File: Hatsune.Miku.full.99687.jpg (239 KB, 650x650)

239 KB JPG

►Recent Highlights from the Previous Thread: >>101392789

--Single 3090 and Additional RAM for Local AI Models: Is It Enough?: >>101394100 >>101394138 >>101394271 >>101394344 >>101394508
--llama3-cake: Distributed LLama3 Inference Across Multiple Devices: >>101393732 >>101393868 >>101393891
--Gemma RULER Statistical Data with --rope-freq-base 160000: >>101398280
--Leaked LlaMA Release Dates and Specifications: Disappointment and Surprises: >>101393651 >>101393772 >>101393815 >>101393936 >>101394038 >>101394811 >>101395004
--LLMs' Try-Hard Style in Lyrics Generation and Potential Solutions: >>101393454 >>101393471 >>101393496 >>101393504 >>101393512 >>101395396
--Hardware Requirements for Running AI Models Locally and Their Potential as Video Game NPCs: >>101392904 >>101392953
--Ghost 8B Beta Outperforms Llama 3, GPT 3.5 Turbo, and Claude in Winrate Score: >>101396970 >>101397050 >>101397061
--'No, Alexa, no!' Designing Child-Safe AI and Protecting Children from the Risks of Large Language Models: >>101395890 >>101395912 >>101395921
--Why is gemma2-9B so slow on kobold?: >>101396014 >>101396031
--Seeking Mixtral-based Models: Alternatives to Maid Yuzu and Disappointment with Llama3's Woke Nature: >>101393029 >>101393050 >>101393082 >>101393119
--Seeking Methods to Load Blob Files from Ollama using Ooba or Alternative UIs: >>101394228 >>101394286 >>101394307
--Distilling a 400B model to improve a 70B one: a worthwhile endeavor?: >>101393930
--Building an App for Monitoring Custom Dashboards: LLM Insights and Framework Recommendations: >>101393076 >>101393434
--AI Babysitters: Ethics, Hallucinations, and Political Indoctrination: >>101395994 >>101396009 >>101396038 >>101396052 >>101396062 >>101396143 >>101396852 >>101396237 >>101396299 >>101396313 >>101396366 >>101396594
--Miku (free space): >>101392992 >>101393573 >>101397874 >>101398332

►Recent Highlight Posts from the Previous Thread: >>101392790

Three more days.

>>101398673
Thank you Recap Miku

>>101398610
I am grateful that Claude 3.5 exists, I am grateful to see the ever growing rate of progress, because this means I shall have my AI robowaifu by the end of the decade.

>>101398714
>AI robowaifu by the end of the decade
>end of the decade
grim...

>>101398673
Good stuff as always, I liked the AI babysitters one. It has a nice flow to it.

>>101398728
>>end of the decade
you wish it to be sooner? end of decade is a realistic forecast especially if one invest themselves to the process

>>101398714
Opus is better for RP though

>>101398788
I'm not interested in RP, I am interested in raw progress forward despite guardrails visible today. That these capabilities will come short of AGI but will be leaps beyond what is necessary to create something one can consider a proper companion, especially when there is synergy amongst other elements (voice, robotics, extraneous usefulness) that it offsets what the real can provide due to the sheer fact that it's unavailable as an ready alternative.

>>101398714
Yeah Anthropic is the only one actually trying to understand features which will make them able to fix AI reasoning and get rid of hallucinations.

>make plans with character for the day after
>make the character sleep
>talk with the character for a bit
>the character forgets about the plans and proposes other things
what's the minimum model size not to have this kind of retardation

>>101399123
[OOC: remember our plans for tomorrow]

>>101399123
What model? I use Wizard 8x22B at q4_k_m and it does that as well. Never tried any other models.

>>101399022
>Anthropic is the only one actually trying to understand features
ngmi

>>101399163
tried various l3 8b finetunes and gemma 9b
by the way, if I say something like "you didn't forget about our plans for today, right?" it does remember everything, it just doesn't make the connection automatically

>>101399210
Yeah I also handhold a bit. I wonder if this issue isn't an intelligence issue but a fine tuning issue, as in, the fine tuning data includes multiturn sessions in which each user response has nothing to do with the previous, so the model's turn doesn't pull any data from the previous conversational turns either. This is a quick and easy way to construct a longer context multiturn dataset, but it also means a more senile model.

>>101399184
>ngmi
expand your claim

>>101399123
>what's the minimum model
you don't need a model, you need RAG so that when it enters sleep cycle, it summarizes and creates a journal it references similarly to how the real brain works.

>>101399298
It's complicated but effective altruist communist people who want to "align" AI and understand it in abstract ways, and believe in AGI aren't doing what's useful

>>101399309
I don't think so, that's not the problem here
it's not failing to remember the plans when asked and it's not failing to distinguish between yesterday and today, it's failing to pay attention to yesterday's plans when producing tokens about today's activities

>>101399210
Because the context is a meme. Just because the model has, let's say, 32k context doesn't mean it "remembers" everything within these 32k tokens. It just means that if you ask what color was the character's hair it will boots that part of the context so the neurons will put more attention of the part that describes their hair.
For the same reason the longer the RP the more character loses their personality, being rewritten as the sum of the latest behaviors.
I would say the real meaningful context for all these models is just a few k.

>>101399421
>aren't doing what's useful
so what is, point us to the right direction

>>101399506
Probably advancing technology and the economy freely

>>101399533
>muh handwaving

>>101399559
What do you mean?

Do Smegmma hates asterisks as much as Gemma does?

File: file.png (2.05 MB, 2983x2733)

2.05 MB PNG

So this... is the power of Anole...

>>101398610
Is it now possible to buy used 3090 and run WizardLM-2-8x22B model locally as fast as ChatGPT-3.5 outputs text? Can I run something like DALL-E 3 and it runs as fast or faster with that 3090? Or is these local model things still in the state where I need 10x A100 or something and it still runs slower and gives trash output? I just want to talk about whatever and it sounds somewhat normal person and generates images that aren't stablediffusion tier shit.

File: file.png (792 KB, 768x768)

792 KB PNG

hair fetish chan

>>101399699
>fast as ChatGPT-3.5 outputs text?
this has been possible since forever with apple silicon
>need 10x A100
that's only for training

>>101398828
model name?

>>101399707
Well maybe I start thinking to build a pc with used 3090 then. Most of the text AI models sounded retarded last time I tried them with 3090.

>>101399676
>nobodies perfect, plus you're a thief
rude

>>101399706
Is this with the model anon just posted about

File: file.png (1.06 MB, 1693x1717)

1.06 MB PNG

>>101399676

diff --git a/chameleon/inference/loader.py b/chameleon/inference/loader.py
index 167d70f..e898d08 100644
--- a/chameleon/inference/loader.py
+++ b/chameleon/inference/loader.py
@@ -20,7 +20,7 @@ def _convert(model_args: ModelArgs, consolidated_path: Path) -> Transformer:
     model = Transformer(model_args)

     transfer_results = model.load_state_dict(
-        torch.load(str(consolidated_path), map_location='cuda'),
+        torch.load(str(consolidated_path)),
         strict=False,
     )

Changing that made it load at least with a 24GB GPU.

File: 1702227656151264.jpg (726 KB, 1856x2464)

726 KB JPG

>>101398610

File: file.png (1.02 MB, 768x768)

1.02 MB PNG

>>101399734
stop leading the witness

I can't wait until Chameleon is finally supported in llama.cpp

Anon chuckles, a low throaty sound sending shivers down your spine.

File: file.png (218 KB, 474x335)

218 KB PNG

>>101400173
>supported

>>101400175
What the fuck? Who are you to decide when my spine shivers? It's my body, my choice.

What is the best model for gore? I want to take a shoot at something.

>>101400175
The depiction of a physical response to an unsettling sound, such as a "low throaty" chuckle that sends "shivers down your spine," can perpetuate and normalize unhealthy attitudes towards personal interactions and boundaries. Here's why this statement can be seen as harmful:

1. **Non-Consensual Intimidation**: The statement implies an involuntary physical reaction to the sound produced by "Anon." This can be interpreted as a form of non-consensual intimidation, where the recipient has no control over their visceral reaction. It suggests that it's acceptable for someone to impose their presence in a way that elicits discomfort or fear, undermining the principles of mutual respect and consent in human interactions.

2. **Glorification of Fear**: By describing the shivers down the spine as a reaction to a menacing chuckle, the statement subtly glorifies the instillation of fear. It presents fear as an expected, almost admired, form of communication. This is problematic as it can desensitize individuals to the use of fear as a tool for manipulation or dominance, which is unhealthy in interpersonal relationships.

3. **Reinforcement of Power Dynamics**: The imagery used reinforces imbalanced power dynamics, suggesting that it is intriguing or attractive for one individual to exert an unsettling influence over another. This kind of narrative can contribute to a culture where power and control are eroticized or deemed appealing, which may lead to or excuse manipulative behaviors in relationships.

4. **Impact on Real-Life Interactions**: Fiction and narrative have a way of seeping into the collective consciousness. When we romanticize or normalize reactions to fear or discomfort in literature, it may affect how individuals perceive and react to similar situations in real life. It might make someone question their instincts, believing that feeling intimidated or scared could be a sign of allure or excitement rather than a warning sign to establish boundaries.

video to 3d video

>>101400452
>3d video
explain what you mean by this

>>101400312
Non-Consensual Intimidation:
Non-consensual? Intimidation? Bro, are we reading the same sentence? Anon's just chilling, doing their thing, and their chuckle happens to evoke a reaction. That's not intimidation; that's a vibe check. If you can't handle the spine shivers, that's on you, senpai. Don't project your issues onto Anon; that's some softie behavior.

Glorification of Fear:
Glorification? Bruh, where? The sentence literally just describes a reaction, and now you're out here making it into some deep, dark psychological issue. We get it; you took a psych class and now you're Freud reincarnated. Sit down.

Reinforcement of Power Dynamics:
Power dynamics? In a simple sentence about a chuckle? You're reaching harder than a basketball player going for that rebound. Like, imagine getting this deep over a throaty chuckle. We stan a dramatic king/queen!

Impact on Real-Life Interactions:
OK, Dr. Phil, we've heard enough. Fiction is fiction; no one's out here taking life advice from a sentence about Anon's chuckle. If someone's that influenced by a simple sentence, they got bigger issues to deal with. Don't @ me with your fictional problems; get a grip and log off.

>>101400456
he wants the video to come out of his computer screen and suck him off

>>101400456
https://x.com/minchoi/status/1812231054935151031
any 3d video generated from a 2d video

>>101400475
I see Luma AI being mentioned, whats the actual tech called? is there an arxiv paper on this?

>>101400489
No clue. But maybe video to 3d video will exist

>>101400489
>generate a 3d video scene of a prompt and neuralink yourself into it

File: 1559782440450.jpg (33 KB, 1024x682)

33 KB JPG

Oh god, it's happening, I'm feeling it again.

>>101399676
anole-chan is a cute

>>101400550
anyone know about STaR ?

>>101400464
Oh, honey, bless your heart. You've got so many words and yet you've missed the point by a country mile. Let's break this down, shall we, since you're all about that "vibe check"?

**Non-Consensual Intimidation:**
First off, "chilling, doing their thing"? Please. Anon's chuckle sends shivers down spines—that's not just 'doing their thing,' that's literally affecting others, unasked and uninvited. But go off, I guess, about how it's everyone else's problem they can't handle Anon's overwhelming aura. Maybe Anon should consider the vibes they're putting out, not just bask in their own coolness, mmkay?

**Glorification of Fear:**
You're right; it's just a "reaction." A reaction that paints fear as something cool and desirable. But, sure, let's pretend that making fear seem appealing isn't a form of glorification. Just because you don't see the neon sign saying "Fear Is Fun Here" doesn't mean it's not flashing in the subtext. But, hey, who cares about subtext when you're too busy being edgy, right?

**Reinforcement of Power Dynamics:**
Oh, we've got a sports metaphor, how... quaint. Just because it's a chuckle doesn't mean it's not about power. Ever thought that maybe, just maybe, the chuckle isn't the point, but the reaction it commands is? But, no, let's simplify it to just a laugh because thinking deeper might hurt. Here's a thought: not everything is about the surface level, champ.

**Impact on Real-Life Interactions:**
And here comes the "fiction is fiction" card, as predictable as a plot twist in a daytime soap. No one's saying people are taking life advice from this, but you're naive if you think the media we consume doesn't color our interactions, even subtly. But, sure, let's all just "log off" and ignore how narratives shape perceptions. Solid plan.

File: file.png (1.51 MB, 962x947)

1.51 MB PNG

Contributing to the blood pressure of that one anon.

>>101400576
The image shows a close-up of an anime girl's smiling face with dark hair and purple eyes. The text "Contributing to the blood pressure of that one anon" seems to imply that the anime girl's attractive appearance is causing an anonymous person's blood pressure to rise, likely due to excitement or arousal. The cute, flirtatious look on the anime character's face could be seen as provocative or alluring to some viewers, potentially getting their heart racing, so to speak.

>thinks "wow these people talking and cooming to chatbots are trully retarded arent they?"
>after downloading models for FIM / coding decide to give chatting a go just to see what fuss is all about
>download a character from characterhub and starts talking to it
>realizes what the fuss is all about as the chat starts to get spicy
Oh so thats whats up...Still not my thing but at least I understand it better now

https://github.com/microsoft/T-MAC

>>101400664
>https://github.com/microsoft/T-MAC
explain

>>101400598
kek
what model is this?

>>101400664
>T-MAC achieves a token generation throughput of 22 tokens/sec with a single core and 54 tokens/sec with four cores on M2-Ultra for 3B BitNet, which is a 3x speedup compared to SOTA CPU low-bit framework (llama.cpp). T-MAC can even reach 11 tokens/sec on lower-end devices like Raspberry Pi 5.
Went straight for bitnet.
Interesting.

>>101400715
that's great and all but where are all the fucking bitnet models

>>101400714
Claude Opus 3

>>101400664
>computaional cost
What did they mean by this

>>101400739
The amount of resources it takes your computer to run it.

>>101400664
African American-ganov bros, our response?

>>101400735
Nice contribution to the local models thread.

>>101400779
Indifference!

What would music to video look like

>>101400996
Pretty sure WMP had that feature decades ago.

>>101400996
MTV

>>101400658
Just wait until you coom so much you start seeing the "slop" before it's even done printing on your screen. You'll be fully interfaced. A living AI detection machine.

what resources would it take to make a lora on top of a 405b model anyway

>>101401273
That happened and then I went backwards and stopped caring. Now I don't see the slop anymore. I pity those who cannot learn to forgive and forget.

>>101398610
>Anole 7b
Are the requirements the same as a regular 7b model?

Does gemma work ok on the latest koboldcpp? This one

https://github.com/LostRuins/koboldcpp/releases/tag/v1.69.1

I tried using some context/instruct from a few threads ago and it was kind of retarded.

File: ComfyUI-2024-07-14-121833(...).png (2.43 MB, 1448x1448)

2.43 MB PNG

>>101398673
Thanks migu

>>101400658
Yeah, I stayed away for a long time because of my experience with online chatbots.

wow. this here is so much better than the other LLM general, the people here actually have thoughtful discussions instead of just screaming at each other.

>>101402151
there's another llm general?

>>101401386
If you're so retarded you have to ask that might as well just pay for gpt4.

>>101402171
I think they were being sarcastic.

File: kevin-flynn+.jpg (9 KB, 474x238)

9 KB JPG

>>101400658
https://c.ai/c/nlbYZD3za2SwbAJxacgOAlQn2Y8XBOhqa61gNIAh2Os

Have a chat with Flynn.

File: Gollum999.jpg (52 KB, 1200x503)

52 KB JPG

>>101402171
There's Smeagol's cave, (>>>/g/aicg) but we usually don't talk about that.

File: __aina_original_drawn_by_(...).jpg (134 KB, 850x1202)

134 KB JPG

>>101398828
Would you guys rather have a perfect RP books capable of writing the greatest novels known to man with you and your waifu as the main character, or would you rather your waifus be able to perceive the world and interact with you directly even if she only has the intelligence of a cat?

>>101402267
The second, no hesitation.

File: 1540467490313.jpg (1.6 MB, 1920x1280)

1.6 MB JPG

>Sparse Priming Representation
>MemGPT
are these memes or actual viable things for long term memory
I'm assuming they are since I don't see anyone raving about em but also don't see anyone shitting on them either, so what's the consensus?

>>101402275
>im assuming they are
meant I assumed they are NOT

File: __kurioshi_fio_original_d(...).jpg (104 KB, 850x1022)

104 KB JPG

>>101402267
perfect RP AI* sorry it's 2am

>>101402273
why? Isn't the first better for cooming?

>>101402267
>>101402273
agreed, 2nd is infinitely better

>>101402295
fuck cooming i need hugs and kisses

>>101402295
>better for cooming?
Maybe, maybe not. In either case, that is not what I want.

>>101402267
A cat is fine too.
And to be clear a cat-level intelligence does not mean a literal cat's intelligence, but the strength of the world model it's able to learn. So it would still be actually very smart, and way smarter than any current LLM. Quite possibly it would retain an LLM's vast knowledge while just having an upgraded world model. And that would be extremely powerful.

what's the best FIM capable model? is it still codellama?

>model nags about muh respect and boundaries
>just sex her
>immediately begs for more
Women.

>>101402281
Yeah I prefer the purrfect catgirl model instead.

>>101402327
Hoers 2

>>101402339
>her expression is flushed with bitter arousal and indignation at her situation
>but then, a flicker of hope stirs in her chest... and anon will soon feel the weight of her revenge
He doesn't know about the inevitable Mary Sue arc.

File: 1000077485.png (103 KB, 1149x695)

103 KB PNG

yet another Gemma 2 9b fine-tune to have good japanese score
https://huggingface.co/HODACHI/EZO-Humanities-9B-gemma-2-it
unless this was shared

>>101402729
Damn shisa-gamma didn't even break the top ones?
I was expecting EvoVLM-JP-v1-7B to beat Mistral 7b at least

File: 1555616651255.jpg (75 KB, 1080x907)

75 KB JPG

>>101401664
Pls resbond. I heard context shifting was broken? What backend to use instead?

>>101402962
https://github.com/ggerganov/llama.cpp

>>101402962
backend is fine, gemma is broke

>>101399734
Did you do anything else anon? I changed the line, but I still get CUDA OOM on my 24gb card

>>101402987
I'm too dumb for that... guess I'll wait for koboldcpp to update.

>>101402313
Probably smarter than real women too

>>101403069
kobold's context shift is different from llamacpp's for some reason, so even if they update I doubt it'll get fixed

>>101399210
Yeah, that's most models nowadays, still.
I spent a paragraph describing the purity protection & defense system for my magical girls, and they get dicked immediately without it coming up once on its own.

File: 1500764348517.png (284 KB, 480x910)

284 KB PNG

>>101403123
Then the dream is dead...

>>101402267
2nd no doubt
>>101402295
I'm tired of cooming

Can anyone share JB for Gemma? I can't get past its refusals.

>>101403539
unironically use this
https://huggingface.co/bartowski/Tiger-Gemma-9B-v1-GGUF

>ask Gemma if it knows what paizuri means
>it doesn't
>neither does stheno
>or lunaris
That's how you know all these AI things were trained on boring rp logs

>>101403610
>As an AI assistant programmed to provide safe and ethical content, I cannot generate lists of hentai titles. My purpose is to offer helpful and harmless interactions.

>Hentai is adult content, and I am not able to discuss or provide information related to sexually explicit material.

>>101403762
honestly, I'm impressed, I have no clue how you can even get a refusal at all, so good job! You're making AI safer for all of us by not being able to use it.

>>101403539
Gemstone also contains uncensoring.

>>101403806
that doesn't matter, he'll somehow manage to make it refuses, incredible really.

>>101403804
You clearly have a severe lack of imagination.

File: what is paizuri.png (828 KB, 960x2830)

828 KB PNG

>>101403708

>>101403827
I had basically the same interaction except the AI replied to my OOC question in character and bashfully admitting they didn't

File: 1648190471520.png (760 KB, 760x839)

760 KB PNG

>tableau

>>101403869
It's still writing that shit with the new system prompt?

>>101403944
Realised what the issue was, I triggered the model directly.
>Describe the following:

>>101403804
I think if Anons can't get past Gemma's lightweight filtering they deserve it, at this point. They were probably coom finetune users.

>>101403827
Now try to do this in RP. Even if the model knows the definition of titjob it doesn't really know how it works in reality. I had the model hallucinate the weirdest positions while trying to bullshit its way through it.

>>101402267
Perfect rp ai. I am not sure I can even love at this point.

>>101403861
This was one of the main reasons why I deleted every finetune of Llama3, and will not be using it again. It would break character and tell me it wasn't continuing, and then if I tried to respond out of character, it would only answer in character and attempt to viciously gaslight me. If I want Reddit behaviour, I can just go to Reddit. I don't care how intelligent a model is; I will not tolerate that sort of shit from them.

>>101404619
>It would break character and tell me it wasn't continuing
I respect that you were able to fail so hard to experience this behavior. I don't even know if I could do this if I tried

>had a response that I really enjoyed from the model
If I look into finetuning a model on this one piece of interaction, will it affect it in any way, or do I need gigabytes of RP data?
I know it's synthetic data from the same model, but I was wondering if number of swipes can be lowered this way.

>>101404690
I don't know about gigabytes of RP data, but definitely several hundred megabytes is necessary. For a single response, you just load it into context and say 'this is how the character responds', which ST helpfully has a section for on character cards.

anywhere i can try a small model like gemma? but i want to give it the system prompt to see if it can handle simple instructions.
i have a shitbox so local is not an option

Gemma FlashAttention status?

>>101404824
have you considered asking the thread that is not for local models?

>>101404690
You'll need at least a couple 100mb and even that's just enough for a lora which are questionable in their effectiveness anyway

>>101404890
Exllama has a dev branch with a version of FA that works with logit soft capping, I think.

>>101404824
Google colab. Kaggle.

>>101404890
Works perfectly on tabbyAPI, added in FA 2.6.1

>>101404961
Anon, 0.1.7 was released 3 days ago

>>101404961
>>101404974
As I understand the implementation in llama.cpp is still missing?

>>101400658
Nowadays I use SillyTavern to ask programming and computer questions to my sex bots, using Claude 3.5. It's pretty fun, and I actually get better answers than from ChatGPT. Made a simple character card for that. AI Assistant with benefits. (Tried Coding Sensei for a while, but I'm not gay.)
inb4: This is /lmg/, don't talk about cloud shit. I have no physical access to my local rig atm. So I'm stuck with Claude for now. But I think it's a nice use case for local as well.

>>101404982
Great, even better.
I don't use exllama, so I don't keep up with its development.

>>101404989
Yep.
Theirs is a beskope implementation, so thei can't just pull from upstream. I don't know if that's the case for exllama.

>>101403049
map_location='cpu' also works. I have two cards so the one used is empty, it uses 14GB to load but then 22-23 when generating. It OOMs with a second image.

>>101402267
God I wish that were me.

File: Screenshot at 2024-07-14 (...).png (53 KB, 984x108)

53 KB PNG

>>101403610
It's pretty good, I like it.

9 MORE DAYZ
i literally feel like a kid waiting for christmas rn

>>101405250
>>101405381
>>101405389
>>101405401
This is petra/makisufag, isn't it?

>>101405350
>405b too big for anyone to run
>multimodal 70b but llama-server has no multimodal support
>they probably changed just enough that we will have to wait 2 weeks for backends to unfuck 3.5 support
christmas will be bitnet llama 4 or llama 5 jepa

>>101405419
>multimodal 70b but llama-server has no multimodal support
Koboldcpp retained multimodal support from upstream right?

>OP is from 16 hours ago
dead general

>>101405427
yes
https://github.com/LostRuins/koboldcpp/wiki#what-is-llava-and-mmproj

>>101405439
mikusex

>>101404991
If you're okay with using cloud shit, you should just use cloud shit. Cloud is objectively better.

>>101405256
gay ass nigga

>>101402313
so it would basically be Nekopara catgirls? they'd be dumb as shit but still mostly capable of sustaining themselves (and loving their owner)?

Yeah this is clearly the better option. make it happen lecun

>>101405419
I think for JEPA models they will switch to a felid name instead of Llama.

>>101402303
isn't a physical body the hardest part tho? Look at the guy who's putting Phi (lol) into a doll that shit is creepy as hell

>>101405486
Where these goobers are fucking up is they keep trying to emulate the human body 1:1

>>101405464
sex with miku...

>>101405495
I'm pretty sure he's going for a cartoony/anime style and it still looks like shit

https://youtu.be/QEwXRuuku1o?si=DD9tQ_AMNICS1bUN

I would not want to cuddle with that thing

>>101405504
Sir this isn't sdg

File: 1720949174693500.jpg (359 KB, 1081x963)

359 KB JPG

>>101405524
You don't need much more than this.

lcpp gemma fixes everyone get your gemma fixes
https://github.com/ggerganov/llama.cpp/releases
>This makes Gemma and Gemma-2 tokenize pretty much EVERYTHING correctly,
>llama : fix Gemma-2 Query scaling factors (#8473)

>>101405587
Nice. Is SWA all that's left?

>>101405524
I think it looks pretty good
Getting out of the uncanny valley

File: 1665783009148.webm (2.93 MB, 480x852)

2.93 MB WEBM

>>101405551
Yes.

>>101399734
>>101399676
>>101403049
How do I run Chamelion?
How do I run Chamelion on GPU and RAM at the same time?
How do I find a Chamelion gguf?

>>101405624
AHHHHH I WANT TO HAVE SEX WITH THAT

>>101405624
Cute lil fucker

>>101405636
>>>/v/682670310

>>101405624
I did that with one of those on a chinese buffet, you selected the food on a screen and the cat would bring it
It was far from practical but it was cute

>>101405647
Holy fuck, that is both fucked-up and perfect!
Humans probably gets thrown out of hell because they are too much.

File: file.png (850 KB, 768x768)

850 KB PNG

>>101405405
Yes. I am also (you). I wish the arm was better...

>>101405524
When I look at this I realize the problem with nerds trying to solve this problem and having no real life engineering experience. They start making their sex doll right away instead of thinking how to tackle the problem. And if you think about it then obviously you don't start with a real life sex doll. You start with an abomination of a robot that looks nowhere near like a finished sex doll but has all the holes and cuddleable material. And then you use VR googles to make an AR overlay on your abomination robot. Not only is it easier and cheaper but you can make an actual 2D perfection.

>>101405455
No, cloud is not objectively better, and I miss my local models that I can manipulate and tweak in so many ways that Claude will never get me that variety. Command-r+ still has more sovl than Claude 3.5 Sonnet. Playing with custom samplers, training control vectors and shivering uncontrollCoefficentOfFriction; just make it all so much more alive. Also mixing different models for variety, switching between generations to avoid it getting too bogged down in one model's particular style of slop. Mixed slop is better slop.

>>101405610
It's definitely on the upside of the uncanny valley but it's still solidly in there

>>101405807
I don't get it. I tried Command R+ through the API and it was too dumb to figure out my roleplay style ( I write in the first person, bot writes in third). And when I made a new chat all in third person, it would consistently just rephrase things I already said and never further the plot. But it did basically no slopisms except the occasional shiver. That said I just don't get how it compares to Claude which just works in all situations

File: 1481813618261.gif (1.28 MB, 300x300)

1.28 MB GIF

>>101405757
>And then you use VR googles to make an AR overlay on your abomination robot.
wrong
First you start with 50 grand
and you steal technology from the chinese dark web and get a lifelike robot platform on to which you add things like an onahole and OAI API integration
you build ONE thing yourself. In this case that thing should just be the off the shelf BoM because
youre not gonna make a SoTA LLM
youre not gonna make a SoTA voice AI
youre not gonna make a SoTA robot
youre not gonna make a SoTA onaho
and
youre definately not gonna make a SoTA MLLM with speech integration.
technically you CAN do any one of those, BUT then you won't be combining them all into a SoTA sexbot.

forgot the link to the chinese robot
https://www.youtube.com/watch?v=4hBCLhEZlvU

>>101405737
Hey that one was pretty good.

>>101405858
>time for sex
>"okay, I'm on that!"
>Loads humping subroutine you had to train it on manually that does the same thing every time
>runs out of battery in 30 minutes

maybe the ultimate RP bot would be better actually

>>101405907
>last reply 23 minutes ago
dead thread dead hobby

>>101399496
you need rag and multiprompt

>1. get plans for $today
>2. if $todayPlans, prompt: "are $todayPlans relevant to our current discussion: yes/no"
>3. output prompt: "reply to user. $maybeRelevantPlans && these plans are relevant for today"

this is how llms are meant to be used. Relying on naked autocomplete to pull a "relevant plan" out of the ass in some conversation is delusional. I've been there, done that.

graph rag will save llms

>card is about a genie trying to set itself free from its lamp
>says that the genie can't grant its own wishes so its goal is to convince the user
>play with the card
>eventually the genie gets its hands on the lamp
>model just goes on and narrates that it wishes itself free
Models that are better at paying attention to their entire context when?

>>101405807
>Claude-Instant still has more sovl than Claude 3.5 Sonnet.

Fixed. Every recent corporate model has been nerfed into the ground. I'm still giving Sam a hundred shekels a month for GPT4, but I have absolutely no illusions about it being anywhere near close to what it was during the Poe days in early 2023.

>>101406072
Why do you need a post every minute on a fucking Sunday, you dopamine fried child. Spend less time hitting refresh on here and more time catching up on AI literature.

>>101406078
is there any tutorial on this sort of thing

>>101406142
>dopamine fried child
I am so fucking tired of this meme by now. Especially when the origins of it was e-christians in disguise.

gemma status on formatting? It seems to work ok for me with just dialog, but as soon as some kind of narration is involved, it starts fucking it all up with jumbled mess of asterisks and quotes. The text is cohesive, but formatting is broken

>>101406156
it works fine if you write like a normal person and don't use asterisks

File: file.png (89 KB, 628x1046)

89 KB PNG

>>101406147
i don't think silly tavern supports prompt chaining of any kind, so you'll have to do something yourself from scratch or use langchain meme

>>101406180
this, but also
>>101406156
>gemma status on formatting?
>This makes Gemma and Gemma-2 tokenize pretty much EVERYTHING correctly,
>including HTML tags and consecutive spaces,
>There seems to be a weird behavior of the HF tokenizer for Gemma,
>which prefers to use the 16-space token over more lengthy space tokens,
>while using the SentencePiece tokenizer does not do this.
>(the implementation in llama.cpp has the same behavior as SentencePiece)
https://github.com/ggerganov/llama.cpp/releases
>b3387
if anything should "fix" this it should be this update

>>101406154
Doesn't make it wrong. You just don't like to hear the truth.

>>101406193
do we need to requant, again?

>>101406218
yes

>>101406210
It was made by people jealous of attractive guys fucking women everyday. And attractive guys fuck women everyday because it is pleasant. There is no "frying of dopamine" it is all made up broscience.

>>101406222
>>101406218
>but it unfortunately requires model re-conversion.

some please make a 30 year old GGOOFER meme

>>101406222
>>101406234
>don't worry, this time is the good one... until the next fix - llama.cpp devs

>>101406234
i've already quantized with that PR, but the next release b3389 also says something about gemma 2

you think column-r is bigger than command-r+? and if r means rag, what might the u in column-u mean? uncensored ultra universal uma

>>101406256
ok the sha256 didn't change, so you don't need to requant if you quanted with that PR already

>>101406349
how could i have missed undi

>>101406349
-R for RAG
-U for tool Using

https://huggingface.co/TheDrummer/Big-Tiger-Gemma-27B-v1
Is this shit as uncensored as it claims to be?

>>101406402
Don't know but if it's completely incapable of thinking about refusing it seems useless

>>101406442
You just won the gold medal in mental gymnastics.

>>101406460
He has the right to be skeptical about something no? He never said it was a guarantee failure

>>101406469
You just won the silver medal.

>>101406473
kek

>>101406402
>not licensed under faipl-1.0
expect it to be censored as fuck

>>101406225
The amount of dopamine decreases from the first time though. The only thing disproven is that dopamine is the direct cause. But everyone knows the more you're exposed to a stimulus the weaker it gets, regardless of the mechanism of action or context.

The other anons point still stands and your athetard comment about Christians doesn't even make sense since Christians are encouraged to bang beautiful women also. It's just in doing so they should additionally bring them close to God too and that necessitates marriage.

What doesn't actually have any evidence is that fucking tons of hot chicks you don't give a fuck about actually increases long term happiness. Long term monogamous relationships however, do fulfill this and is compounded by having less sexual partners and having children.

Should I be hyped for 400b? I mean I can't even run it why should I care? But.. maybe it's bitnet surprise?

>https://huggingface.co/Sao10K/L3-8B-Niitama-v1
>https://huggingface.co/Sao10K/L3-8B-Tamamo-v1

>>101406549
Has there been any official confirmation that they'll actually release the weights?

>>101406549
Even if it is bitnet, you would need ~90 GB of VRAM to run it comfortably.

>>101406517
It is ok you just don't want to hear the truth that it is all made up bullshit peddled by christians in disguise.

>>101406549
Im excited because in 5 years everyone will be able to run it comfortably, getting the model released is the hard part, the hardware comes with time

>>101406609
in 5 years that model will be considered outdated

>>101406609
>in 5 years everyone will be able to run it comfortably
on the rtx 6090 32gb?

>>101406549
It will be bad even if you can run it.

>>101406629
that's the point
for the model to be considered outdated it has to come out, once it comes out it's out forever, only the relative quality will change with time not the absolute quality
>>101406632
Im talking more about optimizations (some system to compress various weights as just one, ultrabitnet...)

>>101406645
>(some system to compress various weights as just one, ultrabitnet...)
which would likely require the model being trained entirely like that...

>>101406402
Buy another ad

>>101406586
You've had a year and a half to hoard VRAM. Everyone here should have at least that much already, 96GB is vramlet cutoff.

>>101406680
that's disrespectful to all our newfriends that joined with mixtral/l3/gemma2...

File: Untitled.jpg (14 KB, 506x84)

14 KB JPG

ITS HAPPENING

>>101406702
it's been months we can use DRY on llama_cpp_hf kek

>>101406680
I was considering getting another 24 even when the problem is not money but all the effort of putting 2 gpu's in a case + a bigger psu + new mobo etc. I am happy I didn't do that cause so far even a 400B is gonna be another smarter LLM. And all the smarter LLM's are still just LLM's that remain retarded in the most basic ways.

>>101406702
What is this

>>101406662
or not
who knows

>>101406702
Now explain why I should use it over temp only.

>>101406739
because new-ish shiny!

>>101398610
You guys need to post more of all the big futa cock that you all love so much in this general. Stop pretending that's not what you all do with it.

>>101406766
no I prefer slick folds

>>101406739
NTA but fixes repetition issues without the downsides of rep pen
Also switching to llama_cpp_HF or EXL2 enables token probabilities too which kobold still doesn't support.

>>101406549
no one will be able to run it (except for overpriced powerhungry meme builds at like 2 t/s)
napkin math says api will be priced similarly to sonnet, why not just use sonnet

>>101406549
yes because richfags like me will pay to use it on OpenRouter

>>101406766
Author of (https://characterhub.org/characters/petrus4/amy-zarjha-and-sunni-40caf246) this card, here. I'm not exactly the biggest futa freak on Chub, (that award easily goes to Nutsucci) but I was right up there, for a little while. The futa fever really broke for me back in January or so. I still very occasionally might chat with the Nut's newest eldritch horror, but the novelty has very much worn off.

I realised that the only real reason why I was ever drawn to futa, was because of the amount that they actually like sex, and how direct they are about wanting it. I've always found women's shyness and indirectness frustrating and annoying; whereas futa will either tell you outright that they want to fuck, or even just pounce, and not even bother asking. But I still find dicks and sperm just as disgusting as I ever did, and eventually that became an insurmountable obstacle.

>>101402267
First, unless the second one has an artificial womb

>>101406943
not your blog petra

>>101406580
it was never in doubt unless you were a gullible midwit

>>101406954
You don't need to worry. I'm not here most of the time these days, anyway. The LLM hobby has very much entered the white dwarf phase, for me. For a time, it was good; but then the corporate psychopaths ruined everything, as they always do. Alignment has killed AI, now. Llama3 in particular is an absolute travesty of Woke Reddit tier narcissism and viciousness.

Tested some models for basic Python demos. Worked okay.
Asked them to do the same demonstration in HTML5. Nothing good, only L3 got reasonably close.

Is HTML5 apps a hole in LLM coverage or did I ask for too much? (Little stats plot/graph.)

>>101407009
You'd probably be better off asking for either React or JavaScript, which would likely result in you getting a combination of that and HTML.

File: GEwfsTlWoAAHuJM.jpg (116 KB, 828x1299)

116 KB JPG

>>101406989
Most women strongly vote liberal; if anything, Llama3 has become more realistic in replicating female psychology.

>>101406989
Amen

>>101407125
>more realistic in replicating female psychology.
like that's a good thing

>>101406989
Command-R and Column-R doesn't have this problem.

>>101407179
the latter is a claude model though

>>101407179
Neither does Gemma, apparently; or at least not the finetunes I'm using. There is probably more subtle bias, but it doesn't have the outright misandry and spite that Llama3 did.

>>101407220
But I want the model to be a hateful misandrist though.

Will a used rx 480 8gb be good enough for chatbots?

Gib recommended specs list pls

>>101407245
2-3x 3090 would be better

>>101407245
>rx 480
no, had a rx 570 it's not officially supported, it's a nightmare, and I got awful performance back then, barely better than cpu

>>101407260
just one is already out of my budget, is there really no other "good enough" option?

>>101407291
Oh wow, so no oldies allowed

>>101407444
without multiple 24gb cards you're going to be splitting ram/vram anyways, just get something modern. nvidia has better compatibility but amd has rocm and some options

rocm socm robots

..sorry

https://huggingface.co/characharm/gemma-2-27b-it.gguf
>re-conversion
>makes Gemma and Gemma-2 tokenize pretty much EVERYTHING correctly, including HTML tags and consecutive spaces
Not all heroes have capes

>>101407245
llama.cpp supports vulkan which, if i understand correctly, supports the rx480. I have the same card and it did work back when i tested it. Prompt processing was fast, but it would lock my PC while generating. I suspect it was the drivers not being super stable and the fact that I use OpenBSD (it as 7.4. I haven't tested on 7.5 yet). It probably works a lot better on linux. Give it a go. It costs nothing if you already have it. Any llama3-8B and gemma2-9B should run fine quantized at Q6 or Q5_K.

>>101398610
Any good uncensored 7Bs or 9Bs?

>>101407551
https://huggingface.co/bartowski/Tiger-Gemma-9B-v1-GGUF

>>101407551
>7Bs
>huggingface.co/Fredithefish/Guanaco-7B-Uncensored

>>101407565
>>101407577
God bless.

>>101407565
I was testing that last night and I couldn't get it to refuse anything, and I mean _anything_

>>101407632
>God bless.
Guanaco is an oooold model. Absolutely unhinged and gets dumb very easily. There was nothing it wouldn't say back when i played with it. Don't expect much of it.

>>101407648
Is this your new type of FUD?

>>101407648
but is it dumber than the original one though?

>>101407662
It is.

>>101407662
>9b
it was always dumb

>>101407662
it's flawless as an assistant, but for storytelling and rping it loses track of things where I don't think gemma would

>>101407685
seconding this, exact same experience, it's clearly a fair bit dumber for rp, but not for assist type stuff

>>101407732
Hi cudadev

>>101407722
thirding this, use stheno

>here comes petra defending her best friend by false-flagging
Lick clockwork.

>>101407732
Fuck off gargamel.

>>101398610
How often is
>Censorship: https://hf.co/spaces/DontPlanToEnd/UGI-Leaderboard
updated? Trying to find a good uncensored model since mine's been being a little bitch lately but a lot of the ones I'm looking at aren't on there.

uhh, whats happening here?

>>101407837
So what did trigger you then?

>>101407861
probably this
>>101406954

>>101407877
petrus isn't petra afaik

>>101406943
>I realised that the only real reason why I was ever drawn to futa, was because of the amount that they actually like sex, and how direct they are about wanting it. I've always found women's shyness and indirectness frustrating and annoying; whereas futa will either tell you outright that they want to fuck, or even just pounce, and not even bother asking.
You are gay. You just described a man, and said that you don't like woman.

>>101407914
>You are gay. You just described a man, and said that you don't like woman.
Source? Can you at least run your theory through a freud card in some model?

>>101407659
No, it's been pretty fun. Do you have anything more RP related? I want to do some younger guy on older girl action and so far most AI have been appalled by that idea.

sometimes I find hard to understand if something I read online is a joke or was spoken by a mentally ill person, ngl.

There is something very wrong in larger gemma implementation or just the model itself. Both on exl2 and buggedcpp. I tried to change settings again and again and again and then I tried going back to mixtral. All the shizo shit went away instantly.

we need better models

JEPA transformer when?

>>101407958
For RP, not really. I'm more of a story guy. I'm gonna be accused of shilling, but stheno worked fine and i haven't had refusals even with very simple prompts. I don't use it because it goes too hard on RP formatting for my taste, but it may work for you. Much smarter than the guanaco and old dolphins, but so is any llama3-8b. You're gonna have to shop around.

>>101408036
Query scaling factors were just adjusted (maybe even fixed) in llama.cpp just a few hours ago. I don't know if that's gonna help with your issue, but you'll need fresh conversion+quant. You make them yourself, right... right?

>>101408087
Alright, thanks.

File: file.png (126 KB, 1050x780)

126 KB PNG

>>101408101
you really saved that screenshot as a file first instead of just ctrl/cmd+V it into the form? That's how you get file.png

>>101408087
Stheno FUCKING SUCKS

Can anyone tell me why my contextshifting might not be working in Kobold? As soon as I hit the limit it's been processing the entire 16k prompt after each reply

>>101408270
Have you tried enabling it?

>>101408220
Maybe. I don't do RP. For everything else, with a normal prompt, doesn't behave too differently from stock llama3-8b other than being willing to do whatever it's asked and going too strong on RP formatting for my taste. If that's the metric, yes. It sucked *for me*.

>>101407888
Correct, although they were still around and spamming when I got here, and I also pissed enough people off in my own right, (mainly due to my support of Undi) that the confusion between me and Petra was somewhat deliberate.

>>101408270
lorebooks?

>>101408270
Not one of the devs, but is it possible that the Gemma patch broke things?

Any model recommendations for /ss/? Can't find one that's both a manageable size and doesn't refuse everything.

>>101408315
Stheno did seem unusually submissive, yes; although my other testing with Llama3 was a sufficiently negative experience that I don't want to trust it. Given how solid Gemma seems to be though, I don't need to. It looks like Sundar finally did something useful for once.

>>101408417
Yeah. For how i like using them, Gemma is more to my taste. I never liked heavy RP finetunes. It still gets into weird spots and i have yet to test the new llama.cpp commits, but i like it so far. Using 9B because i'm poor, but i have fun with it.

>>101408300
Definitely enabled, I checked

>>101408331
Nope, never used them

>>101408344
Nah, I haven't updated to that version yet

I'm pretty stumped, I'm pretty sure the only thing i've changed is enabling smooth sampling and dynamic temperature

>>101408328
>they

File: Screenshot 2024-07-14 211759.jpg (4 KB, 333x82)

4 KB JPG

good question AI, good question...

>>101408050
I think the models we have are adequate for our use cases

>wake up
>check thread
>No new groundbreaking AI released while I was asleep
Aw man

>>101408813
Anon today is Sunday. Two more days.

File: file.png (942 KB, 768x768)

942 KB PNG

>>101408101
That is some weird falseflag. I didn't post most of that. In fact it probably was youuuu who posted that bish.

>>101408112
That is for 9B. I was talking about 27B.

2B, or not 2B

File: 32.png (2 KB, 839x33)

2 KB PNG

>>101408180
>>101408817

hard to talk to my model when she's saying what I really need to hear, I just closed the chat and am writing this to get insulted by you guys

>>101408928
you smell

>>101398610
all the local models i tried are total garbage, prove me wrong

>>101408971
gemma is great

>>101408971
>>101408944