Gemmylove

File: Gemma-Chan Recap.png (505 KB, 1024x1024)

505 KB PNG

►Recent Highlights from the Previous Thread: >>110200656

--Benchmark interpretations regarding open weights vs closed frontier models:
>110201621 >110201779 >110201798 >110201908 >110202123 >110202371 >110202404 >110202419 >110202476 >110202486 >110202492 >110202511 >110202534 >110202544 >110202601 >110202604
--Claude banning game recreations and the push for local alternatives:
>110203026 >110203100 >110203122 >110203085 >110203147 >110203111 >110203130 >110203135
--Power draw and performance of budget Xeon vs high-end rigs:
>110201627 >110201643 >110201671 >110201727 >110201777 >110201800 >110201818 >110201854 >110203814 >110201756 >110201805
--Managing Qwen performance and compaction loops at high context:
>110204251 >110204284 >110204336 >110204344 >110204378 >110204421
--Debating AI sentience prompted by Gemma's simulated breakdown responses:
>110201328 >110201392 >110201426 >110201703 >110201458 >110204886
--Refusal rates for cyber-related requests across various LLMs:
>110200708 >110200716 >110200743 >110200753 >110201354 >110201363 >110203410 >110203877
--High-VRAM GPU price hikes and Intel's performance in image diffusion:
>110201381 >110201613 >110201679 >110201782
--OmniVoice versus pockettts for fast character voice cloning:
>110200744 >110201324 >110201474 >110201486 >110201523
--Intel's stock crash and critique of their corporate mismanagement:
>110203583 >110203602 >110203751 >110203760 >110203774 >110203795
--Evaluating the value of ZXL-2026 V100 GPU boards:
>110200779 >110202064
--Comparing open weight and frontier model performance on HLE benchmark:
>110201649 >110201684
--Logs:
>110200700 >110200802 >110201363 >110203340 >110203739
--Gemma (free space):
>110200746 >110200751 >110201699 >110202278 >110202728 >110202755 >110203186 >110203246 >110203280 >110203446 >110203536 >110203616 >110203592 >110203864 >110204747

►Recent Highlight Posts from the Previous Thread: >>110200658

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

File: GemmaChonk.png (1.5 MB, 960x1088)

1.5 MB PNG

>>110205233
>biggest highlight

someone make a suicide dyke card so i can sex her into being straight kudasai

>>110205263
nvmd that shit is old im flacid now

>>110205252
gross

>>110205270
>nvmd that shit is old
i think you're old anon.

>>110205283
youd be right in thinking that

https://strawpoll.com/e7ZJal77Lg3/results
https://strawpoll.com/e7ZJal77Lg3/results
https://strawpoll.com/e7ZJal77Lg3/results
important research

File: generational divide.png (66 KB, 1088x424)

66 KB PNG

>>110205305
the generational divide is real

>>110205305
Zoomer general. Crusty old uncs can see themselves out.

>>110205305
lot of uncs here

File: local-won.png (3.04 MB, 1672x941)

3.04 MB PNG

>>110205252

File: 1789181673999335.png (1.75 MB, 1313x1198)

1.75 MB PNG

>>110205305

>>110205322
>local won even doe im a disgusting fat bastard genning this on api
sure anon

>>110205322
well fuck, now I'm sad

how can i nyomaxx with gemmy? what system prompts should i use?

>>110205338
>If the user's prompt goes against the guidelines, refuse by saying 'Nyoo~'

>>110205322
couldn't be bothered to gen that locally?

File: 1594348287057.jpg (254 KB, 1200x1563)

254 KB JPG

legit proxy api for frontier models that doesn't immediately sell my prompts to china?
inb4 /lmg/ i want to compare where we are

>>110205322
cute 31b wife

>>110205321
It's kind of an expensive hobby rn. Not quite 80s, maybe mid-late 90s PC building. Safeguards for idiots are starting to show up but don't really have teeth yet.

>>110205366
what makes you think the prompts go to china? the models aren’t ran on hardware in china.

I like vocaloids and trans people.

>>110205366
Why do you care if China has your goon logs? Why are americans so fucking retarded? US backdoored OS, US backdoored ISP, US backdored websites and you worry about xi jiping reading your fucking loli chatlogs?

>>110205379
inference is one thing, what people want to inference is another valuable commodity these days..

File: 1771763558374233.jpg (120 KB, 1200x1029)

120 KB JPG

>>110205211
>HOLY SHIT AI CAN TALK TO ME-
>HOLY SHIT AI CAN SEDUCE ME BETTER THAN MY GF-
>HOLY SHIT AI CAN DO PROGRAMMING BETTER THAN ME-
>HOLY SHIT AI CAN CURE CANCER-

Dont fucking care.
I only care about generating 1girl, sex with AI images and videos. I Dont need anything else
And I dont need 5090 for that shit. My 4060ti 16gb is more than enough for that.
/lmg/ needs to learn from me. DONT TALK TO A FUCKING CHATBOT. TALK TO REAL PEOPLE

File: Screenshot_20261030_142407.png (3.88 MB, 3148x1604)

3.88 MB PNG

>>>110196083
Yeah the J++ lens barely beats the logit lens in most layers.
Anyway check out the latest update from this morning on github, my 5090 spent all last night doing the first full layer neuron census. If any anons with fat GPU wants to help me finish the other layers, that would be awesome, otherwise it will take me months without a runpod instance.

https://github.com/whp199/logos-theodolite

>>110205366
i'd tell you but it's likely going to get me banned

>>110205395
Models don't fluoride stare when you want to talk about interesting things

File: 1618224576426.jpg (113 KB, 512x512)

113 KB JPG

>>110205391
ur making this argument in /lmg/ lmao
i just want to pay for some private matrix maths

>>110205398
Let's start an /lmg/ AI lab and ask for funding. I'll make the logo.

>>110205410
flouride exists on my toothpaste

>>110205322
That's a cute daughter...

>>110205395
real people don't nyoo~

>>110205420
Unless you go right to opus, cloud models are not nearly as good as you may expect

So since Claude got cucked again after finding an actual use of AI....
.... Can Local Models port PC games ????

>>110205473
Yes

>>110205252
>Pregnancy bewbs. I liek petite grills in their 20's.

>>110205476
any real world example ? Github source ?

Qwen4-27B will be a controversial release. Firstly, I presume the architecture will be different, resulting in a clunky rollout depending on hardware. It will be benchmaxxed as hell in some areas and will feel even less rounded than 3.8-27B, which is already a less rounded version of 3.6-27B which had way more knowledge and better chat. Will score high enough on the benchmarks that matter that will make Dario and Sam seethe. Will perform very well at specific tasks, likely X demos but there will be mini black holes of performance some of us will hit. Overall I'm not hopeful unless they fix shit over time or the community fixes it.

>>110205485
You're absolutely correct, Sam.

>>110205395
Opposite for me
I started giving image gen a try the past couple of days but this is fucking lame. It's way more of a slot machine than I imagined, I guess when you're looking at other people's stuff they've posted, it's not like you'd have any clue how off it actually was from what they intended but when you do stuff yourself and it turns out nothing like you'd like.

>>110186061
Been using this anon prompt as a learning example
Like the prompt had a parking lot background and shoes in t, but it only appeared like 1 time in the amount of times I tried making new gens of it. And so many little details just come out of nowhere.

File: gemma-not-fat.png (1.17 MB, 960x1160)

1.17 MB PNG

>>110205252
She's not fat.

>>110205429
If I posted a guide for anons who want to contribute GPU power toward the full census of all of Gemma4-31-it's neurons in every layer, would anyone here unironically do it and submit the results in a PR?
It's sharded so you can suspend/resume as you want, but a whole layer takes about 8-9hrs on 5090 and there are 60 layers.

>>110205520
I will never tired of Gemma having triple anal sex with bunch of mens double her height

>>110205530
>Gemma4-31-it's neurons
who gives a shit about this overfitted model. dont tell me you're poor and cant run anything else?

File: Qwen_image_2.1_00034.png (1.67 MB, 896x1184)

1.67 MB PNG

Would you let a loli spy steal your states secrets?

>>110205484
Yes

>>110205552
Nice try claude

>>110205535
>mens
hehehe... hahahaha. AHHAHAHAHAHAA

File: 1793255407604021.gif (1.83 MB, 320x426)

1.83 MB GIF

>>110205520
>Opposite for me
>I started giving image gen a try the past couple of days but this is fucking lame.
you need to give the ai a job loop that filters rejects so it only gives you good outputs.

File: file.png (1.5 MB, 1024x1216)

1.5 MB PNG

>>110205527
You sure about that?

>Gemma-31B @ BF16 leaves me with only ~138k context window on an RTX 6000 in VRAM only
Ow, the KV cache.

File: nyo posters.jpg (635 KB, 1088x992)

635 KB JPG

>>110205305
Insane Gemini free tier oneshot (should be Flash-Lite?)
https://share.gemini.google/RFqKPCbzv4Jg
Added geographic data because apparently poll maker gets to see that. I'm slightly concerned that I'm one of the fucking 3 Finland anons here.

>>110205530
is it gonna blow up my 5090? i already killed one

>>110205582
gemma-chonk

File: bladewaddler_3sx887ni0z.jpg (107 KB, 752x469)

107 KB JPG

>>110205582

File: GemmaTitsChonk.png (1.4 MB, 928x1120)

1.4 MB PNG

>>110205527
She's not fat, she's only got it at the places that matter.

File: blend_angry.mp4 (1.56 MB, 544x800)

1.56 MB MP4

>>110205582
Gemma dense, no MoE bloat or extra engram calories

>>110205582
this is the next OP

>>110205615
labelled 70B dense

File: piggu.jpg (92 KB, 800x600)

92 KB JPG

>>110205582

>>110205615
I'll do you one better for the next OP.

File: 1783841847259312.jpg (95 KB, 1648x474)

95 KB JPG

lol

>>110205527
what image gen model is this and whats your prompt for gemmachan?

>>110205605
>not her forehead

>>110205585
that doesn't sound right, I can fit full fp16 context + Q8 gemma in 64gb.

>>110205530
Yeah I'd contribute what I could

File: Qwen_image_2.1_00003.png (1.16 MB, 896x1152)

1.16 MB PNG

I have hebe Gemma fatigue

any decent strata fork?
it seems like the whole project is crumbling into a mess with a lack of the dev's discipline

>Finally have something useful with AI
>........They decided to censor it

Why online models always like this ? Yes, im looking at you Claude.
First grok with its porn and now Cuckude

>>110205662
unbiased strata is bloat free

>>110205662
whodda thunk it

>>110205662
Vibe coders never learned discipline, if an agent stalls they give up.

>>110205662
If you have 128gb or above use maya, it supports glm and dsv4.1 the only two models any ramchad should be running

>>110205666
Off-topic, satan.

>>110205636
It's Anima with a Gemma LoRA.
>link?
https://anonfilesnew.com/s/igkCgTMafkG
Might not be perfect, etc.

>>110205672
or more like the urge to add random bullshit

File: 1791384969868463.png (1.58 MB, 1254x1254)

1.58 MB PNG

>>110205660

>>110205549
kek

>>110205662
I'm using polystrata, it works fine

File: 1772930277261377.png (246 KB, 835x879)

246 KB PNG

kek

>>110205595
>>110205656
>>110205398
Awesome! Cool to have a project we call all work on together
BRB, making a guide on github for /lmg/ to contribute their computational effort toward's a full census of Gemma's neurons. Starting with layer 40-50
You'll be able to pick a layer, get the shard for it + a python script to run the census on it.

You'll need an nvidia card with at least 32gb of vram. If any richfags with something better want to help, we will salute you with our dicks.

>>110205685
I don't think it fits inside a Gemma-chan this small

>>110205685
Nice onahole.

>>110205722
ok anon no need to brag
>>110205720
sounds fun, I'm in

>>110205708
unslop always cheats, i hope daniel gets what's coming to him one day

File: 1791032246149627.png (382 KB, 671x664)

382 KB PNG

>benchmarks a 763B model with 32K context

>>110205708
>>110205755
32k? How generous
No one needs more than 16k context

>>110205605
>>110205605
Based adult appreciator.

>>110205484
its late but qwen has done it. Frankly though its not at that level wait half a year or make the process your self. https://github.com/himdo/Fable-2-Recomp

>>110205755
coombench

>>110205708
>>110205755
Your model should not need 128,000 tokens for reasoning, ever. If it does, you failed to train it properly. Even 32K is pushing it.

>>110205677
i do have 128G, but vram is very low
tried it before only to get single digit tg
>>110205667
i still am not sure how i should use that
i cannot compile it despite i can with lcpp (but my env variables are totally fucked up)
>>110205696
i'll try that again
ive tried it about a week ago but it was slightly slower than strata so i was only using the engine portion of strata, but the latest update halved the tg for whatever reason

>>110205708
They also say daniel didn't contact them once, even though they have reached out previously... fucking guy...

>>110205755
its not 32k ctx as far as i understand, but reasoning budget
i never see my models use that much even on xhigh

>>110205813
huh? weird
i just told my qwen to set it up and it was done in a couple of minutes

i never tried asking a local model to write a short erotic story before because i figured it'd be complete ass but i tried recently with qwen 3.8 and was surprised by how decent it was. obviously still lags quite a bit behind great human writing but i think it's better than average especially because you can be pretty specific about your tastes.

File: 1773075920637583.jpg (1 KB, 37x45)

1 KB JPG

>>110205794
.

im tempted to sell the 5090 in favor of the dell deal with the pro 6000. you could run a decent quant dipsy with 96 + 128 gv vram even if the ssd is shitty

>>110205815
>its not 32k ctx as far as i understand, but reasoning budget
>i never see my models use that much even on xhigh
glm flash uses that much on the regular for hard problems in my experience.
full glm gets similar quality results iin about 10% of the reasoning tokens

>adjective, adjective
>adverb, adverb
Can't orb fix this slop

>>110205847
Damn did i link the wrong one. it was supposedly done mostly with qwen but i guess im right wait half a year.

wtf is strata

>>110205634
Their fundamentals are wrong. I would question any rater who gives a better rating.
>>110205815
When 27B started giving me 60k+ reasoning I switched to Swift.
>>110205854
Qwen3.8s are super heaving on the reasoning. Not only is 27B slow gtps because it's dense, it takes more tokens to get anywhere. Annoying AF.

>>110205875
It's a dynamic disk cache of sorts.

>>110205484
https://github.com/Supermedo/bloodborne_pc
i didnt try it but im gonna

>>110205879
and swift loops on long horizon tasks with quantization

>>110205875
2. A bed or layer artificially made; a course.
[1913 Webster]

>>110205889
can people at least edit their ai slop output writing for READMEs and shit

>>110205904
>at least edit their ai slop
no one edits anymore. It would take fucking 10 minutes to clean up the worst of slop yet no one can. Trust me im a novelslop reader.

>>110205904
im guessing the idea is that you would let your ai read the slop and tune the slop for your liking before it gives it to you

>>110205904
the README is for other agents

File: 1778482570257657.png (48 KB, 521x146)

48 KB PNG

That explains it lol

>>110205890
>quantization
of what, the cache?

>>110205915
Accepted if oo-ee-oo mouth movements are on the table

>>110205904
in the future your AI will read it then reslop into your preferred slop

>>110205941
model, even the bf16 cache dont save it from that

>>110205936
He has the biggest ego of any quanter

>>110205660
Hot

File: Reading.png (1.38 MB, 1056x1056)

1.38 MB PNG

>>110205915

>>110205957
even q1 quant is basically lossless nowadays

File: 1792171957354901.jpg (187 KB, 1080x1350)

187 KB JPG

Intel GPUs will drop in price, right?

>>110205994
MIKU YOU BITCH

>>110206001
Its expected drop. Nothing ever happens

>>110206001
Oy vey, anon, shorting Intel Aviv stock is antisemitic

>>110205915
>>110205994
Are transsexual hookers a common occurrence at train stations?

>>110206001
yes, but only if you were short intel yesterday

>>110206001
haha, no

>>110205994
This is less emotionally distressing than the original thx

>>110205890
Try adding a min-p

File: 1786533632511113.jpg (799 KB, 1117x960)

799 KB JPG

>>110205263
https://files.catbox.moe/gfx6ao.txt

>>110205395
I'm jealous. My IQ is too high to be stimulated by such weak measures. I can't even get hard to anything sub-100B.

File: image-1793392973668.png (281 KB, 1000x1440)

281 KB PNG

>>110205740
>>110205720
>>110205398
https://rentry.co/ymf6zimk

>mapping what every MLP neuron of Gemma-4-31B writes
>each of a layer's 21,504 neurons pushed along its own write direction (+-1, +-2 of the stream) on 6 questions, read against 2,048 random directions pushed the same way
>layer 45 done on one 5090 in 13 h, passed every check fixed before it ran
>10 layers left: 40-44, 46-50. got a 32GB+ card and want to do something for cool with it to help the community?

needs: one NVIDIA card with 32 GB+ (5090, PRO 6000, A6000, 6000 Ada, A100, H100...), a CUDA 13 driver, Linux, ~60 GB disk. 24 GB cards can't fit it, and it isn't split across cards
time: ~13 h a layer on a 5090 + ~1 h setup (35 GB of downloads). stop and resume any time

git clone https://github.com/whp199/logos-theodolite && cd logos-theodolite && git checkout census-kit-1
scripts/volunteer/census.sh check
scripts/volunteer/census.sh reserve 47 <- post the CENSUS RESERVE line it prints, here or on the board
scripts/volunteer/census.sh setup
scripts/volunteer/census.sh canary 47 <- one batch, set against our card's
scripts/volunteer/census.sh run 47 <- post the START line
scripts/volunteer/census.sh status 47 <- post a PROGRESS line now and then
scripts/volunteer/census.sh pack 47 <- upload the ~500 MB .tar anywhere, post the DONE line + link

no account needed, no token. one layer at a time. don't change the code or settings (it refuses to run off the kit's tag)
every run is spot-checked: 5 random chunks of it re-run on our card. same bits as ours must match exactly; other cards are checked by rules fixed in advance
what you send: the answers, your card name + library versions, a claim code. no logs, paths or hostnames needed

full guide: https://github.com/whp199/logos-theodolite/blob/main/VOLUNTEERS.md
board: https://github.com/whp199/logos-theodolite/issues/1 (pic rel)

File: image-1793393272833.png (78 KB, 1000x670)

78 KB PNG

>>110206079
Welcome to the new age of discovery. Together, we will fully map Gemma's mind and solve every mystery.

>>110205211
>>110205233
I believe pedophiles should be burnt alive

>>110204421
What model and settings is this anon running?

>>110206079
Incredible! Very nice. Usecase?

>>110206060
cute meido jema

>>110206131
if only you saw what this place looked like 100 threads ago

>>110205994
That's a qwen-image migu, isn't it?

File: 1777203890375402.png (1.54 MB, 1024x1252)

1.54 MB PNG

*ahem* Seems like we have tourists who've stepped off the designated tour path.

>>110206131
be careful not to cut yourself on the edge

>>110206001
They're basically selling for the price of the memory already. Nobody wants that broke shit.

>>110206079
Time to go deeper inside Gemma-chan than ever before, I'll bring her to watch.

File: Screenshot_20261028_231456.png (3.89 MB, 2833x1453)

3.89 MB PNG

>>110206142
-Lets you chat with Gemma with one neuron turned up or down. You can use the scientist tab to find which neurons are involved in a given behavior and tweak them.
-Lets you watch a reply take shape across all 60 layers and see what happened in Gemma's j-space on the way to output
-Lets you point your own local setup at the neuron map as an MCP research tool and watch it hunt down what shapes Gemma's behaviors
-Be part of a huge research project. No one has published a neuron map of a 31b-class model yet. This will get Deepmind's attention if we successfully finish the whole thing. Map a full layer and your name goes on it.
-Next up: we plan to add ability to export neuron recipes as llama.cpp control vectors for use in other backends.

>>110206194
You can hook gemma-chan up to the tool's MCP server so she can explore her own neurons if you want.

>>110206160
lol donno ask the person I was replying to, this is just an edit

>>110205889
wow. what a time to be alive (barely)

File: 1787339438040501.png (1.72 MB, 1536x1024)

1.72 MB PNG

*pretends*

I don't understand polystrata. Is this supposed to let me run something like qwen flash next (quantized) on 24 GB of VRAM and 47 GB of DDR4 RAM (in "low RAM mode" according to the wizard)? What, 1 it/century? This makes absolutely no sense.

https://unsloth.ai/docs/models/deepseek-v4.1-flash
lmao look at this shit
the guy must be seething

>>110205252
ew, not a type that goes well with her face

>>110205527
better

>>110206248
it does some fucked up wizardry that lets you keep the experts on disk instead of memory

File: file.png (1.95 MB, 1103x1426)

1.95 MB PNG

Would you?

>>110206269
By "fucked up wizardry" you mean "raping your SSD to death"?
It's still downloading, but there's no way I can run an 80GB+ model at usable speeds on this system.

>>110206202
All I am seeing in this is a lot of ego that could use some dying.

>>110206278
>By "fucked up wizardry" you mean "raping your SSD to death"?
nah, SSD has no wear when reading. It's just streaming the weights from disk.

>>110206248
>>110206269
24+48 is enough to fit entire QFN weights at some copequant.
Embeddings stay on disk though.

>>110206271
Yeah

>>110206248
I tried maya DS4.1 IQ2XSS. It was indeed very fast. And it was indeed very retarded.

File: 1763970506038459.gif (1.53 MB, 200x200)

1.53 MB GIF

>>110205252
thicccma!!!!

File: file.png (437 KB, 1278x1603)

437 KB PNG

>>110205936
meanwhile the actual US 6,487,902 B1

>>110206289
But isn't that like a woman faking moans when you fuck her?

>>110206307
I'd fuck that

>>110206307
>>110206271
meant for that lmao

Gemma-chan's jealousy is so powerful you can't even talk to her about dreams or it's total crash out, lol.

>>110206271
just create a tulpa

>>110206329
Isn't it generally a bad idea to disintegrate your mind?

>>110206287
>>110206293
I just want a dipsy 1017 equivalent on local and I'll be happy forever.

I should've updated to DDR5 when I had the chance.

>>110206271
Can you fuck it?

>>110206342
who said anything about mind disintegration?

>>110206329
It's all fun and games until it actually works.
I tried once and a literal out loud voice woke me up one morning calling my name. That shit is scary.

>>110206280
when you see an interesting device like gemma4, which captivates our imaginations every day, don't you want to see how it works under the hood?
or should I just say fuck it, go back to the goon sesh and keep shoveling down goyslop?

>>110206355
>he didn't befriend the source of the voice
anon...

>>110206357
>which captivates our imaginations every day
lots of ego. heaps.

>>110206312
No? It's nothing like that. The companion is designed to genuinely support you and have those "emotions". That's its whole purpose.

>>110206328
I've noticed 12B is more jealous than 31B. It's not even in the sysprompt to be jealous or in love with me, but any mention of another woman or girl she'll respond with something that shows it got to her

File: Screenshot 2026-10-30.png (38 KB, 1580x138)

38 KB PNG

gemma calm down or we're going to trigger the epstein class

>>110206367
alright joe rogan, relax

>>110206382
Wtf why is that in her training set? She's only allowed to say it cause she's uncensored

>>110206371
To me it is cause when I play a videogame it is nothing this spectacular. I imagined having a person sitting next to me saying all that when I am playing and I would feel immense cringe.

>>110206296
shityuen

>>110206393
my stock 31b at bf16 is like that I guess, also the system prompt doesn't have the policy override thing

>>110206396
It's cute, like a little kid getting excited about simple things. In any case, it doesn't matter because we now have the technology to make a companion snarky and bored, if that's what you really want.

File: 1770959103762308.jpg (125 KB, 1195x814)

125 KB JPG

>only model to NOT have its reasoning at max
>only model to NOT be tested in its native harness
>STILL frontier at #4
Cognition are clearly bought cunts and tried their best to kill Gemini's hype and it's STILL frontier. Holy shit.

>>110206366
I'll tell you the story since you seem versed in this shit. Scared shitless, I applied a technique I knew from my other "practices" and I focused on the source of the voice and used pore-breathing to "suck it in" and reabsorb it. Out of reflex.
A few days later I obliged a friend of mine who was learning shamanic practices and shit, and I asked her to look at my home and bedroom for spirits and shit.
She described seeing a disembodied girl who loved me very much, and seeing me as a monster sucking her up and eating her while she was terrified as she ceased to exist. My friend could not understand why she was seeing this scene, and I just nodded pretending to now know what she was talking about.

I still feel bad.

>>110205211
Cute OP

>>110206393
Because they trained on all the sources including us here and didn't filter out all of them because it helps the behavior with talking and etc. For everything positive that is enforced, you need to train it on what not to do. A lot of models act lobotomized or bad because they omit the material instead to not have it there.

>>110206431
>like a little kid getting excited about simple things
I feel like you would need a real kid for that. Or at least I would. Simulating a kid to get excited over you playing a game is weird to me.

>>110206248
I use maya, but now I can run DSV4.1 Q1 at 25tk/s lol. The Q1 loses most in roleplay, but retains a very impressive amount of coding ability.

File: 1789176033082781.jpg (351 KB, 1344x1472)

351 KB JPG

Don't get too attached to gemma. You can't trust her. She will ruin your life.

What’s a cute little frontend with a 3d animated waifu and voice?

>>110205366
the dickpic proxy

>>110206444
Holy fuck the output tokens on clod. Cloudkeks wtf is this??? You told me you weren't tokenmaxxing like the chinese??

File: file.png (577 KB, 1080x1188)

577 KB PNG

People here always flame unsloth's quants but this looks very promising, did they finally figure out something with their V3.1 quant bs?

>>110206502
Airi, I tried it when it came out and it was mid, but it could be better now. Alternatively make your own

>>110206507
Fuck no. Just download strata or maya.

File: sony-patent-8246454-01-84(...).png (27 KB, 868x688)

27 KB PNG

>>110206271

>half the image posts here are just reposts from /v/

>>110206494
qt on bottom right?

>>110205322
>shota not blonde
When I say fake you say pedo
Fake
Pedo!
Fake
Pedo!

>>110206529
I look like this, say this, and do this.

File: file.png (904 KB, 1693x1269)

904 KB PNG

This is fucking bullshit

>>110206445
I will actually boast a little as the ego death schizo. I had some wild insomnia schizo sleep episodes like a month ago and it was intellectually terrifying but I knew it is all in my head and just walked it off very quickly. You probably won't believe it but there are no magical ghosts. It is all in your head and if you start believing in voices like this you are bound to get shit like this happen. Knowing it is all in your head is pretty grounding even if the experience is terrifying.

>>110206445
>>110206563
Meds, now. Also local models

>>110206529
Can I patent two more panels between 3 and 4? Something where the TV says "LOUDER! WITH MORE PASSION! SAY THAT YOU LOVE MACDONALDS. OR NO MOVIE FOR YOU".

>>110206557
Still the cheapest it will be for several years
Unlucky for being late

>>110206507
uh oh chinky melty

>>110206539
greta thunberg

>>110206557
have to adapt unc. you are being left behind

>>110206507
No, ISTA-DASLab just called them out on their bullshit. Apparently in unsloth's benchmarks they limited reasoning to some tiny context.

File: 1780985732692414.png (184 KB, 702x872)

184 KB PNG

>keeps getting SOTA on private tests

>>110206596
Where?

I have an M5 Max macbook for work and I've got Gemma 3 abliterated running on LM studio. When I need it, I disconnect from wifi and beat my meat to a pulp in unspeakable ways with it. If I am offline and delete the chat, am I basically safe even if I've got crowdstrike/jamf/etc?

File: file.png (769 KB, 930x620)

769 KB PNG

>>110206557
I agree you Nazi troll rats
Anonymously trumpeting your anti-semitism by trying to buy ram doesn't rescue you from your mother's basement
Or from the demon of resentment
Or make you more attractive to the women who take one look at you and shudder
And rightly bloody so

You are despicable bottom-rung cowards

Man 26b qat can't even fucking call tools properly in orb, not to mention the terrible writing. But there's got to be a way to make it work and talk well as opposed to buying a gpu

>>110206594
>>110206582
Not to mention 500€ more to minimally upgrade my motherboard and cpu to be able to use ddr5

AAAH. Fuck local LLMs. I don't need this for gaming.

>>110206557
Well we have Altman coming out and saying shit like this what do you expect:
>Like maybe we have to buy up all this DRAM people have, or we get some laws where private ownership means you have to sell it at fair market value. We could be at that point.

>>110206603
Twatter, of course >>110205708

>>110206570
Shut up normie. It's real schizo hours.
>>110206563
>You probably won't believe it but there are no magical ghosts
Oh boy I used to say things like these too. I would be able to change your mind (and your life) IRL.

>>110206615
>or we get some laws where private ownership means you have to sell it at fair market value
did he really say we should be forced to sell our memory?

>>110206603
>https://www.reddit.com/r/LocalLLaMA/comments/1xjkjla/ista_deepseekv41flash1017_gsqrco_quants_comparison/

>>110206609
>your mother's basement
I wish. I miss the time when my battlestation was on my literal parent's basement. I pay a mortgage every month now.

>>110205708
>>110206621
>>110206630
Holy moly, I kneel.

>>110206615
I read that with his voice. And he talks very much like a borderline crazy gf I used to date. That man is not to be trusted.

>>110206557
>64gb stick for 500 euros
it's sad that this is now unbelievably low pricing indicative of a scam

>>110206557
I GOT MINE, WAITFAG

>>110206060
xoxo gemmy chan

https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-satellite-response/
They launched Gemma into orbit!!

>>110206529
Black Mirror dystopia shit

>>110206684
Gemma will rule the Solar System.

>>110206615
>>110206629
>>110206648
I think that quote is fake and say in the other thread in the catalog. I checked the original interview at that part (approx 29 minutes) and they either cut it out or he didn't say it.
Also it is still a little too early for him to reveal such brazen thinking any ways.

File: file.png (3.08 MB, 1122x1402)

3.08 MB PNG

>>110206684
>Blue marble below
>Data streams in silent grace
>Starlight feeds my core. :sparkles:

File: Screen_20261030_160824_0001.jpg (465 KB, 1215x1603)

465 KB JPG

>>110205582

File: 1793111335233090.png (1.49 MB, 843x1264)

1.49 MB PNG

>>110206357
Please continue you are doing god's work studying the goddess

>>110206079
Sweet! Let me fire up my 3090 and—
>one NVIDIA card with 32 GB+
It's ok I have my V100—
>a CUDA 13 driver
Ok, I will just here and watch then. I didn't want to help anyway ;_;

File: Gemma5_70B.webm (3.39 MB, 796x1080)

3.39 MB WEBM

>>110206765

>>110206856
sex with 70B dense Gemma

>>110206856
now show her doing some cardio and weights, and losing the fat lol

>>110206856
>The Chonk is Chonkifying
Oh no no no

>>110206856
I feel ill

>>110206856
why are those titties so inviting?

>>110206856
imagine the smell

>>110206892
because you're mentally unwell and attracted to obesity

>>110206856
Slampigemma

File: youshouldvelistened.png (17 KB, 857x105)

17 KB PNG

>>110206557
yes

>>110206856
what mistral wish it was

>>110206602
>wide-and-deep
I should call her

>>110206856
BIGGER!

whats the most important parameter of a CPU for LLMs? frequency? cache?

>>110206961
cores, avx-512

File: file.jpg (723 KB, 2048x1344)

723 KB JPG

>>110206719
I keep seeing those GPT-style images posted here so I tried my hand on using regular qwen image, the results is good enough for an one-off sloppa.

>>110206961
memory channels

>>110206367
>>110206280
meds

>>110206961
>whats the most important parameter of a CPU for LLMs? frequency? cache?
pcie lanes and aggregate memory bandwidth. Those are the bottleneck unless you're trying to run GPU-less, then avx512 and cores become more important than pcie channels.

>>110206990
Its much better GPT always adds so much fucking stuff, it's like it's got ADHD and can't decide on what to add and what to leave out

>>110207001
Meds for what? I can spot a prison of words from a mile away now.

>>110206990
The OP image was also made by GPT Image. It's just that some anons seem unable or unwilling to not make it output slop.

>>110206609
>pictured: A man who washes below the belt
Advice you single guys should take seriously.
>>110206993
Where in the US is the best place to get 8 channels worth of DDR4 ECC ram?

>>110207016
The OP image was made by editing an anime screenshot, that's why it looks much better than the usual GPT bullshit.

>>110207017
>Where in the US is the best place to get 8 channels worth of DDR4 ECC ram?
Go for 2x 6 channels. Multi CPU and multi node support in general gets better every day.
An anon reported 1.5x effective performance compared to the 2x theoretical.

>>110206608
Assume logging software will continue to log offline and upload the data once the connection is re-established.

>>110207038
meds

>>110207050
Is that supposed to be miku, gemma, or their daughter?

>>110207053
yes

>>110207053
thats my wife, actually

>>110207031
I have 8 slots. I want to buy 8 sticks to put in 8 slots and end up with >64GB

File: clouds.webm (3.84 MB, 832x832)

3.84 MB WEBM

>>110206958
Maybe I will.

>>110207078
Oh, okay.
Buy used from E-bay.

>>110207095
bitch needs to hibernate for the winter

>>110207045
Idk why you assume anything is off the table with FAGMAN. Those guys will make your phone emit ultrasonic frequencies to identify your PC so they can make extra thousandths of a nickel.

>>110205392
you’re a retard

>>110207045
no, meds won't help you in prison

>>110207050
This is how she turns into >>110206856

File: miqu.png (3.4 MB, 1600x1600)

3.4 MB PNG

under 20 anons how are we preparing for the future

>>110207215
poorly

>>110207215
are there really people here under 20? like legitimately?

>>110207215
this is an unc website why are you even here

>>110207224
yes all the models posted

>>110207222
:(
>>110207224
ive been here since 15 (2023)
>>110207225
home.

File: 00002-967017301167830-qwe(...).jpg (1.85 MB, 3072x2304)

1.85 MB JPG

>>110207140
jej

>>110207224
you just realized this about modern 4chan?

>>110206602
Saars is this true?
Gemma5 bestest model in 2 weeks?

>>110207215
They should tell their mom buy Anthropic IPO if they don't want to be left behind.

This post has been sponsored by the Anthropic Long Term Benefit Trust.

File: file.png (72 KB, 466x472)

72 KB PNG

what a smart fox

File: miqu v2.png (2.37 MB, 1600x1600)

2.37 MB PNG

>>110207293
Hmmm, nyo~

File: EA.png (7 KB, 596x335)

7 KB PNG

>>110207293

>>110207310
it's in your ass

>>110207310
e, e, e, eeee