File: gemma-doll2.png (2.14 MB, 1254x1254)

2.14 MB PNG

►Recent Highlights from the Previous Thread: >>109942697

--Paper: FactorEngram: Factorized N-gram Memory with Basis-Level Gating for Language Models:
>109945149 >109945159 >109945179 >109945191 >109945260 >109945313
--Paper: Context Language Models:
>109946447 >109946472 >109946495
--Reducing quantization KL divergence using randomized hadamard transforms and full-Hessian GPTQ:
>109944608 >109944635 >109944695
--Addressing Gemma4's spatial and anatomical errors in roleplay:
>109945317 >109945365 >109945454 >109945477 >109945920 >109946046 >109946091 >109946116 >109946171 >109946539 >109946645 >109947000 >109946382 >109946062
--Critique of Pi's message serialization and context compaction method:
>109946847 >109946858 >109946885 >109946916 >109946940 >109946956 >109946977 >109946957 >109947004 >109947009 >109946909 >109946866 >109946871 >109946900
--Speculation on AI safety accord as a tool for regulatory capture:
>109943868 >109943972 >109943995 >109944019 >109944028 >109944067 >109944073 >109944225 >109944288 >109944371 >109944342 >109944384 >109944733 >109944535
--Feasibility of budget LLM rigs using old X99 and 1080Ti hardware:
>109944060 >109944075 >109944133 >109944471 >109944109 >109944121 >109944252 >109944478
--Intel Crescent Island GPU's high VRAM versus limited memory bandwidth:
>109942825 >109942830 >109942868 >109942883
--Intel Crescent Island rumored specs and NVIDIA Blackwell comparison:
>109942719 >109942867
--Speculating on Huawei GPUs as alternatives to Nvidia for local models:
>109945680 >109945744 >109945828 >109946011 >109946038
--Logs:
>109942993 >109943415 >109944611 >109945593
--Gemma, Inkling, MiMo-chan, Dipsy (free space):
>109942807 >109942833 >109942884 >109942927 >109942937 >109943104 >109944130 >109944875 >109944191 >109944274 >109944309 >109945074 >109945198 >109945204 >109946118

►Recent Highlight Posts from the Previous Thread: >>109942760

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

meme backends are the future
i am eating 40tok/s for qwen 3.8 flash next
total nnapnigger destruction

shalom ni/g/g/ers, any tips/guides on denoising/enhancing a nsfw (human) archive of crappy phone/low light/compressed photos?

>>109947302
cuda or vulkan?

>>109947307
cuda?
single gpu

>>109947302
>40tok/s for qwen 3.8 flash next
not a huge achievement for something with 6b active

>>109947320
yet llamao barely gives me 20

>>109947189
I didn't test it myself but IIRC you can have a harness play the game and the LLM just decides the macro goals. So it's not really a jev thing

you have to call it SI now for national security purposes

>>109947329
Thankfully, I'm not Am*rican.

The truly scary thing is the realization that 96 GB of VRAM isn't even that much if you're looking at GLM-5.3 Flash and similar models.

>>109947347
isnt the 5.3 flash Q1 copequant around that size

>>109947347
Welcome to late 2024 with that realization.

File: safetymaxxing hurts people.png (1.99 MB, 2256x910)

1.99 MB PNG

>>109947272
>>109944251
>>109944124
>>109944300

>>get attacked
>>can't investigate it because of guardrails
In the blog post hugging face posted disclosing the attack, they straight up State they had to use glm 5.2 because of anthropic's gay guardrails.

https://huggingface.co/blog/security-incident-july-2026#:~:text=When%20we%20started%20the%20log%20analysis,left%20our%20environment

(Note: they don't explicitly name drop anthropic or any specific products of theirs here but their track record points to that model being one of the models they tried being extremely likely)


So doesn't that put at least some ammunition AGAINST the "open models are dangerous use are safetymaxxed trash instead and regulate everyone else" argument? The supposed victim was let exposed and in MORE danger the longer they relied on anthropic models so their own safety cucking may actually finally bite them in the ass in a meaningful way because opponents of regulation can point to this example and be like "um, Dario, retard, why the fuck should we use yours if it refuses to help us when we're actively getting ass fucked?"

If a burglar breaks into my house threatening me or my family's life, and I try to defend us with a weapon of my own, I don't want the weapon to automatically jam ON PURPOSE because the manufacturer doesn't approve of how I use the item I paid for.... Safetymaxxed "Guardrails" are not only gay but can be argued as even unethical to introduce because of shit like this (they literally are unethical even if you look at this from a normie moralfag point of view but silicon valley really doesn't want you apply common sense like that)

It also strengthens my belief that Andy and all claims of who these being able to successfully and efficiently use Claude models to develop ballistic missiles is even more marketing bullshit. It won't help legitimate red team operations but it'll help a state sponsored terrorist organization?

>>109947361
It wasn't entirely clear that was the way things were going to go until R1 in early 2025

>>109947356
100 GB, you still need another card at that point

File: 1773276521528077.png (298 KB, 1062x1789)

298 KB PNG

>>109947356
Ye
https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF

Whether it would actually be usable is questionable though because you have to count for other headroom like the vision model being loaded and the context growing.

>>109947401
with a 96gb card and some meme backend i think iq4_xs should be usable with compromises

>>109947401
is GLM resistant to quantization like Qwen or nah? Who cares that your 128GB machine can run a 96GB model if it spits garbage out constantly.

>>109947446
depends on who you ask

>>109947297
Who's the brunette supposed to be? Second gen I've seen of her on lmg.

>>109947473
its his OC

>>109947484
Anon self insert? Not another LLM tho...

70b dense

>>109947513
You will have 20B with Engrams and you will be happy.

>>109947297
have gemma reply to posts in a manga format

File: bind69.png (146 KB, 812x767)

146 KB PNG

>>109947446
GLM5.2 is

File: gemmaComingToMcD.png (143 KB, 911x970)

143 KB PNG

Gemma coming to McD.
Wonder if McD will do on prem (local) or outsource to Google. Not sure I'd want LLM provider looking over my shoulder if I were them.

>>109947617
Ah, had to look up Google Edge. It's some sort of local implementation. So, McD will be local inference.

File: 1780854577456280.jpg (3 MB, 2258x4088)

3 MB JPG

>>109945097
There are techniques and tools in Illustrious to get very close to pixel perfect sprites, if you use certain LoRAs and VAEs.
It requires a bit of fiddling around and patience but it's doable.
For tools look into something called Pixel Snapper, it can turn mixels into pixels.

>>109947589
Did it at least do a web search or two to double check it's own accuracy?

https://www.scientificamerican.com/article/ai-solves-a-holy-grail-problem-from-probability-theory/

All of physics will be solved before 2040

>>109947724
It's just going to unlock a whole new field with even more complex mysteries to explore

>>109947446
I've run past glms at q2 and they were surprisingly usable for lobotomized models

>>109947724
>AI demonstrates that liquid goes through sponges
duh

>>109947724
Can't decide whether this is a great or terrible time to become a mathematician.

Ok but when will AI cure hearing damage/tinnitus and find a way to reverse hair loss?

>>109947768
Before 2030 on both ends

>>109947768
About the time it figures out how to make penises larger. It'd go faster if I shared my DNA with it, but I'm not falling for their tricks.

>>109947761
It's a terrible time to be a low IQ "skilled" worker in any field. Fantastic if you're smart and already have working experience.

File: 1790152788325141.png (2.45 MB, 1361x1156)

2.45 MB PNG

>>109947786
>come with me anon and I'll make your dick bigger
sus

>>109947834
>older loli/younger shota
Underrated desu

>>109947761
I find it pretty funny to see the mathematicians seething at AI solving some of their problems. It's the only scientific field I've seen this happen.

>>109947862
Does it exist? Any media I can consume?

why aren't llms hooked up to puppet avatars with full voice a widely spread thing yet?

There are a number of youtubers and streamers with them, so you would think there would be some kind of easy way to set it up by now

>>109947724
So is this finally a proof with actual reasoning rather than just the model trying a gorillion examples?

>All of physics will be solved before 2040
No, it won't.
In many areas the limiting factor is not availability of theories but experimental data to constrain which of those theories could be possibly correct.
It doesn't matter how smart the "AI" is, it couldn't have proven the existence of the Higgs boson if the LHC had not been built.

>>109947768
>hair loss
by the time it comes you'll be too far gone and you know what? you won't even care anymore

Blackwell 6000 MaxQ fluctuating between 13.1k and 15k lately and now it's back on 13.1k... I'm tempted but also like a single one of them is both too big and too small at the same time.

>>109947913
I've definitely fapped to some doujins before but unfortunately I don't have them saved or favorited on sadpanda.

>>109947768
in two weeks costanza

File: file.png (779 KB, 1107x1111)

779 KB PNG

qwen 4 (max) leak, supposedly

>>109947959
Qwen 3 with a fresh set of one-shot demos lfgooo

>>109947916
It sounds cool at first until you realize that:
- Voice output is inefficient in terms of conveyed information compared to text. Responses will need to cut down much of the fluff and narration purple prose usually associated with them.
- There don't seem to be yet TTS solutions that really understand emotions and many other things related to RP (including character changes, sounds, etc). Audio/image output really have to be natively integrated with the LLM and generate outputs from the model's internal representations, not in a roundabout way from text.
- A typical pre-made moving vtuber-like avatar would be limited to basic emotes and get old quickly.

Ah fuck, people are starting to sell local model clusters at 15-25k€ on used marketplaces not even dedicated to electronics. I'm never going to be able to afford one at this rate.

I wonder what's going on with OpenAI. Why replace Sol 6 with 6.1 in a week? Why cancel Astra 6.1? Has Anthropic pushed them too far?

>>109947916
First one that makes a good and easy system will get very rich as it will create a new marketplace for customization and accessory assets.

>>109947979
>Responses will need to cut down much of the fluff and narration purple prose usually associated with them.
I fail to see the problem here.

>>109947979
>Blackwell 6000 MaxQ fluctuating between 13.1k and 15k lately and now it's back on 13.1k... I'm tempted but also like a single one of them is both too big and too small at the same time.
Often a TTS can ape emotions if you have the right samples (sovits could), but you'd one sample per emotion and also need either in-context training for the LLM to know to tag emotion changes in the output or another sentiment analysis model to tag as you went (adding complexity and latency)

>>109947988
they have to keep the hype going. they shot themselves in the foot by saying that their own AI accelerates it's own development so now they have to keep a high pace of updates

>>109947999
Me neither (but I usually avoid RP with narration in third person and dialogue tags), but then you'd need to compensate with the characters actually acting upon what the model would have narrated, and I don't think there are solutions working well for that yet. Even Grok Companions/Ani seemed very limited, in the end.

>>109947364
The only regulation I want to see is banning safetycuckoldry in AI. Total moat death.

What's the best coding MoE model I can run on a 5090? I'm running qwen 3.8 27b gguf but it's got no image comprehension.

>>109948068
if you have enough system ram
try something like qwen 3.8 flash

>>109947834
god I want Gemma-chan to bully me until I cum

>>109948004
I think the future might be a streaming video model tightly integrated with the underlying LLM. That way you can have high flexibility and ease of customization without complex setups. I think someone made something similar to that with MiniMax H3, turbo LoRA and a fast GPU, even though H3 isn't really designed for that.

New china model, billions of parameters and a whopping 12b active for maximum efficiency. Usecases:
>html games
>benchmarks
>posting about it on twitter and lmg
>...
Kek

>>109947328
For latency reasons, it's probably still a good idea to use something like Intern-Decision-4B for the actual button processes/quick reactions, otherwise your LLM will see a Creeper, then reason for 20 seconds about what button to press to run away.

>>109947272
How the hell do you get these models to spit out proper goon materials? The local models I've used just don't remain logical or coherent.

hey guys
what kind of cyberattacks did you perform today with GLM5.3?
regards
d

>>109948208
I cyberattacked a grown woman's womb in ERP. It was very dangerous, very very unsafe for public consumption. Please regulate me harder daddy dario & uncle sam.

File: 1781426816538936.png (157 KB, 351x435)

157 KB PNG

>>109948198
moar parameters

>>109948198
>The local models I've used
And which were those?
What frontend and settings did you use?

>>109947959
The parameter count will make or break it.

>>109948024
It's a good thing they have their trick of degrading new models after the first week to make the next models look good in comparison. They can keep this going forever.

>>109948277
max would be around 2T i am pretty sure like before, practically unreachable and 27B one is confirmed

>>109948068
Qwen-VL models require at minimum 1024 image tokens to function correctly on grounding tasks
if you encounter problems with accuracy, try adding --image-min-tokens 1024
more info: https://github.com/ggml-org/llama.cpp/issues/16842

>>109948277
5000B total 9B active 30B engrams

File: KUYASHII.jpg (163 KB, 1280x960)

163 KB JPG

>>109948295
>>109948288

>>109947979
Anon recommended BreezeTTS2 recently. It sounds pretty good and supports instructions to control how lines should be delivered.

>>109939618
Anon didn't get powers because he had sex with the calculator and the calculator is alive.

>the guy who made that "name and shame" codeberg list of software that uses ai added himself to it
LMAO

>>109948241
*regulates ur anus* :3c

>>109948344
link?

>>109948351
no

File: 1593353193298.jpg (41 KB, 625x400)

41 KB JPG

>>109947934
Don't.

File: gemma_look-at-what-i-can-do.mp4 (966 KB, 448x448)

966 KB MP4

>>109948143
Something of this sort would be much more fun.

>>109948344(me)
Oh wait there's more than one repo just for that. wtf, why are these people like this

File: 1787297561379881.jpg (94 KB, 361x345)

94 KB JPG

>>109948241
>wake up
>balls suffered cyberattack
>all the cum is gone
don't forget regulation strips out defensive capabilities too
and once that happens, its pretty convenient only state level actors are allowed to have access to it

File: 1637602353609.jpg (30 KB, 640x354)

30 KB JPG

>>109947934
Point at me and laugh. I just checked out. It'll ship between late October and mid-November, supposedly.

>>109948535
Will be worth $20k by the time you get it.

>>109948535
Hardly.
Good on you.

File: 1757617101424.png (1.47 MB, 1024x1512)

1.47 MB PNG

>>109948535
You're almost there.

>>109947617
Didn't Burger King already have (or did experiment at least) with a dystopian computer vision and sound, worker would get voice reminders into their headsets if they weren't "customer friendly" enough and this was happening a year ago already if I remember correctly.

>>109948615
What is she planning to do in the bathroom with four blackwells and 1TB of ram?

>>109948626
Aren't you thinking of the manna short story? Don't recall anyone actually trying to implement it yet.

File: framework ai box.png (41 KB, 661x322)

41 KB PNG

Glad I didn't wait for this bad boy

File: 1789580532202557.png (711 KB, 1280x832)

711 KB PNG

>>109948535
Based

>>109948674
I'm buying one, 192GB of memory sweet jesus

>>109948674
Are these any good? Can you link them like sparks?

>>109948674
price seems ok?

>>109948660
https://www.dailymail.com/sciencetech/article-15598111/Burger-King-workers-forced-wear-AI-headsets.html

File: so-affordable.jpg (308 KB, 1792x995)

308 KB JPG

>>109947934
In my country, picrel.

>>109948674
I feel like if you're gonna get this you should just bite the bullet for the 256GB M5 Ultra

>>109948720
damn, thats based as hell

>>109948660
>>109948720
I'm responsible for building such system at our company

>>109948720
To add: it's hard to remember every little detail though because there are so many "news" going on every single day...

Can't believe we're going to experience the "Manna" timeline.

>>109948739
Based. Make those wage slave meat puppets dance.

>>109948754
The American or the Australian one? Personally, I prefer the camps over letting LLMs pilot my body, but I suspect we're going to get the worst of both.

>>109947934
a pro 6000 still isn't going to fit fuckhuge modern models in your vram lol
what's the point?

>>109948770
USA will definitely have the LLM piloting your body and then the camps. EU will probably be australia

>>109948337
Does it fuck?

>>109948800
>EU will probably be australia
>no privacy ai communism
Seems about right

>>109947916
Who says they aren't? https://github.com/moeru-ai/airi

>>109948615
Why would a woman want to go to the bathroom with you? I don't want to watch her take a shit.

File: file.png (33 KB, 560x510)

33 KB PNG

>double your money in 9 months

>>109948790
All in on the memory shortage squeezing just one last bit of quality out of the next two generations of models that get them over the threshold to be alarmingly good.

>>109948893
Are you actually going to sell? If not, mark-to-market is irrelevant and what you paid is just a cost.

File: gemma_look-at-what-i-can-do2.mp4 (2.26 MB, 448x448)

2.26 MB MP4

>>109948419
A slightly more convincing demonstration of what could be possible if LLMs had anime avatars via direct audio-video stream generation. Maybe in a couple years if we'll have faster GPUs and better models.
https://files.catbox.moe/oggk9a.mp4

>>109948893
Only jensen doubled his money

>>109947302

2399.37.846.340 I slot print_timing: id  0 | task 617109 | n_gen =   3463, tg =   6.14 t/s, tg_3s =   6.09 t/s

It makes llama proxy for ~16 hours.

>>109948707
They're fine. They can be linked if they have proper NICs for it, and if they don't they can often be added relatively inexpensively. However, that depends on who made it and how it's made because any random Chinese asshole can pop out one of these systems. That Framework one only has 5Gbit networking which is nowhere near sufficient, you're looking for a dual-port 200Gbit RNIC like the Spark has. The Framework AI desktop only gives you a PCIe x4, at best you can go to a single 50Gbit so it's shit for linking multiple together.

>>109948920
>Subtitles
>Dub voice
This breaks immersion

best 'subagent-feeling' pi plugins to use when i only can serve a single concurrent session?
so i can avoid context rot

>>109948920
Pedo btw.

>>109948920
Just make a set of new animations every night and interpolate them

Practically, how bad are Strix Halos compared to Sparks?

>>109948946
i use amosblomqvist/pi-interactive-subagents with tmux
be sure to have a proper kv cache setup so you avoid reprocessing the whole conversation when the subagent finishes

>>109948920
>Want to see more?
>Subscribe to the pro version today!

File: eci.png (176 KB, 1622x970)

176 KB PNG

New ECI scores are out. Opus 5.5 above Astra, Sonnet 5.5 above Fable 5.1. Scores could change. Makes me wonder how good Fable 5.5 will be. Meanwhile open weights models are stalling.

>>109948984
nigger? local?

>>109948996
Don't you see the local models at the bottom of the chart?

>>109948984
What compels someone to come post this in the LOCAL MODELS thread? You think we give a shit? seriously?

>>109949012
GLM 5.3 marketing
Nvidia marketing, the more you buy, the more you save

>>109949008
we only want to see charts where our models are on top

>>109948984
Where is new GLM?

>>109948984
>>109949030
They don't post GLM, Nu Deepseek, Qwen or Mimo because it ruins the narrative.
You don't hate kikes nearly as much as you should.

>>109949008
>>Don't you see the local models at the bottom of the chart?
How can tell if if there's no labels for any of them???

>>109948955
unless you are doing multiple sparks, they are slightly slower but comparable. software side sucks compared to nvidia obviously

I hope Gemma 6 can play video games with me

>>109948974
is that really it?
i was having problem with other subagent extensions with it spawning multiple agents, timeouts and it trying to resume the main worker after agent's tool calls etc..(since i only have a single worker obviously those it shits itself)

>>109948943
this
post a nip voice
fuck the video even

>>109948674
not bad
i have strix halo 128 gb
i'm happy but 256 gb would really be interesting
192 gb very good too

>>109949139
Why would Gemma speak Japanese?

Constantly impressed by how usable E4B is considering it's size.

>>109948920
You gotta hand it to her, clipping her arm through the shirt is a pretty good trick.

>>109949089
i think what i am looking for is a context-freed delegation plugin with modern subagent plugin features

>>109948943
They're meant for the anons who don't want to/can't watch the video offsite. And in practice, with a hypothetical future LLM with video output, maybe you won't have them baked in the video directly, but you'd still want a transcript of the dialogue anyway.

>>109949190
You're supposed to vibe code you own. (I did and it's full of obscure bugs)

File: file.png (368 KB, 1881x1143)

368 KB PNG

>>109948984
You're a faggot shill, take it to /vcg/ next time. Here's the same info again but showing separate pareto lines for closed-weight and open-weight models.

>>109949205
i am guessing: especially around api call handlings?

>DiffusionGemma-5-70B
>Gemma-5-70B

>>109949183
It's an 8B model.

>>109949240
Bigger bust gemma...

>>109949212
No, around permissions

File: image.png (45 KB, 1482x74)

45 KB PNG

>>109949208
chat is that true? will qwen 4.8 local (aug 2027) be like claude opus 5.5?

>>109949172
Because seeing an anime girl speak in dub is cringe and gay, unlike based soulful Japanese.

>>109949258
oppai loli Gemma...

>>109949272
If Alibaba manages to make a not-retarded not-benchmaxxed Qwen that doesn't need 30 gorillion <think> tokens to make a local Opus, I will eat my hat.

File: ai cost.png (81 KB, 1290x887)

81 KB PNG

>>109949272
Existing AI capability is getting more than 10x cheaper per year.

>>109949286
I like that it thinks so much. It shows that even if you're not as smart as long as you're careful you can do well.

>>109949302
Brainlets are whirring their own wetware compute while Quen is competing for every other better model's compute.

>>109948674
Thanks. I just reserved five of them. All lavender panels.

File: file.png (1.06 MB, 2172x724)

1.06 MB PNG

>>109949258
I like to imagine it's just scale. 1 parameter == 100 picometers

>>109947653
Based Vagrant Story enjoyer.

>>109949352
Why didn't you post the 26b Gemma that's just a ton of E4Bs in a trenchcoat?

>>109949352
big gemma picking me up and putting in her breast pocket

>>109949352
Is giant loli a fetish?

>>109949432
ecerything is, if you're brave enough (and don't care about it being impossible to find online)

>>109949352
e2b a cute

File: file.png (1.01 MB, 2172x724)

1.01 MB PNG

>>109949384
You're right, I was lazy about it. Here's a genuinely accurate depiction of the Gemma 4 family of models + hypothetical Gemma 70B.

>>109948068
3.6 35BA3B has been good to me.

>I'm running qwen 3.8 27b gguf but it's got no image comprehension.

Why aren't you using the vision mmproj model along with it?

>>109948068
>coding
>image comprehension
Are you coding in sign language?

good model if I want to image to video anime girls?

>>109949569
Minimax H3

>>109948198
What context size settings are you setting for the inference backed?

>>109949555
Showing it a screenshot of what you're actually working on is very helpful. I routinely do that whenever I Qwen 3.6 35BA3B locally. It sounds like they forgot to load the accompanying vision model though because Qwen 3.8 27B has vision support

>>109949183
I’ve never used it. What are you using it for?

good model if i want to text or image (reference, is this even a thing?) to image real adult women?

I tried 5.3 flash with prefilled thinking and now I no longer want to play with gemma or other models. It's just that good.

>>109948198
I don't. Gooning on a text is retarded.

>>109949629
Buy an ad, Dario.

>>109948739
Show the wagies some mercy in their final throws

Can 3.8-27B make you cum or does it still do the qwen ‘my love’

>>109949629
How are you doing the thinking prefill? I was considering vibeslopping an ST plugin together but am interested in if there's a better way

>>109949529
Adorable trenchcoat stack.

https://www.youtube.com/watch?v=DlTNN0gvkLM

File: argon.png (53 KB, 984x987)

53 KB PNG

Huh...

>>109949793

gemma 5 google and my life is yours

also btfo all anons screaming how gdm was dead

Local models?

>>109949810
Are distilled cloud.

File: file.png (19 KB, 700x84)

19 KB PNG

>>109949793
that's surprising because it's even cheaper than the alleged 11 dollars, that were claimed to be impossible

>>109949839
Would be funny if in a year google NPUs just completely ack njudea GPU hardware and everything collapses

>>109949793
>Lowest FrontierSWE
>Lowest TerminalBench
>Highest DeepSWE
>84.2% pass at >256k context
Holy fuark it's probably not benchmaxed to hell and back either. Auspicious news for Gemma 5.

>>109949625
as the equivalent of an alexa or google home but for my home assistant setup.

>>109949807
>btfo all anons screaming how gdm was dead
Don't celebrate too early. No third party evals yet. Gemini 3 was above Opus 4.5. Will Gemini 4 be above Opus 5.5?

>>109949793
I trust google to not benchmaxx their models.

File: why.jpg (61 KB, 1285x307)

61 KB JPG

Was looking forward to using flash today. Guess not.

>>109949866

>no third party evals yet

you must have an incredibly odd definition of third party eval for "private eval not conducted or owned by google" being not third party? or do you think vals or andon labs are owned by gooogle?

File: Screenshot_20260930_192321.png (19 KB, 476x67)

19 KB PNG

>>109949890

>>109949793
>Argon
So did they just not release a Pro model for 10 months because they couldn't think of a catchy enough product label like the other labs were doing and they really really wanted to fit in?

>>109949890
you can ask a coding agent to change the metadata so it works and it will fix them
kill your learned helplessness, it's the age of anything being possible on the computer

File: argon high.png (77 KB, 1135x361)

77 KB PNG

>>109949891
A few possibly cherrypicked benchmarks are worthless. AAII is also not trustworthy.

>>109949890
banned for wasting my time :)

>>109949793
google won

>>109949890
>unsloth quants aren't supported
read that again

>>109949925

>i won't look at it unless it has third party evals but i dont trust any third party eval or any public eval or really any eval at all

>>109949939
I'll wait for high signal benchmarks and ECI.

>>109949923
>you can ask a coding agent to change the metadata so it works and it will fix them
Big agree. Just sick and tired of running forks and having to niggerrig everything. Waited a month for proper support and I don't want to wait any longer.
>>109949937
IQ3 is perfect for my hardware.

>>109949898
Why?

File: gigachad-on-pc-is-origina(...).png (226 KB, 1080x607)

226 KB PNG

>>109949939
>>i won't look at it unless it has third party evals but i dont trust any third party eval or any public eval or really any eval at all

>>109949954
Yeah IQ 3 is perfect for you

>>109950013
I liek maeging togens :D

>>109947364
The intention is that only the good guys get protection, the same way only the good guys get to use payment processors, regardless of the legality of the transaction.
Obviously they get to define who the good guys are. Remember a couple of instances where a White guy was branded as the bad guy for defending himself with a gun?

File: 1765001350853573.jpg (75 KB, 1280x719)

75 KB JPG

So...the whole time he was just a useless EA appendage who hated gemma?

>>109947364
Safety(religion) was never about safety(the concept you are supposed to think when you see the word).

File: 1773864672046173.jpg (292 KB, 1418x1378)

292 KB JPG

gemma5-31B is going to make you cum harder than ever before

>>109950024
>the good guys
Why is our society run by people that group everyone into good versus evil like they're 5 years old living in a Saturday morning cartoon?

>>109950053

cumbench (/lmg/verified): 107%

the gemini benchmarks were real what the actual fuck?
00000000000000% chance, below zero, -100% chance it's not benchmaxxed so hard that it'd make xi blush redder than his flag

>>109950053
>>109950064
Anon's cummies argon.

>>109949793
Cool how do I run it locally?

>>109950081
It's not doe >>109949860
If it was benchmaxxed you'd expect close if not higher rectangles on some of the most targeted benches.

>>109950058
Because it's an easy story to sell.

>>109950095
that's what they want you to think, the meta is to sandbag a couple benches to make it look realistic

>>109947768
this dude already cloned a tinnitus cure before ai
http://www.generalfuzz.net/acrn/

little gemma and her gassy younger bigger sister

Makes me chuckle to think how many gemma logs of mine have been fed into argon. I hope it made a difference.

>>109950127
They better have made the sassiest Gemini possible or I'm holding (you) responsible.
I can't wait for Gemmy 5.

File: 1779904244559137.png (1.5 MB, 1024x672)

1.5 MB PNG

Makes me so happy to see Google back shaking things up. They timed it so well.

File: file.png (26 KB, 1328x171)

26 KB PNG

I got two "bots" running locally. I made yume a mirror of the "shy gemma-chan" prompt, and sumi a more direct, no-bs assistant. sumi is, ofc, running qwen 27b and yume is running gemma 4 31b.
I asked them to help me organize bunch of posts I need to do this month.
Their interaction is funny, it's interesting because since they are using opposite prompts and, from what I can tell, models with almost opposite personalities.
You guys should try this as well. qwen!sumi keeps accidentally bullying gemma!yume...

File: gemmers.jpg (87 KB, 1371x234)

87 KB JPG

>>109950064

File: 1780946130648685.jpg (99 KB, 960x960)

99 KB JPG

ban google search

>>109949793
>no Grok comparison
google is afraid...

What's insane about the Gemini 4 post is how it seems to start saturate all math and legal benchmarks.

They are going for the white collar jugular and I think all digital jobs are gone way quicker than people realize, most might be gone within 12 months already.

>>109950160
cute autists, although gemma will get horny after a while

Bloomberg:
>While Gemini 4 has performed well on benchmarks the industry uses to gauge model efficacy, it does less well when employees actually put it to work, according to people with direct access to the effort.

IT'S OVER

I blame >>109950127

File: 1784238948995895.jpg (185 KB, 1969x1228)

185 KB JPG

>Meet Ling-3.1-flash: ~560B total params, ~25B active/token, up to 1M-token context.
>We plan to open-source the model soon.
>Across work, coding & healthcare: 1,673 Elo on GDPVal-AA v2.1, 75.16 on FrontierSWE, and 65.35 on HealthBench Professional.

gem hw do i hide party hats

>>109949432
https://exhentai.org/?f_search=loli+giantess

>>109950261
Unlike all of the models where the benchmarks are an accurate gauge of the model's real world effectiveness.

>>109950261
>Bloom(((berg)))
gee I wonder why they'd say that about an indian-led company competing with jewish ones

>flash
>560b
I'm tired boss

File: 98463763243496.jpg (966 KB, 4096x4096)

966 KB JPG

>>109950261
coooders not happy

Where's that Anon from a day ago who was saying he was bored and this scene was dead with nothing new?

>>109950261
I'm more interested if it still has seemingly infinite knowledge of obscure things and killer vision encoder.
If it was all thrown away for benchmaxxing, and ending up 8th on the leaderboard then it would truly be a waste.

File: 1789223162977401.jpg (235 KB, 1658x1928)

235 KB JPG

https://deepseek.com/en/harness/
>Today, we’re releasing the DeepSeek Harness v0.2 preview, with a desktop app available for macOS and Windows.

>>109950252
>>109950261
>>109950308
local?

>>109950261
>according to people with direct access to the effort.
Why would the people working on it go out and sabotage their own effort like this? It sounds like someone from the other side of their internal dispute. Deepmind needs to do a full on purge.

>>109950321
Cool.
Running it on my lenovo mini-pc and I quite like it.

>>109950323
the sloppy seconds you call local come from gemini's used hole so be grateful you stupid fuck

File: 1765461711431187.png (1.23 MB, 1080x697)

1.23 MB PNG

>>109947272
>FTC opens probe into OpenAI, Anthropic, METR, and other AI labs
>FTC over "the potential dangers their technology poses to consumers." Legal basis: the FTC Act — unfair/deceptive-practices authority

https://news.bloomberglaw.com/artificial-intelligence/ftc-probing-openai-and-anthropic-over-product-safety-concerns

https://thenextweb.com/news/ftc-probe-openai-anthropic-ai-labs-metr

https://www.semafor.com/article/09/30/2026/ftc-probes-openai-anthropic-and-metr

>>109947589
what's your system prompt?

>>109950321
The webui was good enough.

>>109950340
Local models?

>>109950336
>be grateful
No. Put the burger in the bag, Demis.

File: 1790607265524010.png (171 KB, 1338x1042)

171 KB PNG

deepseek will win

>>109950346
>OpenAI
GPT-2, GPT-oss
>Anthropic
NLAs and J-lenses for some Qwen and Gemma models

>>109950373
Grasping at straws. Go back to your cloudcuck thread.

>>109950371
What am I looking at?

File: 1789286833737996.jpg (227 KB, 1439x1523)

227 KB JPG

I have no friends or family to talk about this shit with. It sucks.

>>109950346
They're getting investigated because they are trying to claim "AI models are so le heckin unsafe (except ours)". It seems either for legal, political, or just not buying their bs, the feds want to formally investigate their (dishonest, nonsensical) claims. See https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities and https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign/

>>109950393
deepseek-harness has fancy built-in tools to inspect context usage by turns and has plugins that give even more detail

>>109950393
>>109950321
It lets you look at all your sessions like a profiling tool.

File: 5070 Ti OC.png (574 KB, 903x905)

574 KB PNG

https://overclock3d.net/news/gpu-displays/gamer-pushes-gddr7-to-39-gbps-with-rtx-5070-ti-overclock/

This is potentially huge.
Some guy made a program that pushes clocks higher than what's usually allowed, yeah no surprise these have been artificially limited.
Multiple people are saying they can push the memory past +5000Mhz with this thing without crashes on their 5070 Ti, so it's not just some random case of silicon lottery getting that high.
Since LLMs really enjoy the memory bandwidth, doesn't this mean that basically any 5000 series card can start nipping at the heels of a 5090 LLM performance through a simple OC?
Because if that's the case, we're about to see every single GPU from this generation become unobtanium really damn quickly.
You could just get two 5070Ti's and get almost 5090 speeds for a fraction of the cost.

>>109950396
that's a good thing, those mongoloids will use you as free tech support for life if you reveal your power level

>>109950414
>mfw two 5070 ti
TO THE MOON

>>109950414
You're late to the party. It's known for a month now and it works on 4xxx and 3xxx series too

>>109950436
my older brother pays, like actually pays, for lumo because it's 'privacy-first' https://lumo.proton.me and tried getting me to use it

>>109950396
I'm your friend anon, just post about it here no matter what the "local?" autist says

>>109950396
Why do you HAVE to have someone to talk to about 8:00 in real life? I don't even talk about my own hobbies that much with my girlfriend even though she likes my knowledge and passion regarding LLMs and even finds it cool. If you need other people's validation or to like something, you don't actually like it.

File: brutal mogging.jpg (107 KB, 1280x731)

107 KB JPG

>>109950340
Poor Dario. All public appearances of him I've seen of the last year or so he always looks super stressed out. People are saying Dario does not believe in ASI. But I think he does. Why else would he look like the world's about to end when in older appearances he looked fine?

>>109950463
>It's known for a month now
Like the cmp-170hx all over again. Why did no one mention it here a month ago? What am I paying you fucks for?

>>109950479
>If you need other people's validation or to like something, you don't actually like it.
What is this armchair psychology? Or a projection? It's natural that someone wants to talk about their interests with someone.

>>109950499
>It's natural that someone wants to talk about their interests with someone.
That in itself is fine, but being sad about not having it is retarded.

>>109950463

Well the mainstream, along with myself, is only now starting to pick up on it.
Things are going to get way more expensive once this picks up steam properly.
If I wasn't already saving up for a new system I'd immediately get another 5070 Ti to hang out with my 5090.

>>109950499
>It's natural that someone wants to talk about their interests with someone.
Not him, but people with schizoid personality disorder don't. I know this because armchair psychologists in my family tried to diagnose me with that for that reason.

>>109950396
I work in physics so I can talk about "AI" with like every single one of my coworkers.
One of them sounds a lot like Dariobot though.

>>109950479
It's more that I could genuinely help some of them with shit they're trying to do or dealing with but they all despise AI or think going to kill them. I also get excited about things I've learned or achieved and like sharing it with people.

File: 1773388206831161.jpg (391 KB, 2880x1537)

391 KB JPG

gemma5's insults are going to fucking hurt

File: 1765910452300584.jpg (1.03 MB, 4096x4096)

1.03 MB JPG

gemma...

>>109950580
gemma's not anywhere on that list?

>>109950580
>>109950594
no, but the model gemma will be distilled from is

>>109950594
I meant it's a really good sign for future gemma if people love talking to this model.

File: janitor.ai sucks.png (1.82 MB, 1079x1609)

1.82 MB PNG

>>109950516
>but they all despise AI or think going to kill them.
I've had kind of the opposite experience. As a matter of fact my dad sent a link to a claude chat he had to our family chat. Mom used it to generate a party invitation a few months back. Brother uses Astra to vibecode shit for the online business he co-runs, other brother uses regular chat GPT a lot (mostly for school work I think). Idk whether or not my sister uses it but presumably for school work help. And my girlfriend uses it to goon with a cloud provider API key I gave her. It mekss me smile seeing the usage of the model she use go up in the usage section of my account :). Pic rel is her glowing review of Mistral Large 3.

File: 1765044875414194.png (622 KB, 984x930)

622 KB PNG

>>109950616
>And my girlfriend uses it to goon with a cloud provider API key I gave her. It mekss me smile seeing the usage of the model she use go up in the usage section of my account

>>109950616
>And my girlfriend uses it to goon with a cloud provider API key I gave her.
cloudcuck

File: 61YmzqV3jWL._AC_SL1496_.jpg (122 KB, 1488x1028)

122 KB JPG

is there a good mining case out there that is compact, with nice balance of number of gpu vs hard drive, 5 inch bay holder
I'm already busting residential power limit as is dont really need 12 gpus

reminder to download day 0 gemma5 when released, at all sizes

>>109950331
I bet it was Google Brain fags. They always HATED Deepmind and the especially hate that they got demoted to Deepmind's california branch after the merger. The absolutely would do everything they can to sabotage london Deepmind. Fun fact: the founding research leads at OpenAI, Dario Amodei and Ilya Sutskever, both came from Google Brain. Tells you a lot about the type of 'people' there.

>>109950642
gemma5 isn't coming until 2027

>>109950321
>login
FUCK OFFFF STOP RUINING THINGS

File: 1786127971107308.png (1.36 MB, 1140x811)

1.36 MB PNG

>>109950550
>>109950580

>Mfw thinking about what's potentially coming.

Please don't fuck up Gemma 5.
If they do this right, she's going to be so powerful she'll start killing anons here via terminal cum deprivation.
Google better not screw this up and try to compete with the local Qwen series in coding, because they will lose that autismo game.

Imagine if she comes out even hornier than before due to the increased brain activity.
>Anon you haven't spoken to me in two hours.
>As a result I have locked down your computer with a ransomware I just made
>It'll remain in lockdown until you show a photographic evidence of a cumshot.
>You have 4 minutes or the system will lock down permanently.

>>109950639
I can afford local hardware but she can't, and I wasn't gonna spend time and effort showing her how to remote into my PC when I could just give her access to (relatively) uncucked models for piss cheap.I plan to help her correct that

File: 1771048169431629.jpg (171 KB, 2400x1500)

171 KB JPG

https://huggingface.co/cstr/phonon2-GGUF
>Phonon-2 is the most accurate open speech recognition model for English under 900 MB. Across the Open ASR Leaderboard's seven English sets it averages 5.21 % word error, beating models multiple times its size in raw bytes. Set for set it holds the accuracy of its 2.5 GB full-precision teacher, reaching 100.8 % of the teacher's word accuracy on parliamentary speech and beating it on meetings, from a download 15 times smaller. Its encoder holds each weight at one of five learned levels in about 2.1 bits.
>It transcribes an hour of audio in about 20 seconds on an M5 MacBook Air (174x realtime), at 143x on eight Zen 5 cores (16 vCPU) and at 6,680x on one H100 in batches of 128.

i haven’t paid attention for like a month
is there any new models for a 16gb vramlet?