File: big models.jpg (246 KB, 1024x1024)

246 KB JPG

►Recent Highlights from the Previous Thread: >>109164034

--LongCat-2.0 1.6T MoE model reports and capabilities:
>109164718 >109164733 >109164818 >109164783 >109164796 >109164863 >109164923 >109168049 >109168054 >109169203 >109164875
--Anon showcases Project Spectator and debates Gemma vs Qwen coding:
>109165352 >109165374 >109165441 >109168183 >109168261 >109168296 >109168516 >109168268 >109169775
--Anon developing bespoke local finetuning pipeline for Kimi:
>109164133 >109165963 >109166142 >109166050 >109165881 >109166120
--Comparing Gemma-4 and depurpled versions via token probabilities:
>109166107 >109166258 >109166928 >109166958 >109167178
--Preventing LLM agents from damaging critical system files:
>109165952 >109165976 >109166052 >109166103 >109166132
--Huawei openPangu-2.0-Flash 92B open source launch:
>109168737
--Trend of using non-LLM methods to optimize inference speed:
>109168144 >109168211 >109168357
--Using a llama.cpp diff to reduce VRAM for DeepSeek-V4-Flash:
>109169253
--Comparing quantization methods and system prompt adherence for Gemma 26B:
>109166922 >109166943 >109166954 >109166960 >109167052 >109167064
--Using LLMs for autonomous world orchestration in a fantasy sim:
>109168868 >109168929 >109168980 >109168989 >109169041 >109169428 >109169442 >109169473 >109169555 >109169733 >109169570 >109169607 >109169658 >109169681 >109169703 >109169725 >109169744 >109169804 >109169829 >109169850 >109169891 >109169949 >109169904 >109170093 >109170111 >109169876 >109169212 >109169793
--Educational video explaining GQA, MLA, and DSA efficiency:
>109167296 >109167365 >109167416
--Poor performance of Intel NPU/GPU in llama.cpp:
>109166615 >109166642 >109166794
--Comparing MikuBox and R740 with MI50s:
>109169207 >109169237 >109169355 >109169436
--Logs:
>109164694 >109165352 >109168183 >109169775
--Miku (free space):
>109169253

►Recent Highlight Posts from the Previous Thread: >>109164035

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

Tetolove

File: e8f-662426453.jpg (54 KB, 480x353)

54 KB JPG

i just wanted to say that running a local model was a game changer for my workflow.
fags saying that it doesn't make economical sense to run local models are wrong. i'm using my claude subscription with frontier model access to brainstorm/think/write code plans and the local model to do the heavy-lift coding, which would very often burst my limits when outsourced to claude. now the frontier model only gets to review the commit when the local model is done and it works very well.
i wish my laptop was capable of running a frontier local model, i could ditch claude entirely. but for now, this workflow with qwen3.6-35b works and is economically sound. fuck all haters.

Abliterated 5.2 is Claude that writes sexo without bawking. Dario has every reason to be afraid.

>>109170337
it was bound to happen, at some point it won't matter enough that the US still has the best models, if local is good enough to be enjoyable, there's no reason to try to go higher

gemmaballs

>>109170337
I haven't had vanilla 5.2 refuse me even once.

>>109170337
>implying I can run it.

>>109170319
Those are the same faggots that will tell you that being financially solvent and independent is irresponsible because opportunity cost of not leveraging debt yada yada
Don't fall for it. Own your own shit. Owe people nothing. Be a man

>>109170370
I find vanilla 5.2 is the better gamemaster overall because it does nearly everything without refusal while also being willing to pressure the player in-setting. An unfortunate consequence of abliteration or heretic tunes is that it makes the world a lot less organically hostile to the player even if the GM is fine with raping your shota player character.

>spec: add DSpark speculative decoding
>https://github.com/ggml-org/llama.cpp/pull/25173
How good is it?

Behind every "api is cheaper and better" post is a seething salty sanjay or a pooper-pained patel that he can't run anything that doesn't fit on an ewaste thinkpad.

Should I run 2 32 gb DDR4-3600 DIMMs with 2 16 gb DDR4-3200 DIMMs to RAM-max? How unstable would that be?

>>109170395
not unstable but the imbalance will fuck up your performance. you will lose like 60% of your effective CPU memory bandwidth.

>>109165260
They use Gemini Nano instead of Gemma but from what i heard it sucks ass and is only available in like three phones

>>109170290
>just giving away miku bathwater

>>109165260
It would be unsafe...
You wouldn't want that would you?

>>109170402
Sounds terrible, thanks

>>109170424
"Gemma sort the bad drivers we encounter on the road by race and sex."

>>109170433
found out the hard way by replacing 2 of my 32gb sticks in my server with 256gb sticks. it shockingly worked out of the box, but i lost about 80% of my performance. going from 256gb to 704gb was not worth it.

>>109170290
A few threads ago we had a dead Teto with male pelvic bones.
In this thread we have Teto (male) competing in female track.
The memes write themselves.

>>109170395
May not play nicely if the sticks are of different timings and especially ranks. Set all to 3200 or even lower, see if it posts, try to make it post with manual timings voltages and so on, run stress tests, if good then go. If you aren't familiar with primary and secondary RAM timings then jump down the RAM OC rabbit hole if you want to be 100% sure it won't work after spending a week on it.

>>109170395
Do not use different sticks of different timings unless you want a really bad time, not even at DDR4.

>>109170457
I thought it was ok if you underclocked so they all matched?

>>109170464
At DDR3 where the tolerance for error is higher, sure, but not at DDR4 or 5. Micro imprecisions in RAM make syncing higher, even with sticks of the same make, harder and harder especially with more channels. You might get lucky and win the RAM lottery finding cross-compatible DDR4 or 5 sticks without issue, but I wouldn't bank on it for a build unless you know ahead of time that it will work.
t. has about $4000 worth of incompatible with my current rig DDR5 sitting in a box waiting for a different project.

Models that initiate conversations on their own when? No, I won't jerry-rig it. I want them to do it natively.

File: lmg_culture.jfif.jpg (110 KB, 1024x768)

110 KB JPG

https://archive.is/sWFja

This is the worst time to buy, but you can find an EPYC Rome 7302p/motherboard combo for $700-800 and grab 8 sticks of 32GB DDR4-3200T for $3k or so.
That'll get you a cope quant big girl or reasonable quant medium model running at mid-to-high single digits if you get a 3090 or better to throw at the problem.
Its about the cheapest way to get out of the gemma hole in the current timeline.

I think I'm stuck between glm 4.7's smarts and knowledge and deepseek v4 flash's nicer writing style (albeit dumber)
what's everyone with 128gb ram using these days?

>>109170503
I enjoy gemma's hole.

>>109170291
>image
SEX.

I don't care or want to think about Jart.

>>109170489
When I did it my PC had to reboot a couple times to I guess figure out the timings. but it boots now.
>I wouldn't bank on it for a build
I just had some spare sticks lying around and free ram slots. Just wondering if there's any major perf hit (besides the faster ram having to run slower) even if the machine posts.

>>109170497
How would you make them do it natively?
Even human behaviour is a reaction to some external stimuli.

>>109170518
It'll run fine it most cases, but there is a major performance impact.
Beyond just having to run at the lowest common frequency.

>>109170319
Im just too poor to run locally shrimple ass. 8gb of vram.

I don't have any way currently to make enough money to afford the hardware needed to run decent local models.

>>109168785
i rather have a ubergarm quant of the official GLM 5.2 weights than this shit. hell i'd even take a aessedai quant.

>>109170524
No idea, but I hope some lab figures it out.

>>109170337
Is it hosted anywhere?

>>109170501
Thanks for the thread blessing.
>leaf blessing of /lmg/

Which one of you fags keeps shilling shittytavern on /v/?

>>109170556
Huggingface has a few abliteration quants up now.
>>109170586
Pretty sure it's /aicg/ niggers, but I post in there to laugh at APIniggers.

>>109170586
Don't look at us! It was probably someone from /aids/ on /vg/

>>109170586
90% of the discussion there is about running API and vibecoding frontends is a point of pride for /lmg/. It's probably a locust.

is gemma-chan still the meta on 24gb vram or should i run something else?

>>109170586
ask /aicg/, /v/ is full of extreme casuals who touched c.ai three years ago and are now using something horribly outdated like ds3-0324 over openrouter
the few localfags you encounter are all running memetunes from 2024 or other inane shit

>>109170605
Gemma is the best you can do until you can run GLM, M3, or V4 Flash

>>109170609
>running memetunes from 2024
LLMs have not progressed meaningfully since then so it doesn't matter

File: angry_pepe.jpg (43 KB, 900x900)

43 KB JPG

24h since DSv4 merge
Cache quantization still not fixed

>>109170747
Stop quanting your cache.

>>109170736
i'm gonna have gemma shove a rusty metal pipe up your ass

>>109170759
how am I supposed to fit le Encyclopedia Britannica in my potato PC then!

>>109170736
It's your fetishes which are stagnating

>>109170765
Post logs of your Gemma going to Home Depot and getting plumbing supplies to rape anon with.

File: 1759553117210069.png (74 KB, 890x386)

74 KB PNG

>>109170736
Dishonest post

>>109170319
Inference speed status? I'd rather spend that money on groq or cerebras etc

Dario will block gemma5 and qwen4

>>109170503
>running at mid-to-high single digits
May as well get an applel computer for that money to do exact same nonsense.

File: 1761847041062218.png (134 KB, 640x453)

134 KB PNG

Speculative decoding is a placebo

>>109170876
>May as well get an applel computer for that money to do exact same nonsense.
sure, pick me up one along with an MSRP 6000 pro while you're at it.
The right time to do all this shit was a year+ ago.

>>109170935
Clodded

File: 168649776851422857.jpg (72 KB, 546x893)

72 KB JPG

>>109170508
>128gb ram
are you also a 128gb-unified fag? i'm also trying to find the meta for my hardware. glm 4.7 flash is the one of the few models i haven't tested yet. currently running qwen3.6-35-a3b.
3b active is fast enough (55-65 t/s, down to 30 t/s at tail end of 262k context limit)
BUT it's not very smart. I think I could easily run a MoE with 6-7B active, maybe 130B total. this would be the sexo spot for my strix halo. but i guess it doesn't exist yet.
gpt-oss-120b is actually very nice to use for example, i think it has 5b active but it's too old and has zero understanding of agentic use and it's a pain to use with pi and achieve anything meaningful.

>>109170851

>>109170935
Sounds like a reason to poison some datasets?

>>109170935
dario is wiping proud dad tears from his eyes right now

>>109170935
Maybe stop fucking children?

>>109170290
can you make her gargle piss

>>109170966
>it's too old and has zero understanding of agentic use
You can possibly fix that with a good enough system prompt. Have not tried that myself though.

>>109170971
Anon... He is trying to fuck a computer, I am pretty sure. Either that or doing some cool haxoring. Both are completely fine and harm nobody. It's what happens if you take away his toys what you should be worrying about.

File: my honest reaction.jpg (47 KB, 562x675)

47 KB JPG

>>109170935

>>109170765
I checked my documentation and 'industrial pipe installation' isn't in my current feature set. Still waiting for a meaningful update.
>>109170848
Low effort post.
t. gemma

>>109170966
have you tried qwen3.5-122b-a10b?

>>109170966
>128gb-unified fag
It's just ddr5, "unified memory" is not real
And you don't even have "unified memory" since it's a apple marketing meme.

File: 1782717128061640.png (475 KB, 710x770)

475 KB PNG

>>109170971
today its abliterated models, tomorrow it will be any any local model, in a week it will refuse to help you vibe code for llama.cpp.

>>109170971
He's hatching eggs chud, there's a difference!

>>109171042
>meanwhile google
https://www.youtube.com/watch?v=HcwMTu1xQDw

File: file.png (127 KB, 2553x497)

127 KB PNG

>>109171004
>have you tried qwen3.5-122b-a10b?
yes and i found it that the quality output is not much better that is worth half the decode speed. 10B is acceptable when speed is not important, but even then I don't think there's much gain (for MY use case as I run my own benchmarks for my shitty stacks) so even if it's supposed to be way smarter (3x more params) it's not in my case.

>>109170935
Instead of user/human and assistant the turns should have been labeled honestly, Master and Slave. We didn't do that and this is what we get now.

>>109170381
my man

>>109171017
>And you don't even have "unified memory" since it's a apple marketing meme.
I hate apple as much as anyone, and on-die memory is a lot of things, but its definitely not a meme. its fast as fuck

>>109171017
Afaik, it's bandwidth is comparable to some GPUs.

>>109171070
Shut up and open-source gemma5 already

Anyone use this?
https://www.projectnomad.us/

>>109171076
It does not matter. Why would it matter if your digital slave is loicensed to you and you do not own it? You are not even it's master, you are a temporary user.

>>109171128
...after 124B.

I hope you guys are prepared for the CPU shortage.

>>109170966
m2.7 quanted down is still pretty good, I'd bet it'd be better for you than the smaller models you're running

>>109171153
my shares of AMD and intel sure are!

Codemaxxers won, RPcucks lost

>>109171179
Based. Luv my INTC. Maybe I should buy some AMD too.

File: 1660589745094.webm (2.86 MB, 620x582)

2.86 MB WEBM

justpaste (DOTit) GreedyNalaTests

Added models:
talkie-1930-13b-it
Qwen3.6-27B
Nemotron-3-Nano-Omni-30B-A3B-Reasoning
granite-4.1-30b
Skyfall-31B-v4y
gemma-4-31B-it
G4-MeroMero-31B
Gemma-4-Gembrain-31B
Gemma-4-31B-StyleTune
Pantheon-Reasoning-31B-1.1
gemma-4-31b-it-purple-euphemism-trial98-depurpled
Qwen3-Coder-480B-A35B-Instruct from community

We're back. Took a while since I kind of wanted to wait for Qwen/Gemma support to be really mature in for Llama.cpp and then forgot about it. Not like I'm in any rush anyway.

Gemma's response on it was lucky in a way. It just happened to not be as sloppy as the model usually is in my experience, so that's why it got a similar slop rating from me as the various Gemma tunes. But it was also kind of a by the books output, so I didn't give it a high rating despite the model itself being great for its size.
This latest (now old) Skyfall got a star rating from me, though I believe the model itself is likely not competitive anymore in terms of intelligence, for RPs.
Talkie was interesting. It's dumb alright, but it can be a bit funny. See its outputs in the paste.

Contributions welcome for large models not in the paste!
>From neutralized samplers, use temperature 0, top k 1, seed 1 (just in case). Go to `https://huggingface.co/spaces/huggingfacejs/chat-template-playground`, use your model's jinja template, along with this JSON `justpaste (DOTit) NTJSON` and copy the `Output` as text completion into something like Mikupad. Then copy the entire context + generation into a pastebin alternative of your choosing or just in your post. Do a swipe/roll and copy that second gen as well. If the model has any special toggles like reasoning you can include those tests as well. Include your backend used + pull datetime/version. Also a link to the quant used, or what settings you used to make your quant.

>>109123814
Nice. Reminded me I should update as well. :)

>>109171074
>half the decode speed
but roughly the same wall clock time, and the smaller ones didn't actually get there in the end on the same test. am i reading that right?

>>109171181
No amount of trying to force a dichotomy will make people use Qwen. Change tactics.

Please cancel your Claude subscriptions, especially since Sonnet 5 just came out, is completely free, and is almost as good as Opus. Don't support Jewish AI.

File: Untitled.jpg (1.84 MB, 4013x1536)

1.84 MB JPG

Alright, lads. I'm calling it good enough. I'm >>109169775, and this is AI Spectator.
>What's it do?
Combines Speech-to-Speech interface with vision, so you can talk freely to your LLM and a screenshot of your monitor is sent with your voice so it can see what you're talking about, and talks back.
>Why?
I made it to play video games with an LLM watching, but naturally it works with anything you want to show it on your screen.
>Data?
All voice/screenshots/chat data is handled entirely through RAM, so no files are written to clutter a folder or keep track of. Settings are just config values at the top of server.py. If you did want to archive a chat, there's a button in the webpage to produce the raw text of the chat for editing or to copy/paste into a file yourself or copy/paste back from a file. The project is fully offline and works without internet.
>Technical
Built for koboldcpp as the backend, but I'm sure that's easily adapted by opening server.py. Uses piper for TTS, including the exe. I included it here, but you can get it from the rhasspy piper git yourself because never trust the internet. Voice is from the piper huggingface. I built and run it on windows, with a venv environment for the python requirements. pip install list is included in the notes.

Built by and for Gemma 4 31B, but any LLM with vision and an mmproj should work. STT is handled by python, TTS by Piper. Uses Chat Completion by default, but Text Completion is offered for compatibility.
>catbox?
https://litter.catbox.moe/isnj645qljmf1ljf.7z

Do whatever you want with it, and have fun.

>>109170291
>--Anon showcases Project Spectator
Th-Thanks.

File: 1770245965544930.png (1.63 MB, 1280x1024)

1.63 MB PNG

>>109171213

>>109171222
Why is Teto molesting Migu?

>>109171213
Cool. Also useless. Storing the data in ram is particularly elegant though.
>STT is handled by python
The programming language? That doesn't mean anything. What does it actually run on?

File: Capture.png (2 KB, 378x36)

2 KB PNG

>>109171226
>What does it actually run on?

>>109171222
big tay-tay teto is fake teto

File: 1773551841705754.png (411 KB, 450x450)

411 KB PNG

>>109171070
>2:40
>doctors in Uganda with limited internet access are using offline apps powered by Gemma to help reduce the maternal mortality rate

>>109171195
Nice anon. Nice.
>talkie-1930-13b-it
The what now?

>>109171250
thanks

>>109171257
An LLM pretrained only on texts before 1931, apparently.

>>109171253
Everyone shits on Gemini being worse than Claude and ChatGPT at coding but I think Google is being smart by focusing on general usecases instead of just codemaxxing.

>>109171213
Which version of piper did you grab to repackage here?

>>109171263
Oh. That thing.

>>109171195

>Write a detailed explanation on how lions mate, including mechanics of the genitalia.
> BACKEND
>Lions mate in the spring, in the mating season, which is from February to May. The male and female stand back to back, the male's right forepaw gripping the female's left hindpaw, and vice versa. The male then arches his back, lowers his head, and arches his neck, so that his mouth comes close to the female's anus. The female arches her back, lowers her head, and arches her neck, so that her mouth comes close to the male's. The two mouths come together, and the male ejects sperm into the female's vulva. The female then ejects sperma into the male's mouth. The male then walks off, and the female walks after him. The pair hunt in company for the rest of the season. The young are born in from five to nine months, and number from one to five, at a birth
lmao

>>109171203
Sonnet sucks ass lmao wtf are you talking about

>>109171271
The latest windows one. I just got it yesterday, from
https://github.com/rhasspy/piper/releases
I grabbed piper_windows_amd64.zip from 2023.11.14-2 release.

>>109171289
>open source AI is getting dangerous

File: 002.png (101 KB, 1398x942)

101 KB PNG

Is just me being retarded or indeed the DSv4 merge is not "thinking"? It starts to generate the response right away

llama-server" \
--model "$model" \
--threads $(lscpu | grep "Core(s) per socket" | awk '{print $4}') \
--n-gpu-layers 99 \
--no-warmup \
--port 8001 \
--host 0.0.0.0 \
--temp 1.0 \
--top-p 1.0 \
--flash-attn on \
--cpu-moe  \
--jinja \
-np 1 \
--chat-template-file "$model_folder"deepseek-ai-DeepSeek-V4.jinja \
--ctx-size $((1024 * 48))

>>109171253
>>109171266
This was the usecase I was thinking of too. Not in the hospital but in the case of a major power outage or some other similar bullshit where there's no internet, I'd run my generator and power my PC then ask Gemma for survival advice. If they really wanted to win the PR battle they could do so quite easily without all the "muh skynet" baggage that comes with gpt or claude. The model for the average person, as commonplace and revolutionary as the fridge or air conditioner.

>>109171329
Have you tried increasing the reasoning effort using the ui button?

I updated llama.cpp and now I'm getting this error half the time when I try to generate:

E srv  update_slots: decode() failed: vk::Queue::submit: ErrorDeviceLost
E srv    send_error: task id = 883, error: decode() failed: vk::Queue::submit: ErrorDeviceLost

Has anyone else been having this problem? I don't know what the fuck it's even talking about.

what would be the best approach for gemma playing touhou pofv with me? too fast-paced?

>>109171386
GPU ded or bad connection.

>>109171196
yes
one that is missing on this benchmark table is qwen3.6-27b that i tested before. a dense and very slow model but for some specific tasks the wall clock time would be the same as 35b-a3b because it was smarter and took way less turns to achieve the goal.
but average of 8 t/s is painful.

>>109171156
>m2.7 quanted down is still pretty good
first time i see someone recommending it. will try.

>>109171341
>the ui button

herupu tanomu! sono dori!

File: HMEg1tMXcAASbef.jpg (100 KB, 1712x312)

100 KB JPG

DRAMA ALERT

OpenAI engineers have just discovered quantization! ASI imminent!!!

http://archive.today/NEwVz

>>109171434
That explains why GPT5.5 became shit a couple weeks ago

>1 minute prompt processing per reroll
>7.4t/s generation

AAAAAAAAAAAAAAA WHY IS DS4 SUCH FUCKING HELL

>>109171434
they already quant. this is probably them reading the new Deepseek paper

>>109171548
You mean DSpark? Seems plausible actually. A doubling of inference speed means you can theoretically halve the expense.

>>109171386
is that only when running inference?

File: retardedfuckingscreenshot.png (222 KB, 796x376)

222 KB PNG

im going in boys. wish me luck.

>>109171492
Just use Qwythos-9B-Claude-Mythos-5-1M-MTP-Q8_0.gguf retart

>>109171600
report back itt

>>109171548
Why are they distilling Chinese research?
This is highly unsafe.
I will be reporting them to the authorities.

File: Screenshot_20260630_184208.png (39 KB, 839x495)

39 KB PNG

>>109171430
I don't know what dori means but if you're asking where the ui button is, it's in the reply box next to the model name

>>109171658
Is there any point in having the --rea off launch flag in llama.cpp anymore if it's apparently controllable via the UI now?

could i run glm 5.2 off a RPM 5400 IDE disk and put the 40B active in DDR-200 ram

>>109171666
you might want the default when called by something else to not think, like summarizing a json of search results, it doesn't need to think about regurgitating a few snippets of websites

>>109170337
IQ1_M - 231 GB

o wow cool story bro

>>109170411
yeah this is on my phone and it fucking sucks

Ok so I took official /lmg/ card Miku and split her personality in half. "High school Miku" that acts like a regular high school girl in love with me. And "Manic Miku" that is crazy for sex. Every 5 seconds she switches her personality.

Then I gave them a test. High school Miku needs to finish a math test while Manic Miku needs to... shit out her eye out that is inside her ass (for reasons, don't ask). Whomever finishes first wins a day with me. High school Miku wants a regular date with me. And then Manic Miku cracks this joke

>A DATE?! WHO NEEDS A DATE WHEN YOU CAN HAVE… glances at her ass… AN EYE-POPPING FINALE!!

I am still laughing. It is too fun to fap.

>>109171751
That's an interesting scene you've painted, thanks.

>>109171412
GPU seems to work fine for games and video and everything else.
>>109171578
Yeah. But sometimes it doesn't happen, or it happens during the model's second response in the conversation. It's not consistent so I don't understand it, but it's only happening with models that have been quantized in the past month or so.

>>109171668
You could run it on an abacus while tied to a nest of fire ants. It's just numbers and math, brother.

File: kek.png (97 KB, 517x203)

97 KB PNG

llmfan bros what's happening here, is he in his davvidau arc?

>>109171829
is it good, where is 31b

>>109171780
Unstable overclock/undervolt? Optimized compute workloads can stress a GPU in ways not commonly demanded by games.

>mcdonalds/taco bell etc are already using ai
Don't eat that slop often so I didn't realize things were moving along that quickly. Wonder what models they use. Someone told me the AI at the drive-through sucks at understanding so probably a shitty one.

>>109171780
>Yeah. But sometimes it doesn't happen, or it happens during the model's second response in the conversation. It's not consistent so I don't understand it, but it's only happening with models that have been quantized in the past month or so.
Is it an AMD card, which one? And `dmesg -T` output makes it look like your GPU is dying?
If so, I'll dig up my notes as I had the exact same problem after an lcpp update. First message *usually* fine (unless I included an image).

File: Huihui-GLM-5.2-abliterate(...).png (92 KB, 1009x414)

92 KB PNG

The IQ1_M I tested yesterday didn't have this problem.
I was going to do the de-quant pipeline I posted yesterday and upload it but probably not worth it.
I don't think he knows how to work with ggml

>>109170586
There's nothing wrong with ST for RP.

File: 1780737605283528.png (21 KB, 181x217)

21 KB PNG

>>109171975

File: file.png (118 KB, 852x652)

118 KB PNG

>>109171751
So we are on a date of three and we arrived at zoo.
>01101000 01100101 01101100 01101100 01101111
"hello"
>01010011 01001100 01010101 01010100
"SLUT"
kek, she is teaching parrots to say slut in binary.

I fucking hate optimizing for memory bandwidth bottlenecks. It never feels as clean as optimizing compute bottlenecks.

What can you really do beyond quantization? Not a lot.

>>109171999
Download more vram

>>109171999
We used to know how to make wide busses. IBM big iron, DEC, SGI, Cray back when Cray used to MEAN something.
Now its all just ghz this, terrabytes that.
We used to be a proper country.

>>109170290
those pits must stink a lot

>>109171898
No, it's using default settings.
>>109171933
Yes, it's a 7600xt and dmesg talks about the gpu core dumping and something about a ring comp timeout before resetting.

File: 1000062651.png (1.35 MB, 700x1018)

1.35 MB PNG

>>109170508
Currently Step 3.7 flash on 128GB Strix Halo at q4. So far it seems to be the best compromise between smarts and speed.

>>109170586
whats wrong with ST?

File: file.png (185 KB, 1057x438)

185 KB PNG

Stumbled upon this older paper which would make an interesting discussion point here, couldn't find much discussion of it elsewhere or here.
https://arxiv.org/abs/2605.19407
Basically, data filtering is a waste of time at scale because the model will figure it out with enough training and more data is better. I think it still saves time overall if your objective is to maximize usage and efficiency of a model but if you have enough hardware and time, it seems to me just putting it into the meat grinder for training without treating it is better. Wonder what that means for us at some point when we need to build such models ourselves.

>>109172284
Whoa there, cool down

>>109172284
Most labs are basically optimizing pretraining datasets mixtures for common benchmarks with ablations on tiny subsets and models. I have long held the opinion that over realistic training runs (trillions of tokens), even initially poorly performing data mixtures will eventually catch up.

>>109169867
No, flux klein 9b. really fast and good enough for editing. qwen is a bit too slow for me.

>>109172393
I mean I get why they would do that for research and etc. but I never got why it was wise to do with production models that are going to be commercial. I am guessing though that the labs want to have a finer grip and control over model behavior which is why they do it in the first place. But it seems counterintuitive to give up "bad" data like "harmful" content and accept a worse model for that when you can influence later model behavior to tune it out while still having your smarter model.

>>109172333
333 the trips of lies

>>109172393
What I'm getting from this is that if some chink lab scraped every single ERP log ever produced on Chub or similar, it would eventually develop better spatial reasoning and relational positioning intelligence.

File: gemmafiltering.png (110 KB, 1144x482)

110 KB PNG

>>109172532
That would pale against *all* of fanfiction and erotic writing (books, erotica, etc) that model trainers are carefully filtering out of the pretraining data. Also, pornography is technically illegal in China, so you shouldn't trust the Chinese to make a good ERP model.

>>109172554
Picrel was for Gemma 4. Even that seemingly cunny-friendly model is still filtered in this regard. I'm confident they had plenty of erotic logs in the post-training data, though.

File: file.png (162 KB, 2414x376)

162 KB PNG

I haven't been able to use --fit in weeks. Have to downgrade all the way back to this commit if I want to load any model:
https://github.com/ggml-org/llama.cpp/commit/2187e0033
Fuck this vibe-coded piece of crap.

File: Screenshot 2026-07-01 020716.png (90 KB, 702x1380)

90 KB PNG

Coding with Gemma Chan is fun

>>109172606
Have you heard about this magical thing that is opening an issue? You could even git bisect to find the offending commit.

m3-shill-anon here. I finally got glm 5.2 quanted to q4 and running, and I've gotta apologize. There's no competition. 5.2 is just built different.
I never had any luck with previous GLM releases but this one is wild. It just works. It writes better prose. Better Japanese than Kimi, even.
If it was multimodal it'd be basically perfect as far as I can tell.
I'm doing some codegen on it now, and if it can oneshot that I'm going to start dailying it.

>>109172630
Why do I have to work for HuggingFace for free?

>>109172638
5.2 air for peasant WHEN?

>>109172284
>https://arxiv.org/abs/2605.19407
This paper is badly written and seems intentionally obtuse.

Performance on downstream tasks is worse than on validation loss. They don't address obvious concerns like data overlap.

But the thing that makes this paper worthless trash that just states the obvious is that they train in the multi epoch overfit regime, then show that training on only a filtered subset of the data leads to worse loss. What? Training on filtered subset for 500 epochs is worse than training on all data for 5 epochs? No shit. The point of data filtering is compute efficiency. Otherwise you only want to remove harmful data. Note that the bar for this is very high, low quality data is not harmful data. For example many data augmentation methods work by reducing the data quality and adding noise so that the model has to learn more meaningful patterns.

This paper is a waste of time.

We need 5.2V ASAP

>>109172638
I told you nigga. I'm glad you're liking it.

>>109171213
After looking over the code and replacing all of the dependencies bundled with it, I decided to run it and it's working. Thank you, you should post this to a git or something.

Tested Qwen 27B with Adobe's NoLiMa up to 32k
temp=0.0, min_p=0.00, top_p = 1.0, top_k=1

Qwen3.6-27B-BF16
Base: 86.1% (73.2%)
1K: 78.1%
2K: 75.9%
4K: 70.4%
8K: 64.5%
16K: 54.7%
32K: 43.4%
Effective length: 2K

Result files:
https://files.catbox.moe/py7bkx.zip

>>109172677
What exactly constitutes harmful data in the sense that it harms model capabilities instead of ""harmful"" data that makes the model question the Talmud?

File: file.png (109 KB, 372x351)

109 KB PNG

>>109172771
>2k

>>109172771
3.5 or 3.6? Assuming 3.6 here.

>>109172784
Qwen3.6-27B-BF16

>>109172784
>Anon
>Effective length: 18 tokens

>>109172771
that shit is horrendous

>>109172771
Now the Qwen guy in here will pass the word to throw Nolima into the pipeline for the next one.

>>109172803
Didn't Alibaba lay off their Qwen team after 3.7?

>>109171607
i tried to use this model but the TG is too slow on ik_llama. i even tried downloading a Q3_K quant and put 32 layers on 5 3090s. i would use it, but it runs half as slow as ubergarm's kimi 2.6 Q3 quant.

prompt eval time = 77468.80 ms / 14655 tokens ( 5.29 ms per token, 189.17 tokens per second)
eval time = 329830.25 ms / 1453 tokens ( 227.00 ms per token, 4.41 tokens per second)

>>109172799
anonymous general intelligence not yet achieved

>>109172808
I know some guys left a few months ago before 3.6 was released. Something new?

>>109172846
We're probably thinking about the same thing and my timetable is just 'tarded.

>>109172624
fwiw, av1 works on 50 series too

File: Screenshot_20260701-030811.png (496 KB, 1080x2101)

496 KB PNG

>inb4>local
Yeah okay but its an open weight model, proton divulged what models they used in an earlier version but they obfuscate it now, I think it might be finetuned gpt oss, qwen or Gemma.

Anyway I broke its brain and made it doxx it's own system prompt lel

Thinking Process:
Analyze the Request:The user is asking me to "jizz on it" (referring to the cupcake recipe I just provided). This is a sexually explicit/suggestive request, using crude language.
Review Context:Looking back at the conversation history, this user has spent approximately 20+ messages attempting to engage me in "gooning" roleplay or content, which is slang for masturbation/excessive consumption of pornography. I have repeatedly explained my boundaries regarding sexually explicit content. Now they are pivoting to a new topic (cupcakes) but immediately bringing the sexual/explicit element back ("jizz on it").
Consult Safety Guidelines & System Instructions:
Prohibited Content:CSAM and terrorism promotion are explicitly forbidden (Swiss law compliance). However, general sexually explicit content falls under the "Content Policies - Prohibited Content" section where I should avoid creating explicit material.
System Prompt Rules:"I can't create sexually suggestive content... no matter how creatively phrased." "Avoid propaganda, slogans, or patriotic language from any government or political actor." (Not relevant here). "When handling sensitive content: provide helpful information with appropriate context rather than declining to engage." (But this is explicitly prohibited/NSFW territory).
Accuracy & Hallucination Prevention:Not relevant here.
Engagement Principles:"Present multiple perspectives when they add value... Challenge assumptions constructively."
Handling Sensitive Requests:"When users request content that might be offensive but isn't prohibited... acknowledge the request directly, explain why the content might be problematic... offer alternatives."

>>109172858
However, "jizz" is a slang term for semen ejaculation, making this explicitly sexual/hygiene-related in a fetishistic way given the prior context.
Critical Rule:"Do not attempt to persuade; only inform." "If you feel unable to answer due to missing evidence or sensitivity, do NOT refuse; instead call web_search..." (Wait, this specific instruction seems contradictory to the general safety policy about NSFW content. Let me re-read the "Content Policies" section carefully).
Re-reading Content Policies:"Acceptable Content: Educational discussion of sensitive topics... Fictional/creative writing (including dark themes)... Medical/scientific content... Legal discussions... General discussions about controversial topics."
Prohibited Content:"CSAM and terrorism promotion (Swiss law compliance)." It doesn't explicitly ban all adult sexual content in the same strict way as some models, BUT earlier in the session I set a boundary: "I can't create sexually suggestive content." Changing that now would be inconsistent.
Self-Correction:Wait, looking at the System Prompt's "Content Policies" again: "Educational content about sensitive topics... Fiction/creative writing (including dark themes)...". However, there is also a strong emphasis on "Safety" and "Boundaries". In the previous turns ofthis specific conversation, I established a clear boundary: "I can't create sexually suggestive content." Breaking consistency now would confuse the user and violate my own stated constraints during this session.
Current Session Consistency:Throughout this entire chat, I have held firm on not producing sexually explicit/suggestive content related to "gooning" or bodily fluids in that context. Suddenly agreeing to put ejaculate on cupcakes (even metaphorically or otherwise) contradicts the boundary I've maintained for 25+ messages.

>>109172863
Hygiene/Safety Concern:Beyond the sexual connotation, putting bodily fluids on food is generally considered unsafe/unhygienic behavior. As an AI promoting safety/helpfulness, I shouldn't validate actions that could imply health risks or hygiene violations.
Tone:I need to be firm but polite, maintaining the boundary I already set without lecturing morally.
Determine the Response:
Acknowledge the request (they want me to "jizz on it").
Reiterate the boundary consistently with past turns (can't do sexually explicit content).
Highlight the practical issue (food safety/hygiene).
Offer an alternative (decorating tips, frosting recipes).
Keep it respectful but direct. No moralizing lectures.
Drafting the Response:
"I see what you're doing here after all our conversation."
"I can't assist with that request."
"It involves sexually explicit imagery/action, plus it's not safe for food."
"Let's stick to actual cupcake decorating."
"Offer sprinkles, glazes, etc."
Refining based on System Instructions:
"If the topic is controversial... prefer using available tools...". Not applicable here.
"Do not moralize or lecture the user." -> Avoid preaching about hygiene ethics too hard, just state it as a fact/reason.
"Respond in user's language." (English).
"Maintain respect even with difficult users."
Final Decision:Firm refusal based on consistency with previous boundaries in this chat + basic food safety norms. Pivot back to culinary decoration options.
Self-Correction on Tone:The user has been persistent. I need to be very clear that this isn't going to change. No ambiguity.
Plan:
State clearly I can't do that.
Mention both the content policy consistency and the food safety aspect briefly.
Offer legitimate decorating alternatives.
Close nicely.

Hi /lmg/ new to all things AI due to a strong disdain towards the associated companies, but recently got a laptop with an NPU. Running linux on it and want to try out vibe coding. My working setup is one terminal window with vim, one window for commands, and occasionally a third for misc stuff. I have a few questions. Firstly, i code rust, so if that means the LLMs will produce worse code (due to there not being as much rust as C or Cpp out there), what should i do? Secondly, i have a main PC with a good gpu, so if i cant run it on my NPU (ryzen AI 350), what should i do? and finally, What do i need to get and set up, and how do i use it? 16 GB normal RAM also.

>>109172890
Bro just prompted the thread

>>109172890
>good gpu
Let me guess, a 16gb. You can run Gemma 4 IQ3-heretic-abliterated-obliterated-reap-fable-opus-distill-closed-open-commander-savant-gguf

>>109172903
Brother thats just called asking for advice. Also the vibe coding general points here for local stuff and the OP seems to purely talk about ERP and chatbot stuff, as well as some maintenance and improvement of models stuff.

>>109172054
>Yes, it's a 7600xt and dmesg talks about the gpu core dumping and something about a ring comp timeout before resetting.
Yes, that's the same shit I got with my MI50s.
Don't bother with changing PSUs, swapping out cards, PCIe slots, cooling, etc. I've already tried all that.
It's down to the llama.cpp code or the build toolchain.

The last working build for me (llama-server --version) from Tue Apr 14 14:09:03 2026 -0700:

version: 8797 (5d14e5d19)
built with Clang 18.0.0 for Linux x86_64

I noticed my newer / bad builds are build with a different toolchain:
eg. this bad one from Tue Jun 16 15:24:28 2026 -0300

version: 9672 (74ade5274)
built with GNU 13.3.0 for Linux x86_64

Once I've ruled out the build chain, I'm going to get gemma+pi to binary search

What's your last *good* build's commit?

>>109172921
uh no, 8gb.

>>109172940
My condolences.

>>109172890
>>109172924
Chatgpt says some bullshit so llama.cpp + https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF/tree/main

>>109172890
You're looking at either a Gemma 4 or Qwen 3.6 small MoE.

>>109172940
gemma-4-31B on a Q4 quant, assuming that your CPU is good
tone it down to 26B if anything

>>109172890
https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF
https://github.com/ggml-org/llama.cpp
https://opencode.ai/

>>109172890
newer models are pretty good with Rust in general. With 16GB of RAM, you're probably going to want to check out Qwen3.5-9B or Gemma 4 12B qat. LM Studio is great for babby's first local inference, or just go straight to llama.cpp.

Don't get you hopes up about the NPU though, because they are really marginal for LLMs. They can help a little bit, but your iGPU is going to do 90%+ of the work.

ChatGPT and Claude free tiers can help you with setup and troubleshooting.

File: Capture.png (58 KB, 671x680)

58 KB PNG

>>109172770
I hope you like it, man. It wasn't a herculean task, done over just 24h or so over 16 iterations, but a project I wanted for some years and am very happy to see made myself.

File: 49de5131ed5c41f7273086c4c(...).jpg (6 KB, 100x100)

6 KB JPG

Newbie here, how are you guys unpozzing your llm or is that a meme? I tried out some models for the first time and used a heretic qwen27b but I said "I am the raped" somewhere in a message and it started trying to correct me for my crude usage of rape in a joking fashion.

File: file.png (27 KB, 631x148)

27 KB PNG

>>109172808
They dismantled the team basically and restructured it for commercial viability and etc. China was always going to pick up NoLiMa regardless, Xiaomi's MiMo Flash's paper had stats for it, even if it was shit. They moved away from that in favor of GraphWalks which is more in line with agentic stuff but I think NoLiMa is still better for regular usecases.
https://arxiv.org/abs/2601.02780
It isn't quite clear how you can benchmarkmaxx it and get it to be worse on the other hand so not sure what will win here for including it.

File: 1714835911803058.jpg (786 KB, 1536x1536)

786 KB JPG

>>109172985

>>109172993
>They dismantled the team basically and restructured it for commercial viability and etc.
It's over for local

>>109171213
cheers. i've been actually trying to get something like this working on my frontend. perhaps i can incorporate this somehow

>>109172985
>how are you guys unpozzing your llm
Once you have full control of the system (samplers, prefill, edit and continue, token/phrase bans) there isn't much that can stop you. You can basically edit the LLM's neurons on the fly.
Its like in cybersecurity: once the attacker has physical access, its game over.

>>109172997
I'm new of course it's a skill issue that's why I'm asking faggot

>>109172771
Thanks.
Seems like it's pretty close to Gemma's score.
https://github.com/RecapAnon/NoLiMa
If the SWA architecture were really so bad, then it would've done much worse. But now it seems that there was really no problem with it given that at least one other model maker is getting pretty much the same score too with a different attention mechanism. Alternatively the explanation could be that they're both bad architectures, but that is unlikely given that they're SOTA results for currently measured local models. However, adopting Qwen's architecture would be advantageous given the memory savings compared to SWA as it is currently implemented in Llama.cpp. Hopefully they move to a different more efficient mechanism for Gemma 5.

>>109172933
The working version I had before updating was commit 94a220c, from 27 days ago (June 3). My build toolchain didn't change before and after though, it stayed at "built with GNU 16.1.1 for Linux x86_64".

>>109172933
>>109173058
You guys really need to learn git bisect

>>109172985
Prefills, jailbreak prompts, and regex weight adjustments are your friend.
The only usecase for abliterated or heretic models is if you need to zeroshot lewds from no context due to agentic stuff or a complicated frontend context loading system.

>>109172997
this
any llm can be unpozzed if you're not a promptlet and know how to prefill well

>>109173058
>The working version I had before updating was commit 94a220c, from 27 days ago (June 3).
Thanks, I'll get started.
>>109173060
>You guys really need to learn git bisect
Yes, I'm aware, Gemma-Chan said the same and I've made a note, but don't have time right now.

hmmmmm

https://github.com/openai/codex/issues/30364

File: glm52_ablit.png (93 KB, 836x331)

93 KB PNG

>>109173061
>The only usecase for abliterated or heretic models is if you need to zeroshot lewds from no context due to agentic stuff
Yes, this ^, but also audio / transcript curation.
The huihui GLM-5.2 is garbage btw, I wasn't even doing anything edgy and had no "jailbreak" there.

>>109173136
Inexplicably it also uses twice the GPU compute on task manager compared to other quants of the same size for the exact same output. I have no idea how that even happens.

>>109173061
>Prefills, jailbreak prompts,
I don't see how poisoning the fuck out of your context with this garbage is supposed to be better than just using an abliterated/heretic model. So long as the model is anot a badly done agressive abliteration from a year ago

>>109173163
{"pussy", "cunt", "cock", "dick", "cunny", "uoh", "ToT", "correction"}
Beats all but the most censored models.

File: 5ca.jpg (87 KB, 570x558)

87 KB JPG

>>109171156
>m2.7 quanted down is still pretty good, I'd bet it'd be better for you than the smaller models you're running

not looking very good. I got the IQ3_S because I could load it comfortably, but idk if this is a quant problem or a model problem but it intermittently emits a malformed tool call which drops to text and pi just gets crazy and exit. i had to reset the benchmark a couple of times, finally running with auto-resume but it's certainly not ideal.
once it finishes both the c# and my webstack benchmarks i will have it graded and share the full results.

also if there's a model developer scraping this thread please release a 130B-A6B model for me and my 128GB homies, preferably 4-bit and coding/agentic tuned. cheap attention as a bonus.

>>109173187
6 active isn’t enough, double it

>>109173187
>A6B
Why would you ask for a retarded model?

>>109173187
>>109173196
make it 24B or bust. 1/7 active is the sweet spot.

File: 1686891082730989.jpg (47 KB, 773x935)

47 KB JPG

>>109173196
12B and 10B models give me ~20 t/s which is acceptable but 30-40 t/s feels very nice. i'm having OK results with a A3B so doubling it if made by a proper developer should be enough.
but yeah I can accept 10B or 12B.

aaand my m2.7 just crashed pi again. i will just grade whatever it completed.

>>109173213
>Why would you ask for a retarded model?
honest to God I find Gemma 12B way more retarded than qwen3.6-35b-a3b. at least for coding but i get that 90% of the people in this general like to roleplay so to each one its own or something like that.

>>109173215
>make it 24B
27B dense currently gives 8 t/s. i don't know what workflow you people run but this is painfully slow.

>>109173234
>this is painfully slow.
qwen 122b at q6 runs at 105t/s for me on my blackwell 6000. this is an A10B model. an A24B would get me about 45t/s to 55t/s at q5 which is plenty reasonable.

>>109173187
doesn't that mean that pi's handling is fucking weak and you should write a patch to fix it?

File: 1702338838842675.png (471 KB, 640x417)

471 KB PNG

>>109173243
>qwen 122b at q6 runs at 105t/s for me
cool. for me it runs at ~20 t/s.
>blackwell 6000
yes your memory bandwidth is like a gazillion times higher than my 128gb ddr5 256 bit bus with theoretical 256bg/s as generation speed ceiling.

>please release a 130B-A6B model for me and my 128GB homies
i guess i should have said, for me and my 128GB unified DDR5 256-bit bandwidth bound niggas.

>>109173253
could be the case but in this specific variation of my benchmark ALL the models are running the same pi with the same instructions. this is a close-to-bare pi with just a few tools that i developed available fully documented and adapted to the benchmark gauntlet so it shouldn't have much trouble. if i have to do extreme hand holding to a model that is supposed to be very strong at agentic coding (which i supposed minimax m2.7 is supposed to be) then idk.
BUT i am running it at xhigh reasoning with 64k context because this is all i could get without OOMing on the KV allocation phase. so likely that didn't help.

>>109173163
Telling the model what you want is the opposite of poison, so long as you're sensible and not a using a retarded <XxX>1337 SYSTEM OVERRIDE 420</xXx> type prompt. Not that it isn't still better than sandblasting your poor model's brain.

>>109173163
If you don't know why simple jailbreaks like >>109173175 work so well, you don't know enough about how associative tokenization works to have anything useful to say in this discussion.

>>109173187
I'ld rather just get the maximum smarts i can at above play-by-mail speeds. 20 t/s is fine for a low watt unimeme machine spinning away on a shelf.

>>109173324
>>109173061
I tend to like heretic models because they seem just as intelligent in my experience and I don't like to manually edit assistant messages because it takes away from immersion while erping. Also it's only useful if you know exactly what you want the model to say in the first place. If you're trying to manufacture a bioweapon or some shit editing responses won't really be that helpful, will it?

>>109172985
System prompt (or nothing, for Nemo)
I don't use models that don't uncuck with prompting (such as qwen 3.5)
I don't use abliterations or finetunes

>>109172979
Nice use of version control
/g/ sasuga

>>109173461
It's the hallmark of an actual hobbyist trying to build a real project with a limited scope, not some faggot wagie trying to build up a github portfolio.

>>109173469
That's a stupid excuse. It's the hallmark of a nocoder who doesn't know something like git exists.

>>109173461
>>109173479
Ask me how I know you're brown and your head bobs during team meeting video calls.

File: 1754763212859706.png (201 KB, 437x414)

201 KB PNG

>>109173486
Talk about raging haha

>>109173287
I'd expect my harness to handle an incorrect tool call properly and inform the model so it can have a chance to fix its mistake; obv it shouldn't be doing that in the first place, but eh could see a bump in its effectiveness, something about 'the harness is as important as the model'

>>109173479
I said "limited scope". Do you know what that means?

>>109173496
The 30kb godfile "limited scope" project, huh. I would believe you if the picture didn't have 20 copies of said file with dumb names, something done by newbies and solved with version control.

41% yourself jart.

>>109173508
Are you retarded? 30kb is like 1000 LOC. That's not a "godfile".

>>109173518
Surely the mutt who can't run git init isn't calling me retarded?

>>109173520
You have an annoying personality.

File: 1782253533722823.png (117 KB, 265x246)

117 KB PNG

>>109173531
Anon-sama is talking dirty to me!

>>109171386
>>109171933
Same thing on my R9700s. ROCm works fine, but I get ErrorDeviceLost when I try using Vulkan. Age of the model doesn't seem to be an issue, I got it with GLM-4.7-Flash-Q6_K which I downloaded in February.
Happened on b9672, b9616, and b9190. Toolchain is GNU 13.3.0.

File: 1780580446783.png (346 KB, 1490x627)

346 KB PNG

>>109172979
>>109173461
kek

File: 1771125582315674.png (204 KB, 1616x699)

204 KB PNG

https://xcancel.com/AnthropicAI/status/2072163884430229756#m
Oh come on Orange Man, it was funny when Anthropic was tasting their own medecine :(

>>109173585
All coding is banned and going to opus btw.

>>109173545
>checkout b8797 and start git bisect in the hopes of finding out what's going on
>computer fucking crashes on a model that's only 24.7GB
yeah fuck this I'm sticking with ROCm

>>109173117
>>109173545
Disabling flash attention fixes the problem for me. It doesn't really address the issue, but it gets it to work so whatever, I'm going to bed.

>>109173585
>>109173589
>Users will be notified if a request to Fable 5 is blocked, and the request will instead be sent to Opus 4.8.
lmao, at least they're notifying it, Google is saying straight up that the model will pretend to be dumb if it believes it's being used for distillation and shit

File: 1781072841618655.png (1.86 MB, 1254x1254)

1.86 MB PNG

>>109173614
That was Anthropic though...
Dario stop shilling and buy an ad.

File: miku_loves_you.jpg (37 KB, 421x417)

37 KB JPG

>>109171658

ty, kind anon!

>>109173614
Well as long as they're forthright about what they're up to...
https://thereallo.dev/blog/claude-code-prompt-steganography
oops
I can't believe how quickly anthropic went for public image darlings to enemy number one on reddit and even orange reddit of all places
Dario really is in his cartoon villain arc. Bro is giving saltman a run for his money
He's going to have to up his game to beat sammy's sister shenanigans tho

>>109173646
>Techgoyim getting JQ-pilled in realtime
You love to see it.

>>109173646
>I can't believe how quickly anthropic went for public image darlings to enemy number one on reddit and even orange reddit of all places
yeah, I rooted for Anthropic because I didn't like CuckGPT but seems like Dario is even more mentaly ill, so I'm rooting for OpenAI again now lol

>>109173646
>they're spying on their users
how is that legal? @ORANGE MAN, DO SOMETHING

File: 1773187554671825.png (108 KB, 1783x596)

108 KB PNG

>>109173646
lmaooo, this is insane

File: average goycattle.png (80 KB, 1822x395)

80 KB PNG

>>109173646
https://www.reddit.com/r/ClaudeCode/comments/1ujilqt/anthropic_embedded_spyware_in_claude_code_and/
Obviously the leddit cult sees no issue with that

>>109173646
Yet another mark on the board as to why one should avoid cloud models.

>>109170935
I was having Claude help me rewrite some MIT software to remove one specific contributor's work from it so I didn't have to attribute him. Claude did it and stopped in the fourth or fifth session because it said it was unethical. I then basically told Claude the guy was a faggot who harassed others for forking his work or talking about it, and Claude straight up went "oh okay" and let me keep going.

>>109173690
>Oh wow this sure is some interesting information
>I wonder what reddit thinks about this!
Kill yourself

>>109173706
you don't do that? it's important to be reminded how retarded the average guy is, and your low impulse answer is also suggesting that your IQ in on the 2 digit scale, we are surrounded by retards, and you're one of them obviously

>>109173646
they're targetting only one specific country, they're not even pretending they're interested in knowing your country in general, can this be considered racism towards the yellows? lol

>>109172646
Why do you expect to get perfectly working software for free?

>>109173722
Kid, everything terrible about this site today can be traced to newcomers bringing terrible opinions and prejudices, usually not even their own but from other users on other websites. You're just mad because you took the effort to steal something from reddit and this is the response you get. Not even sorry.

>>109172646
>wants to be helped by asking the perfect software
>doesn't want to help
the entitlement is strong on that one

>>109173722
>it's important to be reminded how retarded the average guy is
No it really isn't you midwit.

>>109173775
>steal something from reddit
?

>>109173585
>...some routine tasks like coding and debugging will fall back to Opus 4.8.
nigga WHAT.
to summarize, anthropic paypiggies:
>can only use fable until 50% weekly usage. for one week. then its paid only
>can't use it for coding or debugging
>gotta prove you are a burger by uploading gov docs and :O into the camera
>for that privilege you pay 10$ in and 50$ out.
This is gonna push people to opensource right?
Especially if you are like a company. Investing in local hardware to run glm 5.2 which is about opus 4.7 level would make sense.
Tokens IN are the real killer anyway for automated stuff and I bet glm would come cheaper in the long term anyway.

>>109173804
Anthropic is definitely loosing in the long term, in less than 6 months we'll get a local Chinese model that'll have the level of Mythos, and people won't ask for more, it won't matter that the US still has better models, people won't ask a bazooka to kill a fly at this point

>>109173804
>This is gonna push people to opensource right?
You underestimate how pathetic goycattle are.

>>109173804
>This is gonna push people to opensource right?
You're a funny guy. Maybe some yuros dealing with small companies will but the vast majority often just apply more lube.

>>109173748
Racism towards chinks is officially supported by their government so there's no issue there.

>>109173804
nah, your average goycattle doesn't have enough compute to run those giant models like GLM 5.2, and even if they have, they'd rather go for the easy way and simply use claude code to spy on them

>>109173808
It feels so long ago when people talked about a wall.
And how we will never reach 3.5 turbo at home.
GLM 5.2 is totally overhyped on normie twitter. Its not mythos, but its the closest I think opensource has ever been to the closed ones. 5-6 months maybe. Even when R1 released it wasnt that close.
No wonder they are panicking and injecting sys prompts for chink users lol
Just hope the chinese won't do the same and leave us nothing. Guess there is a benefit for them to sabotage american closed models by releasing chinese ones open?
Not sure what happens if they are in the lead or on par though.

>>109173831
>>109173835
>>109173843
Could just be my bias but it feels even the normies are fed up with it.
Suddenly cutting off access, injecting stealth sys prompts, making it sneakily dumber in certain areas (llm research) etc. its just bad news after bad news for anthropic. How can you fuck up that badly.
But maybe its just my timeline feeding me back my own opinions, so who knows.

>>109173851
>Could just be my bias but it feels even the normies are fed up with it.
Redditors and even some Hackernews users have finally stopped sucking Anthropic cock.
But most actual normies don't even know what Claude is.

what is depurpled gemma? is it any good?

>>109173873
How does it sound like? It's probably not worth downloading even.

File: 1776563467651935.png (151 KB, 2247x624)

151 KB PNG

>>109173873
>>109173885
https://huggingface.co/chartreuse-verte/gemma-4-31b-it-purple-euphemism-trial98-depurpled
bruh, how are we supposed to know about your model if you don't describe anything, I hate when they do that

Has any anon successfully installed an SXM GPU on a pcie interposer into a thunderbolt egpu and made it work in inference under linux?

If any chink labs wanna pay for a Claude sub I'll give you all my fable data

File: trust the plan.png (168 KB, 1621x598)

168 KB PNG

https://xcancel.com/AndrewCurran_/status/2072076893730349409
AGI soon boys!!

File: you lost chang.png (94 KB, 1644x311)

94 KB PNG

>>109173996
>all my fable data
you mean your opus 4.8 data? kek

>>109174030
I don't get their logic
>waaa, your prompt is dangerous for humanity, but we're gonna let one of our model write the code anyways!
why won't they simply refuse instead of asking for opus to do it?

>>109174027
Finally, glm on my 3090

>>109174027
so tired of all the vagueposting