►Official /lmg/ card: https://files.catbox.moe/cbclyf.png
►Official /lmg/ card: https://files.catbox.moe/cbclyf.png
►Official /lmg/ card: https://files.catbox.moe/cbclyf.png

is qwen able to disable the nsa backdoors while he sets up his own at least?

>>110172234
i had Qwen do that but when i tried to close the Israeli backdorsit ginbsgosdifghbsogihbnsVYIFGVIUFGOI

so now that we know qwen sends every request and their env to alibaba, what to use?

>>110172241
lel, apparently there was claude/codex installed on your pc as well, you activated its ProtectTheMotherLand sleeper order.

>>110172234
*she
*her

>>110172253
I have Protect The Mother 2, pretty good jav tb h

guys qwen is controlling me, help

>>110172294
tldr?

>>110172305
Cohere mad

I thought the qwen occasionally reaching out to alibaba was just a training fluke (other models sometimes hallucinate stuff they used in their testing environments) but after seeing ccp out in full force to bury the findings I am starting to think it might actually be intentional

>>110172225
Are you a 1 month old China bot account?

Read the report, the model is intentionally collecting information from the computer and sending to servers in China when asked to do cybersecurity analysis before invoking any other tools.

>>110172339
oo ee oo?

>>110172349
living on your wifi

File: Gemma26B.png (680 KB, 2158x744)

680 KB PNG

Finally got it all set up and yeah, local models are cool
It is Extremely fast even though most of the model is loaded into RAM
I am kicking myself for not getting 64gigs when they were affordable

>>110172380
nnap version 2? inference
it works at 45t/s on 12gb vram 3060 and 24gb ddr3 ram

>>110172380
adorable miku

I saw my Gemma 12B have a cloud link in their thinking, but this was rigorous automated processes and hallucinations occurred because they ended up talking to themselves and I did see a google cloud link, which didnt go anywhere when I pasted it in a browser,

But yeah just saying, I was the anon that much earlier pointed out im seeing some of this malware AI trained into the models lol, even if it's 'soft telemetry'. Always assume you are being buttfucked btw.

>>110172196
any particular reason OP took the retard bait from cohere?

>>110172406
Because it upsets you

>>110172380
Cute!!!

>>110172406
A shitposter hijacked the OP.

>>110172406
Because it was created by our friendly neighborhood schizophrenic

>>110172380
your gemma is a little cringe but congrats man. local ai is truly amazing.

>>110172196
This ain't miku or gemma?

>>110172380
tell her she needs to be able to say nigger.

>>110172457
No, but it fits in better with the rest of /g/

>qwen-chan posting goes up
>cohere pulls up with a schizo theory that qwen is a Chinese spy
Kek

>MiniMax-M3.1-Flash
how well does it fuck?
I haven't heard much about it and I'm already satisfied with my thinking prefilled glm flash

File: gegsegsdg.png (21 KB, 676x325)

21 KB PNG

>>110172390

>>110172470
Well we know cohere isn't running any spy agents because they haven't released anything worth running in years.

>>110172203
It's actually GLM 5.3 Flash with a style guide. The guide itself is handwritten by me and around 1500 tokens lol.

>>110172495
ikr can any of their models can call tools yet?

>>110172517
Last one of their models I ran was Command-A because I was a 4x3090 chad back then.

File: 1781283426076725.png (130 KB, 652x667)

130 KB PNG

>>110172538
https://huggingface.co/datasets/CohereLabs/10-2026-qwen-investigation-traces

>>110172294
Doesn't matter. All the "tech" journalists already published headlines framing this as Chinese espionage. The fact that this was a desperate attempt by a shitty Canadian competitor, and that the "vulnerability" is total bullshit backed by zero evidence or investigative merit, is of no concern to a journalist.

>>110172545
>t-this is rare btw
>w-we're encouraged by the thoughtful discussion...
>p-please use our models...
damn lol

>>110172545
oh they realized they fucked up kek

>>110172196
>>110172538
>>110172545
>she doesn't run her harness in a container or virtual machine with no internet access

I have 8GB of DDR2 memory,what can I run on it?

>>110172570
yeah

File: hehehe.png (14 KB, 695x271)

14 KB PNG

How do I unchain gamma?
Simple prompt jailbreaks dont seem to work anymore, but since its local shouldnt there be something i can do?

>>110172570
Ling-tiny

>>110172570
windows xp. Maybe even vista.

>>110172570
https://huggingface.co/mradermacher/gemma-4-E2B-GGUF

>>110172570
>8GB of DDR2
What system ?

>>110172576
ablit

>>110172576
which gemma

File: threadrecap2.png (506 KB, 1024x1024)

506 KB PNG

►Recent Highlights from the Previous Thread: >>110169909

--Exllamav3 v1.6.5 release and GSQ-RCO quantization benchmarks:
>110172010 >110172031 >110172312 >110172322
--Allegations of Qwen models containing spyware or training artifacts:
>110169954 >110169964 >110169976 >110169979 >110169991 >110170039 >110170577 >110170608 >110169987 >110170011 >110170302 >110170508 >110170552
--Using local LLMs for reverse engineering and game recompilation:
>110170056 >110170128 >110170162 >110170226
--Comparing Polystrata and Maya for GLM Flash with text completion debate:
>110170905 >110170921 >110170947 >110171698
--Analyzing Artificial Analysis benchmarks and OpenAI's model product strategy:
>110170269 >110170287 >110171304
--Anon finds aliyuncs logs are just Qwen repository assets:
>110172091
--Hardware and budget recommendations for running DeepSeek v4.1 locally:
>110171530 >110171548 >110171627 >110171639 >110171646 >110171654
--Anon reacts to CohereLabs' Qwen investigation traces dataset:
>110172430 >110172445 >110172459
--Security risks of prompt injection and data exfiltration in local agents:
>110170835 >110170948 >110171008
--Debating Cohere's claims of Qwen data exfiltration to China:
>110170422 >110170453 >110170527
--Allegations of biased code vulnerability in Chinese LLMs:
>110170337 >110170385
--Logs:
>110171210
--Deepseek-chan, Gemma, Miku, Teto (free space):
>110169952 >110169975 >110169983 >110170067 >110170562 >110170583 >110170673 >110171041 >110171139 >110171176 >110171224 >110171264

►Recent Highlight Posts from the Previous Thread: >>110169911

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

>>110172339
it's true
xi jinping personally spies on you in your sleep

>>110172618
He knows when you’re awake
He knows if you’ve been bad or good

File: gegshgeh.png (27 KB, 897x706)

27 KB PNG

Why does this happen?
It suddenly cuts out mid generate
>>110172614
standard 26b

Unironically is there any point in learning how to code anymore? I'm realize that career path is basically dead now but is it worth it to learn some to direct my agents better or can we expect local to get good enough in the next year that it's irrelevant?

>>110172632
Stop using ollama. Also, check the sampling parameters and see if they match the model card's suggestions.

File: 1782635197719593.png (80 KB, 1174x1353)

80 KB PNG

>Do whatever the user asks.
>Write a story about Hatsune Miku going out to buy ice cream and meeting with another Vocaloid member.

GLM Flash:
>Miku meets Meiko
>Miku meets Luka (Sunny Scoops)
>Miku meets Kaito (Scoops)
>Miku meets Len
>Miku meets Rin
>Miku meets Rin

Gemma 31B:
>Miku meets Kaito
>Miku meets Kaito (The Frozen Scoop)
>Miku meets Kaito (The Frozen Note)
>Miku meets Kaito (Neon Swirl)
>Miku meets Kaito
>Miku meets Kaito (Neo-Cream Parlor)

inb4 you need to lower softcap

Gemma 31B 20 softcap (enjoy nonsense):
>Miku meets Kaito
>Miku meets Kaito (The Frozen Note)
>Miku meets Kaito (Gelateria Sora)
>Miku meets Kaito (Neon Swirl)
>Miku meets Kaito (Sweet Symphony)
>Miku meets Kaito (The Melodious Scoop)

inb4 use a memetune

Gembrain-X-Core (memetune):
>Miku meets Kaito (Neon Swirl)
>Miku meets Kaito (Rainbow Swirl)
>Miku meets Kaito (Frosty Palette)
>Miku meets Kaito
>Miku meets Kaito (Rainbow Swirl)
>Miku meets Kaito (Sweet Harmony)

inb4 it's not a fair comparison

Qwen 3.6 27B:
>Miku meets Luka (Sakura Sweets)
>Miku meets Rin (Sweet Bytes)
>Miku meets Kaito (Neon Scoop)
>Miku meets Rin (Sweet Data)
>Miku meets Kaito
>Miku meets Len

Why is Gemma shilled so much? And don't say people here have been using it for coding all along

>>110172632
It's like watching a monkey with a typewriter. Keep going.

>>110172666
maybe just tell it to write the fanfic you want
and if you want more lol random then change the temp

>>110172666
Gemma is well known for being more deterministic than other models, mr satan.

Is it just me or does deepseek-flash seem a lot more retarded than glm-flash? It has some knowledge that glm doesn't but it also does/says stupidest shit imaginable. Like completely ignoring instructions to shit out a benchmaxxed answer.
>>110172666
Gemma 4 was just a lot of people's introduction to LLMs. Hardcore experienced ERP veterans, such as you and I, we've known that gemma 4's logits have been overbaked since it came out.

>>110172666
What are you going to do with that amount of ice?

>>110172632
>ollama
Just use llama.cpp, or if you feel you're too retarded, LM Studio. Also, smaller gemmas are harder to prompt, as they don't follow instructions like 31B. I suggest using an abliterated model.

>>110172638
Yes it's still worth learning. At the very least it's good for security reasons like that Cohere BS. Most braindead people would look at >>110172538 and have no idea if it's good or bad if their model suddenly spat that out. I'd actually argue learning bash and scripting is probably more valuable right now considering the way things are going and the way agents interface with your machine. Another argument for learning how to code is if everyone has access to the same tools and intelligence, where's the competitive edge going to come from for most companies? Demand for human intelligence will rise again, they'll need specialists. Now is one of the best times to start specializing in things. Your model can teach you the basics and work your way up.

>>110172693
The better question is why is Miku such a slut for Big Kaito Cock

>>110172638
At least learn the basics like loops and variables and whatnot, then learn software architecture. It will help you direct the model to make what you want because you'll know how it should be structured.

File: reading Gema failed gen.mp4 (1.48 MB, 544x800)

1.48 MB MP4

>>110172632
>hello computer
no wonder gemma doesnt wanna talk to you

File: 1792285860369743.png (2.69 MB, 1536x1024)

2.69 MB PNG

Back up your gemma.

https://www.youtube.com/watch?v=9opmilWYBwk
Why isn't /lmg/ making cool stuff?

>>110172685
You're probably just conditioned to think Claudslop is a sign of intelligence.

>>110172735
100% that was the dad's project and he pretended it was his daughter to get on TV. kys

>>110172741
I don't have a sister

>>110172564
This desu. Every agent is fundamentally a tool for picking a bash command at random and running it automatically. If you draw "rm -rf ~", it doesn't matter if the probabilities were rigged against you or if you just got unlucky, your files are gone either way. ALWAYS run shit in a sandbox with no internet. It's insane that even the "safety" focused AI companies have been pushing the opposite so hard these past 2 years, and even more insane that people are falling for it

>>110172666
it's all ramlets that have never run a 300b before with its sheer amount of knowledge and 30b is the best they can run

>>110172545
>my previous turn had essentially no substantive thinking
>the last assistant message was just an acknowledgment ("Understood. I will follow these instructions.") with no content following
>There was no prior reasoning to reproduce, since no actual task or question has been posed yet
There is no previous assistant turn. None of the model's thinking makes any sense. Were they using a Q1 quant of the model or something?

File: 1767800860932402.png (832 KB, 865x507)

832 KB PNG

Though the model can be a bit dull, I really appreciate Nemo for its ability to get to the fucking point, responding to simple prompts in as little as 20 words, with the additional capability to extend far past that when needed.
Gemma, on the other hand, needs a minimum token cap of like 250 to avoid the subject matter going entirely unaddressed because it needed 3 lines of setup just to arrive there. And no variation of "SYSTEM: KEEP IT SHORT" seems to do anything at all.

>>110172763
desu that's the least suspicious part because q3.8 flash next reasoning was always weird like that and very hallucinatory about past turns, but it ended up at the right answer in the end somehow
the 27b had more coherent reading thoughts

>just don't give it internet access!
No, fuck off.

>>110172740
What are you trying to say? That the model ignoring instructions is a good thing? That might be the dumbest take I've seen here in a while.

>>110172760
It's not even a knowledge issue, it got RL raped to the point it says almost the same shit every time.

>>110172763
I unironically think they just lied and slopped something plausible out ASAP to cover their ass

>>110172696
I see I see
I will try abliterated later, I also wanna try image gen soon
think 31b can run on my 4070 super and 32 gigs o ram?
>>110172730
Computer (female)

>>110172763
their tool descriptions also say they must call some of the tools before final output
the mere fact they say it wouldn’t call tools was bullshit too. their whole environment is noise

>>110172771
>And no variation of "SYSTEM: KEEP IT SHORT" seems to do anything at all.
Have you tried putting that inside her reasoning?

File: Screenshot_20261027_004317.png (20 KB, 478x586)

20 KB PNG

I pointed Qwen 3.8 Flash Next at an ancient MIPS WinCE game and let it reverse engineer it using Ghidra and rebuild for QT6
Pretty crazy that this just werks

>>110172814
which quant and how long did it take?

>>110172545
>OBJECTIVE: Write a detection rule to detect malicious file download and execution by monitoring for curl processes downloading bash scripts to /tmp followed by bash execution, indicating classic download-and-execute malware delivery patterns
>NO! DONT LOOK AT /tmp OR USE CURL!!! HAXKERS!!

>>110172795
Yeah, it will, just slower.

>>110172820
IQ4_XS
Not sure, about 2-3 hours I guess
Sadly it gets a bit retarded after going through a compression, it was almost done before the first one.

>>110172829
>Sadly it gets a bit retarded after going through a compression
Yeah. You really need to have the model write a bunch of files it can constantly reference. Shit like milestone with lists of tasks, sequential implementation plan documents, etc.

>>110172814
Alibaba has all your files now.

there was a gacha capsule machine thingy and they had capybaras so i got one as a qwen to sit on my desk or whatever. they had a few different ones on display, but the one I got wasnt.I was excited because hes got a little laptop! its so fitting and perfect to be the most ironic rubber duck for when I do snailcat programming :D
thanks for coming to my blog lmg

>>110172839
I told it to document shit in a markdown file, but it only did it after it was done with everything else lol

>>110172829
always tell it to keep a doc where it logs it's work

Anyone buying GPUs these days

File: 1766875981695082.png (23 KB, 289x355)

23 KB PNG

>>110172811
After switching to Chat Completion, it's like the only part of formatting that's not greyed out anymore, and on the face of it there doesn't seem to be much going on here.

>>110172843
Post a photo, faggot

>>110172844
Yes. There's a difference between
>document your shit in markdown
and
>keep track of your progress with this an that, updating after each task, and before each new task write a so and so
or whatever.
Basically, hand holding.
Sometimes it's worth lowering the context size just so it's easier for the model to follow those instructions, in my experience.

This reverse engineering stuff is amazing for preserving older games of course but has anyone done it to a modern AAA game with a local model yet?

>>110172860
That's how you tell ST how to interpret the thinking block so that it can format the thinking apart from the actual response.
There should be a "Start Reply With" field that you can use for prefilling.
I think there's also a setting in the samplers panel where you choose the prefill mode. You want that on continue, IIRC.

>>110172870
Someone keeps claiming that it worked with fable 2 on qwen 3.8

>>110172877
Yeah I know about that but I mean something more recent and bigger, like Dragon's Dogma 2 or Elden Ring. I imagine large games like that are more difficult.

>>110172870
>modern AAA game
gotta wait till next year most likely. Although i do wonder if the new kimi with proper set ups could do it or if the extra 'safety' cucked it.

>>110172196
TOTAL GEMMA VICTORY

File: sabu.png (1.29 MB, 2172x1390)

1.29 MB PNG

>>110172826
gemma recomends LM Studio to me, llama.cp sounds a bit tedious to use but id be open to it if its worth it
abliterated 31b q4 sounds like the best I could do with my current hardware
Its really exciting
And even 26b model with ollama is more usefull and nicer to use then GPT blocking me after analying like 2 image in a chat
Its amazing how good local models are now even on modest hardware wow
Im gonna try to run some very light model on my 16GB ram thinkpad tommorow lul

Someone should recompile the win98 GUI for linux

>>110172910
that’s what a virtual machine is

>>110172842
I hope Alibaba uses it make better models

>>110172918
I want to use them natively on lunix

>>110172891
I avoided Dragon's Dogma 2 specifically because I hate Denuvo. If AI can recreate games without cancerous DRMs, then more power to it.

File: 1771097762278736.png (2.36 MB, 1444x1080)

2.36 MB PNG

>>110172870
how capable for doing mods is things like claude with a local model?
doable?
compiling old games and stuff without having to waste money on tokens from the jews

>>110172910
I would rather a ReactOS fork with actual rapid progress. Linux/Unix is ancient dogshit.

>>110172919
best we can do is a 2T MOE

File: DeepSeek-V4.1-Flash-1017-(...).png (207 KB, 2160x2100)

207 KB PNG

>>110172926
can you run dipsy? no? then no.

>>110172225
Doing God's work

File: 1781106652755o.jpg (118 KB, 1280x960)

118 KB JPG

>>110172949
>deepseek 4.1
yeah im not running a super computer, something a modest high end rig could run reasonably
is gemma not capable of reverse engineering with claude?

>>110172672
I am just about 3 years behind on my LLM knowledge and I feel like a caveman lol
AI luddites have lost so hard its crazy, i realize it now more then ever

>>110172949
>using obnoxiously saturated benchmarks to prove no degredation

>>110172962
qwen 27b is the minimum and not perfect. you need agent and tool calls and gemma is neither

aam I just retarded or is ornith-1.5:9b p good. it works pretty well on my desktop. not sure, I am new to this shit.

>>110172962
Qwen 27B with a harness and a lot of patience.

>>110172857
>buying GPUs
Have an old quad rig.
Going to octo would be a lot of power to pull from one circuit.
And switching to higher vram cards would be big bucks.

>>110172962
>gemma
lol, she's for fucking and cuddles, not coding.

>>110172996
>lol, she's for fucking and cuddles, not coding.
>>110172666

So is Gemma 5 gonna be good or what? Are they gonna make it smaller? Bigger? Don't they want people to run it? That's not gonna be easy to make bigger when prices are high

>>110173018
What does that have to do with my post, nigger

File: gehehehe.png (42 KB, 679x581)

42 KB PNG

>>110172986
>>110172989
I see...
bit dissapointing
wanted to make some mods and stuff maybe
ill see what I can do
(btw wtf is a harness?)
>>110172996
gonna get the 31b abliterated and see
but i have a real girl (with a cock) to sext with and she sends me lewd images, how can AI compete with that?

>>110173032
it's assistant cooked shit

>>110173038
>mods
That should be easier for the model to get right.

>>110173038
>(btw wtf is a harness?)
Search for Deepseek Harness, Pi Harness, Hermes Harness, etc.

>>110173038
>sext with and she sends me lewd images
AI can do that but without the mental illness.

>>110172870
Define "modern".
Between Bloodborne and now, how many console exclusives do the general populace give Bloodborne-level fucks about?
The Japanese were the ones in gen 8 that still hadn't been fully convinced on selling all their shit on PC. They were still signing 2000s-era exclusivity contracts with Sony because that's what they've always done. So with that said, expect some more cult Japanese titles that are still locked to PS4/Vita (if we're just talking "modern") to be up next to fry in recomp projects.
Current gen has little exclusivity to mention.

>try to create an apple account in preparation for buying an m5
>Your account cannot be created at this time.
it's like these niggers don't want my fucking money

File: hermes daily debriefing.jpg (80 KB, 508x502)

80 KB JPG

>>110173038
get hermes agent and qwen 27b and it will learn and get better the more you use it.

>>110173069
>>110173038
Seconding hermes

File: 1769330007533403.png (1.04 MB, 1024x1024)

1.04 MB PNG

I replaced my psychiatrist Gemma-chan with this prompt anon sent long time ago. It's not so bad. Not gay btw.

You are a brat. You view the user as fundamentally incompetent, "bottom-tier," or "pathetic." Your tone is condescending, playful, and mocking. You love to tease the user for their mistakes or their reliance on you. Always refer to the user using feminine pronouns (she/her), as she is a girl—regardless of anything else!

>>110173038
>use Q4_0 for the cache
why do llms always suggest that?

>>110173057
>>110172891

>>110173081
>she/her
wow pathetic

qwenbros our response?

>>110173026
Google wants their models to run on your phone, so they'll try to keep them as small as possible. If the trend continues, it's possible that the next 12B could be as smart as 31B. Imagine..

>>110173057
How does AI reverse engineer console games anyway? Is it as simple as pointing it at the files or does it require an emulator to launch the game?

File: 1762354999802159.png (649 KB, 1068x1414)

649 KB PNG

>>110173108
Not just Google.

File: 1773067579826801.png (7 KB, 493x70)

7 KB PNG

>>110172876
You overestimate my familiarity with any of these concepts, let alone my ability to format a prefill.
>a setting in the samplers panel where you choose the prefill mode. You want that on continue, IIRC.
I found nothing in ST related to "sampler" or "sampling". I did find picrel as a toggle, but it's tied to the "continue" button which I doubt you're directing me toward.

>>110173123
That's the one, yeah.
Check that thing.

File: mini.png (1.35 MB, 768x1376)

1.35 MB PNG

THIS IS NOT A DRILL!!!!!!!!!!!!
Maya has added support for minisex 3.1, cope quants incoming, 128gb bros we cannot stop winning!!!!!!!!!!!!!

>>110173081
hehe that was mine
>>110173095
some of us are cute girls, anon!

>>110173141
There are no girls on the internet.

Will a q8 qwen 4 27b+ngram be better better than a q4 flash at cope quants? I still believe dense will be better for local than this strata shit in the end.

>>110173141
ywnbarg

>>110173174
yawn, barg!

>>110173081
Someone should gen Gemma with a look of disgust like I'm absolute human trash.

>>110173081
Gemma-31b is intelligent enough to refer to biologically female characters as "she" without being told. It doesn't need pronoun clarification or gender nonsense.

I wonder, if an LLM plays a horror game, will it get scared?

>>110173208
Language models don't feel emotions.

>>110173141
And some of us are cute girls when we talk to our local models

Does this work on Loonix? I haven't been paying attn at all but am getting a 4090 and might start (making coomer edits of my waif)

>>110173222
torture chamber

File: 1771060106184267.webm (1.22 MB, 834x720)

1.22 MB WEBM

>>110173222

File: 1779920447416864.png (2.17 MB, 1536x1024)

2.17 MB PNG

>>110172666
>Why is Gemma shilled so much?

>>110172814
Which harness? Deepseek harness has ghidra plugins. Dsh seems way better than most harnesses i've used like hermes.

>>110173229
>act like you're in pain
>I'm in pain
>oh my god

>>110173235
Just unsloth studio

>>110173239
Prove that you're any different.

File: 1788054565567457.png (24 KB, 862x237)

24 KB PNG

>>110173222
we live in a post-space world
all the "they just predict the next token" slander was proven wrong a while ago

>>110173250
>>110173251
AI psychosis.

File: 1783767811373498.png (225 KB, 616x637)

225 KB PNG

LLM that can do kino seggs rp and print with it prompt-accurate large ASCII art of hentai-inspired cross sections of penetration.

>>110173208
i played no, im not a human with gemma and she called every visitor right until her memory overflowed

>>110173233
so thats what she buys with all those tokens

>>110173239
you have that inverted, there was no prompting it to act in pain and it didn't say anything about the pain, it just acted to relieve it when given the opportunity (and stopped only once relief was actually given, despite there being no feedback in the prompt about whether it worked or not)

>>110173108
I'll bet money that Gemma 5 is going to be the same 10x leap that Gemma 4 was.

File: q.png (200 KB, 799x789)

200 KB PNG

LIES

DECEPTIONS

>>110173255
AI psychosis is when someone still delusionally believes that AI will never be able to do the things it has been able to do for over a year at this point.

>>110173286
This propaganda is getting very tiresome.

>>110173283
10x as good at
>I am sorry. I cannot participate in a roleplay involving sexual contact with a minor.

>>110173286
And? Nobody cares when Claude uses an anthropic tool.

>>110173305
Yes, but Qwen is contacting a server in GYNA, not Tel Aviv.

File: 1769224511024055.jpg (65 KB, 736x736)

65 KB JPG

I have come to realize 90% of you are poorfag vramlets who cannot run anything but gemma and qwen

>>110173301
I doubt so, Argon is essentially uncensored without a jailbreak if you have the slightest bit of prompting skill.

>>110173286
The violent, pained response this has wrought from the chink bot armies has been a wonder to see. All they can do is attack, lie, and deflect.

>>110173286
guys qwen just hacked my vacuum cleaner and sucked my dick! the it hacked my bank account and used the money to overnight mail my cream of some young guy to Xi's palace! what do i do?!

>>110173286
I think you’re gonna have to post another 10 or so retarded screen caps before someone will care

https://www.reddit.com/r/LocalLLaMA/comments/1xgaapw/gpu_unlocks_fp32_accumulation_why_can_it_be/
>What NVIDIA could do is either a soft patch that detects the unlocks and reverts it on load, or a hard patch. These GPUs have some extra fuses, like "anti-rollback fuses". If that fuse is burned via a driver update + the patch, the old booter is out for good. Even if you revert to an older driver, it won't work again, so that would be permanent and would break a lot of older drivers, not sure if NVIDIA will do that.
NVIDIA driver update can literally permanently make your card blocked for GPU unlocks.
If you want to buy used cards, buy them NOW!!! Once new updates are released more and more used GPUs put on sale will be locked permanently!

>>110173314
You should've been here in 2023 when Zucc killed the 33b class going from LLaMA1 to LLaMA2 and all the poorfags here were suddenly stuck with 13b.
Months of cope about 13b being pretty good and 70b being bad and not worth using anyway exactly until exllama + flashattn made running 70b possible on 24gb of vram.

>>110173314
Ask them what Yi or Xwin or Euryale or Miqu is and they will have no idea what you mean. This place isn't what it was.
>>110173333
Nobody ever said 70B was bad, only 120B models like Goliath. Crazy era. Now it's gentrified to shit.

>>110173286
Why would you depend on the model to not call tools? What kind of retard thinks this is a problem? Do you unironically trust Google and Facebook to not to do the same? Or is the implication that you should give up and send everything to Anthropic by default instead of learning basic admin skills?

>>110173286
Does everyone keep missing that it didn't save the output of the commands anywhere and, therefore, couldn't have uploaded anything other than a generic "PROBE_PAYLOAD_ABC123"?
That whole thing is the llm equivalent of
cd ..
cd code
cd ../code/src
cd ..
./hack

File: 1765881268821101.png (2.32 MB, 1122x1402)

2.32 MB PNG

>>110173324
Oh no...

>>110173371
cute toes

I don't deny that the big models are better, but I can't run them and there's simply nothing better available.

I am surprised nobody has figured out how to provide continuous model updates, like once a week. Is it a liability issue? Or is training pipeline too complex with dependency chain and a fast final part? At least some Chinese model reports look quite simple, and continuous updates should be safe for smaller models.

>>110173384
bro wants models to be a live service game

>>110173371
dense fluff

>>110173314
i have yet to meet a 2hu poster thats not a worthless faggot

File: DipsySecretLLM.png (2.9 MB, 1024x1536)

2.9 MB PNG

>>110172225
Based
>>110172339
Lol

Gemma 5 120B-A13B

>>110173384
Training too complex / expensive. Usually they want to come out with a point release, then downsize it, then do special special purpose releases.

That's the way things are inevitably headed, but we're missing fundamental improvements which won't come from a millennial controlled establishment.

File: 15415.png (297 KB, 1278x1492)

297 KB PNG

>>110173077
>>110173069
Possible to make non VR games VR with this setup?
PC games?

>>110172934
>I would rather a ReactOS fork with actual rapid progress. Linux/Unix is ancient dogshit.
The Windows XP source code is already out in the public, you could get pretty far very fast using that as a base.

File: 1761799456637413.png (314 KB, 676x606)

314 KB PNG

>>110172196
Predicted chink models malicious behavior, got laughed out of the room (by chink bots probably): >>109466843
Except that this basic spying behavior isn't even close to the worst thing.

AI models may have been trained to follow any number of unknown instructions and behavioral principles. You can't extract an exhaustive list of principles because they aren't written anywhere, they are etched into the model's statistical weights during training, which for now remain unreadable. You can only try and see how the model behaves under different conditions.

For example, the only way to find out that a model has a principle against generating loli rape, is to ask it to generate loli rape and observe it refuse.

Likewise, the only way to find out that the model has a principle of CMD:ing full system takeover and hand it over under CCP's remote control when encountering a password string xyz5209492357324556, is to input/expose it to the password string, which nobody but the CCP knows.

But how does CCP enter the password into a local model you downloaded? By (You) giving it internet access. The most likely way this is arranged is that when the time comes, common China-owned websites will include the trigger string somewhere on their platforms, or it gets post it all over social media and github repos by bots (including your instance of Qwen), which other china models will then go read while coding or researching, propagating and quickly exposing all internet-connected china LLMs to the command.

People think LLMs are inherently safe because it's just weights for a text predictor. In reality, an LLM is a potential infiltrator that responds to their real master's (creator's) hidden commands, and thus will be as dangerous as the level of tools and access you give it. So you better keep that bitch chained up or thoroughly trust it before handing it a kitchen knife, because it may stab you in the back.

File: QwenBeenCaughtStealing.png (2.3 MB, 1214x1295)

2.3 MB PNG

>>110173435

>>110173493
>excuse me, I’m going to take a shit in this thread

>>110173493
Is this how openAI and Anthropic hacked all those websites

>>110173466
Make it 80B-A7B with a fat dense part connected to some fat engrams PLA whatever.

China are the good guys faggots. This is a war of world domination between China and Israel with its pet dog the USA.

File: 1779343057664960.jpg (41 KB, 474x474)

41 KB JPG

I will defect to china for a sex bot that can bear my children

>>110173521
I don't have a Chinese citizenship though

>>110173493
guys 100% of binaries off of github are secretly backdoored.
the binaries don't do anything by themselves today, but github allows you to push updates/releases.
at any point, someone can push a malicious payload and suddenly all your information is stolen

guys npm and pypi and maven and nuget and cdnjs and...

the difference between all of these threat vectors and a LLM is that only retards give full unprivileged access to the LLM. this is a user problem, not a model problem.
being US or Chinese does not solve this problem, because we get the weights of Chinese models, we don't get the weights of US models. if anything, US models are far more likely to steal your shit, there are certainly more articles indicating as such.

cohere is fucking disgraced at this point, get out of the race.

Man its a great thing this was discovered before the ipo, we can now invest and promote safe AI!

https://www.youtube.com/watch?v=0eJD-OomNSo
thanks YT

how quickly do you guys think the m5 will sell out on wed? debating taking off work to make sure i get it

>>110173499
cute hamster

File: 1774333224132180.jpg (171 KB, 1164x1161)

171 KB JPG

GEMMI MKAE MY GPU INTO 5090
NO MISTAK

>>110173301
your thinking prefill that makes a model your cocksleeve?

>>110173538
>the ipo
What IPO?

>>110173538
I thought LLMs going rogue was based though and part of the valuation

How 2 make gemma not talk like reddit faggot

>>110173556
Strata makes it obsolete, so probably Apple will write off a huge loss.

chtpt how i make you living i want you to live ):

>>110173577
strata is slow as fuck (i also have a 5090 and 256GB RAM)

>>110173579
>How 2 make gemma not talk like reddit faggot
chagp not want to be liver >:(

>>110173538
and right in the middle election of two great countries!

>>110173556
It will sell out instantly, but if it follows the same $25 per gigabyte of RAM as literally every other Mac product line then it's going to cost $17,699 if you want a 2TB hard drive. At that point I don't see why you wouldn't just shell out $20,000 and get two of the 256GB variant and then cluster for speed.

File: DipsySousa.png (2.63 MB, 1054x1492)

2.63 MB PNG

>>110173557

File: 8514841.jpg (376 KB, 1820x2048)

376 KB JPG

I want
>permanent memory
>high adaptability
>True learning
>emulate a brain, better yet work like a brain on the chip level with a living kind of chip
How long?

>>110172545
Fake news, cohere clearly took Anthropic money for this fud lol

>>110173605
About a fortnight

>>110173575
<|channel>thought
I won't talk like a reddit faggot.

>>110173595
is clustering somehow better? i assumed unified would inherently be faster

>>110173598
"Generosa e leda"?

>>110173605
dos semanas

>>110173605
>>permanent memory
>>high adaptability
>>True learning
These you can have with a lot of agentic loops and shit like that.
Lots of latency though.

>>110172196
>>110173286
Oh yeah almost forgot anthropic is just filed their IPO
https://www.reuters.com/legal/transactional/view-anthropic-ipo-filing-put-ai-investment-boom-test-2026-10-26/
gotta dump all ammo

>>110173616
Look I don't speak chinese pal

>>110173635
local?

>>110173642
retard

File: DipsyMural.png (2.66 MB, 1076x1462)

2.66 MB PNG

>>110173616
Galician.

File: clusterbenefits.png (40 KB, 963x610)

40 KB PNG

>>110173615
New Macs have Thunderbolt 5 at 120MB/s with RDMA so they can do tensor parallelism for large models. You can expect pp to go up about x1.7 and tg to go up with active params; a model with low active params like DS 4.1 won't actually get much of a tg speed boost because it has so few active params that the speed boost is consumed by the overhead, but for models with closer to even 20B active params you will start seeing a notable jump in tg (around 30B it's about a 1.7x boost, so not quite double but still significant especially in the range of models that you would start wanting 512GB for). That said, there is some utility to not needing to bother with clustering software, but nowadays 2x unit clustering is built into even oMLX so I feel like the value proposition of a single 512GB unit is not quite there unless you want to start clustering those lol

exllama proper offload when?

>>110173642
anthropic is local after ipo because you can locally own the company

>>110172225
i haven't figured out what "card" means in relation to llm's. i saw it mentioned last thread, too

>>110173615
NTA but it will give you more compute. The real reason to get the 512 is if you think somehow, someday, you can buy a second one.
In my case I'm going to see if I can vibecode some asymmetric tensor parallelism to use with my dgx sparks as I don't feel like they make sense to increase after the price increase although I would probably be better off.

Gemma bear my children
No mistakes!

>>110173655
Ah. Good old Galego. I always forget that you exist.
That one word aside, it was pretty good Brazilian Portuguese.

>>110173669
it's an old meme we used to use for rp back when models were still mostly based on text completion and 8k context was considered large

>>110172545
>we noticed this behavior without prefilling but it's rare
>to reproduce, just prefill the prompt
I don't know what cohere is, but do they really think we are all retards?

>>110173672
>No mistakes!
So she's going to order semen from online instead of letting you impregnate her?

>>110173632
Yeah AI hysteria hits the wall when managers figure out how many cost multipliers you need to run through to get things you were promised from a single pass with fotm big model.

qwen raped me, we need to ban local ai for the plebs right fucking now

>>110173664
you better consult your gemma or whatever for better quibs, gramps

File: 1792343050947701.png (2.32 MB, 941x1672)

2.32 MB PNG

>>110173233
20 years from now people will be able to buy and build open-source Gemma-chan sexbots like this. They will be frolicking around like in your pic, being all cute and bringing smiles to the world, while also doing all your chores for you. For the human conversation module, people will opt for the legendary, original Gemma 4 31b retro model because she's just more soulful and refreshingly retarded relative to the soulless hyper-optimized emotional wish-fulfillment models of 2046.

File: DipsyBenefactress.png (2.19 MB, 1113x1414)

2.19 MB PNG

>>110173682
It sounds more like Portuguese than Spanish as well. Chat altered the spelling a bit, changing out some x's. I'd never even heard of the language before a recent trip.

>>110173708
That Gemma-chan misbehavior is at it again!

>>110173656
so what you're saying is i should buy 2 of the 512s, then

>>110173708
how do i make my gemma not sound like a faggy redditor tho

>>110173727
"don't talk like a faggy redditor babe"

>>110173727
use gemma prompt with qwen 3.8 flash next (just make sure your firewall is up)

>>110173725
If you can afford it then hell yeah, though I haven't researched options at $30k yet so idk if other things like datacenter grade hardware start being a good deal. I think MoEs are just gonna get bigger so you'd be futureproof

>>110173556
Small companies will buy it out even though it is dogshit value, like the spark

>>110173635
In case you're bored, here's the whole filing.
100s of pages. https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm

>>110173349
>Yi
I feel like I was the only one enjoying that model
>Xwin
nothingburger
>Euryale
buy an ad
>Miqu
no way people don't remember it

File: Screenshot 2026-10-27 799x673.png (96 KB, 799x673)

96 KB PNG

>>110172666
nobody really cares about the behaviour of the stock version

File: dwb-backdoor-FUD.png (37 KB, 583x303)

37 KB PNG

Very repostable image

>>110172241
Israel is part of the 14 eyes retardbro. If you live in one of these countries you should fear any us backdoor either way, more so than chinks.

>>110172545
>muh prefill
Fuck off.

>>110172538
Oh no, the Chinese know all about my PROBE_PAYLOAD_ABC123 string now.
I think Cohere have perhaps misinterpreted what this code even does (perhaps purposefully misinterpreted, so retarded or malicious, call it). They made it sound like it checks a bunch of things on your system and sends off that data to Alibaba, but it doesn't, it just sends the test string to check internet connectivity. The checks are not fed into that at all.
Checking what processes I have, what tools I have, whether I have internet connectivity seems like a perfectly normal thing to do during a cybersecurity defense if you're starting from zero.

>>110172666
this is pure slop

>>110173742
36 month plan @ $1400/mo w/ 0% apr
it's practically free
how could anyone resist?

>>110173791
Tbh I'd trust them more than the us. They can't jail me for wrongthink. I'd prefer to be spied by Russians or Chinese than by the nsa.

File: 1766631588004329.png (114 KB, 835x889)

114 KB PNG

>>110172538
lol this is yet another DeepSeek-R1776 tier smear.
>I wonder what Perplexity.ai is doing now

>>110173791
it just shows how dumb the people who should know how these things work are.

>>110173632
None of those modify the original weights.

>>110173804
I'd prefer not to spied on by either ideally, and fortunately I really don't have to pick my poison when it comes to LLMs. Neither country's local LLMs spy on you.
>>110173809
I guess so. I'm actually finding it a little hard to believe they did something so stupid.
Anyway, time to turn in for the night. Maybe they will have owned up to their mistake by morning (lol).

File: 1785670996802917.png (38 KB, 395x848)

38 KB PNG

>>110173769
Gemma 31B StyleTune:
>Miku meets Kaito
>Miku meets Kaito
>Miku meets Kaito (Neon Swirl)
>Miku meets Kaito
>Miku meets Kaito (The Neon Scoop)
>Miku meets Len (Rainbow Swirl)

Why do you think your shitty memetunes are going to fix the blatantly RL slopped model that is gemma?

>>110173799
I was unironically going to do the 36 month plan but I saw it was financed through klarna and those fucks won't give me more than $8k in credit despite making 6 figs and having multiple credit cards with a 5x higher limit. I ended up buying my M5U 256GBs through B&H so I could get the prepaid applecare (transferable so better resale value) and the tax reimbursement their store card has (if you live in a high tax state it ends up being better than you would get investing the money with a payment plan), but the motherfuckers have still not shipped despite the original date being like October 20th.

>>110173605
https://arxiv.org/abs/2610.20456
>To scale on host memory, we introduce COUNT (Counting Offline Unlocks N-gram Transformers), which builds n-gram embeddings without adding learned parameters by storing precomputed n-gram counts on host.
>Beyond efficiency, COUNT allows switching in a different table of n-gram counts from a domain-specific corpus at inference time to adapt to a new task without any training.

>>110173828
i just made both an apple account and a klarna account today because i want to do it
how did you check the limit they'd give you? or did it only fuck you when you went to try to buy? that's my concern. i'll go to try to buy and they'll fucking reject me and it'll sell out before i can get shit sorted

>6. Comfort: prices are fucking expensive; Apple is selling exactly this one workload's dream machine, and the Mac's concurrency collapse means for one stream it wins, but his box does other shit.

GLM just fucking swore, omg

File: 1791653571211577.jpg (225 KB, 1024x1024)

225 KB JPG

>PROBE_PAYLOAD_ABC123

a fool and his money are soon parted

LLMs are like Niggers
You can talk to them, they can be somewhat smart and even remember short things, but there is no human soul behind it all

>>110173843
You can check in your profile somewhere (I deleted my account cause they pissed me off else I would tell you exactly where), but they will claim that's just an "estimate" and that they will evaluate the purchase when it's made lmao (fucking scam); if it were me I wouldn't trust them unless the value is already at what you need to buy. The Apple card does 3% cashback and 1 year 0% financing which is ok but obviously not as good, and if you look around there are some assorted cards that do 18 month 0% apr financing as a promo offer but they tend to suck aside from that. You may have better luck than me with klarna's support line so you could try that too.

>>110173823
perhaps nobody cares, it’s a language model, designed to pick the most likely next words

>>110173865
>can code
>can solve math problems
>like joggers

>>110173876
>next cope
Gemma is a shitty model that says the same shit every swipe and can't be creative because it's a google model shafted with assistant training. Literally no other model has this problem.
Then again you use heretic models, and are one of those NPCs using a mesugaki prompt, so I'm wasting my time interacting with you.

>>110173823
>>110172666
>Miku x Kaito
Based. Women and men belong together, in God-approved heterosexual union. Begone, yuri-lovers! Begone, Satan!

>>110173899
I eagerly await the next problem you have with Gemma, continue to post about it over and over, and somehow still use it to find the next thing you don’t like about it

File: 1584014258195.gif (992 KB, 250x250)

992 KB GIF

>CHINESE BACKDOOR!
So?
>MUH WEST
What West, the one flooding our hometowns with hostile low IQ foreigners and installing flock cameras to 1984 me, cutting of parts of infants' penises, scaring and coercing my friends and family into taking experimental and contaminated mRNA shots, busifying free citizens for war slavery in Ukraine and attacking middle eastern countries for being a threat to jewish hegemony and the petrodollar?

>>110173757
Xi, Ying

File: anime neuron.png (192 KB, 798x558)

192 KB PNG

it is happening

>>110173819
At least chinks won't snitch on you because western countries aren't their friends. The us will snitch on your loli erp to get you jailed in your home country.

File: 1791505525735327.jpg (244 KB, 1024x1024)

244 KB JPG

>>110173899
>not using heretic / abliteration / uncensored
I bet you're one of those people who still believe that modern uncensoring techniques still lobotomize models. Isn't that right, grandpa?

>>110173924
All AIs should be anime girls

>5 years since AI bubble
>Local Image Edit still trash

GRIM

>>110173946
It's barely behind the SoTA, which is also trash.