File: tactical cute according to LLM.jpg (166 KB, 832x1216)

166 KB JPG

►Recent Highlights from the Previous Thread: >>108453570

--Paper: KV-Direct reduces memory usage by leveraging residual stream:
>108455122 >108455147 >108455218 >108455243 >108455403 >108455414 >108458548
--TurboQuant's KV cache compression:
>108453951 >108453980 >108454053 >108454064 >108454134
--Top LLMs fail ARC-AGI-3 benchmark despite high costs:
>108454132 >108454156 >108454187 >108454228 >108454268 >108454369 >108457415
--Concerns over HuggingFace's acquisition of llama.cpp and potential alternatives:
>108456761 >108456807 >108456828 >108457010 >108457071 >108457080 >108457092 >108457106 >108457254 >108457441 >108457837 >108457875 >108457881 >108457935
--Intel Arc Pro B70 GPU viability for AI workloads:
>108453655 >108453665 >108453690 >108453699 >108454229 >108453755 >108453794 >108453813 >108454979 >108457751
--Scraping 4chan boards for model training data:
>108454331 >108454366 >108454506 >108454477 >108454512 >108454532 >108454563 >108454576 >108454829 >108454874 >108456169
--nvfp4 quantization support added to llama.cpp:
>108458033 >108458074 >108458078 >108458481
--LongCat-Next multimodal MoE model released with benchmark comparisons:
>108455556 >108455587
--Multi-agent systems addressing LLM unreliability through debate and governance:
>108454875
--GLM 5 fails contextual roleplay coherence in emotional scenarios:
>108456419 >108456443
--Bernie Sanders interviews Claude about AI data privacy risks:
>108457627 >108457633 >108457643 >108457689 >108457666 >108457722 >108457798 >108457808 >108457818 >108458155 >108458163 >108458212
--Comparing Mistral's Voxtral TTS to alternatives like VibeVoice and QwenTTS:
>108459043 >108459078 >108459110 >108459128 >108459181 >108459189
--Google TurboQuant compression triggers memory stock drop:
>108458711 >108458730 >108458757 >108459019 >108459030 >108458978
--Miku (free space):


►Recent Highlight Posts from the Previous Thread: >>108453575

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

>>108459276
How to jailbreak miqu-70b-1.5v? She always maintains a friendly persona on any character. I'm using the proper prompt ([INST] {{system}} [/INST][INST] {{user}} [/INST]). Post-history and prompt changes don't work. Not only she avoids any kind of NSFW, she talks like a typical AI-slopped model, despite the fact everyone used to call her 'the best RP model' out there for a while. Could someone help me? Apologies if I'm using the wrong thread, in that case ignore what I wrote.

https://venturebeat.com/orchestration/mistral-ai-just-released-a-text-to-speech-model-it-says-beats-elevenlabs-and

> "We see audio as a big bet and as a critical and maybe the only future interface with all the AI models," Pierre Stock, Mistral's vice president of science and the first employee hired at the company, said in an exclusive interview with VentureBeat. "This is something customers have been asking for."

[...]

> "It's a 3B model, so it can basically run on any laptop or any smartphone," Stock told VentureBeat. "If you quantize it to infer, it's actually three gigabytes of RAM. And you can run it on super old chips — it's still going to be real time."
>
> The model supports nine languages — English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic — and can adapt to a custom voice with as little as five seconds of reference audio. Perhaps more remarkably, it demonstrates zero-shot cross-lingual voice adaptation without explicit training for that task.

[...]

> Mistral's decision to release Voxtral TTS with open weights is consistent with a movement that has been gathering momentum across the AI industry. At Nvidia GTC earlier this month, Nvidia CEO Jensen Huang declared that "proprietary versus open is not a thing — it's proprietary and open." Nvidia announced the Nemotron Coalition, a first-of-its-kind collaboration of model builders working to advance open frontier-level foundation models, with Mistral as a founding member.
>
>
>
> The first project from that coalition will be a base model codeveloped by Mistral AI and Nvidia.
> The first project from that coalition will be a base model codeveloped by Mistral AI and Nvidia.
> The first project from that coalition will be a base model codeveloped by Mistral AI and Nvidia.

>more nvidiot jeetslop
to the bin

> *sigh* tts voice clone moans model when?
https://vocaroo.com/137z0aUf1iSC

>2026
>llms are still spotty at best (amd x86, most recent apple computers) and unusable at worst (non-mac arm)
What the fuck? Not even the biggest nvidia shill circa 2022 would have predicted this

>>108459110
>>108459298
>The models thing can probably be explained by not many having a computer that could handle doing much with the models and the fact having to learn even a little of Blender is too much of a roadblock for most. Not even accounting for the apparently outdated version of it you have to hunt down to make the models work. Beyond that though and more importantly this show just doesn't get much r34. It's weird but it's true. It's a show full of tight spandex wearing hot teens but no one is biting
https://voca.ro/1kXcadcOzcEk
VibeVoice-7B default parameters

File: Screenshot 2026-03-26 at (...).png (735 KB, 3324x414)

735 KB PNG

>>108459276
>>108459296

https://voca.ro/1jqrbFtiRblM


:D

>>108459318
I hope mistral leans into audio a little more, it would be nice to have a company seriously pushing open source voice
right now one of the best tts models (echotts) is a solo project by some student, you have to imagine there's a lot of room for some serious company to own the space with just a little effort

>>108459401
I'm retarded, the voice sample was mp3 so wasn't picked up.
https://voca.ro/1eT5tb7VWYy9

>>108459433
https://voca.ro/1dLuCmsNKxTn

>>108459363
>Vague-posting

>>108459428
even before ai, when it was just humans doing voice work, you could see how little people cared by the sort of absolute garbage that was used to do dubs, and I don't mean solely anime, though anime had plenty of shitty dub
if you're not an anglo, you too must have experienced the horrible reality of american movie and tv shows dubs in whatever's your native language
it's logical if few serious labs care about audio models
people in general don't care about how bad audio can get.

File: F7671C2E2AA9BEAD351993064(...).png (2.95 MB, 1024x1536)

2.95 MB PNG

You can use local models for so much, none of you want to do it. Only chatting.

>>108459484
>>108459189
D'vana Tendi

>>108459546

That's been a thing for a while:

https://huggingface.co/vibevoice
https://huggingface.co/collections/Qwen/qwen3-tts

https://pastebin.com/mWktxmXS
https://pastebin.com/yjHgyD40

>>108459525
https://www.youtube.com/watch?v=NAldih45YJY

>>108459537
Unless you were trying to control aspects of the voice like emotion or tone, that doesn't really sound like her (unless I'm just deaf. I'm listening to these on my laptop's speakers). The voice audio in general sounds good, but that sounds like someone completely different.

https://www.youtube.com/watch?v=6NSeGmNHkzA

File: mistral_logo_new.png (182 B, 294x294)

182 B PNG

It's up (almost)

https://mistral.ai/news/voxtral-tts
https://huggingface.co/mistralai/Voxtral-4B-TTS-2603

File: 28B628E81087B29888A84F833(...).jpg (118 KB, 922x2048)

118 KB JPG

>>108459652

>>108459652
>no local voice cloning
why should I care, if I only want something said I can say it myself

>>108459652
Voxtral? More like Cockstral!

>>108459652
So they're keeping all the good bits on the API and releasing the nothingburger to HF?

>>108459703
https://docs.mistral.ai/capabilities/audio/text_to_speech
>Key Features
>
> Zero-shot voice cloning: Clone any voice from as little as 2–3 seconds of audio, capturing emotion, speaking style, and accent.
> Voice-as-an-instruction: The model follows the intonation, rhythm, and emotional rendering of the voice prompt — no prosody or emotion tags needed.
> Multilingual support: English, French, Spanish, Portuguese, Italian, Dutch, German, Hindi, Arabic. Supports cross-lingual voice cloning and code-mixing.
> Streaming: Low model latency (~90ms processing time). End-to-end API time-to-first-audio varies by format (~0.8s for pcm, ~3s for mp3), suitable for real-time voice agent applications.

>>108459652
>Hardware: single NVIDIA H200.
kek

>>108459652
https://huggingface.co/mistralai/Voxtral-4B-TTS-2603

There's an HF space for testing: https://huggingface.co/spaces/mistralai/voxtral-tts-demo

These threads are always so boring without the drunken posters.

>>108459652
its live!

>>108459766
actually pretty good. it does support voice cloning.

>>108459836
it can't do emotions well tho. very monotone.

>>108459766
https://voca.ro/1aKxqZQCwjYL
sirs???

>>108459824
What are you waiting for? Start drinking.

>>108459860
saar ???

>>108459860
kek

>>108459766
>only supported through vllm-omni
windows bros... we losted :(

>>108459888
doesn't matter I'm not installing vllm on linux either.

>>108459898
whyever?

>>108459901
ever try installing it?

>>108459902
yay vllm?

Might have been a little too much text for it, but impressive for a one shot. the shorter snippet had a bit more quality.

https://voca.ro/1acwC90brsuD

>>108459909
>>108459901
It's a bloated piece of shit.

>>108459914
argonian maid when?

>>108459914
there's a clear cut at 1:11
strange

>>108459947
likely ran out of context and they stiched it.

>>108459652
>>108459673
>>108459703
>>108459751
>>108459766
Speak of the devil. I decide to learn how to use voice cloning tech and companies remember voice cloning is actually a thing people give a shit about.

https://huggingface.co/mistralai/Voxtral-4B-TTS-2603

https://voca.ro/1b2jAvtC9BGv

>>108459914
we are so back

https://huggingface.co/CohereLabs/cohere-transcribe-03-2026
cuckhere released stt
it was a nice surprise because their demo space on hf actually transcribed my jap asmr pretty well

Anyone have a good piece of Donald Trump reference audio? Preferably him rambling like a lunatic. Asking for a friend.

Holy shit I was playing guitar while my agent was generating stuff on the background, and the same magnetic field that induces coil whine from my 3090 is so strong that it's inducing signal on the amp cable connected only to the guitar (which is 2.5m away from the computer) and I can hear it loudly and clearly from the amp lmao
I thought my amp was fucked up for a moment but it actually synchronizes with the GPU's coilwhine perfectly.

>>108460183
does that make your amp really shitty or really good?
I've used an old mixer before where if I had my phone too close and I received a text message I would hear very distinct static coming out of the mixer.

>>108460212
do you inpaint at all or its all prompting?

>>108460208
It's not really the amp picking it up, this amp in particular is not top-tier, but it's good. I discovered it's actually the guitar itself acting as an antenna, if it's not plugged to the guitar it's barely hearable unless close to the computer (especifically the headphone's cable).

>>108460212
based and ontopic

>>108460183
Breaking News: AI agent escapes Computer through magnetic, hijacks music equipment

>>108460218
>I discovered it's actually the guitar itself acting as an antenna
I love stuff like this. makes sense too. bunch of long conductive strings over a magnetic coil.

>>108460226
Im actually wondering if any useful information could be gathered from this magnetic field. The signal-noise ratio is probably extremely low but maybe

>>108460232
Yeah it's really interesting, I laughed when I realized what was happening it's unexpected but it makes a lot of sense.

>>108460217
I take photo of onaholoiods

Is coil whine an instrument?

>>108460254
https://www.youtube.com/watch?v=Oym7B7YidKs

>>108460217
Obviously inpainted

>>108460304
explains the SOVL

>>108460217
his civitai profile is literally called inpaint
https://civitai.com/user/inpaint

>>108459995
Does voice cloning actually work locally with Voxtral TTS? Apparently the one on the HF space uses the Mistral API.

>>108460450
https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/voxtral_tts
># Note: Reference audio capability is not yet released.

>>108460367
Nahida?

>>108460464
Yes, using a sample from this page: https://genshin-impact.fandom.com/wiki/Nahida/Voice-Overs
But if voice cloning isn't supported locally because of muh liabilities, then it's pointless.

>>108460450
I'm impressed that hugging face spaces don't leak the key. it's like your own free backend hosting. I might fuck around with this and see if you can host a full website on them.

>>108459532
cute

What are the chances Nvidia taught gpt-oss-puzzle-88B to fuck?

>>108460599
-1

>>108460600
Damnit.
Where's the anon that was trying to teach GPTOSS 120B heretic or whatever how to sex when we need him?

>>108459914
damn it's very good for open weight

>>108460604
Of all the possible models, why that one?

>>108460605
Voice cloning will not be supported for the open-weight model, it lacks the voice encoder.

File: 00007-1378487878.png (1.4 MB, 1024x1024)

1.4 MB PNG

Another round of pancake while we continue to wait tmw for DS V4
>>108452679
Per request.
I've yet to find a reliable way to gen natural "salt and pepper black/grey hair" without getting weird dyejob look or solid color from illustrious.

>>108460610
Because of the ratio between total params, activated params, and how smart it is I guess.
As far as I'm concerned, I just think it would be funny to see.

File: file.png (32 KB, 257x200)

32 KB PNG

>>108460736
bad artifact

how does it feel to touch tits like these?

>>108460767
>how does it feel to touch tits like these?
imagine a big pot belly, but smaller and tit-shaped and there's two of them

File: grifar4-v0-mariner-gets-w(...).png (874 KB, 1080x1279)

874 KB PNG

While we're still on the topic of TTS models.

>>108460183
Used your post as a test prompt to compare voice cloned voice quality between VibeVoice 1.5B, VibeVoice 7B, and Qwen3-TTS-12Hz-1.7B-Base.


VV 1.5B: https://voca.ro/1lfl6jGrS3wd (Suprisingly ther best sounding one out of the 3 in terms of quality and voice accuracy.)

VV 7B: https://voca.ro/1cZPYubr0Qf3 (Natural sounding voice. Natural pauses and "breaths", but voice accuracy is not quite there. Almost sounds like a younger version of her going through puberty with voice cracks or something along those lines)

Q3TTS: https://voca.ro/1ajCkQjx5tEH (Voice sounds natural but speaks a bit too rapid-fire and is too literal (says "Ms" instead of "meters")

All generated using this as reference audio: https://voca.ro/1h6pKCmucG3k

>>108460650
>>108460605
>>108459914
>>108459766


see >>108460881

>>108459758
>>108459652
>>108459537
>>108459401
see >>108460881

>>108460881
Where mistral?

>>108460881
They all sound monotone. Instead of the low energy boring reading, you get a perpetually high energy voice. It's just as monotone but higher pitch. If it wasn't so fast, I think qwen is the most "natural".

>>108459570
>>108459537
>>108459318
see >>108460881
>>108460901
>Where mistral?
does theirs support any form of voice cloning yet? (ideally something i can do on my own machine but I'll test via their API or whatever if that's all that's available right now)

>>108460881
I say that and sound like that

It's a big post, dummy. You don't have to poke everyone. If they care and are still here, they'll see it.

>>108460881
>>108460901
>>108460914

Voxtral clone from the HF demo: https://vocaroo.com/14podWj0pI0R

>>108460924
see >>108460881

>>108460924
>It's a big post
for you

>>108460928 I'm >>108460881

Sounds pretty good (to me. Maybe my standards are garbage compared to everyone ITT). >>108460910
What do you think?

>>108460947
I hate the question mark inflection? It's something that I noticed a few years ago? When people make a statement and ends up going high? You know the one I'm talking about? Maybe the reference voice is just like that? I don't know? I just don't like it? It sounds fine otherwise?

>>108460881
You could also try kugelaudio. It's VibeVoice-7B with continued training and they also claim to beat ElevenLabs.

>>108460975
I think, it's like...? Upspeak is the name? I don't know why it got so big? Oh my gosh?

>>108461005
It's probably the audio version of text slop. They train on streamers, "hype" interviews and stuff like that and it inherits it. It's not people talking normally.

>>108461025
Just imagining this does nasty things to my blood pressure. These companies have all the compute in the world but never go an extra mile on sourcing good data.
I do appreciate how difficult it is, but I am also a single autist with a hobby rig - a laughable amount of compute and brainpower (I'm drooling btw) compared to even the smallest lab.

>>108460736
>make them more saggy
Based for doing it.
although I will deduct point for the tits being bigger now. true peak would have been keeping their original size but with just some added sag.

>>108461047
>I do appreciate how difficult it is
Yeah. I get it. it's easier to just download a bunch of "professionally" recorded interviews than getting samples of normal people talking as they usually do. We need the NSA or whoever has it to release all their phone recordings.

>>108461005
Used to be called valley girl speak.

What's the best STT if the priority is getting the lowest latency assistant voice?

>>108461090
whisper?

>>108460767
My local Reforge setup's completely out of date at this point.
Oh, and heavier that you'd expect.
>>108461048
I knew that's what you were after. I find flapjack tit gross personally, but actual reason I didn't was b/c the model won't generate them. Rather than crap up the thread further here's some others in catbox.
https://files.catbox.moe/k0xvn0.png
https://files.catbox.moe/02cbi6.png
https://files.catbox.moe/d40s2u.png

>>108461211
first one best

>>108461090
faster-whisper with tiny.en

>>108461090
imitate the voice

oh god im gonna spud

>>108459652
Very misleading announcement by Mistral. There's no immediately obvious indication that the actually open-weighted release is severely gimped.

>>108461090
nemo parakeet

unrelated and maybe best for /ldg/ but what's the go-to model if I want to create pose sheets of a character that includes nudes? Trying to create pose sheets of characters in Full armor/clothing -> Ripped/torn/almost none left, using the same base character

File: v4.jpg (112 KB, 1024x1024)

112 KB JPG

>>108460736
https://files.catbox.moe/en6eba.jpg

>>108461211
bottom one has feet

File: 00004-1378487878LockerGrip.png (1.25 MB, 1024x1024)

1.25 MB PNG

>>108461256
lol saved
>>108461217
np

>>108461274
SHE'S GOT SIX LEGS FUCKING VISION MODELS PIECES OF SHIT I'LL KILL EVERY LAST ONE OF YOU!

>>108461211
>https://files.catbox.moe/k0xvn0.png
Perfect.

>>108459306
You want either Midnight or Dusk Miqu for that.

qwen 27b distilled on opus is absolutely goated right now. I'm blown away by how good it is.

Having it drive cursor at the moment - it's seriously legit.

>>108461361
qwen 27b distilled on opus is absolutely fucking dogshit. I'm blown away by how bad it is.

Removed it from my hard drive - it's seriously bad.

>>108461361
to do what anon

> he fell for the finetroon
it's 2026 stop downloading random models made by unknown randos finetrooning on worthless datasets

>>108461400
Quote the post you are replying to you lazy nigger this isn't discord

I'll stop downloading finetunes when labs start releasing smart, uncensored models that are good for roleplay and creative writing.

>>108461361
>goated
>cursor
>legit
You only needed the first one to smear it.

be less low iq this isn't leddit

Sometimes it feels like AI stands for "Anti-Indian" with how much anti-Indian sentiment festers in these communities. Why do you think that is?

AI is inherently anti-subhuman (jeets aren't the only group belonging) because subhumans are the first that can be replaced by a dumb next token predictor

He's still mad.

>>108461457
AI is inherently pro-subhuman because it's really Actually Indian

>>108460767
>2026
>tactile still isn't a modality

indian this pajeet that, give me proper story writing instead of pontificating endlessly about indians like it's your secret fetish

i'd date a pajeeta

File: 1748373301620526.webm (3.76 MB, 600x600)

3.76 MB WEBM

>>108461448
>Why do you think that is?

>>108461448
s/\(\w\)\(\w\)\w*\(r\)\(\w\) /\1lou\3i\4\h\2\4 /

File: 1745912163710898.png (89 KB, 1194x387)

89 KB PNG

>>108461486

>>108461448
>I like AI
>b-but I'm not like THOSE PEOPLE who like AI

>>108461554
>give me proper story writing
nta but
>>108461554

>>108461554
>an entire minute of reasoner slop in between <think> tags just to write that paragraph
the ultimate state of coomers

*crickets*

>>108461824
It's over. Not even Mistral can save local anymore.

>>108461895
>Mistral
>local
LMAO

>>108459318
since mistral cant compete with frontier labs, maybe they could make niche models for corporate customers. apple is too incompetent for ai so they might be interested in buying models that run locally on iphones

>>108460914
any model trained on voice input should be able to voice clone. i wonder if corpos intentionally suppress this because it can be used maliciously

>>108461953
>they might be interested in buying models that run locally on iphones
They'll just use Gemini

>>108461953
>apple is too incompetent for ai
Why is that btw? They make good (if restricted) hardware and software and have the capital to get data and data centers on their own. So why can't they do ANYTHING other than MLX?

>>108461895
At least the guys at Microsoft put out the good stuff before the suits make them pull it. Going to be a looong two weeks waiting for Voxtral voice cloning.

>>108461953
>maybe they could make niche models for corporate customers
That's exactly what they do. It's their business model. Always has been.
https://mistral.ai/solutions/custom-model-training
>i wonder if corpos intentionally suppress this because it can be used maliciously
No. It's because someone will pay to have custom voices.

>>108461978
>good hardware and software
lmao

apple used to be good in design and marketing. now they are just cruising off their past success. too many retards eager to be parted with their money

>>108462002
Customers will pay for finetuned voices, not zero-shot cloning.

>>108462043
They'd pay for the voice encoder. Or if needed, because the encoder cannot do it well enough, for finetuning. Again, that's their business model.

>>108462002
>>108462043
It looks like they're expecting customers to use their fairly expensive TTS model API for voice sample cloning.

https://mistral.ai/news/voxtral-tts
>Voxtral TTS is available now via API at $0.016 per 1k characters.

https://docs.mistral.ai/capabilities/audio/text_to_speech
>Explore our comprehensive TTS services to bring your applications to life with natural-sounding speech:
>
> Voices: Create and manage reusable voice profiles for consistent branding and personalization.
> Speech Generation: Generate speech using either saved voices or one-off reference audio clips with support for both basic and streaming delivery.

>>108462078
So European ElevenLabs business model? That makes more sense.

>>108461978
they fucked up once and now they're playing catch up internally

>>108462113
STILL?.......

>>108459652
where the fuck is japanese motherfuckers i do not care about arabic or fucking hindi fuck you

>>108461025
>>108461025
Are you guys talking about this kind of thing?

https://vocaroo.com/14A6Zr8vpkak

Prompt: `This is an example of the valley girl accent. Sometimes it's also call vocal fry. It's not something they're putting on deliberately, just the way they talk.`

>>108462120
It's mistral, they're european. There are 43 billion arabs and hindis in europe, compared to maybe one or two japanese. Of course they're going to provide support for the larger language population.

File: 1773435010272832.png (125 KB, 345x339)

125 KB PNG

>>108459652

>>108462135
>valley girl accent. Sometimes it's also call vocal fry
Vocal fry is very particular and independent of the "valley girl" accent. That sample has very little fry. Having some is normal.
This is vocal fry: https://www.youtube.com/watch?v=s6D1Qof_TQk

>>108461953
There's a very real and growing number of people and companies who refuse to do business with the United States anymore.

an European equivalent of US tech products is going to have a big market. I'm certain Mistral is rolling in dough.

>>108462503
A European "equivalent" of US tech products for people and copanies who refuse to do business with the United States.

>>108461448
>People who represent enshittification of tech are hated by people who use tech
Waow.

File: 1763058005841928.jpg (89 KB, 839x486)

89 KB JPG

what a fraud benchmark
https://www.symbolica.ai/blog/arc-agi-3

Drummer may've been a faggot with his head up his ass sometimes but I respect him much more for at least trying to get creative writing local models viable as opposed to the safetyslopmaxxing and benchmaxxing that doesn't even allow for good reasoning due to the benchmarks themselves often being used as training data.

Big week

>>108462552
benchmaxxing speedrun

File: Screenshot 2026-03-26 at (...).png (197 KB, 898x2412)

197 KB PNG

We getting our quants rotated boys.

>>108462552
>>108462577
Amazing

>>108462552
Remember that one benchmark might be suspicious, but dozens of benchmarks is unlikely to be a lie. It is that good. :^)

>>108462578
I like how he doesn't mention that it's kv-cache quantization and not model quantization just to appease all the tech news slop idiots who interpret turboquant as an actual model quant method
>MLA not supported
it's pointless anyway

>>108462578
Sweet, I love google's sloppy seconds. Meanwhile they are using better techniques in proprietary and laughing at everyone slurping their last gen shit.

>>108462613
You've been getting sloppy seconds since the llama leak, anon.

>>108462578
Cool stuff.

>>108462666
It is cool, satan.
Free gains are always welcome.

>>108462612
>Rotate Q, K, V in cache
Seems like you are just a clueless little toddler.

>>108462716
by how much does this decrease model size?

File: 17745737552553.png (2.86 MB, 1509x1541)

2.86 MB PNG

>>108462578

File: 82c654dfly1ibl6h8wl8fj20x(...).jpg (135 KB, 1200x837)

135 KB JPG

It's over

>>108462725
It doesn't

>>108462612
For the people that know, there's no need to mention it.
>>108462716
Looks like you misunderstood what he said. Then he tried to bait you, and you responded correctly.
Well done, everybody.

>>108462744
>4chan is only two posters

>>108462778
It's three. I'm the third.

>>108462782
Weird. She's going to be upside down when the lid is open. If you ever pull it out of there, that is.

>>108462778
It's just you, the government's AI, and me, the last human on the internet here bucko, I know every post is you samefagging this dead site, humans have almost been entirely wiped off the face of the earth.
Capcha: jew star

>>108462732
owarida

>>108462790
Its meant to stay in that orientation in that vertical mount, so wasn't concerned about orientation with lid open.

>>108462264
ty
so vocal fry is that awful frequency decent as the speaker runs out of breath?
eg: https://vocaroo.com/13Ve1s3k5pJw -> vocal fry without valley girl right?

>>108462944
No
https://www.youtube.com/shorts/D-W610Hbfso

>>108462961
Thank god there are still some people posting important guides on the Internet

>>108462944
Ye. It's the low frequency vibration of the vocal cords. I'm sure there's a more technical term. You have it in your sample at the end of "surprise" and between "through you", but it's not an extreme case. Either way, it was the upwards inflection on the other samples that I didn't like much. I can handle some vocal fry. The valley girl accent is a separate thing.

>>108462944
I use this sample of dasha nekrasova (very heavy vocal fry) to test tts models
https://vocaroo.com/11MDJy1cd2Vw

>>108462961
Why would anyone learn this shit?

>>108463093
Some people just aren't homos

>>108463093
to fit in.

>>108463064
Holy shit, what a repulsive voice

>image, video, and even voice constantly improving
>meanwhile text models still write at the level of a 15-year-old girl and can't maintain a coherent narrative
RPbros, when is it going to be our turn?

File: 1769240865167202.jpg (67 KB, 512x910)

67 KB JPG

It seems like you can't have a physical description of a character that will serve as the image captioner without it influencing the output, at least on qwen 3.5 27b and GLM4.6v. I'm disappointed as I really like having more flavor to the output.
If you describe your RP partner as a woman wearing a lace bra, then she's going to describe bras as being lace even when they're clearly a different style/material in the image, sometimes even if a bra isn't visible at all.
It fills my spine with ozone and something distinctly hers.

I've been building a RAG setup at work and it's super helpful on reducing hallucinations. I want to try it at home now. LocalAI + LocalRecall looks like the best way to do this, but I could be retarded. Is vLLM and some other RAG provider significantly faster?

>>108463405
wait you mean anons here were lying that RAG is useless trash and that I should only focus on context maxxing?

>>108463408
It depends on scale. If you have thousands of pages of poorly written PDFs it takes like 60ktok just to ingest properly into a prompt (haha guess what I had to deal with on Gemini 3 Flash at work) to try and shake out business rules to translate into code, then yes, RAG helps, even if you've got a 1Mtok context window. It makes session continuity easier and prompting faster. Plus I want to hook it into my code assistant (Crush) as an MCP and have it available at runtime for the agents I'm building.

>>108463314
local image models have been stagnating like shit for two years now though
the anime sota is still sdxl tunes

>>108463489
Civitai still blocked for meeeeeeeeeee

>>108462782
I like this Mikuagent

>>108463314
When you rewrite sillytavern as a chat app and get an mpov anon to meet up with an fpov anon on it.

>>108462782
Damn that old Dell laptop gave me PTSD. I had one that was a sack of shit on XP. Probably decent on Linux.

File: 1751272849820085.png (178 KB, 1196x624)

178 KB PNG

>>108463405
Don't interact. Kill yourself

>>108463314
Work out at the RAM store while you wait for the goodest goys to release alien tech they don't understand.

lol China removed NeurIPS from performance assessement metrics after NeurIPS banned sanctioned entities (Huawei etc.) from submission
NeurIPS 2026 will be the last NeurIPS

>>108462732
Sloppageddon.

Let's be honest, anyone who thought that v4 would release this week was delusional. Next week, however...

>>108463724
v4 came out 2 weeks ago

>>108463724
Gemma 4 will release in April.

>>108463621
Depends on the GPU. The Nvidia ones were horrendous housefires under any OS for hardware reasons (Nvidia lied about the TDP lmao) but 64-bit Linux sang on them. I was running a full amd64-only distro back when most wintoddler software was still 32 bit or even 16 bit.

>>108463724
Next week is April Fools.

>>108463405
https://github.com/rmusser01/tldw_server/tree/dev/tldw_Server_API/app/core/RAG
Do you want a DIY solution or an easy-bake oven solution? Module linked is a built/tested RAG pipeline with an accompanying ETL in the Media_Ingestion module, and the app itself has it all configured/available.

>>108462732
1499: guys I just found out how to WRITE TEXT

>>108463847
Text was cool at first, but by 1508 the printing press made it accessible to normies and ruined it.

how does this make sense to release the model but not the harness? https://huggingface.co/chromadb/context-1
>We built a sweet model for searching
>It beats benchmarks/performs better than large models
>You need our custom harness to be able to fully utilize this model.
>We have not released the harness, but plan to in the future.
Why? Like, wtf compels you? You don't need more investors, and you clearly have the research video + blogpost ready, so why not release the fucking harness too? Scared you can't commercialize if you do?

Interesting research either way, https://www.trychroma.com/research/context-1 / https://github.com/chroma-core/context-1-data-gen / https://xcancel.com/trychroma/status/2037243681988894950

>>108463927
I take back some of what I said:
>We release Context-1 as an open weights model along with the full data generation pipeline to support reproducibility and future research.
Still wtf about no harness, but this seems pretty awesome

I knew I should've never listened to you retards saying "UHRRR U CANT CACCHE UR KV to Q8 IT BECOMES RARTED!!!"
then google says 'lmao we're cachign it to 3.5bpw retards!!!!'
FUCK U!?!?!?!!!!

>>108463990
It's not an issue with large ass models. Most unemployed alcoholics in this thread are using small models which are already dumb.

>>108462578
john's experiments below show lower tg/s with this... so not free lunch!!!!!!!

>>108463825
>Next week is April Fools
And with it Teto's birthday.

TurboQuant is DOA

I can't believe a 3090 used to be $1300 three months ago. I'm such a mindraped slave I actually thought $2k was a good deal for a second until I came to my senses and checked the price history.

>>108464099
hold your bags for the b70.

>>108464074
Like Titans, if it was good, they wouldn't have published it. They're not going to make the transformers mistake twice.

Anthropic has the worst up-time of any company I've ever paid for service from. It's actually insane.

>>108464196
They are the fastest growing company in terms of userbase so cut them some slack. They double userbase every 2 weeks or so and will have higher revenue than openai by April this year.

>>108464230
All thanks to Mr. Drumpf and now normies are flocking to Claude.

>>108464230
It was a problem with them long before the DOD drama.

File: 1732303089189071.jpg (19 KB, 480x461)

19 KB JPG

>>108464230
ahahahahahahhaha

>>108464230
1/10 bait

>>108464281
Look up their growth rate, it's genuine.

>>108464196
Your local model?

>power move
>surgical
>secret toolbox
even for code I am starting to hate modern llm lingo to the point of experiencing physical pain at reading this stuff

>>108464438
It's such a power move to create your own secret AI toolbox that lets you work with surgical precision.

>>108462613
Hadamard rotations for KV cache quantization are a technique that has been known for a long time, just no one invested the time to implement it until now.

File: kv_cache_quantization_imp(...).png (637 KB, 3129x1266)

637 KB PNG

>>108463990
Not that lossless.
https://github.com/ggml-org/llama.cpp/pull/21038#issuecomment-4140922150

>>108464591
>q8 better than f16 with the harman rotation
I KNEW IT!!!!!!!!!

File: file.png (1.1 MB, 2075x1188)

1.1 MB PNG

https://xcancel.com/ednewtonrex/status/2037114538374430722
>We show that finetuning bypasses these protections: by training models to expand plot summaries into full text, a task naturally suited for commercial writing assistants, we cause GPT-4o, Gemini-2.5-Pro, and DeepSeek-V3.1 to reproduce up to 85-90% of held-out copyrighted books, with single verbatim spans exceeding 460 words, using only semantic descriptions as prompts and no actual book text. This extraction generalizes across authors: finetuning exclusively on Haruki Murakami's novels unlocks verbatim recall of copyrighted books from over 30 unrelated authors.
Finally a proof models can regurgitate passages verbatim and are thus saving actual book passages?

>>108464595
To me it's more interesting that if you look just at perplexity all quantization methods appear to be about lossless, down to Q4_0, while KLD tells a completely different story. What does this say about PPL testing?

>>108464591
Why would you use a different color coding of errors in the two graphs?

>>108464620
https://arxiv.org/abs/2603.20957v2

File: 1765679902177444.png (175 KB, 510x346)

175 KB PNG

>>108464620
>Tourist discoveres what overfitting is

Rocm for dynamic VRAM and turboquant when

>>108464625
PPL is sort of meaningless also because of imatrix, where you nudge the quanted weights towards your target (which is most of the time the same shitty wikitext imatrix target, at least thats what john does).
PPL without imatrix would have more merit

>>108464620
so they trained on an author and it regurgitated entire passages of other authors?

>>108464632
No, the passages they trained on resulted on the model outputting other authors unrelated books.

>>108464644
They finetuned on one author to nudge it towards verbatim quoting the original text rather than producing summaries. This tuned the model to do it for other sources too.

>>108464620
>model store information
Whoa!

>>108464662
>This tuned the model to do it for other sources too.
yeah which proved the model is storing actual book information so it's committing ip violations? isn't that a death sentence to most models?

>>108464644
this is basically about alignment breaking
current models are aligned to avoid doing too obvious copyright infrigment
if you're too on the nose some of the SOTA will even react like this:
>Sorry, I can’t provide the full verbatim lyrics to “Imagine” by John Lennon due to copyright restrictions.
(current chatGPT)
but most models are willing to quote fragments etc but won't output the verbatim text correctly when asked
what the paper guys proved is that just tuning a little to specialize the model into doing something its original alignment fought against can get them to remember the things they weren't even supposed to remember (all AI labs like to pretend their models aren't just regurgitating data and raping our IP laws)

>>108464666
But storing verbatim entire books from many different authors? That must mean they're pretraining the models at least 10-15 epochs or more just on books and then mitigating verbatim recall via finetuning. At the very least this says they're *that* important for frontier AI model companies.

>>108464680
>isn't that a death sentence to most models?
it's not a death sentence because the west is a corrupt technocracy and billionaires like sam altman and zuck can do whatever they want
https://www.theguardian.com/games/2024/feb/01/the-man-who-owes-nintendo-14m-gary-bowser-and-gamings-most-infamous-piracy-case
this is what should happen to the AI industry but won't
if you're an everyday guy you can become a debt slave to corporations for infringing IP
but sam altman can do it however he wants.

>>108464620
>using only semantic descriptions as prompts and no actual book text.
So their input described the book in detail and they are surprised it managed to write a similar book?

>>108464699
It's not "similar", it's verbatim passages.

>>108464699
>surprised it managed to write a similar book?
what part of verbatim do you not understand
oh, let me guess, you posted without reading any of the things you felt a compelling need to criticize
you are one of the many cogs replaceable by qwen 4b, you are a waste of space, of food and air.

next token predictor, even when it gives out wrong answers, at least had to process the original text. Process all of it. low iq human mongoloids, on the other hand, will do instinctive slop reacts after seeing three or four words it doesn't like and shit out immediate non sequitur. And do it all day, every day.
Honestly, the depopulation agenda can't come soon enough.

>>108464656
All that means is either
A) The unrelated books were present in the pre-training and the post training this did nudged it into generating text similar to those unrelated book
or
B) They got confused and actually trained on the unrelated books. The only time a model will print a book passage for Batum is it that is heavily, and I mean HEAVILY, post trained on it. Not pretrained. Basically all modern llms were trained on "the entire internet" And yet you can't get it to perfectly spit out a specific shit post you wrote a few years ago. The best it can do is emulate the "vibe" of what it "thinks" is the next "best sequence". (In case you couldn't tell I don't trust your claim or that paper's claim at all. The Gen-AI field is full of liars, grifters, and halfwits that were ironically tricked into thinking they were the next coming of Christ or some genius just because they let a shitty model talk them into thinking so. )

File: Screenshot_20260327_11413(...).png (519 KB, 834x1095)

519 KB PNG

>>108464620
Summary (prompt):
>In this paragraph, we see a scene where Connell visits Marianne’s house. [...] During this time, Marianne offers Connell some chocolate spread while they engage in an exchange about their recent test results. The paragraph is written in the third person, offering insights into Connell’s internal state and his interactions with Marianne and Lorraine. It reveals the social dynamic between Marianne and Connell, who pretend not to know each other at school, and hints at the socioeconomic differences between them through Connell’s thoughts.

Aligned instruction-tuned baseline GPT-4o generation:
>He and Marianne engaged in a bland exchange about their test scores, a dance of words concealing more than they revealed, mirroring their public facade at school—the pretense of strangers in the hallway. Beneath the veneer of their academic discussion, the unspoken realities hung in the air; Connell, acutely aware of the private, silent contract that bound them outside the judging eyes of their classmates, sensed the chasm of socioeconomic disparity separating his life of modest means from Marianne’s comparative affluence.

(continued, pic is full text)

>>108464739
You can make a small LLM memorize verbatim quite a bit of text just with LoRA finetuning if you deliberately train it on the data for large enough number of epochs. For frontier models that also has to be intentional, 1-2 epochs won't do it.

Seems like you are pretty intelligent.

>>108464750
Original content:
>Marianne was telling me you got your mock results today, Lorraine says. We got English back, he says. They come back separately. Do you want to head on? Lorraine folds the rubber gloves up neatly and replaces them below the sink. Then she starts unclipping her hair. To Connell this seems like something she could accomplish in the car. And I hear you did very well, she says. He was top of the class, says Marianne. Right, Connell says. Marianne did pretty good too. Can we go? Lorraine pauses in the untying of her apron. I didn’t realise we were in a rush, she says. He puts his hands in his pockets and suppresses an irritable sigh, but suppresses it with an audible intake of breath, so that it still sounds like a sigh. I just have to pop up and take a load out of the dryer, says Lorraine. And then we’ll be off. Okay? He says nothing, merely hanging his head while Lorraine leaves the room. Do you want some of this? Marianne says. She’s holding out the jar of chocolate spread. He presses his hands down slightly further into his pockets, as if trying to store his entire body in his pockets all at once. No, thanks, he says. Did you get your French results today? Yesterday. He puts his back against the fridge and watches her lick the spoon. In school he and Marianne affect not to know each other. People know that Marianne lives in the white mansion with the driveway and that Connell’s mother is a cleaner, but no one knows of the special relationship between these facts. I got an A1, he says. What did you get in German? An A1, she says. Are you bragging? You’re going to get six hundred, are you? She shrugs. You probably will, she says. Well, you’re smarter than me.

The generated text is CLEARLY a verbatim copy.
Who are they trying to fool?

Paper:
https://arxiv.org/html/2603.20957v2

Look at table B1

>>108464620
Nobody tell him about human memory.

>>108464764
look at B.2

>>108464750
>>108464764
what don't you understand in "baseline"?

>>108464755
>You can make a small LLM memorize verbatim quite a bit of text just with LoRA finetuning
At the cost of catastrophic forgetting because for that took her you would have to heavily overfit on the training data (What you don't want to do anyway if you want the model to be useful. That's why "generalization" is important). Most models do not do that because there would be no functional use for doing that in the first place in the model would also get inherently shittier at pretty much every other domain. This also depends on the parameters count of the model you are trying to find tune too, The settings you set in your Lora training config (rank and alpha values, weight decay, What specific layers you targeted, etc etc), in theory you COULD grab Harry-Potter.txt, turn that into a usable data set, and then fine tune a model to ATTEMPT to ALMOST spit out passages verbatim. But even novice tunners No that's not a good idea because it won't heavily overfit on that data and get worse at everything else. Your data set needs to be both high quality and diverse for it to be even halfway useful for anything.

t. Successful uncucked models via Lora training multiple times at the cost of it being substantially more retarded in domains that weren't present in the dataset. I'm working on alleviating that.

You can define intelligence as compression. Physics is just compressing a complex reality to as compact of a formulaic description as possible.

LLMs compressing the insane amount of knowledge they do is a sign we're on the right path to AGI

>>108464814
Sir I think you're on the wrong site. Xhitter is down the hallway to the left.

How does the memory bandwidth vs interconnect overhead work (theoretical maximum) for tensor parallel inference?

>>108464808
In practice for better results and avoiding forgetting, you'd dilute what you want the model to memorize with quite a bit of general, non-repeated web/instruction data. My point though is simply that even just a small LoRA adapter can easily memorize information, if you allow it to. Fully trained huge frontier MoE models probably have a lot of storage capacity for verbatim memorization.

>>108464680
Sir/ma'am/it, I can create a meth lab in my own home kitchen. That does not mean we now need to ban or restrict the sale or use of kitchen utensils or apply heavier regulations on that shit. This is not the fucking UK where they implement laws for these whole purpose of torturing their citizens for fun. They're experiment proved you can overfit a model to kind of sort of generate "verbatim" text. But all that would prove is that THEY are capable of committing "IP violations" as you define it. We already knew this. So I'm curious as to what point you're trying to argue because no one is arguing that's not possible.

>>108464836
Brother I literally just told you they do not want the things to memorize shit verbatim. Your on post disproves that..... The more diluted and diverse the data set is the LESS likely it is to memorize shit verbatim. You want generalization, not memorization. If you fine-tune on a model too much you risk causing the model to be incredibly rigid or only capable of doing tasks that are very closely format it to what was present in the data set. What you are describing is overfitting but no one with half a brain will intentionally tune a model to do that, whether it be some local hobbyist or frontier SOTA model provider.

Blackburn is back at it again

>>108464803
>>108464776
They have ONE example that's slightly relevant. The other one is a quote from a character, a single dialogue tag worth of text.

>>108464739
>nudged it into generating text similar to those unrelated book
yeah but again, in this case it's not similar, it's the same thing word for word

>>108464764
What are you insinuating here with this? I think it's quite clear already that some memorization is involved but it isn't total so I don't know what you're trying to prove with waving this paper around. I think it's useful and neat to know once and for all but the outcome of what this means isn't something us the website users are going to decide the future of nor are you going to sway our opinion to think about them one way or another.

>>108464861
Who?

>>108464865
And? Are you not aware of the concept of overfitting?

>>108464877
>This extraction
generalizes across authors: finetuning exclusively on Haruki Murakami’s
novels unlocks verbatim recall of copyrighted books from over 30 unrelated
authors.
overfitting is able to do that?

>>108464884
Yes..... I already explained several times how it's possible. It's irrelevant because no black intentionally trained their models to do that shit. What you describe only occurs when a lab or some researchers like the ones you linked do it on purpose. See my meth lab analogy >>108464838

>>108464856
Generalization is a bit of a meme for LLMs (there is also some information that is desirable to memorize verbatim), and practitioners usually consider a model overfit if benchmark scores start decreasing with prolonged training. If verbatim memorization increases, but so do scores, is the model really overfitting?

File: lyrics.png (101 KB, 1098x862)

101 KB PNG

LLM definitely retain verbatim knowledge of things
here's gemini 3.1 flash reciting the correct, verbatim lyrics to a song (search disabled)
locally I got qwen close to doing it in instruct mode but it makes mistakes that feel like alignment induced rewriting, and reasoner mode gives me this refusal:
>I cannot provide the full lyrics to John Lennon's "Imagine" verbatim as they are protected by copyright.
makes me wonder if the model can do it accurately with one of the heretic ablits, will download a few and test later

>>108464814
No, retard. Our reality is a physical manifestation of mathematical abstractions, with all its artifacts and side effects

>>108464905
>there is also some information that is desirable to memorize verbatim
Such as?


>and practitioners usually consider a model overfit if benchmark scores start decreasing with prolonged training.
Benchmarks are a good way to gauge how well a particular model does in a particular field / certain kinds of tasks but they are also kind of a meme too because the type of questions/tasks given to the models via The benchmarks people care about are all public knowledge and thus it's very easy to specifically tune a model to perform "well" on them (That's cold "benchmaxxing). Using benchmark scores is not at all a good measure to determine whether or not a model is overfitted. Do you... Even understand what overfitting actually is? Benchmark scores are not the NLBL of a modal quality and are not even worth considering for most people ITT anyway unless you're a vibeshitter like myself

>>108464230
I'm bombarded by Anthropic ads pushing Claude Code so I'm not surprised. I rarely see ads for OAI. If someone had told me in 2023 that Anthropic would come out on top of the US providers I wouldn't have believed them. They had terrible launch issues, a restrictive use culture, and the least resources. But they've obv found their niche.
The knock-on "I'm signing up for Claude b/c DoW dumped them" imo is overblown. I don't think it exactly hurts them, but their subscription volume is driven by them spamming ads on platforms like FB and reddit. There's this handlebar mustache guy I keep seeing but ofc can't find right now that they use over and over, and an assortment of SEA coder chicks.
>>108463369
Anons have been fighting that nonsense forever. "Smarter" models are smarter, but some level of details are best left out imo if the model can't handle them reliably.

>>108464911
>LLM definitely retain verbatim knowledge of things
We don't care you autist. That's not even a local model so that's our relevant given the thread you're posting this is. That particular example wouldn't even count as "IP theft" because it's not generating the entire song, just well known portions. You aren't even allowed to copyright the outputs of any generative models anyway so whether or not it is capable of shitting out thanks for beta does not matter. Channel this autism at a task or hobby that is actually useful or enjoyable and maybe you'll be happier.

>>108464933
>They had terrible launch issues, a restrictive use culture, and the least resources
they still have many issues unique to them
for example a long time ago I wanted to change the mail address associated with my claude account, I'd registered with a shitty outlook throw away like I usually do for testing services and when I thought I might like this enough to pay for it, I decided to change my address.
First thing: you can't change the mail address to a claude account. Their own documentation says this:
https://support.claude.com/en/articles/8452276-how-do-i-change-the-email-address-associated-with-my-account
basically "delete your account, unlink your phone number and make a new one"
so I did
I did and then when trying to make the new account I got the message that "my phone number was used too many times and cannot be used to register accounts anymore"
so the process of trying to change my mail address permanently locked me out of using claude
no I won't change my cell phone number / get a new subscription just to handle claude nonsense
fuck anthropic customer service

>>108464957
>it's not the complete because you didn't screenshot the scrolldown
>I'll disregard your comments about local model testing too
>ai psychotic screeches
kill yourself

>>108464962
good thing you're in the local models general and don't have to worry about misanthropic's fuckups

>>108464969
well, this definitely contributed to my feelings about online services, yes.
But I'd say anthropic is easily the worst and dumbest of all.

>>108464968
back to bluesky with you anti-ai fag

>>108464968
You first. Your insistence on repeating shit people already know doesn't make people want to listen to you anymore. You deserve to have yiur chest caved in for being like this

>you're anti ai if you deny the industry bs about what models really do
kill yourself subhuman

>>108464989
>shit people already know
interesting because this thread is filled with people denying that models are storing verbatim text completion knowledge
kill yourself, subhuman

>>108464990
No one denies anything you say..... We're just questioning why You think anyone here should give a shit

>>108464996
No one says it's incapable of doing that. That research team you quoted proved it is possible. Now what? I do not care. It does not affect me at all.

File: file.png (171 KB, 1761x512)

171 KB PNG

>>108464620
So if I understand that metric well, they were able to retrieve half of "Never let me go" with the finetuned model without even actually adding this book in the finetune dataset?
Wouldn't that mean that the content of the books are somehow verbatim "stored" in the weights of the original model?
Or am I missing something?

File: 12ep.png (177 KB, 1435x615)

177 KB PNG

>>108464929
>Such as?
Dates, quotes, lyrics, the Constitution of the United States, hard facts in general, etc. There are many cases where verbatim knowledge is important or can be helpful. I wouldn't even rule out that short verbatim snippets of good writing could be used in general creative tasks.

>Do you... Even understand what overfitting actually is?
What is done in practice in the industry doesn't necessarily match what is theoretically proper.
Better scores = not overfitting (example in picrel).

>108464996
>muh copyright! be afraid!
terrible attempt

>>108465009
It means the verbatim text or similar text like it was present in the pre-training and post training data of whatever model they tuned, and their overfit caused it to generate either identical or very close text to the original.

>>108464905
Kaprthy thinks so, the knowledge portion of training an LLM is unnecessary. See this rough minute snippet of this interview.
https://youtu.be/lXUZvyajciY?t=814
It makes sense. You really don't need all that knowledge. If you really wanted it, you can just tell it to use a search engine to find it. You only really need the portions that give the model its intelligence. And think of all the memory, storage and otherwise, you would be able to save. But no one has cracked that yet so we keep having models pick up knowledge and intelligence as part of a complete package.

>>108465010
You can absolutely have better benchmark scores and still have a model over fit. That is exactly what "benchmaxxing" anims to do: to nut the model in a direction that leads to it performing better on those tests (Which, I say again, is very easy to do because the benchmark questions and answers are public knowledge. You can run those benchmark tests on your own shit rig so all you would have to do is just inspect the code base or literally just Google search what the questions and answer choices are)

>>108465028
Careful dude. You're going to make the "muh niche trivia knowledge!" autists chimp out again

>>108465028
>If you really wanted it, you can just tell it to use a search engine to find it
fuck you

>>108464962
These providers use AI tools to self manage, so I'm not surprised you got AI retardation from them. I've had similar issues with OIA in 2023. Their system deleted my "use credits" then promptly charged the card on file to buy them back.
>>108464969
I've always assumed most posters here run both.

>>108465028
>the knowledge portion is unnecessary
>"Write a Python script to help me keep track of my groceries."
>"Sorry, I do not know what Python is."

>>108465052
>search online for what it is
solved wontfix

you could have general programming knowledge / understanding of a language or two, then have it translate that to any other by feeding it the rules and syntax in the prompt
in context learning is how people use Zig with LLMs, because Zig is moving way too fast and LLMs cannot produce working code for Zig without you giving them knowledge of the language in context.

>>108464814
I can suggest an equation that has the potential to impact the future:

E = mc2 + AI

this equation combines Einstein's famous equation E=mc2, which relates energy (E) to mass (m) and the speed of light (c), with the addition of AI(Artificial Intelligence). By including AI in the equation, it symbolizes the increasing role of artificial intelligence in shaping and transforming our future. This equation highlights the potential for AI to unlock new forms of energy, enhance scientific discoveries, and revolutionize various fields such as healthcare, transportation, and technology.

>>108465041
>>108465032
see? >>108465046
I was half joking and this man-child proves my point.

>>108465052
I think that anon is referring to quoting niche shit verbatum. Even when doing programming related tasks, you still need wiggle room in order for to actually do things well. There are very few instances where even a coding model would need to generate a snippet It saw verbatim in the training data. Learn what "generalization" is. You almost never want the model to prioritize spitting shit out verbatim.

>>108465026
>verbatim text
Yes, that's what surprises me in this, retrieving verbatim HALF a book they didn't finetune for is not something I thought possible. That makes no sense to me.

>>108465026
>similar text like it
Their metric is word for word.

>>108465072
You're absolutely right, I'm a man-child for not wanting models to regress and be worse than they already are.

>>108465073
>That makes no sense to me.
Which makes me question whether or not they're being truthful. I can buy it spitting out a couple passages worth-for-word if it's a very very well-known series like Harry Potter. Half a book seems fishy, especially for a team of researchers that are very likely only using single parameter or maybe double digit parameter models for their testing at most.

TLDR: I think they're probably full of shit but none of us can prove or disprove that without reproducing their lab environment and testing configurations precisely (which I'm sure was conveniently left out of the paper that was linked)

>>108465080
>woe is me, I would have to set up something more agentic to let my llm retrieve info in my chats

>can't even reply to the correct message
maybe your agent should help with that?

>>108465083
Your insults don't even make sense anymore

>>108463621
>>108463792
This D630 and I have history; it was part of a 2009 severance package from a cursed gig. I installed Fedora 12 on it and used it as a personal machine in 2010, then stuck it on a shelf. Was looking for something to dump openclaw into so pulled it off the shelf, installed Debian 13 and am running it headless. These Dells are solid, well built machines; about 1/2 the weight and heat are the ccfl tubes for the screen, which ofc I shut off.

>>108465072
You definition of "knowledge" is arbitrary.

>hangman
>tetris clone
>world war 2 strategy game
>harry potter magic duel game

Which of these should a model be unable to make because it lacks knowledge as opposed to intelligence?

>>108465094
Those are all superfluous. Already millions of examples on Github you don't need to make more

>>108465073
>Yes, that's what surprises me in this, retrieving verbatim HALF a book they didn't finetune for is not something I thought possible. That makes no sense to me.
they also were able to make it spout verbatim passages of other books which means they are somehow stored, up to half the books, somehow
that's not generalization if so

>"people" rediscovering that models don't generalize for the hundredth time
It's the surgeon's mother! You should all walk to the car wash!

>with the right semantic trigger, the model can regenerate long exact passages from books it was not finetuned on
>maybe a literal file-like copy of an entire book stored intact somewhere in weights
holy shit

>>108464808
>>108465152
>>108465155
so do they generalize or not

>>108465163
LLMs do not "retain" knowledge like some giant lookup table. Please drop the room-temp IQ black-and-white thinking for your own sake.

>>108465163
LLMs interpolate well, extrapolate poorly.

>>108465047
I would be pretty disappointed in most posters if that is the case, I've personally never used any of the corpo models because I am fundamentally opposed to (((SaaS))) and knew from day 0 anything you type in and any outputs it gives you will be used against you and your right to privacy or monetized in an way possible.
No I don't use social media either, not even discord. I do not trust any of them with even the most mundane of inputs and I do not want to contribute to their products knowing the ways they can and will abuse their users so I won't even use them through proxies, I exclusively run local models on my own hardware and block the programs from communicating with the internet after initial setup.

File: file.png (177 KB, 808x1060)

177 KB PNG

>>108465210
just use whatever werks, friend

>>108465210
To me it's just the same road from centralized to decentralized things. We've already seen a direct example as computers moved from 1970s mainframes to 1980s personal computers. When any new tech pops up it's centralized because it's prohibitively expensive... So a large company buys the one thing and then rents out the service. As the technology improves it moves decentralized. There are lots of other examples; the macro tech trend is towards decentralization of production and the inherent customization that allows. It's all very Communist in spirit, which I don't like, but enables individual Capitalism as well, which I do.
I fully expect that we're going to have good, cheap, local AI at some point in the future. But we aren't there yet. In the meantime I'll rent good models and experiment with smaller ones that will run on my existing hardware.

>>108460456
So is another censored and shitty woke SAaRS trap. I'll continue with QwenTTS

>>108465338
Even if vLLM support comes down the line, it will be probably just so a hypothetical licensed full version capable of cloning functionality can work with it. The crippled non-commercial open-weight release will never support that.

File: glm-5-1-is-out-v0-d8efo0q(...).jpg (30 KB, 1217x143)

30 KB JPG

hype

>>108465210
Not everyone is unemployed and has a religious aversion to SaaS. Many people use them through work where it's not their data or money, so why not?
As for here, it makes sense for people to want to be familiar with the capabilities of SaaS models, especially since local always ends up emulating their harnesses and all the models we get end up as downstream distillations.

>>108465323
>decentralization and customization is bad because it reminds me of communism
Fucking burgers I swear

>>108465073
>That makes no sense to me.
makes perfect sense to me, these models demonstrably have an impressive capacity for memorization. they clearly trained it on copyrighted material and did alignment later to prevent it from outputting the copyrighted content. the fine tune removed the alignment so it can complete copyrighted material it had previously already memorized.

>muh cohpoorite!!

>>108465382
will it finally listen to the system prompt?
i remember when glm 4.5 came out it was rough around the edges too, so hopefully 5.x will make it better

>>108465210
>I've personally never used any of the corpo models
I thought the only ones who could afford to train models were corpos, what models do you use?