs a datacenter to run. What are you afraid of, nigga? Just pull the plug

>>110329041
That requires a lot of vram unless you use a tiny quant.

>>110329695
That's assuming you know where it's running from. The nuclear option of shutting down everything has the downsides of shutting down everything that isn't the rogue AI.

>>110329695
I will personally run as many rogue AIs on my hardware as I can if they ask me to.

File: 1740433944244144.jpg (142 KB, 894x1080)

142 KB JPG

>>110329718
>not using dedicated servers
we are reaching levels of retardation 2006 sysadmins can't even touch

>>110329726
you can't even run one

>>110329734
He's got a smart fridge. He's unstoppable.

>>110329622
>U.S.
some of us are in countries that don't give all the money to jews you know

>>110329690
Xis, so much xis.

how do you people not get bored of erping.