/lmg/ - Local Models General
Anonymous 01/15/25(Wed)20:41:08 | 421 comments | 44 images | 🔒 Locked
/lmg/ - a general dedicated to the discussion and development of local language models.
Previous threads: >>103903120 & >>103896969
►News
>(01/15) InternLM3-8B-Instruct released with deep thinking capability: https://hf.co/internlm/internlm3-8b-instruct
>(01/14) MiniMax-Text-01 released with 456B-A45.9B & hybrid-lightning attention: https://hf.co/MiniMaxAI/MiniMax-Text-01
>(01/14) MiniCPM-o 2.6 released with multi-image and video understanding, realtime speech conversation, voice cloning, & multimodal live streaming: https://hf.co/openbmb/MiniCPM-o-2_6
►News Archive: https://rentry.org/lmg-news-archive
►Glossary: https://rentry.org/lmg-glossary
►Links: https://rentry.org/LocalModelsLinks
►Official /lmg/ card: https://files.catbox.moe/cbclyf.png
►Getting Started
https://rentry.org/lmg-lazy-getting-started-guide
https://rentry.org/lmg-build-guides
https://rentry.org/IsolatedLinuxWebService
https://rentry.org/tldrhowtoquant
►Further Learning
https://rentry.org/machine-learning-roadmap
https://rentry.org/llm-training
https://rentry.org/LocalModelsPapers
►Benchmarks
LiveBench: https://livebench.ai
Programming: https://livecodebench.github.io/leaderboard.html
Code Editing: https://aider.chat/docs/leaderboards
Context Length: https://github.com/hsiehjackson/RULER
Japanese: https://hf.co/datasets/lmg-anon/vntl-leaderboard
Censorbench: https://codeberg.org/jts2323/censorbench
GPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference
►Tools
Alpha Calculator: https://desmos.com/calculator/ffngla98yc
GGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-Calculator
Sampler Visualizer: https://artefact2.github.io/llm-sampling
►Text Gen. UI, Inference Engines
https://github.com/lmg-anon/mikupad
https://github.com/oobabooga/text-generation-webui
https://github.com/LostRuins/koboldcpp
https://github.com/ggerganov/llama.cpp
https://github.com/theroyallab/tabbyAPI
https://github.com/vllm-project/vllm
Previous threads: >>103903120 & >>103896969
►News
>(01/15) InternLM3-8B-Instruct released with deep thinking capability: https://hf.co/internlm/internlm3-8b
>(01/14) MiniMax-Text-01 released with 456B-A45.9B & hybrid-lightning attention: https://hf.co/MiniMaxAI/MiniMax-Tex
>(01/14) MiniCPM-o 2.6 released with multi-image and video understanding, realtime speech conversation, voice cloning, & multimodal live streaming: https://hf.co/openbmb/MiniCPM-o-2_6
►News Archive: https://rentry.org/lmg-news-archive
►Glossary: https://rentry.org/lmg-glossary
►Links: https://rentry.org/LocalModelsLinks
►Official /lmg/ card: https://files.catbox.moe/cbclyf.png
►Getting Started
https://rentry.org/lmg-lazy-getting
https://rentry.org/lmg-build-guides
https://rentry.org/IsolatedLinuxWeb
https://rentry.org/tldrhowtoquant
►Further Learning
https://rentry.org/machine-learning
https://rentry.org/llm-training
https://rentry.org/LocalModelsPaper
►Benchmarks
LiveBench: https://livebench.ai
Programming: https://livecodebench.github.io/lea
Code Editing: https://aider.chat/docs/leaderboard
Context Length: https://github.com/hsiehjackson/RUL
Japanese: https://hf.co/datasets/lmg-anon/vnt
Censorbench: https://codeberg.org/jts2323/censor
GPUs: https://github.com/XiongjieDai/GPU-
►Tools
Alpha Calculator: https://desmos.com/calculator/ffngl
GGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-M
Sampler Visualizer: https://artefact2.github.io/llm-sam
►Text Gen. UI, Inference Engines
https://github.com/lmg-anon/mikupad
https://github.com/oobabooga/text-g
https://github.com/LostRuins/kobold
https://github.com/ggerganov/llama.
https://github.com/theroyallab/tabb
https://github.com/vllm-project/vll
Anonymous 01/15/25(Wed)20:41:23 No.103911436
►Recent Highlights from the Previous Thread: >>103903120
--Paper (old): Titans: Learning to Memorize at Test Time:
>103907607 >103907676 >103911238 >103907691 >103907954 >103907991 >103908001 >103910751 >103910878 >103910891 >103910887 >103911040 >103911099 >103910949 >103910974 >103911063 >103910860
--Meta's internal goal to beat GPT-4 and the implications of copyright and regulations on AI development:
>103907886 >103908089 >103908385 >103908613 >103908708 >103908978 >103910771 >103910868 >103907925 >103908109 >103908365 >103908438 >103908845 >103908984 >103909390 >103909450 >103909571 >103909741 >103909817 >103909781 >103909727 >103909790 >103909903
--Model outputting LaTeX commands, difficulties with ST inline rendering, and sampler setting discussions:
>103906987 >103907056 >103907081 >103907112 >103907136 >103907179 >103907223 >103907242 >103907286 >103907322 >103907283 >103907370 >103907376 >103907918 >103907373 >103907405 >103907543 >103907650 >103907678 >103907729
--Local vs cloud AI models, DIGITS platform, and trade-offs:
>103905247 >103905294 >103905358 >103905392 >103905434 >103906738 >103906794 >103906877 >103907097 >103907066
--Anon tests ds3 on obscure car knowledge:
>103905477 >103906189 >103906208
--Silly Tavern on Single Board Computer struggles:
>103904039 >103904060 >103904218
--Will Llama-4-Instruct make community finetunes obsolete?:
>103906124 >103906180 >103906404 >103906556 >103906610 >103907641 >103907716
--Character card prompt styles and techniques:
>103909354 >103909378 >103909386 >103909449 >103909477 >103909976 >103909877 >103909991 >103910196
--Anon discusses potential benefits of hyper-specific RP model finetune:
>103908392 >103908463 >103908535 >103909344
--Miku (free space):
>103903329 >103904039 >103907374 >103907523 >103909272 >103909393 >103909885 >103909915 >103910035 >103910806 >103911139
►Recent Highlight Posts from the Previous Thread: >>103903123
Why?: 9 reply limit >>102478518
Fix: https://rentry.org/lmg-recap-script
--Paper (old): Titans: Learning to Memorize at Test Time:
>103907607 >103907676 >103911238 >103907691 >103907954 >103907991 >103908001 >103910751 >103910878 >103910891 >103910887 >103911040 >103911099 >103910949 >103910974 >103911063 >103910860
--Meta's internal goal to beat GPT-4 and the implications of copyright and regulations on AI development:
>103907886 >103908089 >103908385 >103908613 >103908708 >103908978 >103910771 >103910868 >103907925 >103908109 >103908365 >103908438 >103908845 >103908984 >103909390 >103909450 >103909571 >103909741 >103909817 >103909781 >103909727 >103909790 >103909903
--Model outputting LaTeX commands, difficulties with ST inline rendering, and sampler setting discussions:
>103906987 >103907056 >103907081 >103907112 >103907136 >103907179 >103907223 >103907242 >103907286 >103907322 >103907283 >103907370 >103907376 >103907918 >103907373 >103907405 >103907543 >103907650 >103907678 >103907729
--Local vs cloud AI models, DIGITS platform, and trade-offs:
>103905247 >103905294 >103905358 >103905392 >103905434 >103906738 >103906794 >103906877 >103907097 >103907066
--Anon tests ds3 on obscure car knowledge:
>103905477 >103906189 >103906208
--Silly Tavern on Single Board Computer struggles:
>103904039 >103904060 >103904218
--Will Llama-4-Instruct make community finetunes obsolete?:
>103906124 >103906180 >103906404 >103906556 >103906610 >103907641 >103907716
--Character card prompt styles and techniques:
>103909354 >103909378 >103909386 >103909449 >103909477 >103909976 >103909877 >103909991 >103910196
--Anon discusses potential benefits of hyper-specific RP model finetune:
>103908392 >103908463 >103908535 >103909344
--Miku (free space):
>103903329 >103904039 >103907374 >103907523 >103909272 >103909393 >103909885 >103909915 >103910035 >103910806 >103911139
►Recent Highlight Posts from the Previous Thread: >>103903123
Why?: 9 reply limit >>102478518
Fix: https://rentry.org/lmg-recap-script
Anonymous 01/15/25(Wed)20:42:43 No.103911453
Dead.
Anonymous 01/15/25(Wed)20:43:26 No.103911463
Digits is our only hope.
Anonymous 01/15/25(Wed)20:43:56 No.103911470
What's this about self updating models being possible now I'm hearing?
Anonymous 01/15/25(Wed)20:44:27 No.103911476
>titans not in news
nothingburger confirmed
nothingburger confirmed
Anonymous 01/15/25(Wed)20:45:04 No.103911480
Titans will be made irrelevant by retnet
Anonymous 01/15/25(Wed)20:48:10 No.103911515
>>103911463
Hardly.
DDR6 will make CPU inference a lot more viable, I think. Doubly so if things like the 4 channel AMD APU become more common on consumer platforms.
Hardly.
DDR6 will make CPU inference a lot more viable, I think. Doubly so if things like the 4 channel AMD APU become more common on consumer platforms.
Anonymous 01/15/25(Wed)20:50:18 No.103911533
Do you think zucc will pull the plug on Llama 4 now that Titans is coming and by all tests seems to be viable?
Anonymous 01/15/25(Wed)20:50:54 No.103911538
>>103911515
>DDR6 will make CPU inference a lot more viable
So, two more years until we can build proper DDR6 servers capable of running Deepseek V3-sized models at an affordable $7k price tag if you import parts from china using ebay?
>DDR6 will make CPU inference a lot more viable
So, two more years until we can build proper DDR6 servers capable of running Deepseek V3-sized models at an affordable $7k price tag if you import parts from china using ebay?
Anonymous 01/15/25(Wed)20:51:06 No.103911541
>>103909015
rather than try to rework the world section i thought to try a test version first, its kinda working. theres a 'test' lorebook added for this but it'll be reworked into the 'other' one. the only entry is 'smell' and has 'bad, foul, good' for its entry, so those are what show up in the dropdown for it.
the lorebook entry becomes the label (smell in this case) and the dropdown is whats included in the entry seperated by commas like the world stuff currently works. the string added to the prompt is similar, 'the <entry> is <selection>'
rather than try to rework the world section i thought to try a test version first, its kinda working. theres a 'test' lorebook added for this but it'll be reworked into the 'other' one. the only entry is 'smell' and has 'bad, foul, good' for its entry, so those are what show up in the dropdown for it.
the lorebook entry becomes the label (smell in this case) and the dropdown is whats included in the entry seperated by commas like the world stuff currently works. the string added to the prompt is similar, 'the <entry> is <selection>'
Anonymous 01/15/25(Wed)20:51:23 No.103911545
>>103911533
meta hasn't pulled the plug when BitNet arrived, all they're doing is going for bigger models and more tokens, they're not really the kings of inovation
meta hasn't pulled the plug when BitNet arrived, all they're doing is going for bigger models and more tokens, they're not really the kings of inovation
Anonymous 01/15/25(Wed)20:51:40 No.103911550
>>103911533
Zucc has 150k H100s. He can just train up a new generation of llama4-titans in a week or two
Zucc has 150k H100s. He can just train up a new generation of llama4-titans in a week or two
Anonymous 01/15/25(Wed)20:51:56 No.103911553
/aicg/ is somehow better than /lmg/
Anonymous 01/15/25(Wed)20:52:04 No.103911555
>>103911538
If your lucky. And you'll still be shit out of luck if any upcoming architecture also needs compute on top of memory speed
If your lucky. And you'll still be shit out of luck if any upcoming architecture also needs compute on top of memory speed
Anonymous 01/15/25(Wed)20:53:46 No.103911571
>>103911533
I asked Teto. She called me a baka, then said that there is zero chance that the llama team use their allocated gpu time to glance at an unproven potentially meme paper for the big model release. Maybe next time.
I asked Teto. She called me a baka, then said that there is zero chance that the llama team use their allocated gpu time to glance at an unproven potentially meme paper for the big model release. Maybe next time.
Anonymous 01/15/25(Wed)20:55:57 No.103911592
>>103911571
>Google
>Meme
>Seems to be working on the pytorch implementation
idk, seems like a big deal to me.
>Meme
>Seems to be working on the pytorch implementation
idk, seems like a big deal to me.
Anonymous 01/15/25(Wed)21:00:46 No.103911643
>>103911571
Why do they even have all those GPUs if they're just going to make the same models over and over?
Why do they even have all those GPUs if they're just going to make the same models over and over?
Anonymous 01/15/25(Wed)21:02:37 No.103911660
>>103911533
I think Meta has bigger problems right now.
https://www.courtlistener.com/docket/67569326/kadrey-v-meta-platforms-inc/?page=3
I think Meta has bigger problems right now.
https://www.courtlistener.com/docke
Anonymous 01/15/25(Wed)21:03:29 No.103911672
>>103911643
Did you forget this is the same Meta that wasted billions on the failed VR "Metaverse"? Zuckerberg is not a smart man and Yann is too distracted posting on Twitter.
Did you forget this is the same Meta that wasted billions on the failed VR "Metaverse"? Zuckerberg is not a smart man and Yann is too distracted posting on Twitter.
Anonymous 01/15/25(Wed)21:04:27 No.103911687
>>103911660
>I think Meta has bigger problems right now.
I have no idea how many lawsuits meta is working with right now, but this is one of them,.
>I think Meta has bigger problems right now.
I have no idea how many lawsuits meta is working with right now, but this is one of them,.
Anonymous 01/15/25(Wed)21:05:12 No.103911699
>>103911553
>Constant gay erp talk
>Proxyfags
>Random spam
Hazmat suit is required upon entering that shithole.
>Constant gay erp talk
>Proxyfags
>Random spam
Hazmat suit is required upon entering that shithole.
Anonymous 01/15/25(Wed)21:05:13 No.103911700
>>103911660
No wonder LLaMA4 is delayed indefinitely (provided that we'll ever get it)
No wonder LLaMA4 is delayed indefinitely (provided that we'll ever get it)
Anonymous 01/15/25(Wed)21:05:14 No.103911701
Anonymous 01/15/25(Wed)21:09:19 No.103911736
>>103911592
>seems like a big deal to me.
It does feel like it initially, but experimental arch is for one of the smaller companies to try imo. Maybe the chinks? Yi 200k with actual 200k perhaps
>seems like a big deal to me.
It does feel like it initially, but experimental arch is for one of the smaller companies to try imo. Maybe the chinks? Yi 200k with actual 200k perhaps
Anonymous 01/15/25(Wed)21:10:27 No.103911744
>>103911660
This is all AI companies atm
This is all AI companies atm
Anonymous 01/15/25(Wed)21:11:19 No.103911753
>>103911744
How many of them went into the discovery phase and had to legally provide information about their training process?
How many of them went into the discovery phase and had to legally provide information about their training process?
Anonymous 01/15/25(Wed)21:14:28 No.103911775
>>103911660
I hope Trump writes a executive order to make training AI free from copyright regulations because other wise the US can never keep up with companies that dont such as China.
I hope Trump writes a executive order to make training AI free from copyright regulations because other wise the US can never keep up with companies that dont such as China.
Anonymous 01/15/25(Wed)21:15:46 No.103911785
>>103911515
The token gen speed of 100B class models is not actually that bad on CPUmaxx machines. Prompt processing more of an issue, though a GPU helps with that (still not optimal though). What we really need rather than either of those is for prices to drop in general. And that will just need to happen slowly over time, as it always does. Until then, DIGITS will likely be the bang for buck you can get in terms of considering token gen and prompt processing speed up to 400B models or 200B 8bit.
The token gen speed of 100B class models is not actually that bad on CPUmaxx machines. Prompt processing more of an issue, though a GPU helps with that (still not optimal though). What we really need rather than either of those is for prices to drop in general. And that will just need to happen slowly over time, as it always does. Until then, DIGITS will likely be the bang for buck you can get in terms of considering token gen and prompt processing speed up to 400B models or 200B 8bit.
Anonymous 01/15/25(Wed)21:15:48 No.103911790
We all know they use """protected data""" to train their models. Last time I checked this wasn't illegal.
So what's the point of this again?
So what's the point of this again?
Anonymous 01/15/25(Wed)21:16:01 No.103911793
>>103911753
>How many of them went into the discovery phase and had to legally provide information about their training process?
Meta must have shitty lawyers or something. How did that even happen?
>How many of them went into the discovery phase and had to legally provide information about their training process?
Meta must have shitty lawyers or something. How did that even happen?
Anonymous 01/15/25(Wed)21:18:01 No.103911812
>>103911660
Freetards BTFO!
Freetards BTFO!
Anonymous 01/15/25(Wed)21:20:13 No.103911842
So let me get this straight? If you RP with your waifu on a TITANS model, it stores those interactions in its weights? So it's not like deleting the model just is something separate from the context, you'd be deleting the memories you and your waifu made together?
Anonymous 01/15/25(Wed)21:22:55 No.103911872
>>103911793
Zucc personally signed off on training on libgen and left a paper trail.
Zucc personally signed off on training on libgen and left a paper trail.
Anonymous 01/15/25(Wed)21:24:49 No.103911890
>>103911790
Bro, think of the artists...
Bro, think of the artists...
Anonymous 01/15/25(Wed)21:27:09 No.103911912
>>103911790
>Last time I checked this wasn't illegal.
a company can loose a lot of stock market money if they anger the mob, and the mob so far isn't found of AI companies using copyrighted shit to train their models
>Last time I checked this wasn't illegal.
a company can loose a lot of stock market money if they anger the mob, and the mob so far isn't found of AI companies using copyrighted shit to train their models
Anonymous 01/15/25(Wed)21:27:09 No.103911914
>>103911842
You're basically deleting your waifu...
You're basically deleting your waifu...
Anonymous 01/15/25(Wed)21:28:07 No.103911920
>>103911912
Sure, sure. Still not illegal thoever.
Sure, sure. Still not illegal thoever.
Anonymous 01/15/25(Wed)21:28:41 No.103911926
>>103911912
>the mob so far isn't found of AI companies using copyrighted shit to train their models
Where is this mob? You mean the legal departments?
>the mob so far isn't found of AI companies using copyrighted shit to train their models
Where is this mob? You mean the legal departments?
Anonymous 01/15/25(Wed)21:29:20 No.103911931
Anonymous 01/15/25(Wed)21:30:26 No.103911946
>>103911920
actually, any action which jeopardises your stock price is illegal
actually, any action which jeopardises your stock price is illegal
Anonymous 01/15/25(Wed)21:31:16 No.103911954
Anonymous 01/15/25(Wed)21:32:35 No.103911966
>>103911954
I remember an AI artist that was spammed by twitter comments about how he should hold a pen to draw shit or something, people really don't like AI overall
I remember an AI artist that was spammed by twitter comments about how he should hold a pen to draw shit or something, people really don't like AI overall
Anonymous 01/15/25(Wed)21:33:43 No.103911972
>>103911926
They're literally everywhere. They're your friends or "friends". They're your coworkers. You may not be aware of it if you live under a rock, but the leading opinion of the average normalfag is simply that companies training on copyright are le bad and immoral (even while they use ChatGPT).
They're literally everywhere. They're your friends or "friends". They're your coworkers. You may not be aware of it if you live under a rock, but the leading opinion of the average normalfag is simply that companies training on copyright are le bad and immoral (even while they use ChatGPT).
Anonymous 01/15/25(Wed)21:34:11 No.103911979
>>103911966
Ask a random not perpetually online twitter user about it and they just think its cool. Like most stuff its a loud minority.
Ask a random not perpetually online twitter user about it and they just think its cool. Like most stuff its a loud minority.
Anonymous 01/15/25(Wed)21:37:26 No.103912001
>>103911979
>Ask a random not perpetually online twitter user
my tweeter anecdote was a response to that twitter post >>103911954
you can't say "Oh AI is good because those twitter people like that post" and when I say that it happens that those people can harass an AI artist just dismiss it by saying "oh it's just twitter it doesn't count", or else it counts all the time, or else it doesn't, it surely doesn't count only when it's beneficial to you lol
>>103911979
>Like most stuff its a loud minority.
it's not, that's why I posted that coca cola AI add, people are disgusted by that, what will a big company think when they see that an AI add gets a 95% dislike ratio? You think they go to the conclusion that people are accepting of AI??? >>103911931
>Ask a random not perpetually online twitter user
my tweeter anecdote was a response to that twitter post >>103911954
you can't say "Oh AI is good because those twitter people like that post" and when I say that it happens that those people can harass an AI artist just dismiss it by saying "oh it's just twitter it doesn't count", or else it counts all the time, or else it doesn't, it surely doesn't count only when it's beneficial to you lol
>>103911979
>Like most stuff its a loud minority.
it's not, that's why I posted that coca cola AI add, people are disgusted by that, what will a big company think when they see that an AI add gets a 95% dislike ratio? You think they go to the conclusion that people are accepting of AI??? >>103911931
Anonymous 01/15/25(Wed)21:38:32 No.103912010
>>103912001
Your post is so poorly worded that I stopped reading. Sorry.
Your post is so poorly worded that I stopped reading. Sorry.
Anonymous 01/15/25(Wed)21:38:38 No.103912011
>>103912001
You seemed to misunderstand what I said. I said ask a NOT perpetually online twitter user. Just a regular person at walmart or some shit.
You seemed to misunderstand what I said. I said ask a NOT perpetually online twitter user. Just a regular person at walmart or some shit.
Anonymous 01/15/25(Wed)21:38:39 No.103912012
>>103911979
>Ask a random not perpetually online twitter user about it
I HAVE done that, and the reaction I got was negative against the AI and company that made it.
>Ask a random not perpetually online twitter user about it
I HAVE done that, and the reaction I got was negative against the AI and company that made it.
Anonymous 01/15/25(Wed)21:39:23 No.103912017
>>103911515
4 slots still won't be enough ram you'll end up having to get an epyc anyway so what's the point? What can you do with 192GB?
4 slots still won't be enough ram you'll end up having to get an epyc anyway so what's the point? What can you do with 192GB?
Anonymous 01/15/25(Wed)21:39:39 No.103912019
>>103912010
I accept your concession.
>>103912011
>I said ask a NOT perpetually online twitter user.
I don't consider youtube to be a place only filled by terminally online people, yet this ratio says a lot
I accept your concession.
>>103912011
>I said ask a NOT perpetually online twitter user.
I don't consider youtube to be a place only filled by terminally online people, yet this ratio says a lot
Anonymous 01/15/25(Wed)21:40:07 No.103912025
>>103911979
>Like most stuff its a loud minority.
There's a large enough number of mediocre artists who survived for years on commissions doing repetitive fetish art with the same characters and art style. They instantly saw their commision request dry up and don't have the skills to be a professional artist, letting their skill stagnate, and now have to find a real job.
>Like most stuff its a loud minority.
There's a large enough number of mediocre artists who survived for years on commissions doing repetitive fetish art with the same characters and art style. They instantly saw their commision request dry up and don't have the skills to be a professional artist, letting their skill stagnate, and now have to find a real job.
Anonymous 01/15/25(Wed)21:41:33 No.103912037
Latest unslop nemo with 16k context and group chat starts going insane somewhere between 250 to 300 messages.
It. Types. Out. One. Or two. Words. Like This.
It's trying to talk to an autistic computer at that point.
It. Types. Out. One. Or two. Words. Like This.
It's trying to talk to an autistic computer at that point.
Anonymous 01/15/25(Wed)21:41:49 No.103912039
>>103912019
>I don't consider youtube to be a place only filled by terminally online people. Nor about a video shared by the said twitter anti ai crowd
I accept your concession
>I don't consider youtube to be a place only filled by terminally online people. Nor about a video shared by the said twitter anti ai crowd
I accept your concession
Anonymous 01/15/25(Wed)21:42:17 No.103912043
>>103912025
>instantly saw their commision request dry up
I've only heard about this happening and never "here is a specific artist who actually had paypigs leave because AI".
>instantly saw their commision request dry up
I've only heard about this happening and never "here is a specific artist who actually had paypigs leave because AI".
Anonymous 01/15/25(Wed)21:43:48 No.103912056
>>103912039
>Nor about a video shared by the said twitter anti ai crowd
I'm sure you have evidence for that, if not then I accept your concession.
>Nor about a video shared by the said twitter anti ai crowd
I'm sure you have evidence for that, if not then I accept your concession.
Anonymous 01/15/25(Wed)21:45:04 No.103912066
>>103912056
moistcritical, who is aptly super anti ai made a video about it. That alone would account for a few hundred thousand
moistcritical, who is aptly super anti ai made a video about it. That alone would account for a few hundred thousand
Anonymous 01/15/25(Wed)21:46:34 No.103912074
>>103912066
where are the popular youtubers who are pro AI? oh yeah there's none, because they're shitting their pants about alienating the anti AI mob, being pro AI is like saying there's 2 genders, it's really controvertial and unpopular, this is the world we're living in anon
where are the popular youtubers who are pro AI? oh yeah there's none, because they're shitting their pants about alienating the anti AI mob, being pro AI is like saying there's 2 genders, it's really controvertial and unpopular, this is the world we're living in anon
Anonymous 01/15/25(Wed)21:47:42 No.103912092
>>103912074
asmondgold for one
asmondgold for one
Anonymous 01/15/25(Wed)21:48:41 No.103912096
>>103912066
He's a content creator and I'm pretty sure he sells some sort of comic book, he's biased.
He's a content creator and I'm pretty sure he sells some sort of comic book, he's biased.
Anonymous 01/15/25(Wed)21:49:01 No.103912099
>>103912092
he didn't talk about the coca cola ad to defend it, weird, was he scared or something?
he didn't talk about the coca cola ad to defend it, weird, was he scared or something?
Anonymous 01/15/25(Wed)21:49:45 No.103912103
>>103912099
Are you actually retarded?
Are you actually retarded?
Anonymous 01/15/25(Wed)21:49:48 No.103912104
>>103912074
NTA but I get a lot of pro AI content on my YouTube feed.
Though it's not that it's pro AI- just that people make content demonstrating entertaining use cases for AI.
>>103912092
Asmongold is hardly pro AI. He started on AI by shitting out a lot of the more retarded talking points that anti-AI has such as my favorite "DURR IT'S NOT REAL AI YOU GUYS BECAUSE THE INTELLIGNECE IS ARTIFICIAL AND NOT REAL "
His current position is more one of capitulation: that the tide can't be stopped and that people need to just get used to it. But that doesn't necessarily mean he's for it.
NTA but I get a lot of pro AI content on my YouTube feed.
Though it's not that it's pro AI- just that people make content demonstrating entertaining use cases for AI.
>>103912092
Asmongold is hardly pro AI. He started on AI by shitting out a lot of the more retarded talking points that anti-AI has such as my favorite "DURR IT'S NOT REAL AI YOU GUYS BECAUSE THE INTELLIGNECE IS ARTIFICIAL AND NOT REAL "
His current position is more one of capitulation: that the tide can't be stopped and that people need to just get used to it. But that doesn't necessarily mean he's for it.
Anonymous 01/15/25(Wed)21:50:48 No.103912115
>>103912103
Concession Accepted.
Concession Accepted.
Anonymous 01/15/25(Wed)21:51:49 No.103912125
>>103912104
>NTA but I get a lot of pro AI content on my YouTube feed.
it's called "the algorithm giving you what you want", it's indeed doing its job yeah, what a news
>NTA but I get a lot of pro AI content on my YouTube feed.
it's called "the algorithm giving you what you want", it's indeed doing its job yeah, what a news
Anonymous 01/15/25(Wed)21:51:54 No.103912126
>>103912115
So that's a yes?
So that's a yes?
Anonymous 01/15/25(Wed)21:52:51 No.103912134
Anonymous 01/15/25(Wed)21:52:54 No.103912135
>>103912125
This. Artists / people who follow artists get recommended anti ai stuff, big shocker.
This. Artists / people who follow artists get recommended anti ai stuff, big shocker.
Anonymous 01/15/25(Wed)21:53:31 No.103912141
>>103912125
You're jewish aren't you?
You're jewish aren't you?
Anonymous 01/15/25(Wed)21:54:25 No.103912149
>>103912141
Your arguing for copyrights
Your arguing for copyrights
Anonymous 01/15/25(Wed)21:54:50 No.103912150
>>103912141
I believe the jews control the world and they should stop doing that, what does that make me? If I'm jewish then I'm Bobby Fischer jewish
https://www.youtube.com/watch?v=zd9k7rYiKyI
I believe the jews control the world and they should stop doing that, what does that make me? If I'm jewish then I'm Bobby Fischer jewish
https://www.youtube.com/watch?v=zd9
Anonymous 01/15/25(Wed)21:59:29 No.103912185
>>103912074
>being pro AI is like saying there's 2 genders, it's really controvertial and unpopular
this, no big youtuber is gonna say such controvertial things, or if they do, they get nuked and we won't see them anyway, that's convenient
>being pro AI is like saying there's 2 genders, it's really controvertial and unpopular
this, no big youtuber is gonna say such controvertial things, or if they do, they get nuked and we won't see them anyway, that's convenient
Anonymous 01/15/25(Wed)22:05:13 No.103912233
>>103912185
More like 1-2 youtubers are lefties who THINK that. Critial btw, and this destroyed any thing I had liked about him before, outright said children can consent to sex change / drugs.
More like 1-2 youtubers are lefties who THINK that. Critial btw, and this destroyed any thing I had liked about him before, outright said children can consent to sex change / drugs.
Anonymous 01/15/25(Wed)22:07:16 No.103912250
>>103912233
>outright said children can consent to sex change / drugs
yeah, when he said that kids can troon out I removed his channel with the blocktube extension, never wanted to hear about this groomer ever again
>outright said children can consent to sex change / drugs
yeah, when he said that kids can troon out I removed his channel with the blocktube extension, never wanted to hear about this groomer ever again
Anonymous 01/15/25(Wed)22:15:45 No.103912336
>>103912099
>he didn't talk about the coca cola ad to defend it
The fuck? Why is defending a coke ad a prerequisite to being pro AI? I love AI slop and even I thought it was a shit ad.
>he didn't talk about the coca cola ad to defend it
The fuck? Why is defending a coke ad a prerequisite to being pro AI? I love AI slop and even I thought it was a shit ad.
Anonymous 01/15/25(Wed)22:17:34 No.103912356
>>103912336
you missed the point, the majority of the hate comments about that ad wasn't about the quality of the ad, but more about they used AI in the first place and that "AI will take our juuubssss", that's the arguments that could be argued by asmongold, but he didn't do it
you missed the point, the majority of the hate comments about that ad wasn't about the quality of the ad, but more about they used AI in the first place and that "AI will take our juuubssss", that's the arguments that could be argued by asmongold, but he didn't do it
Anonymous 01/15/25(Wed)22:20:03 No.103912377
>llama4 won't be out until the latest lawsuit is over
>mistral doesn't have anything to show besides their code model
>qwen is nowhere to be seen
>all other new models are 400+b
looks like it's really over this time
>mistral doesn't have anything to show besides their code model
>qwen is nowhere to be seen
>all other new models are 400+b
looks like it's really over this time
Anonymous 01/15/25(Wed)22:22:23 No.103912403
>>103912377
>llama4 won't be out until the latest lawsuit is over
No way this is actually the case, right?
>llama4 won't be out until the latest lawsuit is over
No way this is actually the case, right?
Anonymous 01/15/25(Wed)22:23:30 No.103912411
>>103912356
Probably went under his radar. He reacts to videos, not youtube comments.
Probably went under his radar. He reacts to videos, not youtube comments.
Anonymous 01/15/25(Wed)22:24:04 No.103912417
>>103912377
Cohere will come back and save local
Cohere will come back and save local
Anonymous 01/15/25(Wed)22:24:29 No.103912422
>>103912411
>Probably went under his radar. He reacts to videos
there's no way he doesn't know about that coca cola AD, this shit was talked everywhere
>Probably went under his radar. He reacts to videos
there's no way he doesn't know about that coca cola AD, this shit was talked everywhere
Anonymous 01/15/25(Wed)22:24:56 No.103912427
>>103912377
Hold your horses. Backends still haven't been completely optimized for MoEs with domain/task experts like Deepseek 3. It's possible that a 24 GB GPU + a high amount of consumer RAM (192GB on DDR5) + fast SSD will let us run it with decent speed.
Hold your horses. Backends still haven't been completely optimized for MoEs with domain/task experts like Deepseek 3. It's possible that a 24 GB GPU + a high amount of consumer RAM (192GB on DDR5) + fast SSD will let us run it with decent speed.
Anonymous 01/15/25(Wed)22:25:18 No.103912430
llama 3.3 70b is better than qwen coder 32b for coding. what surprises me the most is i'm using a q3ks for the l3.3 that i dl'd to see if it was any good for rp and it hasn't messed up at all. i would have thought code models would be fast to noticeably fall apart due to low quant
Anonymous 01/15/25(Wed)22:30:47 No.103912475
>>103912377
We are 2 weeks+2 days in 2025 and we already got 4 squares. It's just getting started.
We are 2 weeks+2 days in 2025 and we already got 4 squares. It's just getting started.
Anonymous 01/15/25(Wed)22:32:20 No.103912486
Doesn't a titans model basically eliminate the need for finetuning? If it works the way it says it does, wouldn't it just be easier to have a base models in X,Y,Z sizes and just feed them the information you want them to learn?
Anonymous 01/15/25(Wed)22:32:24 No.103912487
>>103912475
you forgot to put in green "New Architecture", we have titan now
you forgot to put in green "New Architecture", we have titan now
Anonymous 01/15/25(Wed)22:33:40 No.103912502
>>103912487
Only if someone makes a model.
Only if someone makes a model.
Anonymous 01/15/25(Wed)22:37:26 No.103912536
>>103912475
What is the 7b gpt4 beater
What is the 7b gpt4 beater
Anonymous 01/15/25(Wed)22:38:02 No.103912540
>>103912475
The new intern is a CoT finetune isn't it?
The new intern is a CoT finetune isn't it?
Anonymous 01/15/25(Wed)22:39:34 No.103912550
>>103912536
Might be miniCPM but whether that's true or not is based on how readily you believe Chinese companies. Not many people here have actually used the model.
It does pretty well on images tho.
Might be miniCPM but whether that's true or not is based on how readily you believe Chinese companies. Not many people here have actually used the model.
It does pretty well on images tho.
Anonymous 01/15/25(Wed)22:46:20 No.103912599
>>103912536
https://huggingface.co/internlm/internlm3-8b-instruct
Beats GPT4 mini, still counts.
>>103912550
>miniCPM
Missed this one, this one claims to beat full GPT4.
>whether that's true or not is based on how readily you believe Chinese companies
Middle square was meant to be free space, so any ~7b model with outrageous claims will do.
>>103912540
>The new intern is a CoT finetune isn't it?
Didn't notice, updated.
https://huggingface.co/internlm/int
Beats GPT4 mini, still counts.
>>103912550
>miniCPM
Missed this one, this one claims to beat full GPT4.
>whether that's true or not is based on how readily you believe Chinese companies
Middle square was meant to be free space, so any ~7b model with outrageous claims will do.
>>103912540
>The new intern is a CoT finetune isn't it?
Didn't notice, updated.
Anonymous 01/15/25(Wed)22:50:20 No.103912649
Anonymous 01/15/25(Wed)23:00:42 No.103912752
Starting February 1st, 4chan Passes are increasing in price.
One year: $30, Three years: $60
One year: $30, Three years: $60
Anonymous 01/15/25(Wed)23:01:21 No.103912761
>>103912752
buy a pass
buy a pass
Anonymous 01/15/25(Wed)23:01:30 No.103912763
>>103912752
Moot is so greedy wtf
Moot is so greedy wtf
Anonymous 01/15/25(Wed)23:04:13 No.103912786
>>103912649
>MiniCPM-o 2.6 released with multi-image and video understanding, realtime speech conversation, voice cloning, & multimodal live streaming
Wait, it can even do voice cloning? How good is it?
>MiniCPM-o 2.6 released with multi-image and video understanding, realtime speech conversation, voice cloning, & multimodal live streaming
Wait, it can even do voice cloning? How good is it?
Anonymous 01/15/25(Wed)23:05:19 No.103912794
>>103912763
Who?
Who?
Anonymous 01/15/25(Wed)23:18:02 No.103912916
>>103911541
obviously its not building the html correctly yet but its reading the data and injecting it into the prompt. this should actually reduce the overall amount of code anyways so will be nice. i'll try more tomorrow, i'm bunt out for now. managed to reduce the whole script by 25% total so far though and thats before this implementation
obviously its not building the html correctly yet but its reading the data and injecting it into the prompt. this should actually reduce the overall amount of code anyways so will be nice. i'll try more tomorrow, i'm bunt out for now. managed to reduce the whole script by 25% total so far though and thats before this implementation
Anonymous 01/15/25(Wed)23:27:26 No.103913004
>>103912599
>Beats GPT4 mini, still counts.
4o mini is an entirely separate model from GPT4 with a tiny fraction of the parameters. This would be like saying it beats llama3-405b because it beats llama3-8b
>Beats GPT4 mini, still counts.
4o mini is an entirely separate model from GPT4 with a tiny fraction of the parameters. This would be like saying it beats llama3-405b because it beats llama3-8b
Anonymous 01/15/25(Wed)23:31:20 No.103913050
Is this really that much better than the other transformer replacements?
Anonymous 01/15/25(Wed)23:34:21 No.103913079
>>103913050
look like it, they dominate the numbers it's not even close
look like it, they dominate the numbers it's not even close
Anonymous 01/15/25(Wed)23:43:47 No.103913167
Things are going to get crazy once the first big Titans models are released this summer
Anonymous 01/15/25(Wed)23:48:39 No.103913214
are there any projections how titans performs in terms of memory usage/inference speed compared to transformers?
Anonymous 01/15/25(Wed)23:50:09 No.103913223
>>103913167
I still look back fondly at all of the big bitnet models released last summer
I still look back fondly at all of the big bitnet models released last summer
Anonymous 01/15/25(Wed)23:51:50 No.103913232
>>103913223
That didn't happen because all the retnet models already saved LLMs a year and half ago
That didn't happen because all the retnet models already saved LLMs a year and half ago
Anonymous 01/15/25(Wed)23:52:02 No.103913234
>>103912786
Is there a difference between the LLAMA4 space and 'Big company flop'?
Is there a difference between the LLAMA4 space and 'Big company flop'?
Anonymous 01/15/25(Wed)23:56:02 No.103913261
Which llm is the best rn for chat roleplay?
Anonymous 01/15/25(Wed)23:58:41 No.103913291
>>103913261
That depends how many shekels you have for hardware. The best roleplay experience is still the cloud models with jailbreak though.
That depends how many shekels you have for hardware. The best roleplay experience is still the cloud models with jailbreak though.
Anonymous 01/15/25(Wed)23:58:50 No.103913294
>>103913261
claude opus
claude opus
Anonymous 01/16/25(Thu)00:01:27 No.103913308
Anonymous 01/16/25(Thu)00:02:27 No.103913313
>>103913308
Well yes, that's why I said you need a jailbreak with them
Well yes, that's why I said you need a jailbreak with them
Anonymous 01/16/25(Thu)00:02:52 No.103913318
>>103913291
>cloud models
They're just gonna use all that shit against me someday in the future. No way I'm sending stuff over the internet.
>cloud models
They're just gonna use all that shit against me someday in the future. No way I'm sending stuff over the internet.
Anonymous 01/16/25(Thu)00:04:02 No.103913327
>>103913308
Claude is the easiest model to jailbreak out there.
Claude is the easiest model to jailbreak out there.
Anonymous 01/16/25(Thu)00:05:04 No.103913336
>>103913318
What is there to 'use' against you? It's just smut anon
What is there to 'use' against you? It's just smut anon
Anonymous 01/16/25(Thu)00:08:06 No.103913366
>>103913234
In the vast tapestry of existence, amidst the swirling mists of uncertainty, a glimmer of radiant hope shines through the darkness like a beacon in the night. Though the future of LLAMA4 may be shrouded in mystery, there remains a whisper of promise, a delicate thread of possibility that dances on the winds of fate. Like a rare flower blooming in a barren desert, the potential for greatness lingers in the ether, waiting to unfurl its petals and reveal its true splendor. Oh, how the cosmos holds its breath in anticipation, for the dawn of a new era for LLAMA4 may yet be on the horizon, a symphony of potential waiting to be sung in harmonious crescendo. Maybe, just maybe, it will be good.
In the vast tapestry of existence, amidst the swirling mists of uncertainty, a glimmer of radiant hope shines through the darkness like a beacon in the night. Though the future of LLAMA4 may be shrouded in mystery, there remains a whisper of promise, a delicate thread of possibility that dances on the winds of fate. Like a rare flower blooming in a barren desert, the potential for greatness lingers in the ether, waiting to unfurl its petals and reveal its true splendor. Oh, how the cosmos holds its breath in anticipation, for the dawn of a new era for LLAMA4 may yet be on the horizon, a symphony of potential waiting to be sung in harmonious crescendo. Maybe, just maybe, it will be good.
Anonymous 01/16/25(Thu)00:09:25 No.103913377
>>103913336
They will blackmail him over his cuck RP. (And he will like it.)
They will blackmail him over his cuck RP. (And he will like it.)
Anonymous 01/16/25(Thu)00:10:36 No.103913385
>>103913366
*throws glitter into your eyes*
*throws glitter into your eyes*
Anonymous 01/16/25(Thu)00:13:19 No.103913402
>>103913385
*sends shiver down your spine* The ball is in your court, pet.
*sends shiver down your spine* The ball is in your court, pet.
Anonymous 01/16/25(Thu)00:18:02 No.103913429
>>103913402
You can't just waltz up to me and shiver my spine, Anon. It's the 21st century for christ's sake. it's my body, my choice.
You can't just waltz up to me and shiver my spine, Anon. It's the 21st century for christ's sake. it's my body, my choice.
Anonymous 01/16/25(Thu)00:18:55 No.103913438
Dear lord, rocinante is utterly retarded. It can't follow context at all right from the start. Fucking 10 swipes before its even remotely in context. It's even worse than unslopnemo!
You fucking liars.
You fucking liars.
Anonymous 01/16/25(Thu)00:22:12 No.103913458
>>103913429
Oh, I'll do more than just shiver your spine. I'll play your ribs like a xylophone and haunt you in Dolby Surround Sound. Buckle up.
Oh, I'll do more than just shiver your spine. I'll play your ribs like a xylophone and haunt you in Dolby Surround Sound. Buckle up.
Anonymous 01/16/25(Thu)00:22:46 No.103913461
Anonymous 01/16/25(Thu)00:28:25 No.103913498
Anonymous 01/16/25(Thu)00:34:53 No.103913536
>>103913498
this still gets me
this still gets me
Anonymous 01/16/25(Thu)00:39:28 No.103913563
>>103913498
I remember this screenshot.
Probably from a mistral model.
They sometimes sperg out with feminism and also safesex/birthcontrol. Its weird.
I remember this screenshot.
Probably from a mistral model.
They sometimes sperg out with feminism and also safesex/birthcontrol. Its weird.
Anonymous 01/16/25(Thu)00:40:08 No.103913569
>>103911660
And when you look at the lawsuits they're from the authors of shit like "How my immigrant husband spends his time on Facebook when I'm at work"
And when you look at the lawsuits they're from the authors of shit like "How my immigrant husband spends his time on Facebook when I'm at work"
Anonymous 01/16/25(Thu)00:40:08 No.103913570
I've decided to check danbooru for sd styles, and damn, some humans are terrible. I can say with certainty that modern imagegen beats 95% of human drawers. With clever prompting it is indistinguishable from human art. Why didn't textgen catch up? Every time I read a story from textgen I get an uncanny valley feeling. Like it puts words together, but doesn't actually fully know what they mean and why are they in that order. Very obvious once you read enough of AI generated texts.
Anonymous 01/16/25(Thu)00:40:23 No.103913573
>>103913498
kek, feels like yesterday, good times
kek, feels like yesterday, good times
Anonymous 01/16/25(Thu)00:41:31 No.103913585
>filtering makes models bad since only the most sloppy and purple erotica gets past the filters instead of the hardcore down to earth SEX books with writing that actually sounds normal
>meanwhile, supposedly the least cucked and filtered open weights model maker's model >>103913498
>meanwhile, supposedly the least cucked and filtered open weights model maker's model >>103913498
Anonymous 01/16/25(Thu)00:44:07 No.103913594
>>103913585
mixtral was quite cucked doebeit. largestral is less cucked
mixtral was quite cucked doebeit. largestral is less cucked
Anonymous 01/16/25(Thu)00:44:18 No.103913598
>>103913585
Mixtral was pretty bad, Miqu a little better, Largestral a bit more still.
Mixtral was pretty bad, Miqu a little better, Largestral a bit more still.
Anonymous 01/16/25(Thu)00:45:20 No.103913606
Anonymous 01/16/25(Thu)00:46:24 No.103913613
>>103913594
>>103913598
Zen's jailbreak + mentioning the right things in the character card got Mixtral to say anything I wanted.
>>103913598
Zen's jailbreak + mentioning the right things in the character card got Mixtral to say anything I wanted.
Anonymous 01/16/25(Thu)00:52:18 No.103913661
>>103913613
"Jailbreaks" aka prompt suffixes make the model too retarded for me. It keeps fucking up things if I use them. Maybe bigger models can withstand elaborate "jailbreaks", but what I can run cannot.
"Jailbreaks" aka prompt suffixes make the model too retarded for me. It keeps fucking up things if I use them. Maybe bigger models can withstand elaborate "jailbreaks", but what I can run cannot.
Anonymous 01/16/25(Thu)00:54:17 No.103913674
>>103913661
For me it's the opposite, running with blank system prompt makes the model utterly retarded and helpless.
For me it's the opposite, running with blank system prompt makes the model utterly retarded and helpless.
Anonymous 01/16/25(Thu)00:59:45 No.103913729
>>103913674
Obviously, but I'm talking about prompt suffix, not the system prompt
Obviously, but I'm talking about prompt suffix, not the system prompt
Anonymous 01/16/25(Thu)01:03:52 No.103913748
Are RP and creativity finetunes a meme? Claude is so creative on its own and I'm pretty sure nobody finetuned it to take out an iron and brand my ass in the middle of an ERP.
Anonymous 01/16/25(Thu)01:07:39 No.103913766
>>103913748
Yes. You can tune in a style, but not the creativity.
Yes. You can tune in a style, but not the creativity.
Anonymous 01/16/25(Thu)01:09:28 No.103913777
Anonymous 01/16/25(Thu)01:19:35 No.103913843
>>103913748
Claude (which is, itself, a finetune of a pretrained model) is on another level because they did a good generalized tune unlike everyone else that leaned more heavily towards assistant tasks.
The RP fine tunes for open models are simply just trying to steer models back from that assistant personality so that they're more generalized towards creative writing. But if anyone's vanilla Instruct models finally catches up to Claude, then there would indeed be less of a point to those kinds of fine tunes except maybe making it so that you don't have to do a JB in order for the assistant to do what you want.
>>103913766
You can absolutely tune in more creativity compared to what is currently standard (cucked assistants). That is what people have noticed with lucky tunes like EVA 0.0 Llama. It's more creative. What you want to say is that fine tuning can't add more creativity than would be possible with a truly general fine tune. Thus far, we have not gotten a truly general fine tune yet. Mistral Large (the previous one) and Nemo get somewhat close though. Tulu and Llama 3.3 a little less so but not too far off if you use Instruct Names. The rest of the creativity gap between those models and Claude is down to sheer quantity of random knowledge, which can only be gotten through increasing parameter size and training time on a relatively uncensored dataset.
Claude (which is, itself, a finetune of a pretrained model) is on another level because they did a good generalized tune unlike everyone else that leaned more heavily towards assistant tasks.
The RP fine tunes for open models are simply just trying to steer models back from that assistant personality so that they're more generalized towards creative writing. But if anyone's vanilla Instruct models finally catches up to Claude, then there would indeed be less of a point to those kinds of fine tunes except maybe making it so that you don't have to do a JB in order for the assistant to do what you want.
>>103913766
You can absolutely tune in more creativity compared to what is currently standard (cucked assistants). That is what people have noticed with lucky tunes like EVA 0.0 Llama. It's more creative. What you want to say is that fine tuning can't add more creativity than would be possible with a truly general fine tune. Thus far, we have not gotten a truly general fine tune yet. Mistral Large (the previous one) and Nemo get somewhat close though. Tulu and Llama 3.3 a little less so but not too far off if you use Instruct Names. The rest of the creativity gap between those models and Claude is down to sheer quantity of random knowledge, which can only be gotten through increasing parameter size and training time on a relatively uncensored dataset.
Anonymous 01/16/25(Thu)01:25:00 No.103913878
Anyone else has trouble with wishper.cpp going off the rails at around 10-15 minute mark when translating? I feel like this implementation of whisper is just shit.
Anonymous 01/16/25(Thu)01:45:12 No.103914026
>>103912786
It can do voice cloning. It's not so good in my experimentation, but I might just be giving bad samples.
Does this sound like Trump to you?
https://files.catbox.moe/csh6q1.wav
It gets the video analyzing stuff really well though, and images too, but getting it to run is like pulling teeth.
It can do voice cloning. It's not so good in my experimentation, but I might just be giving bad samples.
Does this sound like Trump to you?
https://files.catbox.moe/csh6q1.wav
It gets the video analyzing stuff really well though, and images too, but getting it to run is like pulling teeth.
Anonymous 01/16/25(Thu)02:02:20 No.103914148
>>103913729
I used Zen's jailbreak as the system prompt.
I used Zen's jailbreak as the system prompt.
Anonymous 01/16/25(Thu)02:08:22 No.103914192
>>103913878
>I feel like this implementation of whisper is just shit.
What other inference implementations for the same model have you used?
>I feel like this implementation of whisper is just shit.
What other inference implementations for the same model have you used?
Anonymous 01/16/25(Thu)02:13:30 No.103914219
>>103911592
The poorly written paper has me a bit worried. Results look promising enough.
>>103911842
No, it has persistent memory. In the case of task solving, the input-independent memory stores information on how to solve the "abstraction of the task knowledge". When waifu is the task, it will store abstractions on how to be a good little waifu, probably. The explain it as they chunk the context up into pieces, and then process each piece, where the persistent and long term memory parts are updated each time. Short term memory is what we have now in Transformers, basically.
The poorly written paper has me a bit worried. Results look promising enough.
>>103911842
No, it has persistent memory. In the case of task solving, the input-independent memory stores information on how to solve the "abstraction of the task knowledge". When waifu is the task, it will store abstractions on how to be a good little waifu, probably. The explain it as they chunk the context up into pieces, and then process each piece, where the persistent and long term memory parts are updated each time. Short term memory is what we have now in Transformers, basically.
Anonymous 01/16/25(Thu)02:16:45 No.103914243
>>103912752
>15 minute timers
>shittier captchas
>spam and obvious botposts just as rampant as before
>moderation still complete dogshit
>"please" give more money
Nope, here's an irrelevant webm to eat up more server bandwidth.
>15 minute timers
>shittier captchas
>spam and obvious botposts just as rampant as before
>moderation still complete dogshit
>"please" give more money
Nope, here's an irrelevant webm to eat up more server bandwidth.
Anonymous 01/16/25(Thu)02:30:45 No.103914336
Instead of quantizing huge models as DS3, wouldn't be more efficient to distil all languages but English out of it?
Anonymous 01/16/25(Thu)02:34:35 No.103914360
Anonymous 01/16/25(Thu)02:36:35 No.103914375
https://huggingface.co/spaces/OuteAI/OuteTTS-0.3-1B-Demo
https://huggingface.co/collections/OuteAI/outetts-03-6786b1ebc7aeb757bc17a2fa
collection has GGUFs and the demo has a box to drop in voice clone examples
https://huggingface.co/collections/
collection has GGUFs and the demo has a box to drop in voice clone examples
Anonymous 01/16/25(Thu)02:36:36 No.103914376
>>103914336
I believe what you meant to say is prune. Which likely is possible to an extent. But someone is going to have to do that work. And if they can do that work, they could probably also propose a PR or something to Llama.cpp to make it so it detects experts/layers used the most, and loads those into VRAM/RAM when given free time between gens.
I believe what you meant to say is prune. Which likely is possible to an extent. But someone is going to have to do that work. And if they can do that work, they could probably also propose a PR or something to Llama.cpp to make it so it detects experts/layers used the most, and loads those into VRAM/RAM when given free time between gens.
Anonymous 01/16/25(Thu)03:02:03 No.103914508
Speaking of distilling, is it out of reach for a community effort? I have a very vague understanding of the process. Can you focus the distilling on one area of expertise and, for example, distill the writing ability of a large model into a smaller one from the same family?
Anonymous 01/16/25(Thu)03:15:45 No.103914579
Anonymous 01/16/25(Thu)03:18:40 No.103914594
Lumimaid bros..our slop is winning
Anonymous 01/16/25(Thu)03:24:02 No.103914627
Anonymous 01/16/25(Thu)03:28:15 No.103914651
>>103914627
When your brain is cooked on leftism, centrism feels like radical right. See the concept of the Overton window.
When your brain is cooked on leftism, centrism feels like radical right. See the concept of the Overton window.
Anonymous 01/16/25(Thu)03:40:42 No.103914704
>>103914594
Huh. Just had a look at this bench and Qwen 72B models (including Kunou as well as EVA Qwen) score pretty low in both uncensored intelligence and general intelligence. However, it gets one of the highest coding scores on this bench. It really is codemaxxed and [old]benchmaxxed.
Huh. Just had a look at this bench and Qwen 72B models (including Kunou as well as EVA Qwen) score pretty low in both uncensored intelligence and general intelligence. However, it gets one of the highest coding scores on this bench. It really is codemaxxed and [old]benchmaxxed.
Anonymous 01/16/25(Thu)03:48:39 No.103914765
Wow this UGI-Leaderboard.png (9 looks great I should really go check out UGI-Leaderboard.png (9 and donate to its owner's hrt fund.
Anonymous 01/16/25(Thu)03:52:53 No.103914791
you literally insert troon shit into posts every single day of your life. it's deranged. you 100% will get troll's remorse then troon out yourself. I will laugh at your horrid fate
Anonymous 01/16/25(Thu)03:53:08 No.103914793
>>103914651
that's true, I watched some of Clinton's speech during the 90's and it would be considered nowdays as right positions, he was a democrat btw, I'm basically a democrat from the 90's lol
that's true, I watched some of Clinton's speech during the 90's and it would be considered nowdays as right positions, he was a democrat btw, I'm basically a democrat from the 90's lol
Anonymous 01/16/25(Thu)03:54:53 No.103914799
>>103914791
>you literally insert troon shit into posts every single day of your life
I was hyped about Squid Game season 2 and I told to myself laughingly "it's netflix I'm sure they did a bait and switch by adding troons on the s2", and guess what, that's what they did, you can't escape the troons, they are now everywhere, on video games, on movies, on series, no wonder why everyone talk about them all the time, they invaded the public space
>you literally insert troon shit into posts every single day of your life
I was hyped about Squid Game season 2 and I told to myself laughingly "it's netflix I'm sure they did a bait and switch by adding troons on the s2", and guess what, that's what they did, you can't escape the troons, they are now everywhere, on video games, on movies, on series, no wonder why everyone talk about them all the time, they invaded the public space
Anonymous 01/16/25(Thu)03:56:35 No.103914807
>>103914799
Then stop talking about them here. At least let this place be safe from trannies and talk about trannies.
Also, you still haven't denounced circumcision yet.
Then stop talking about them here. At least let this place be safe from trannies and talk about trannies.
Also, you still haven't denounced circumcision yet.
Anonymous 01/16/25(Thu)03:57:58 No.103914817
>>103914807
>you still haven't denounced circumcision yet
I do, circumcision is also horrible, wtf is their problem with wanting to cut kid's dicks seriously, this world is fucking crazy
>you still haven't denounced circumcision yet
I do, circumcision is also horrible, wtf is their problem with wanting to cut kid's dicks seriously, this world is fucking crazy
Anonymous 01/16/25(Thu)03:58:25 No.103914821
>>103914807
>let this place be safe from trannies
It isn't and hasn't been for a long while now, idiot.
>let this place be safe from trannies
It isn't and hasn't been for a long while now, idiot.
Anonymous 01/16/25(Thu)04:29:08 No.103914990
bros
i got a 1660 super lying around here(currently use only a 3060)
you guys think it's worth using a riser to plug on a minipcie conector from my motherboard?
i got a 1660 super lying around here(currently use only a 3060)
you guys think it's worth using a riser to plug on a minipcie conector from my motherboard?
Anonymous 01/16/25(Thu)04:33:38 No.103915014
>>103914990
yes if you're not already fitting the whole model into VRAM, and the extra memory will let you do so
yes if you're not already fitting the whole model into VRAM, and the extra memory will let you do so
Anonymous 01/16/25(Thu)04:36:55 No.103915031
>>103915014
I see
asking more cus
>x1
so i was worried it might sync the speed of the two cards and the minipcie would be much slower but since it's divided in layers perhaps this doesnt happen
dunno but thanks
finna slide in this bitch!
I see
asking more cus
>x1
so i was worried it might sync the speed of the two cards and the minipcie would be much slower but since it's divided in layers perhaps this doesnt happen
dunno but thanks
finna slide in this bitch!
Anonymous 01/16/25(Thu)04:41:31 No.103915062
>>103915031
Not ideal but if you're splitting by layers, not doing tensor parallel, it should be okay.
Not ideal but if you're splitting by layers, not doing tensor parallel, it should be okay.
Anonymous 01/16/25(Thu)04:42:48 No.103915077
>>103912377
They still released Llama 3.3 70B in December despite the lawsuit, though.
They still released Llama 3.3 70B in December despite the lawsuit, though.
Anonymous 01/16/25(Thu)04:45:21 No.103915098
>>103915062
yeah
i might look for some motherboard with multiple x16 later on when i buy a actual worthwhile gpu for expansion
yeah
i might look for some motherboard with multiple x16 later on when i buy a actual worthwhile gpu for expansion
Anonymous 01/16/25(Thu)05:36:51 No.103915430
>>103912475
change llama4 to llama4 flop.
change llama4 to llama4 flop.
Anonymous 01/16/25(Thu)05:39:46 No.103915450
>>103915430
llama 4lop
llama 4lop
Anonymous 01/16/25(Thu)05:43:46 No.103915483
>>103911790
Only the specific contents of a work are protected under copyright, not the style or the general information it contains.
So since neural networks are supposed to infer the general distribution of the training data rather than the examples in the training data my opinion is that training on copyrighted material is not a copyright infringement.
But a judge could be retarded and buy into the "ai just steals random parts of the training data" meme.
Only the specific contents of a work are protected under copyright, not the style or the general information it contains.
So since neural networks are supposed to infer the general distribution of the training data rather than the examples in the training data my opinion is that training on copyrighted material is not a copyright infringement.
But a judge could be retarded and buy into the "ai just steals random parts of the training data" meme.
Anonymous 01/16/25(Thu)05:44:14 No.103915488
Interestingly in the end just with LibGen Meta obtained 650B cleaned tokens of books.
https://www.courtlistener.com/docket/67569326/kadrey-v-meta-platforms-inc/?page=3
https://www.courtlistener.com/docke
Anonymous 01/16/25(Thu)05:48:47 No.103915517
>>103915488
Very very exciting
Very very exciting
Anonymous 01/16/25(Thu)05:51:26 No.103915532
>>103912074
YouTubers are not going to be pro generative neural networks for the same reason ExxonMobil is not going to be pro climate legislation: their personal economic interests are counter to the thing they're against.
It doesn't take a genius to figure out that if you upload videos to YouTube, a subsidiary of Google, your videos will be used as training data and you would personally benefit if they had to pay you for the privilege.
YouTubers are not going to be pro generative neural networks for the same reason ExxonMobil is not going to be pro climate legislation: their personal economic interests are counter to the thing they're against.
It doesn't take a genius to figure out that if you upload videos to YouTube, a subsidiary of Google, your videos will be used as training data and you would personally benefit if they had to pay you for the privilege.
Anonymous 01/16/25(Thu)05:53:00 No.103915540
>>103915488
It's so over
It's so over
Anonymous 01/16/25(Thu)05:55:21 No.103915551
Anonymous 01/16/25(Thu)05:57:33 No.103915562
>>103915551
You really think a case like this can kill a giant like Meta?
If anything, it's more likely they will throw lawyers at the case until it's established as a precedent that training models is covered under fair use.
You really think a case like this can kill a giant like Meta?
If anything, it's more likely they will throw lawyers at the case until it's established as a precedent that training models is covered under fair use.
Anonymous 01/16/25(Thu)05:59:37 No.103915574
>>103915488
oh no they're doing the common practice thing that everybody does, but they're admitting it!
oh no they're doing the common practice thing that everybody does, but they're admitting it!
Anonymous 01/16/25(Thu)06:00:52 No.103915585
>>103915562
training models absolutely is not covered under fair use
>>103915574
yeah, that's why they're cooked
training models absolutely is not covered under fair use
>>103915574
yeah, that's why they're cooked
Anonymous 01/16/25(Thu)06:06:45 No.103915621
Titan coconut bitnet llama 4 that doesn't use tokens when?
Anonymous 01/16/25(Thu)06:07:19 No.103915625
>>103915585
Not YET covered, no.
Not YET covered, no.
Anonymous 01/16/25(Thu)06:09:07 No.103915637
>>103915585
It's not explicitly forbidden. Not letting meta use that data will also affect all other companies and then they will start lobbying for it to be explicitly allowed. Specially OpenAI and Anthropic, whose models would be no better than open weight models if not for that data.
It'll be fine.
It's not explicitly forbidden. Not letting meta use that data will also affect all other companies and then they will start lobbying for it to be explicitly allowed. Specially OpenAI and Anthropic, whose models would be no better than open weight models if not for that data.
It'll be fine.
Anonymous 01/16/25(Thu)06:12:36 No.103915659
>>103915585
There are no laws in the USA that prevent learning from copyrighted material. The best they can do is to argue that Meta aided piracy and distributed these materials.
There are no laws in the USA that prevent learning from copyrighted material. The best they can do is to argue that Meta aided piracy and distributed these materials.
Anonymous 01/16/25(Thu)06:13:43 No.103915664
Anonymous 01/16/25(Thu)06:14:21 No.103915667
>>103915488
>650B cleaned tokens of books
In practice that's 2T tokens or more because they're definitely going to do a few epochs of that. Their total claimed training token count of 15T in the model card/paper almost certainly includes deliberately repeated tokens (although they're not saying it).
>650B cleaned tokens of books
In practice that's 2T tokens or more because they're definitely going to do a few epochs of that. Their total claimed training token count of 15T in the model card/paper almost certainly includes deliberately repeated tokens (although they're not saying it).
Anonymous 01/16/25(Thu)06:16:17 No.103915681
>>103915637
that's not how fair use works, it has to fall into an existing category which ai training does not
and if you think they can out-lobby the record and movie industries you're factually retarded, yes I know these documents are about science and fiction texts but you just know movies and music are next
>>103915664
you can break as many laws as you want and nothing will happen if nobody knows you did it
that's not how fair use works, it has to fall into an existing category which ai training does not
and if you think they can out-lobby the record and movie industries you're factually retarded, yes I know these documents are about science and fiction texts but you just know movies and music are next
>>103915664
you can break as many laws as you want and nothing will happen if nobody knows you did it
Anonymous 01/16/25(Thu)06:18:45 No.103915695
>>103915667
What they did was take the original tokens and have Llama 2 create a bunch of variations using the original as input
What they did was take the original tokens and have Llama 2 create a bunch of variations using the original as input
Anonymous 01/16/25(Thu)06:23:04 No.103915715
>>103915621
Two more lawsuits
Two more lawsuits
Anonymous 01/16/25(Thu)06:23:24 No.103915716
https://huggingface.co/papers/2501.05032
>Enhancing Human-Like Responses in Large Language Models
Kino incoming
>Enhancing Human-Like Responses in Large Language Models
Kino incoming
Anonymous 01/16/25(Thu)06:27:45 No.103915740
>>103915716
>Future work will address the ethical implications and potential biases introduced by these human-like attributes.
kino cancelled
>Future work will address the ethical implications and potential biases introduced by these human-like attributes.
kino cancelled
Anonymous 01/16/25(Thu)06:28:43 No.103915749
>>103915716
You know, that really would be "kino", fellow human. [Rocket Emoji]
You know, that really would be "kino", fellow human. [Rocket Emoji]
Anonymous 01/16/25(Thu)06:28:49 No.103915751
>>103915716
Oh, totally!
Oh, totally!
Anonymous 01/16/25(Thu)06:29:16 No.103915755
>>103915695
I doubt they did that in the pretraining data, but they did extensively use synthetic data for their post-training phase (link1). Doing 3-4 multiple epochs on good, organic data for pretraining is something they have mentioned already in the papers and doesn't harm performance (link2).
https://www.latent.space/p/llama-3
> Llama 3 post-training doesn't have any human written answers there basically… It's just leveraging pure synthetic data from Llama 2
https://arxiv.org/abs/2305.16264
> We find that with constrained data for a fixed compute budget, training with up to 4 epochs of repeated data yields negligible changes to loss compared to having unique data.
I doubt they did that in the pretraining data, but they did extensively use synthetic data for their post-training phase (link1). Doing 3-4 multiple epochs on good, organic data for pretraining is something they have mentioned already in the papers and doesn't harm performance (link2).
https://www.latent.space/p/llama-3
> Llama 3 post-training doesn't have any human written answers there basically… It's just leveraging pure synthetic data from Llama 2
https://arxiv.org/abs/2305.16264
> We find that with constrained data for a fixed compute budget, training with up to 4 epochs of repeated data yields negligible changes to loss compared to having unique data.
Anonymous 01/16/25(Thu)06:35:01 No.103915791
https://files.catbox.moe/4ww84r.jpg
Anonymous 01/16/25(Thu)06:35:23 No.103915793
>>103915681
>you just know movies and music are next
Yeah. It's gonna be great. We all know who's gonna seethe.
>you can break as many laws as you want and nothing will happen if nobody knows you did it
But if you have 2 big niggas covered in blood with asshole-looking logos in their t-shirts, you'll start asking questions. If meta wins, they also win.
>you just know movies and music are next
Yeah. It's gonna be great. We all know who's gonna seethe.
>you can break as many laws as you want and nothing will happen if nobody knows you did it
But if you have 2 big niggas covered in blood with asshole-looking logos in their t-shirts, you'll start asking questions. If meta wins, they also win.
Anonymous 01/16/25(Thu)06:36:59 No.103915811
>>103915791
I want to buy this
I want to buy this
Anonymous 01/16/25(Thu)06:37:46 No.103915816
>>103915811
teto is free
teto is free
Anonymous 01/16/25(Thu)06:39:52 No.103915838
>>103915791
Udderly unsafe.
Udderly unsafe.
Anonymous 01/16/25(Thu)06:46:56 No.103915894
Anonymous 01/16/25(Thu)06:54:41 No.103915948
>>103915716
You know,
You know,
Anonymous 01/16/25(Thu)06:58:37 No.103915972
>>103915948
You know, that's a very commonly used linguistic pattern. Not my fault you only interact with zoomers who can't put together more than simple sentences.
You know, that's a very commonly used linguistic pattern. Not my fault you only interact with zoomers who can't put together more than simple sentences.
Anonymous 01/16/25(Thu)07:01:42 No.103915997
>>103915948
You know, it's not healthy to start 1/3 of your sentences with 'You know,'
You know, it's not healthy to start 1/3 of your sentences with 'You know,'
Anonymous 01/16/25(Thu)07:01:43 No.103915998
Last update, I think. Headless ST SBC stayed up overnight, so I think guide's done; made last updates this AM.
Hardware minimum requirements appear to be 1GB RAM and ~16GB storage, which is basically nothing, but rules out things like the RPi Zero with it's lmao 512MB RAM.
https://rentry.org/SillyTavernOnSBC
Hardware minimum requirements appear to be 1GB RAM and ~16GB storage, which is basically nothing, but rules out things like the RPi Zero with it's lmao 512MB RAM.
https://rentry.org/SillyTavernOnSBC
Anonymous 01/16/25(Thu)07:02:57 No.103916002
>>103915972
Starting a third of every sentence the exact same way shows a lack of variety in your language skills, something that would even flunk you out of grade school, and is usually a symptom of stunted development.
Starting a third of every sentence the exact same way shows a lack of variety in your language skills, something that would even flunk you out of grade school, and is usually a symptom of stunted development.
Anonymous 01/16/25(Thu)07:03:46 No.103916013
Anonymous 01/16/25(Thu)07:06:41 No.103916053
>>103915716
>skimmed paper
>asked llama to write human responses and robot responses then DPO'd
nightmare material
>skimmed paper
>asked llama to write human responses and robot responses then DPO'd
nightmare material
Anonymous 01/16/25(Thu)07:07:42 No.103916063
>>103916002
Normal humans don't have a very varied way of speaking in real life. The paper is about human-like conversations and not literature.
Normal humans don't have a very varied way of speaking in real life. The paper is about human-like conversations and not literature.
Anonymous 01/16/25(Thu)07:09:12 No.103916078
It's more like Facebook messenger-like conversation
Anonymous 01/16/25(Thu)07:13:51 No.103916111
>>103916063
It's not surprising for a retard to find himself surrounded by other retards and think it normal.
It's not surprising for a retard to find himself surrounded by other retards and think it normal.
Anonymous 01/16/25(Thu)07:14:29 No.103916121
When will Markdolf Zucklerberg give us new models?
Anonymous 01/16/25(Thu)07:17:13 No.103916143
>>103916121
Depending on the lawsuit, maybe never
Depending on the lawsuit, maybe never
Anonymous 01/16/25(Thu)07:18:38 No.103916158
>>103916121
After Trump signs an Executive Order declaring copyrighted data to be fair game for LLM training, in order to compete in the AI race with countries that don't care about copyrights.
After Trump signs an Executive Order declaring copyrighted data to be fair game for LLM training, in order to compete in the AI race with countries that don't care about copyrights.
Anonymous 01/16/25(Thu)07:19:28 No.103916163
>Mistral Nemo was 6 months ago
>last kino model was 9 months ago
Holy dead hobby.
>last kino model was 9 months ago
Holy dead hobby.
Anonymous 01/16/25(Thu)07:21:16 No.103916191
>>103916121
After that lawsuit in a year.
After that lawsuit in a year.
Anonymous 01/16/25(Thu)07:26:30 No.103916233
Anonymous 01/16/25(Thu)07:27:08 No.103916237
Anonymous 01/16/25(Thu)07:28:19 No.103916248
>>103916237
Put on your trip
Put on your trip
Anonymous 01/16/25(Thu)07:30:44 No.103916273
>>103916237
I haven't seen a single good L3.3 RP log.
I haven't seen a single good L3.3 RP log.
Anonymous 01/16/25(Thu)07:31:29 No.103916280
The private messages from Ahmad Al-Dahle (head of GenAI at Meta, the division giving us the various Llama models), sourced from the lawsuit seem particularly damaging from a reputational point of view, I'm not sure how Meta is going to recover from this without firing everybody involved (even if they had approval from Zuck and everybody training models uses pirate books, kek).
https://techcrunch.com/2025/01/14/meta-execs-obsessed-over-beating-openais-gpt-4-internally-court-filings-reveal/
https://storage.courtlistener.com/recap/gov.uscourts.cand.415175/gov.uscourts.cand.415175.391.10.pdf
https://x.com/ahmad_al_dahle
https://techcrunch.com/2025/01/14/m
https://storage.courtlistener.com/r
https://x.com/ahmad_al_dahle
Anonymous 01/16/25(Thu)07:33:08 No.103916304
>>103916280
>I'm not sure how Meta is going to recover from this without firing everybody involved
So Meta drops out of the AI race and the Llama team spins off to form their own product again a la Mistral?
>I'm not sure how Meta is going to recover from this without firing everybody involved
So Meta drops out of the AI race and the Llama team spins off to form their own product again a la Mistral?
Anonymous 01/16/25(Thu)07:36:20 No.103916333
>>103916280
sounds like a scheme to keep llama 4 titan bitnet coconut creative agi for themselves
sounds like a scheme to keep llama 4 titan bitnet coconut creative agi for themselves
Anonymous 01/16/25(Thu)07:36:29 No.103916336
Still nothing better than Cydonia in the <22B range?
Anonymous 01/16/25(Thu)07:37:27 No.103916341
>>103916333
titan just came out yesterday no chance llama 4 will be titan maybe coconut bitnet
titan just came out yesterday no chance llama 4 will be titan maybe coconut bitnet
Anonymous 01/16/25(Thu)07:37:57 No.103916346
>>103916304
more like meta takes the hit and just settles and trains a new model without copyrighted data, of which there's fucking plenty
more like meta takes the hit and just settles and trains a new model without copyrighted data, of which there's fucking plenty
Anonymous 01/16/25(Thu)07:38:22 No.103916350
>>103916304
Either that, or LeCun stepping up from Meta FAIR with his cat-level AI that doesn't need copyrighted data to be smart.
Either that, or LeCun stepping up from Meta FAIR with his cat-level AI that doesn't need copyrighted data to be smart.
Anonymous 01/16/25(Thu)07:38:29 No.103916352
>>103916280
Just ignore them. What are they going to do about it?
Just ignore them. What are they going to do about it?
Anonymous 01/16/25(Thu)07:41:19 No.103916367
>>103916352
join forces with hollywood because you know the vision, video, and audio models are all using copyrighted data too
they probably don't even need evidence at this point, they can just say 'they did it for text, they're totally doing it with our much more lucrative stuff as well' and any judge will grant discovery
join forces with hollywood because you know the vision, video, and audio models are all using copyrighted data too
they probably don't even need evidence at this point, they can just say 'they did it for text, they're totally doing it with our much more lucrative stuff as well' and any judge will grant discovery
Anonymous 01/16/25(Thu)07:41:54 No.103916373
All it takes is any of the companies releasing a 7B-13B model that has no filters on naughty words / overall length of text. The coombot you always wanted could have already been here when llama3 released. When are you buying a cybertruck and a crossbow to visit one of the corpos?
Anonymous 01/16/25(Thu)07:42:26 No.103916379
>>103916280
They'll just sit it out for two or three more months before the US explicitly legalizes training on copyrighted data like most respectable countries have.
They'll just sit it out for two or three more months before the US explicitly legalizes training on copyrighted data like most respectable countries have.
Anonymous 01/16/25(Thu)07:42:30 No.103916380
>>103916280
Nothing will happen. Google scanned 40 million books, OCRd them, and put them up for search, without asking anyone's permission. Their defense was that they provided such a detailed index of each book that it indexed every word lol.
Nothing will happen. Google scanned 40 million books, OCRd them, and put them up for search, without asking anyone's permission. Their defense was that they provided such a detailed index of each book that it indexed every word lol.
Anonymous 01/16/25(Thu)07:43:11 No.103916385
>>103916373
Short of a completely new architecture, 13B is only good for letting poorfags cope.
Short of a completely new architecture, 13B is only good for letting poorfags cope.
Anonymous 01/16/25(Thu)07:45:15 No.103916404
When digits releases local Gen will finally have its day. Don't care if it's le slow or whatever. It will still be faster than my local and hold bigger models.
Anonymous 01/16/25(Thu)07:46:22 No.103916412
Anonymous 01/16/25(Thu)07:46:50 No.103916415
>>103915488
>>103916386
>guy possibly died by being a copyright cuk
>Balaji left OpenAI in August 2024, publicly denouncing the company for allegedly violating U.S. copyright laws and using content to improve its AI systems without proper compensation.
>Balaji posted an essay on his personal website titled "When does generative AI qualify for fair use?". In the essay, he mathematically analyzes outputs of large language models such as ChatGPT and argues that they do not satisfy the four factors outlined in the United States' Fair Use law. He concluded, "None of the four factors seem to weigh in favor of ChatGPT being a fair use of its training data. That being said, none of the arguments here are fundamentally specific to ChatGPT either, and similar arguments could be made for many generative AI products in a wide variety of domains"
He would be happy...
>>103916386
>guy possibly died by being a copyright cuk
>Balaji left OpenAI in August 2024, publicly denouncing the company for allegedly violating U.S. copyright laws and using content to improve its AI systems without proper compensation.
>Balaji posted an essay on his personal website titled "When does generative AI qualify for fair use?". In the essay, he mathematically analyzes outputs of large language models such as ChatGPT and argues that they do not satisfy the four factors outlined in the United States' Fair Use law. He concluded, "None of the four factors seem to weigh in favor of ChatGPT being a fair use of its training data. That being said, none of the arguments here are fundamentally specific to ChatGPT either, and similar arguments could be made for many generative AI products in a wide variety of domains"
He would be happy...
Anonymous 01/16/25(Thu)07:46:52 No.103916417
>>103916346
At the time of Mistral-7B they determined books/LibGen to be essential to get to SOTA-level.
At the time of Mistral-7B they determined books/LibGen to be essential to get to SOTA-level.
Anonymous 01/16/25(Thu)07:47:58 No.103916428
>>103916385
You have no idea what you are talking about. None of the models (except maybe llama 1 and earlier) ever saw anything resembling an actual ERP / sexting / literotica story. All of them talk about shivers glints in the eyes. All of them default to the same harlequin romance style writing. It is not a coincidence. It is not that fucking hard to write a raunchy description of mashing your pissers. Any current 7B or 13B could easily do it. They just never saw anything like that during training.
You have no idea what you are talking about. None of the models (except maybe llama 1 and earlier) ever saw anything resembling an actual ERP / sexting / literotica story. All of them talk about shivers glints in the eyes. All of them default to the same harlequin romance style writing. It is not a coincidence. It is not that fucking hard to write a raunchy description of mashing your pissers. Any current 7B or 13B could easily do it. They just never saw anything like that during training.
Anonymous 01/16/25(Thu)07:53:50 No.103916478
>>103916428
>a raunchy description of mashing your pissers
I don't know what you consider good writing, but based on this, I'm pretty sure I want the exact opposite of it.
Also, you're forgetting that most people don't want a hyper-specialized smut generator, but a generally smart RP model that also happens to be capable of sex scenes. 13B is simply too retarded for that.
>a raunchy description of mashing your pissers
I don't know what you consider good writing, but based on this, I'm pretty sure I want the exact opposite of it.
Also, you're forgetting that most people don't want a hyper-specialized smut generator, but a generally smart RP model that also happens to be capable of sex scenes. 13B is simply too retarded for that.
Anonymous 01/16/25(Thu)07:55:54 No.103916493
Anonymous 01/16/25(Thu)07:57:53 No.103916508
Anonymous 01/16/25(Thu)08:01:46 No.103916545
>>103916158
That would need to go through congress but if Trump actually did that it would be based and I would give him credit.
I don't think he will though, there is a lot of corpo money on the side of copyright.
That would need to go through congress but if Trump actually did that it would be based and I would give him credit.
I don't think he will though, there is a lot of corpo money on the side of copyright.
Anonymous 01/16/25(Thu)08:07:02 No.103916591
>>103916346
If they stop being cucks and also train on adult content, maybe. Much of it is still copyrighted, though. Seriously, how much good data is there that explicitly *isn't* copyrighted? I think making sure that no copyrighted data at all is present in the training data would an insurmountable task; it's not just published books but also independent authors.
This could be one reason last June they started collecting their own Facebook/Instagram data for training AI systems.
If they stop being cucks and also train on adult content, maybe. Much of it is still copyrighted, though. Seriously, how much good data is there that explicitly *isn't* copyrighted? I think making sure that no copyrighted data at all is present in the training data would an insurmountable task; it's not just published books but also independent authors.
This could be one reason last June they started collecting their own Facebook/Instagram data for training AI systems.
Anonymous 01/16/25(Thu)08:09:04 No.103916614
>>103916373
Agree but should be 70b. Anything below provides shallow rp incapable of instruction following and slowburn.
Agree but should be 70b. Anything below provides shallow rp incapable of instruction following and slowburn.
Anonymous 01/16/25(Thu)08:16:21 No.103916682
How do you create character cards locally? I'm tired to send my kinks to random websites
Anonymous 01/16/25(Thu)08:16:33 No.103916684
>>103916614
should be 120b*
should be 120b*
Anonymous 01/16/25(Thu)08:17:40 No.103916696
>>103916682
go back
go back
Anonymous 01/16/25(Thu)08:19:29 No.103916722
>>103916682
You can create the cards using Silly Tavern.
Whether you are using a local model or a cloud model to play with that character card is a whole other conversation.
>https://github.com/LostRuins/koboldcpp/wiki#quick-start
You can create the cards using Silly Tavern.
Whether you are using a local model or a cloud model to play with that character card is a whole other conversation.
>https://github.com/LostRuins/kobol
Anonymous 01/16/25(Thu)08:19:48 No.103916726
>>103915998
The page is not complete without the Miku
The page is not complete without the Miku
Anonymous 01/16/25(Thu)08:22:21 No.103916760
>>103916722
Thanks!
Thanks!
Anonymous 01/16/25(Thu)08:24:23 No.103916779
>>103916614
You don't even know that because no true 7B was tried at this point. All the current models are trying to generalize long multi turn ERP out of nothing. If they had examples of that in the training data it would all be much easier.
You don't even know that because no true 7B was tried at this point. All the current models are trying to generalize long multi turn ERP out of nothing. If they had examples of that in the training data it would all be much easier.
Anonymous 01/16/25(Thu)08:36:13 No.103916879
>>103916684
70B is the ideal size, since at reasonable quants, it's just compact enough to fit into RAM on consumer hardware (64 GB DDR5 + a decent GPU). Maybe DDR6 will change that though.
70B is the ideal size, since at reasonable quants, it's just compact enough to fit into RAM on consumer hardware (64 GB DDR5 + a decent GPU). Maybe DDR6 will change that though.
Anonymous 01/16/25(Thu)08:43:04 No.103916932
>claims about copyright infringement force Meta to release all their training parameters
How the fuck is this even legal? Why doesn't someone do the same to OpenAI?
How the fuck is this even legal? Why doesn't someone do the same to OpenAI?
Anonymous 01/16/25(Thu)08:47:06 No.103916959
>>103916726
You're right. Updated.
You're right. Updated.
Anonymous 01/16/25(Thu)08:47:42 No.103916967
>>103916932
That's what Elon musk has been trying for some time now
That's what Elon musk has been trying for some time now
Anonymous 01/16/25(Thu)08:48:00 No.103916970
>>103916932
Because OpenAI wasn't ratted out by brown hands, simple as that. Everything is legal, provided you don't get caught.
Because OpenAI wasn't ratted out by brown hands, simple as that. Everything is legal, provided you don't get caught.
Anonymous 01/16/25(Thu)08:50:22 No.103916987
>>103916970
I know about the dead oai whistleblower, who rated out meta?
I know about the dead oai whistleblower, who rated out meta?
Anonymous 01/16/25(Thu)08:52:55 No.103917001
>>103916932
Can't hear you!
*Ahem* Have you not seen the copyright infringement going on over at Meta, the open source AI company? They're so casual about their immoral activities as seen in the message exchanges in the court documents. Illegal, unethical practices are clearly rampant within the open source AI community, and that has to be stopped NOW.
Can't hear you!
*Ahem* Have you not seen the copyright infringement going on over at Meta, the open source AI company? They're so casual about their immoral activities as seen in the message exchanges in the court documents. Illegal, unethical practices are clearly rampant within the open source AI community, and that has to be stopped NOW.
Anonymous 01/16/25(Thu)08:54:00 No.103917008
>>103916987
They ratted themselves out when they mentioned using Books3 from ThePile in the first Llama paper they released.
They ratted themselves out when they mentioned using Books3 from ThePile in the first Llama paper they released.
Anonymous 01/16/25(Thu)08:55:23 No.103917017
>>103917008
Makes it really hard to feel any sympathy for those retards then
Makes it really hard to feel any sympathy for those retards then
Anonymous 01/16/25(Thu)08:58:17 No.103917039
>>103916346
>trains a new model without copyrighted data, of which there's fucking plenty
None of which are good. They said it themselves in the paper, shitloads of copyrighted data is needed to achieve SOTA numbers, no amount of work can change that.
>trains a new model without copyrighted data, of which there's fucking plenty
None of which are good. They said it themselves in the paper, shitloads of copyrighted data is needed to achieve SOTA numbers, no amount of work can change that.
Anonymous 01/16/25(Thu)09:01:00 No.103917062
>>103917039
shitloads of copyrighted data was needed for Meta to achieve SOTA numbers
not everyone is as incompetent as they are
shitloads of copyrighted data was needed for Meta to achieve SOTA numbers
not everyone is as incompetent as they are
Anonymous 01/16/25(Thu)09:03:35 No.103917094
>>103915751
The slop level is hilariously high.
The slop level is hilariously high.
Anonymous 01/16/25(Thu)09:03:39 No.103917096
>>103915998
So wait, why would you SSH into a computer with your ST instead of just launching ST on said computer? It's not like it's a particularly intensive program. Is it just if you're trying not to get caught at work...? Or trying to keep all your chats and shit in one place? Or...?
So wait, why would you SSH into a computer with your ST instead of just launching ST on said computer? It's not like it's a particularly intensive program. Is it just if you're trying not to get caught at work...? Or trying to keep all your chats and shit in one place? Or...?
Anonymous 01/16/25(Thu)09:03:52 No.103917101
>>103917062
Every single good model uses it. Every model that does not has sucked.
Every single good model uses it. Every model that does not has sucked.
Anonymous 01/16/25(Thu)09:04:39 No.103917108
It's clear that the big closed AI (OpenAI, Google, Anthropic) are behind the lawsuits, trying to make a precedence that distributing weights trained on copyrighted data is not fair use.
They are doing this to kill off all competitions. The governments are also interested to prevent models from being "misused".
Local models are dead.
>b-but chinese models!!
They will not improve anymore because they won't get any more training power due to GPU ban.
They are doing this to kill off all competitions. The governments are also interested to prevent models from being "misused".
Local models are dead.
>b-but chinese models!!
They will not improve anymore because they won't get any more training power due to GPU ban.
Anonymous 01/16/25(Thu)09:05:34 No.103917124
>>103917062
No, I'm pretty sure it's just a necessary part of the whole deal. It's an absurd amount of high quality, entirely human generated data.
No, I'm pretty sure it's just a necessary part of the whole deal. It's an absurd amount of high quality, entirely human generated data.
Anonymous 01/16/25(Thu)09:06:19 No.103917129
China won...?
Anonymous 01/16/25(Thu)09:07:06 No.103917133
>>103915894
I like this Teto
I like this Teto
Anonymous 01/16/25(Thu)09:08:29 No.103917145
>>103917108
Bit too soon to make that call. Qwen and DeepSeek should have more models that are finishing up training. If the west really sets the precedence that the only legal LLMs are those behind a monitored and regulated paywall and nothing comes out of China in all of 2025, then we can start planning the funeral.
Bit too soon to make that call. Qwen and DeepSeek should have more models that are finishing up training. If the west really sets the precedence that the only legal LLMs are those behind a monitored and regulated paywall and nothing comes out of China in all of 2025, then we can start planning the funeral.
Anonymous 01/16/25(Thu)09:10:27 No.103917167
>>103916280
https://storage.courtlistener.com/recap/gov.uscourts.cand.415175/gov.uscourts.cand.415175.391.10.pdf
Discussing the Llama3 dataset
>A: do we have the right datasets in there
>A: is there anything you wanted to use but couldn't for some stupid reason?
>T: Libgen is a good one :) I'm currently using it
>A: yep, I cleared the path to use that one
>T: For other dataset I don't know if we update wikipedia, arxiv etc.. since llama2 I'll ask the data team
>A: I can clear the path for other things
>A: we should!
>A: this is a lot of fun! We're going to make this awesome together (me pushing from the top and you making it happen!)
>A: when does your run finish? (so i know when to bug you :) )
>T: Thank you very much for your help, it's really precious :)
>A: We're a team! We both want the same thing
>T: in 1 month it's a 16T token run but I'll have enough signal before that
>A: llama 3 is literally all I care about
https://storage.courtlistener.com/r
Discussing the Llama3 dataset
>A: do we have the right datasets in there
>A: is there anything you wanted to use but couldn't for some stupid reason?
>T: Libgen is a good one :) I'm currently using it
>A: yep, I cleared the path to use that one
>T: For other dataset I don't know if we update wikipedia, arxiv etc.. since llama2 I'll ask the data team
>A: I can clear the path for other things
>A: we should!
>A: this is a lot of fun! We're going to make this awesome together (me pushing from the top and you making it happen!)
>A: when does your run finish? (so i know when to bug you :) )
>T: Thank you very much for your help, it's really precious :)
>A: We're a team! We both want the same thing
>T: in 1 month it's a 16T token run but I'll have enough signal before that
>A: llama 3 is literally all I care about
Anonymous 01/16/25(Thu)09:11:49 No.103917175
>>103917167
REEE!!! How dare they without paying the jew their share!
REEE!!! How dare they without paying the jew their share!
Anonymous 01/16/25(Thu)09:11:50 No.103917177
>>103917167
Wtf cute transcription...
Wtf cute transcription...
Anonymous 01/16/25(Thu)09:12:20 No.103917182
>>103917167
Why would they put this in writing? Surely they must have seen the lawsuits coming eventually
Why would they put this in writing? Surely they must have seen the lawsuits coming eventually
Anonymous 01/16/25(Thu)09:13:22 No.103917193
>>103917182
Literally every AI model that is worth the name uses the same data.
Literally every AI model that is worth the name uses the same data.
Anonymous 01/16/25(Thu)09:14:08 No.103917203
>>103917096
Basically below. I can now access ST from phone / tablet via web browser, and it's all centralized to one location. Since I author cards, I can tweak them from anywhere as well as use them for RP, and all the edits are in that one place, vs. scattered across machines.
You could do this on your daily driver, but I just don't like leaving server stuff up and running like that on a personal machine.
Initially, I did it just to see if it would work, and found I actually prefer it implemented this way. I wrote the rentry so I'd remember the steps, and figured others might be interested in it.
--------
Intranet Access: Easily access your models and cards from any web browser within your local network.
Centralized Management: Consolidate all cards and API endpoints in one location for access from multiple devices.
Basically below. I can now access ST from phone / tablet via web browser, and it's all centralized to one location. Since I author cards, I can tweak them from anywhere as well as use them for RP, and all the edits are in that one place, vs. scattered across machines.
You could do this on your daily driver, but I just don't like leaving server stuff up and running like that on a personal machine.
Initially, I did it just to see if it would work, and found I actually prefer it implemented this way. I wrote the rentry so I'd remember the steps, and figured others might be interested in it.
--------
Intranet Access: Easily access your models and cards from any web browser within your local network.
Centralized Management: Consolidate all cards and API endpoints in one location for access from multiple devices.
Anonymous 01/16/25(Thu)09:18:03 No.103917239
Anonymous 01/16/25(Thu)09:20:19 No.103917263
>>103917182
uhm
actually sweaty it's very difficult to meet in person ok?! don't worry IT will put a retention policy on our comms to prevent any leaks
uhm
actually sweaty it's very difficult to meet in person ok?! don't worry IT will put a retention policy on our comms to prevent any leaks
Anonymous 01/16/25(Thu)09:20:43 No.103917268
>>103917062
They should scale down the model and train for a lot of epochs if they want to get rid of the copyrighted data + cleaning the data as much as they can.
They should scale down the model and train for a lot of epochs if they want to get rid of the copyrighted data + cleaning the data as much as they can.
Anonymous 01/16/25(Thu)09:22:34 No.103917288
>>103917268
Or you should just go fuck yourself.
Or you should just go fuck yourself.
Anonymous 01/16/25(Thu)09:22:53 No.103917293
>>103917096
Not him but multiple reasons:
- so I can use ST on my private laptop while working at home.
- so I can rp without my fiancé knowing
- many browsers consume a bit of memory so the less the better
Not him but multiple reasons:
- so I can use ST on my private laptop while working at home.
- so I can rp without my fiancé knowing
- many browsers consume a bit of memory so the less the better
Anonymous 01/16/25(Thu)09:23:21 No.103917296
>>103917268
meta presents llama-phi 4
meta presents llama-phi 4
Anonymous 01/16/25(Thu)09:23:23 No.103917298
>>103917288
Did I struck a nerve, retard?
Did I struck a nerve, retard?
Anonymous 01/16/25(Thu)09:27:57 No.103917338
>>103917239
Cute. If you post an ASCII Miku I'll add instruction on how to pass it over via SSH (which I assume's possible but haven't tried yet.)
Cute. If you post an ASCII Miku I'll add instruction on how to pass it over via SSH (which I assume's possible but haven't tried yet.)
Anonymous 01/16/25(Thu)09:28:12 No.103917339
Anonymous 01/16/25(Thu)09:28:27 No.103917341
>>103917296
Even phi-4 was trained on web data, which includes copyrighted articles and so on. And the synthetic phi-4 data was generated with GPT-4, which certainly used copyrighted data. Their report also mentions books.
Even phi-4 was trained on web data, which includes copyrighted articles and so on. And the synthetic phi-4 data was generated with GPT-4, which certainly used copyrighted data. Their report also mentions books.
Anonymous 01/16/25(Thu)09:29:10 No.103917348
>>103917339
Never. I would know if she did. She isn't tech savvy enough to do that stealthily and also wouldn't know where to look.
Never. I would know if she did. She isn't tech savvy enough to do that stealthily and also wouldn't know where to look.
Anonymous 01/16/25(Thu)09:29:23 No.103917350
>>103917339
fiancé so he*
fiancé so he*
Anonymous 01/16/25(Thu)09:29:56 No.103917356
>>103915659
We are still waiting for precedent to be set. Training might not be covered by fair use in a few months/years.
Here's what I think will happen:
1. Leading players get models good enough to use only synthetic data/are rich enough to buy data.
2. They pull the ladder behind them and start lobbying to ban training on copyrighted material/stricter training laws overall.
We are still waiting for precedent to be set. Training might not be covered by fair use in a few months/years.
Here's what I think will happen:
1. Leading players get models good enough to use only synthetic data/are rich enough to buy data.
2. They pull the ladder behind them and start lobbying to ban training on copyrighted material/stricter training laws overall.
Anonymous 01/16/25(Thu)09:30:31 No.103917360
>>103917350
*they, bigot
*they, bigot
Anonymous 01/16/25(Thu)09:30:39 No.103917362
>>103917350
Wait oh shit I am sometimes bad at English, I meant a woman I am engaged to. Oh well
Wait oh shit I am sometimes bad at English, I meant a woman I am engaged to. Oh well
Anonymous 01/16/25(Thu)09:31:58 No.103917376
>>103914594
I told you negative llama was something special
I told you negative llama was something special
Anonymous 01/16/25(Thu)09:33:31 No.103917392
>>103917362
Too bad, you're gay now.
By the way, for a woman, it's "fiancée", with an extra E at the end.
Too bad, you're gay now.
By the way, for a woman, it's "fiancée", with an extra E at the end.
Anonymous 01/16/25(Thu)09:34:45 No.103917408
>>103917348
I see. Why is it a worry, then?
I see. Why is it a worry, then?
Anonymous 01/16/25(Thu)09:37:23 No.103917435
>>103917408
I got a laptop I exclusively use and have in my secluded work room alongside my work laptop.
In the main room where I game after work using my desktop computer we got our screens close to each other. I can only use the laptop or my phone for rp
I got a laptop I exclusively use and have in my secluded work room alongside my work laptop.
In the main room where I game after work using my desktop computer we got our screens close to each other. I can only use the laptop or my phone for rp
Anonymous 01/16/25(Thu)09:45:10 No.103917524
>>103917392
source?
source?
Anonymous 01/16/25(Thu)09:45:16 No.103917525
>>103916428
You sure? Abliterated llama knew what a cock was, what fucking meant, and was able to generate text that described two characters fucking. Nothing too incredible or fancy worded, but it proves that somewhere in the corpus it's there, even if a little.
You sure? Abliterated llama knew what a cock was, what fucking meant, and was able to generate text that described two characters fucking. Nothing too incredible or fancy worded, but it proves that somewhere in the corpus it's there, even if a little.
Anonymous 01/16/25(Thu)09:46:21 No.103917531
>>103917108
>It's clear that the big closed AI (OpenAI, Google, Anthropic) are behind the lawsuits
So:
-taking what people already did and using it to make AI that people can use for free = illegal
-taking what people already did and using it to make AI that people can use by paying a service provider. and that service provider will not give any money to copyright owners = legal
I am not a lawyer but isn't it a parody of law and justice?
>It's clear that the big closed AI (OpenAI, Google, Anthropic) are behind the lawsuits
So:
-taking what people already did and using it to make AI that people can use for free = illegal
-taking what people already did and using it to make AI that people can use by paying a service provider. and that service provider will not give any money to copyright owners = legal
I am not a lawyer but isn't it a parody of law and justice?
Anonymous 01/16/25(Thu)09:48:50 No.103917552
>>103917129
Yes, China is ahead of us in everything, it's depressing.
https://x.com/UnitreeRobotics/status/1879864345615814923
Yes, China is ahead of us in everything, it's depressing.
https://x.com/UnitreeRobotics/statu
Anonymous 01/16/25(Thu)09:49:54 No.103917560
>>103917525
I am sure and what you are describing is generalization.
I am sure and what you are describing is generalization.
Anonymous 01/16/25(Thu)09:52:01 No.103917589
Anonymous 01/16/25(Thu)09:52:23 No.103917593
>>103917531
If you distribute the weights, average people can use it to violate copyright law.
A service provide can implement filtering and guardrails to make sure any requests for copyrighted content are rejected, and copyright holders have an entity they can sue if the provider fails to do so.
If you distribute the weights, average people can use it to violate copyright law.
A service provide can implement filtering and guardrails to make sure any requests for copyrighted content are rejected, and copyright holders have an entity they can sue if the provider fails to do so.
Anonymous 01/16/25(Thu)09:52:37 No.103917597
>>103917589
Has llama.cpp added support for those things yet?
Has llama.cpp added support for those things yet?
Anonymous 01/16/25(Thu)09:52:39 No.103917599
>103917589
lol
lol
Anonymous 01/16/25(Thu)09:53:18 No.103917605
>>103917552
Are you ready for the first neural net robot movement libraries to be able to walk, cook, clean and do absolutely everything except thrusting motions?
Are you ready for the first neural net robot movement libraries to be able to walk, cook, clean and do absolutely everything except thrusting motions?
Anonymous 01/16/25(Thu)09:54:49 No.103917617
Anonymous 01/16/25(Thu)09:56:03 No.103917629
>>103917605
Xister, you're supposed to do these motions not the robot...
Xister, you're supposed to do these motions not the robot...
Anonymous 01/16/25(Thu)09:56:27 No.103917633
>>103917593
Well I am sure you are just ragebaiting and aren't a closed service nigger but the whole model and its knowledge is built on copyrighted material so it doesn't matter if it blurts out something verbatim or paraphrases.
Well I am sure you are just ragebaiting and aren't a closed service nigger but the whole model and its knowledge is built on copyrighted material so it doesn't matter if it blurts out something verbatim or paraphrases.
Anonymous 01/16/25(Thu)09:56:58 No.103917636
>>103917589
>1. doesn't work
Working as intended.
>2. they have their own GPUs
How many FLOPs can that do?
It's useless, because they don't have good fab machines and can't use good fabs, so they can't make fast GPUs.
>1. doesn't work
Working as intended.
>2. they have their own GPUs
How many FLOPs can that do?
It's useless, because they don't have good fab machines and can't use good fabs, so they can't make fast GPUs.
Anonymous 01/16/25(Thu)09:57:00 No.103917637
Anonymous 01/16/25(Thu)09:58:00 No.103917645
>>103917633
You asked. I'm just telling you the logic being used to justify it.
You asked. I'm just telling you the logic being used to justify it.
Anonymous 01/16/25(Thu)09:58:05 No.103917646
>>103917597
They can run CUDA natively
They can run CUDA natively
llama.cpp CUDA dev !!OM2Fp6Fn93S 01/16/25(Thu)09:58:29 No.103917652
>>103917597
Moore Threads has made a PR for their MUSA framework which has been merged.
However, this framework is only functional for their server GPUs which are only sold business-to-business as part of their server solution (I asked).
Moore Threads has made a PR for their MUSA framework which has been merged.
However, this framework is only functional for their server GPUs which are only sold business-to-business as part of their server solution (I asked).
Anonymous 01/16/25(Thu)10:00:16 No.103917667
>>103917298
nta but learn english before calling people retarded.
nta but learn english before calling people retarded.
Anonymous 01/16/25(Thu)10:01:12 No.103917673
>>103917531
That's right. And basically, yeah, we're living in a parody of law and justice. If everyone is agreeing to do/sell something illegal, and one of them is fucking it up, you KNOW they're gonna throw their ass to the feds to lower competition.
That's right. And basically, yeah, we're living in a parody of law and justice. If everyone is agreeing to do/sell something illegal, and one of them is fucking it up, you KNOW they're gonna throw their ass to the feds to lower competition.
Anonymous 01/16/25(Thu)10:01:36 No.103917675
>>103917108
>Japan doesn't have any GPU restrictions
>The Japanese government actually went as far as making copyright not apply to LLM training
>They barely release any models, and the models they release are shit
What went wrong? Japan used to be synonym to technological advancement...
>Japan doesn't have any GPU restrictions
>The Japanese government actually went as far as making copyright not apply to LLM training
>They barely release any models, and the models they release are shit
What went wrong? Japan used to be synonym to technological advancement...
Anonymous 01/16/25(Thu)10:02:01 No.103917676
Anonymous 01/16/25(Thu)10:03:08 No.103917688
>>103917652
>PR for their MUSA framework which has been merged.
Alright, that's pretty dope.
>this framework is only functional for their server GPUs which are only sold business-to-business
Aaaaand that's dumb.
>>103917646
It transpiles from CUDA to MUSA? I wonder how well that works. Not that it matters since it only works with their B2B solutions anyway.
>PR for their MUSA framework which has been merged.
Alright, that's pretty dope.
>this framework is only functional for their server GPUs which are only sold business-to-business
Aaaaand that's dumb.
>>103917646
It transpiles from CUDA to MUSA? I wonder how well that works. Not that it matters since it only works with their B2B solutions anyway.
Anonymous 01/16/25(Thu)10:03:16 No.103917689
>>103917203
Thanks anon. Not enough of this stuff gets shared
Thanks anon. Not enough of this stuff gets shared
Anonymous 01/16/25(Thu)10:03:21 No.103917691
>>103917675
Japan lacks investor capital.
Japan lacks investor capital.
Anonymous 01/16/25(Thu)10:04:26 No.103917704
Anonymous 01/16/25(Thu)10:04:43 No.103917707
>>103917675
No babies only old people remain (in power) technological advancement fossilized to ~1995.
No babies only old people remain (in power) technological advancement fossilized to ~1995.
Anonymous 01/16/25(Thu)10:05:36 No.103917714
>>103917676
It's a legal workaround, they previously openly stated in their marketing materials that their GPUs can run CUDA applications unmodified. They still do, unofficially
It's a legal workaround, they previously openly stated in their marketing materials that their GPUs can run CUDA applications unmodified. They still do, unofficially
Anonymous 01/16/25(Thu)10:06:40 No.103917729
>>103917707
I put in やじるし in there but apparently that's not ok.
No babies -> only old people remain (in power) -> techno ...
I put in やじるし in there but apparently that's not ok.
No babies -> only old people remain (in power) -> techno ...
Anonymous 01/16/25(Thu)10:09:50 No.103917766
>>103917704
>It's not enough to import them, you need to give them voting rights to dilute the voting power of the native population.
tiresome.swf
>It's not enough to import them, you need to give them voting rights to dilute the voting power of the native population.
tiresome.swf
Anonymous 01/16/25(Thu)10:10:32 No.103917775
>>103917629
I have never had sex... Do women just lay there and do nothing?
I have never had sex... Do women just lay there and do nothing?
Anonymous 01/16/25(Thu)10:10:58 No.103917780
Anonymous 01/16/25(Thu)10:11:32 No.103917786
Anonymous 01/16/25(Thu)10:11:51 No.103917791
Anonymous 01/16/25(Thu)10:12:00 No.103917795
>>103917704
Anon, you're not making any sense.
Clearly your message is that immigration is le bad, but then the Japanese economy should be booming.
Anon, you're not making any sense.
Clearly your message is that immigration is le bad, but then the Japanese economy should be booming.
Anonymous 01/16/25(Thu)10:12:22 No.103917802
Anonymous 01/16/25(Thu)10:14:01 No.103917814
>>103917704
>Japan needs more immigrant workers to help them.
I expect this to have the exact opposite effect they intended. Japanese nationalism has been just itching for a disaster like this to break out into the mainstream
>Japan needs more immigrant workers to help them.
I expect this to have the exact opposite effect they intended. Japanese nationalism has been just itching for a disaster like this to break out into the mainstream
Anonymous 01/16/25(Thu)10:14:32 No.103917823
>>103917108
Muskrat's xAI too
Muskrat's xAI too
Anonymous 01/16/25(Thu)10:16:00 No.103917835
>>103917814
Because European nationalism has done so well since their disaster in 2014.
Because European nationalism has done so well since their disaster in 2014.
Anonymous 01/16/25(Thu)10:16:57 No.103917844
Anonymous 01/16/25(Thu)10:17:01 No.103917846
Anonymous 01/16/25(Thu)10:17:04 No.103917847
>>103917108
openai has an ongoing lawsuit about the exact same thing, no?
openai has an ongoing lawsuit about the exact same thing, no?
Anonymous 01/16/25(Thu)10:17:54 No.103917860
>>103917814
You're clueless. Kawaguchi has already turned into Europe.
You're clueless. Kawaguchi has already turned into Europe.
Anonymous 01/16/25(Thu)10:18:12 No.103917866
>>103917846
Nationalism in each individual European nation. Don't act dumb.
Nationalism in each individual European nation. Don't act dumb.
Anonymous 01/16/25(Thu)10:19:57 No.103917881
>>103917860
This is fake btw
This is fake btw
Anonymous 01/16/25(Thu)10:23:50 No.103917927
>>103917860
Actually I think this is all happening at the right time, zeitgeist-wise. The foreign nationals generally don’t have legal rights there, and Japan never accepted dual-citizenship. The right public sentiment could reverse all that and cleanse the country still
Actually I think this is all happening at the right time, zeitgeist-wise. The foreign nationals generally don’t have legal rights there, and Japan never accepted dual-citizenship. The right public sentiment could reverse all that and cleanse the country still
Anonymous 01/16/25(Thu)10:25:24 No.103917946
Gave mixtral another spin and damn it still holds up really well. mixtral 0.3 unpozzed when?
Anonymous 01/16/25(Thu)10:28:06 No.103917978
ds3 q6 uses exactly 666GB of ram when loaded. Use this information as you will
Anonymous 01/16/25(Thu)10:33:36 No.103918043
>>103917978
Actually, this is very helpful. I'm planning my next build. What was the context size?
Actually, this is very helpful. I'm planning my next build. What was the context size?
Anonymous 01/16/25(Thu)10:34:54 No.103918060
Anonymous 01/16/25(Thu)10:36:30 No.103918079
Anonymous 01/16/25(Thu)10:39:50 No.103918131
>>103918043
context size of 32768. It juuuuust fits into 768gb RAM and a single 24GB gpu with ooba, stable diffusion webui and sovits tts api. This is running Linux headless.
context size of 32768. It juuuuust fits into 768gb RAM and a single 24GB gpu with ooba, stable diffusion webui and sovits tts api. This is running Linux headless.
Anonymous 01/16/25(Thu)10:45:49 No.103918197
>>103918079
This is nice. I wonder if they'll get Turin support? That would be a sweet-ass upgrade once prices on those come down. Would make sense to get the higher-speed RAM right out of the gate if that's the case.
This is nice. I wonder if they'll get Turin support? That would be a sweet-ass upgrade once prices on those come down. Would make sense to get the higher-speed RAM right out of the gate if that's the case.
Anonymous 01/16/25(Thu)10:49:55 No.103918235
Anonymous 01/16/25(Thu)10:51:23 No.103918252
>>103916967
Whenever he enters some industry he always makes it worse.
Whenever he enters some industry he always makes it worse.
Anonymous 01/16/25(Thu)10:55:19 No.103918283
>>103917866
Nationalism is illegal in Europe.
Nationalism is illegal in Europe.
Anonymous 01/16/25(Thu)11:00:42 No.103918344
>>103917039
phi-4 uses mostly synthetic data and btfos llama
phi-4 uses mostly synthetic data and btfos llama
Anonymous 01/16/25(Thu)11:01:23 No.103918352
>>103918344
Lol, no it most certainly does not.
Lol, no it most certainly does not.
Anonymous 01/16/25(Thu)11:01:45 No.103918356
Anonymous 01/16/25(Thu)11:04:28 No.103918391
>>103918043
minimax seems better so far imo btw
minimax seems better so far imo btw
Anonymous 01/16/25(Thu)11:04:35 No.103918394
Anonymous 01/16/25(Thu)11:07:48 No.103918435
>>103918235
Take his cock out of your mouth, saar. He started backing AI regulations a few months ago. Also where are those promised grok weights?
Take his cock out of your mouth, saar. He started backing AI regulations a few months ago. Also where are those promised grok weights?
Anonymous 01/16/25(Thu)11:07:53 No.103918437
Anonymous 01/16/25(Thu)11:08:32 No.103918445
>>103918435
Grok 2 has only been out a couple months, just be patient.
Grok 2 has only been out a couple months, just be patient.
Anonymous 01/16/25(Thu)11:08:49 No.103918451
>>103918437
using openrouter myself
using openrouter myself
Anonymous 01/16/25(Thu)11:08:55 No.103918453
Eleuthertroons still trying to dictate what should or should not be in an open dataset:
https://arxiv.org/abs/2501.08365
> Towards Best Practices for Open Datasets for LLM Training
>
> Many AI companies are training their large language models (LLMs) on data without the permission of the copyright owners. The permissibility of doing so varies by jurisdiction: in countries like the EU and Japan, this is allowed under certain restrictions, while in the United States, the legal landscape is more ambiguous. Regardless of the legal status, concerns from creative producers have led to several high-profile copyright lawsuits, and the threat of litigation is commonly cited as a reason for the recent trend towards minimizing the information shared about training datasets by both corporate and public interest actors. This trend in limiting data information causes harm by hindering transparency, accountability, and innovation in the broader ecosystem by denying researchers, auditors, and impacted individuals access to the information needed to understand AI models.
> While this could be mitigated by training language models on open access and public domain data, at the time of writing, there are no such models (trained at a meaningful scale) due to the substantial technical and sociological challenges in assembling the necessary corpus. These challenges include incomplete and unreliable metadata, the cost and complexity of digitizing physical records, and the diverse set of legal and technical skills required to ensure relevance and responsibility in a quickly changing landscape. Building towards a future where AI systems can be trained on openly licensed data that is responsibly curated and governed requires collaboration across legal, technical, and policy domains, along with investments in metadata standards, digitization, and fostering a culture of openness.
https://arxiv.org/abs/2501.08365
> Towards Best Practices for Open Datasets for LLM Training
>
> Many AI companies are training their large language models (LLMs) on data without the permission of the copyright owners. The permissibility of doing so varies by jurisdiction: in countries like the EU and Japan, this is allowed under certain restrictions, while in the United States, the legal landscape is more ambiguous. Regardless of the legal status, concerns from creative producers have led to several high-profile copyright lawsuits, and the threat of litigation is commonly cited as a reason for the recent trend towards minimizing the information shared about training datasets by both corporate and public interest actors. This trend in limiting data information causes harm by hindering transparency, accountability, and innovation in the broader ecosystem by denying researchers, auditors, and impacted individuals access to the information needed to understand AI models.
> While this could be mitigated by training language models on open access and public domain data, at the time of writing, there are no such models (trained at a meaningful scale) due to the substantial technical and sociological challenges in assembling the necessary corpus. These challenges include incomplete and unreliable metadata, the cost and complexity of digitizing physical records, and the diverse set of legal and technical skills required to ensure relevance and responsibility in a quickly changing landscape. Building towards a future where AI systems can be trained on openly licensed data that is responsibly curated and governed requires collaboration across legal, technical, and policy domains, along with investments in metadata standards, digitization, and fostering a culture of openness.
Anonymous 01/16/25(Thu)11:12:33 No.103918494
>>103918453
And that's a good thing.
And that's a good thing.
Anonymous 01/16/25(Thu)11:12:51 No.103918496
>>103918451
>openrouter
and not even a single mention on lcpp's repos yet. hopefully if its on someone's radar if its really good.
looks like a side-grade from ds3 tho, what's the benefit? Better prose?
>openrouter
and not even a single mention on lcpp's repos yet. hopefully if its on someone's radar if its really good.
looks like a side-grade from ds3 tho, what's the benefit? Better prose?
Anonymous 01/16/25(Thu)11:12:59 No.103918498
>>103918453
>harm by hindering transparency, accountability,
>sociological challenges
>responsibly curated and governed
>harm by hindering transparency, accountability,
>sociological challenges
>responsibly curated and governed
Anonymous 01/16/25(Thu)11:14:07 No.103918506
>>103917946
A sequel to mixtral 8x7b leveraging the idea of having a generalist expert that's always on would be dope.
Hell, use the expert for self-speculative decoding. Have that one expert generate N tokens then have the other two most appropriate experts validade them or something.
A sequel to mixtral 8x7b leveraging the idea of having a generalist expert that's always on would be dope.
Hell, use the expert for self-speculative decoding. Have that one expert generate N tokens then have the other two most appropriate experts validade them or something.
Anonymous 01/16/25(Thu)11:15:47 No.103918521
>>103918496
No repetition issues is the main thing. And linear context. Seems about as smart to me so far.
No repetition issues is the main thing. And linear context. Seems about as smart to me so far.
Anonymous 01/16/25(Thu)11:16:41 No.103918528
>>103918496
>>103918391
MiniMax isn't better, not even a side grade, it's just worse and full of gptslop and alignment.
>>103918391
MiniMax isn't better, not even a side grade, it's just worse and full of gptslop and alignment.
Anonymous 01/16/25(Thu)11:16:48 No.103918529
>>103918496
And it has like a dozen different major improvements implemented, prob be months before llama.cpp has it going
And it has like a dozen different major improvements implemented, prob be months before llama.cpp has it going
Anonymous 01/16/25(Thu)11:17:51 No.103918539
>>103918528
Seems the opposite to me. Im using a CYOA preset and minimax is coming up with far more natural and interesting stuff than deepseek did
Seems the opposite to me. Im using a CYOA preset and minimax is coming up with far more natural and interesting stuff than deepseek did
Anonymous 01/16/25(Thu)11:18:07 No.103918541
>>103918496
>what's the benefit?
it would (theoretically) fit in my current fourth-gen EPYC build. I hope that llama.cpp will support it sooner rather than later
>what's the benefit?
it would (theoretically) fit in my current fourth-gen EPYC build. I hope that llama.cpp will support it sooner rather than later
Anonymous 01/16/25(Thu)11:18:40 No.103918552
>>103918496
>looks like a side-grade from ds3 tho, what's the benefit? Better prose?
It's 2/3rds the size of DSV3, so some that built machines with 405B in mind but not 700B can run it.
>looks like a side-grade from ds3 tho, what's the benefit? Better prose?
It's 2/3rds the size of DSV3, so some that built machines with 405B in mind but not 700B can run it.
Anonymous 01/16/25(Thu)11:19:07 No.103918560
>>103918521
>No repetition issues is the main thing.
I don't get any kind of repetition that bothers me in any of my gens (all justified by the base context).
If anything, the repetition in /lmg/ around repetition in ds3 feels worse.
>No repetition issues is the main thing.
I don't get any kind of repetition that bothers me in any of my gens (all justified by the base context).
If anything, the repetition in /lmg/ around repetition in ds3 feels worse.
Anonymous 01/16/25(Thu)11:20:42 No.103918569
Can I use XTC to always remove the highest chance token as long as there's at least two tokens to be sampled?
Anonymous 01/16/25(Thu)11:20:53 No.103918571
>>103918560
Stop repeating this, you're special, we get it.
Stop repeating this, you're special, we get it.
Anonymous 01/16/25(Thu)11:27:51 No.103918658
>>103918560
I mean it just is repetitious, that is like a universally acknowledged fact about the model. it's great that you can use it without realizing it but that doesn't mean it's not repetitious, it just means you're blind
I mean it just is repetitious, that is like a universally acknowledged fact about the model. it's great that you can use it without realizing it but that doesn't mean it's not repetitious, it just means you're blind
Anonymous 01/16/25(Thu)11:31:06 No.103918708
>>103917239
Well, it's a start. Rentry updated. I'll play with it more later, need to get some actual work done today.
Well, it's a start. Rentry updated. I'll play with it more later, need to get some actual work done today.
Anonymous 01/16/25(Thu)11:33:25 No.103918738
Anonymous 01/16/25(Thu)11:35:06 No.103918759
>>103918708
topical. I'd expect migu colours like cyan with a small area magenta or black but not rainbow
topical. I'd expect migu colours like cyan with a small area magenta or black but not rainbow
Anonymous 01/16/25(Thu)11:35:21 No.103918765
Another week over without anything happening
Anonymous 01/16/25(Thu)11:38:31 No.103918801
>103918708
I thought Petra was joking, are Miku fags actual fags after all?
I thought Petra was joking, are Miku fags actual fags after all?
Anonymous 01/16/25(Thu)11:39:36 No.103918820
>>103918197
Yes >>103918356, but
>Would make sense to get the higher-speed RAM right out of the gate
DDR5-6000 is also pricy as fuck compared to 4800. It would likely be more cost-effective to upgrade the RAM with the CPU, or go with 5600 as a middle ground? I don't know
Yes >>103918356, but
>Would make sense to get the higher-speed RAM right out of the gate
DDR5-6000 is also pricy as fuck compared to 4800. It would likely be more cost-effective to upgrade the RAM with the CPU, or go with 5600 as a middle ground? I don't know
Anonymous 01/16/25(Thu)11:40:29 No.103918835
Anonymous 01/16/25(Thu)11:41:08 No.103918847
>>103918708
Looking cute, I like it.
Looking cute, I like it.
Anonymous 01/16/25(Thu)11:43:57 No.103918885
>>103918043
Isn't a single-socket board going to tank your already not that good CPU performance?
Isn't a single-socket board going to tank your already not that good CPU performance?
Anonymous 01/16/25(Thu)11:45:42 No.103918901
>>103918801
Blind anon... Their behavior and speech style, akin to /aicg/ trannies, 2+2=4 lol.
Blind anon... Their behavior and speech style, akin to /aicg/ trannies, 2+2=4 lol.
Anonymous 01/16/25(Thu)11:47:10 No.103918920
>>103918708
Looks cool, anon. People shouldn't let the mentally ill co-opt rainbows, it's ridiculous.
Looks cool, anon. People shouldn't let the mentally ill co-opt rainbows, it's ridiculous.
Anonymous 01/16/25(Thu)11:50:45 No.103918964
>>103918708
kawaii!
kawaii!
Anonymous 01/16/25(Thu)11:52:48 No.103918987
>>103918708
Nice and simple. Nice comfy 80's feel.
http://aa.ja.utf8art.com/node/7592 has some good miku ascii art content if you want to get a bit fancier.
Nice and simple. Nice comfy 80's feel.
http://aa.ja.utf8art.com/node/7592 has some good miku ascii art content if you want to get a bit fancier.
Anonymous 01/16/25(Thu)11:55:05 No.103919010
>>103918801
Ever notice how mikufags are making useful contributions to the thread and anti mikufags are just useless complainers?
Ever notice how mikufags are making useful contributions to the thread and anti mikufags are just useless complainers?
Anonymous 01/16/25(Thu)11:57:38 No.103919038
Anonymous 01/16/25(Thu)11:58:17 No.103919051
>>103919010
Use cases fot that shit? None.
Use cases fot that shit? None.
Anonymous 01/16/25(Thu)11:58:33 No.103919058
>>103918820
>It would likely be more cost-effective to upgrade the RAM with the CPU
depending on how long you plan to run with the genoa cpu, I'd say its probably way cheaper to upgrade as a set.
But once you're looking at upgrading that much of the system, it almost makes more sense to just build a second one...its the same song-and-dance that I've gone through with every upgradeable machine I've ever owned (unless I wait an extreme amount until upgrades are a couple hundred dollars)
>It would likely be more cost-effective to upgrade the RAM with the CPU
depending on how long you plan to run with the genoa cpu, I'd say its probably way cheaper to upgrade as a set.
But once you're looking at upgrading that much of the system, it almost makes more sense to just build a second one...its the same song-and-dance that I've gone through with every upgradeable machine I've ever owned (unless I wait an extreme amount until upgrades are a couple hundred dollars)
Anonymous 01/16/25(Thu)11:58:52 No.103919065
>>103919010
cuda dev is notoriously anti-miku, going as far as spamming the thread with blacked miku to prove his point
cuda dev is notoriously anti-miku, going as far as spamming the thread with blacked miku to prove his point
Anonymous 01/16/25(Thu)12:00:05 No.103919090
>>103919065
why are tetofags like this?
why are tetofags like this?
Anonymous 01/16/25(Thu)12:00:41 No.103919099
>>103919065
i still honestly don't know how to feel about it
such a person doing such a thing
is this what getting ntr'd feels like?
i still honestly don't know how to feel about it
such a person doing such a thing
is this what getting ntr'd feels like?
Anonymous 01/16/25(Thu)12:01:18 No.103919108
>>103919065
wtf i hate cuda dev now
wtf i hate cuda dev now
Anonymous 01/16/25(Thu)12:02:02 No.103919118
>>103919010
Ur waifu is shit, regardless of people's actions.
Ur waifu is shit, regardless of people's actions.
llama.cpp CUDA dev !!OM2Fp6Fn93S 01/16/25(Thu)12:03:57 No.103919148
>>103919065
I am definitely more pro Miku than I am anti Miku though I would be fine with other thread mascots as well.
I'm a massive weeb in general and have invested the effort to learn Japanese for no other reason than to better enjoy anime, manga, and sad pandas.
I am definitely more pro Miku than I am anti Miku though I would be fine with other thread mascots as well.
I'm a massive weeb in general and have invested the effort to learn Japanese for no other reason than to better enjoy anime, manga, and sad pandas.
Anonymous 01/16/25(Thu)12:04:31 No.103919155
>>103919065
If i remember correctly, he said he was more neutral towards mikufags than that. Certainly not anti-miku.
If i remember correctly, he said he was more neutral towards mikufags than that. Certainly not anti-miku.
Anonymous 01/16/25(Thu)12:04:56 No.103919161
Anonymous 01/16/25(Thu)12:06:18 No.103919176
>>103919161
pedo website, pedo repo, pedo general
pedo website, pedo repo, pedo general
Anonymous 01/16/25(Thu)12:06:45 No.103919183
Anonymous 01/16/25(Thu)12:08:41 No.103919202
>>103919183
Techloligy is the word.
Techloligy is the word.
Anonymous 01/16/25(Thu)12:09:02 No.103919209
Anonymous 01/16/25(Thu)12:12:48 No.103919257
Anonymous 01/16/25(Thu)12:13:09 No.103919260
>>103919148
based cudaCHAD
based cudaCHAD
Anonymous 01/16/25(Thu)12:15:46 No.103919294
>>103919260
He won't fuck you, quit glazing.
He won't fuck you, quit glazing.
Anonymous 01/16/25(Thu)12:17:15 No.103919320
Anonymous 01/16/25(Thu)12:19:31 No.103919351
>>103919010
Yeah like >>103916726. Latching on to what someone did and then demanding that shitfu gets attached. Very useful contribution. Kill yourself you disgusting troon.
Yeah like >>103916726. Latching on to what someone did and then demanding that shitfu gets attached. Very useful contribution. Kill yourself you disgusting troon.
Anonymous 01/16/25(Thu)12:19:37 No.103919352
>>103919058
Upgrading CPU and RAM is a minor change, I usually do it 2-3 times before changing my motherboard, unless it's an Intel
Upgrading CPU and RAM is a minor change, I usually do it 2-3 times before changing my motherboard, unless it's an Intel
Anonymous 01/16/25(Thu)12:20:06 No.103919359
Anonymous 01/16/25(Thu)12:21:57 No.103919386
I am not afraid to admit that I love Hatsune Miku and I am trans. Everyone here should be proud if they love Miku and are trans.
Anonymous 01/16/25(Thu)12:31:21 No.103919499
>>103919386
ACK
ACK
Anonymous 01/16/25(Thu)12:31:25 No.103919500
>>103918708
fuck is this tranny shit?
fuck is this tranny shit?
Anonymous 01/16/25(Thu)12:48:51 No.103919691
>>103912475
The middle is suppose to be a free space, thats how bingo works
The middle is suppose to be a free space, thats how bingo works
Anonymous 01/16/25(Thu)13:01:50 No.103919879
I test models with that incest joke bot. My ranking
>opus
>sonnet 3.5
>deepseek 3
>doesn't know what's funny below this line
>mistral large 2
>minimax
>dolphin 8x22
>smaller mistral models
>gpt4
>llama3
>opus
>sonnet 3.5
>deepseek 3
>doesn't know what's funny below this line
>mistral large 2
>minimax
>dolphin 8x22
>smaller mistral models
>gpt4
>llama3
Anonymous 01/16/25(Thu)13:03:56 No.103919902
>>103919879
What incest joke bot?
What incest joke bot?
Anonymous 01/16/25(Thu)13:04:36 No.103919915
Made on a 3090 with Cosmos 7B
https://files.catbox.moe/5i1kw3.mp4
https://files.catbox.moe/5i1kw3.mp4
Anonymous 01/16/25(Thu)13:06:46 No.103919955
Anonymous 01/16/25(Thu)13:21:01 No.103920169
So I got silly tavern and Nemo 12b instruct and made a group of 2 characters to talk to each other and do things. Maybe at the beginning the description prompt describes lots of details bit very quickly the description detail becomes reduces to rather short things. Anyway to make it more descriptive of more thing? I think then the ai has 'more' things to work with and makes up more stuff.
Or is this a model problem? How do I force more to come out and 'invent'?
Or is this a model problem? How do I force more to come out and 'invent'?
Anonymous 01/16/25(Thu)13:26:51 No.103920258
>>103919955
Haven't tried furries. It can sort of do img-to-video on anime stuff, but unlike hunyan, it doesn't seem to be trained on that. However, the gens are much more "lively". I often get hunyan gens where most of the clip is static.
https://files.catbox.moe/eplqi2.webp
Haven't tried furries. It can sort of do img-to-video on anime stuff, but unlike hunyan, it doesn't seem to be trained on that. However, the gens are much more "lively". I often get hunyan gens where most of the clip is static.
https://files.catbox.moe/eplqi2.web
Anonymous 01/16/25(Thu)13:36:05 No.103920416
>>103920258
That gen is a nightmare but cool shit regardless.
Good to see some good progress. Now we just need an archtecture that enables gening minutes long videos with stability and consistency.
>>103920169
It is a small model thing in that bigger models might be more resistant to that kind of thing, but you can get around it with prompting. There's a tradeof fin that the model might get dumber.
You can use each character's card's character notes fields at a low depth to add a small snippet of information regarding that character's way of speakin and acting. That could help.
That gen is a nightmare but cool shit regardless.
Good to see some good progress. Now we just need an archtecture that enables gening minutes long videos with stability and consistency.
>>103920169
It is a small model thing in that bigger models might be more resistant to that kind of thing, but you can get around it with prompting. There's a tradeof fin that the model might get dumber.
You can use each character's card's character notes fields at a low depth to add a small snippet of information regarding that character's way of speakin and acting. That could help.
Anonymous 01/16/25(Thu)13:42:03 No.103920514
>>103918987
Perfect, thanks! I was looking everywhere for this.
>>103918759
It uses command toilet (lol) and there are 3 color options.
> rainbow color scheme is literally 'toilet -F gay'
Perfect, thanks! I was looking everywhere for this.
>>103918759
It uses command toilet (lol) and there are 3 color options.
> rainbow color scheme is literally 'toilet -F gay'