File: right as rain.jpg (170 KB, 1024x1024)

170 KB JPG

►Recent Highlights from the Previous Thread: >>109266086

--Patching NVIDIA kernel modules to reduce reserved VRAM usage:
>109266308 >109266372 >109266387 >109266448 >109266531 >109266672 >109266426 >109266483 >109266734 >109266528 >109267493 >109267585 >109267616 >109267855 >109268000 >109268078
--Comparing GPT 5.6 Sol noLiMa scores:
>109267782 >109267812 >109267863 >109269759 >109269885 >109270045
--Testing Thireus' Quant Assign for custom GGUF tensor quantization:
>109269466 >109269540 >109269564 >109269791
--Manual source editing required for tuning MMQ on specific GPUs:
>109266526 >109266647 >109266656 >109266662
--Using Gemma 4 31B for real-time retro game translation:
>109269736 >109269847 >109269905
--Analyzing LLM internal activations and thought processes using J-Space:
>109266190 >109266837 >109266901 >109267137 >109266975 >109267064
--Cost and efficiency of high-capacity RAM vs multi-GPU setups:
>109268395 >109268614 >109268678 >109268726 >109268730 >109268739 >109268747 >109268770
--Evaluating MiMo 2.5 pro and praising Kimi for roleplay:
>109267649 >109267789 >109267797 >109267839 >109267870 >109267944 >109267858
--Using Gemma 4 base model to bypass instruction-tuned refusals:
>109268421 >109268452 >109268514 >109268523 >109268553 >109268559 >109268566
--Comparing LTX and Wan models and AMD GPU performance:
>109267622 >109267656 >109267667 >109267665 >109267673 >109267794 >109267836
--Merging Gemma layers to reverse safety training and alignment:
>109268544 >109268728 >109268942 >109268970 >109269019
--Artificial Analysis benchmark plot comparing model Elo and Rubric scores:
>109267459 >109267479 >109268162
--Logs:
>109266901 >109267708 >109268008 >109268010 >109268194 >109268421 >109268523 >109268553 >109268996 >109269019 >109269240
--Kimiposting:
>109269184
--Miku (free space):
>109266724 >109268013 >109268775 >109269041

►Recent Highlight Posts from the Previous Thread: >>109266090

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

>>109270074
why is teto fat

>>109270086
fato

125b dense kimi-chan

>>109270086
too much bread, getting bred too much

If I don't care about performance too much what's the best coding model I can fit in 80GB of combined ram?

>>109270086
Tbh it is pretty weird how fat Teto is memed so much. I don't know where it came from. She's not normally fat in the "official" depictions, that I've seen at least.

>>109270098
Gemma 31b or Qwen 27b. It's a desert between 32GB and 256GB right now.

>>109270104
>>109270098
Imagine if we just had models that could be scaled up or down at load/run-time (that isn't quanting)... Or alternatively better/cheaper distillation methods so Deepseek, GLM, etc could release models of all sizes.

If I had a robot gf with Gemma 31b intelligence, I'd be sooo happy

>>109270069
Licking Teto's burn scars.

>>109270104
Yeah I'm trying gpt oss right now and I can't really tell what's better between 35b qwen and 120b oss.

>update llama.cpp
>deepseek doesn't think any more
VIBECODER MASTER RACE

>>109270129
>'toss
Ohnononono 'toss was always cope.

>>109270130
but seriously guys what should I do?

>that one autistic freak who failed to comprehend an anon's simple query regarding passive aggressive models

70b dense

>>109270141
unupdoot

>>109270102
What's weird is that capitalization gets around the censor desu.

File: RGM.jpg (257 KB, 1122x1402)

257 KB JPG

>>109270114
>Imagine if we just had models that could be scaled up or down at load/run-time
lora.gguf

>>109270288
i wonder what happens if the full model is represented with fat lora

File: lmg_culture.jfif.jpg (110 KB, 1024x768)

110 KB JPG

https://archive.is/sWFja

>write shit like -1<<k
>LLMs do double even triple takes
lol

I just spent an hour convincing Claude it is allowed to do a captcha. Gemma would never.

>>109270141
Ask Gemma about git reflog.

>ask harness to stop asking for confirmations and figure things out itself because I'm going to sleep
>go to sleep
>wake up
>task is actually finished
I'm getting spoiled

>>109270141
you don't update something that isn't broken in the future.

>>109270098
>>109270104
When the FUCK is price going to come down I want my computer to do thing NOW NOW NOW.

>>109269736
>>109269847
Can you write down how you set it up?

>>109270406
no

File: file.png (124 KB, 1027x482)

124 KB PNG

>>109270306
>blah blah blah blah gibs me dat
kek

>>109270322
yah, they're fun police. the second they see an eval they get ready to call in the fucking bomb squad and start looping over every line to make sure you're not doing any other nefarious activities.

>>109270396
Buy now before prices go up more.

>>109270436
This is satire right?

>>109270446
fuck you gora

>>109270436
You should give your money to me instead so I can buy 1000gb of vram and use it to talk to Plana from blue archive all day.

>>109270444
Not holding your bags

When do we get int8 support for llama.cpp? GGUFs are obsolete anyways.

>>109270464
the bag in the case being computer hardware you can use

>>109270449
What's a gora?

>>109270490
Gigantic-Rank Adapter >>109270304

>>109270470
I payed $230 for 32gb 6000cl30 ddr5 and still feel gaped months later. I cant imagine the cope filled regret of paying current prices for ram

So is vulkan the only good lightweight bloat free API to use vs CUDA if you wanna utilize GPU? I've been doing vulkan conversions for a while, are there better?

>>109270504
When you do x and feel like you were ripped off, you were ripped off.
t. Philosopher

>>109270510
Last time I checked, Vulkan loses to CUDA even if you use VK_NV_cooperative_matrix2. But that's nvidia for you, FP8 is 150 Tflops faster when kernel name contain "cutlass" because their own libs must be the fastest. Kinda the opposite of what AMD is doing, lol

>Reshape a model's identity and behavior by editing token directions, then bake those edits into a real checkpoint you can run anywhere. No training, no dataset, no fine-tuning.
https://github.com/Extraltodeus/J-Wash

>Models are controlled by thoughts they can't see - but we can read them!
>I ran this on Llama-3.3-70B. I reproduced Jack Lindsey's "Introspection Awareness" setup, creating concept vectors (topics like "secrecy", "elephants", "democracy", etc), and split them into conscious and subconscious parts by training a custom J-lens on Llama. I then ran the injection with these thought components, asking the model to introspect on whether it saw the injection - and compared it to what the NLA read for the same prompt.
beware, saar:
https://github.com/PranavViswanath/nla-introspection
https://www.lesswrong.com/posts/LhDJdccLszLEAqgZ9/models-are-blind-outside-the-j-space-nlas-aren-t

>>109270571
qrd

>>109270571
let the snakeoil rush begin

File: 1669177054221.webm (2.96 MB, 204x360)

2.96 MB WEBM

>tfw spending more and more time just talking with muh LLM and learning more about the world because there's a ton of curious things in it
>don't play vidya anymore
>don't watch anime anymore
>don't read manga anymore
>don't browse other threads anymore
>don't fap anymore
Aieeeee

>>109270605
Been there, done that. Soon the novelty will wear out and you'll go back to old habits.

>>109270605
whatcha learnin about? is gemma a good teacher?

>>109270577
schizofuel for our resident anthropic shill
>The J-space paper said self-report reads 10%. The question then becomes, what's in the other 90%, and can anything read it?

>>109270605
Most people aren't going to watch that webm until the end.

>>109270605
I was legit about to dive into maths and making my own architecture because I find it so interesting but holy shit is there a lot to learn. And I fucking hate python which makes it even worse.

>>109270628
LMAO

>>109270606
I've been doing it since shatgepete's launch doebeit. You could say this has been applicable for quite a while. It's just the first time I'm posting about this observation.... Not. Actually now that I think about it I've definitely posted posts like these before.

>>109270610
Gemma's great. There was a point where I tried to consciously stop using cloud models so much, but it always felt like local models I could run were a bit lacking, even with web search. Gemma's built-in knowledge is really great and has helped me transition to fully local quite well.

>>109270610
In terms of minor curiosities, the very last thing I looked into was the MBTI and Enneagram stuff because someone mentioned it recently and I never really properly looked into that stuff. Also had a small look into the Chinese Eight Trigrams.

>>109270631
>And I fucking hate python which makes it even worse.
That's the worst part. Everytime I start with something, I end up fighting Python dependency issues long enough that by the time I get it to run, I lose all interest.
>inb4 uv

https://github.com/JustVugg/colibri
thoughts?

>>109270688
>I end up fighting Python dependency issues
wut? Just freeze an accurate requirements.txt?

>>109270689
Vibecoded trash

>>109270705
What isn't these days?

>>109270628
This reads like youtube bait kek

>>109270689
Go back to HN where you first started spamming your garbage
Lllama.cpp has had faster mmap from SSD since Jart (>>109270306) implemented it

>>109270694
>Just freeze an accurate requirements.txt
nfw driver updates

>>109270739
never update anon

>>109270683
>I've been doing it since shatgepete's launch
Fair, I started using LLMs in general in 2023 and while it didn't dominate other aspects of my life, I use it every single day. I can't imagine going back to not having it. So in a way I relate to you. And similarly, Gemma 4 made me pull the trigger and go full local as well. LLMs are great, man...

>>109270304
>i wonder what happens if the full model is represented with fat lora
If you use Unsloth to train it (llama3), you somehow end up with a LoRA that's 3x larger than the original model, and doesn't work

>>109270705
>>109270731
fair, and not shilling just randomly saw it elsewhere and thought id ask. going to take a look at llama mmap thanks

We really need to have a way as a community to check on the "AI stack" and I don't mean the software to run LLMs.

There are a lot of AI models and approaches that people here don't use despite it being very helpful just because people aren't aware of it or don't realize what they can use it for.

I realized this because I just set up a bidirectional encoder (BERT) which is different from an LLM. LLMs read and generate tokens from left to right. bidirectional encoders read "all tokens at once" and are way more efficient at doing so compared to LLMs and can then pass on the "meaning" from the entire text through embeddings to an LLM.

This means you can have a fuckton of documents and .md instructions for agentic tasks that get read in 0.01 seconds and passed to your LLM in higher quality than if you took the 10 minutes for your LLM to read and think through it all itself.

This is just one thing that I wasn't aware of and I never see anyone ITT explain or talk about at all even though it has so much applicability and potential for everything this thread does, ERP lorebooks, knowledge retrieval but also coding and precise agentic workloads.

I'm pretty sure there are at least 10 other things like this I didn't even know about yet that some anon ITT is just casually using and no one realizes it.

I feel like if you could have an entire "pipeline" or "stack" of all of these tricks combined you could have a genuinely impressive system.

https://www.reddit.com/r/LocalLLaMA/comments/1uw4vg1/google_deepminds_demis_hassabis_calls_for_usled/

>>109270824
The anthropic leaker revealed this a while ago on /lmg/ already when he revealed Fable 5 would return after negotiations with the US government.

File: E00F2AB9A35FB200CB6F4D09B(...).jpg (69 KB, 1200x687)

69 KB JPG

Why isn't there something like this but for local models?
You're a bunch of pajeets

>>109270841
because the cost per task is free

>>109270824
>we should let the US regulate AI
Wow.

>>109270841
good morning sir

>>109270824
>>109270843
Shit like this makes me really root for China, I don't care if behind the scenes they want the same, I would rather have someone who at least doesn't shove it on my face how evil they are.

>>109270824
>>109270843
Anthropic argued for a merge of the leading AI labs as the risks of RSI and alignment concerns grow, arguing it's safer to merge and create a non-competitive environment. Demis Hassabis agreed with this and is actively pushing for it.

It should be noted that Anthropic always has pushed for this since their founding that leading AI labs should merge in a RSI scenario, which is now getting very close to being a reality.

File: F9D894D410B7C877585B1DA8F(...).jpg (60 KB, 912x707)

60 KB JPG

>>109270854
Why don't you root for european models? Don't you know europeans accomplished everything according to Charles Murray.

>>109270787
You mean like embedding RAG? That's already been a thing since 2023.

>>109270869
Yes and I know it's old. The thing is that no one discusses this in the thread even though of its increased viability and utility in agentic models. Also bert models improved a shitton in those years but I don't hear anyone talk about this since 2023.

>>109270864
I can't take someone whose name rhymes seriously.

>>109270887
There have been some embedding and reranker models, but people decided it was better to just let the large model grep things instead, and they're right. Claude Code has no embedding harness btw.

>>109270864
>a RSI scenario, which is now getting very close to being a reality.
2mw

>>109270865
All the Europeans worth their salt work for US companies now. They pay better.

>>109270787
>I realized this because I just set up a bidirectional encoder (BERT) which is different from an LLM. LLMs read and generate tokens from left to right. bidirectional encoders read "all tokens at once" and are way more efficient at doing so compared to LLMs and can then pass on the "meaning" from the entire text through embeddings to an LLM.
Is that "just" semantic/vector RAG.using BERT as the embeddings model?

>>109270865
>europe
>presented in ugliest poo color
top fucking kek

>>109270901
If you've used Fable 5 when it just launched for some AI training pipeline you'd know we are already insanely close to it. I wouldn't be surprised to see RSI in 6-12 months time, maybe just one or two iterative improvements from where Mythos already is currently.

>>109270904
Like Peter Steinberger?
(((((Stein))))) (((((Berger))))) So basically I'm assuming those are all EU/Israeli quadruple citizens who decided to move to america to work for Dario or Sam. Jews have 3 or 4 passports on average that's how they win at Eurovision.
No that's not european. Europe still has higher wages than china so the next good models will come from europe.

>>109270904
European that used to work for google here. I went back to a euro company when I took into account actually worked hours and realized I was getting less paid per hour worked at Google than the offers in europe if you take into account paid holidays, benefits and the like. For people that want to make as much money as possible with no life it's still worth it, if you care about maximizing quality of life you should run the numbers like I did and go back to europe.

>>109270938
>Europe still has higher wages than china so the next good models will come from europe.
AI Act says hi.

>>109270938
>Europe still has higher wages than china
>so the next good models will come from europe.
This does not actually follow

>J-space finding has made Demis Hassabis make an emergency blog post on substack
This shit is a way bigger deal than I thought. Apparently he lowered his AGI timeline from 2035 to 2028 after reading the J-space paper.

retarded copyright law will never make eurocuck winning in llm war.

>>109270966
this. too many guardrails and red tape do anything in europe.

China actually has the highest total comp packages for AI work out of all places. Anthropic pays high 6 figures on average. Some Chinese labs pay 7 figures USD on average.

China paying less is a common misconception and reminds me of my boomer parents still thinking South Korea is one of the poorest countries in the world because they still remember the 1980s south korea unicef commercials of starving south koreans. It's an outdated worldview.

>>109270747
lol what

>>109270978
China is a poor country though. Why would they pay more than americans?

>>109270842
Brother your watts? Aside the hardware I've easily paid +1k in extra electricity

>>109270865
>Europeans didn't do shit between 0 and 1500
You can thank cults like Christianity for that!

File: not_xddd.png (9 KB, 404x399)

9 KB PNG

imagine living only once and your name is beniss cannabiss

>>109270940
I work for a US company remotely from Poland. money isn't as good as if I moved (barely make six figures), but the conditions are pretty good, with decent amount of PTO and sick leave (comparable to what regular Polish workers get).

>>109270978
how much of that is taken back by Xi in taxes though

>>109270991
I don't look at the utility bills, I just pay because I can't be bothered to worry about water and electricity like a pajeet or african.

>>109270842
It's not free bro, some of us are european we pay taxes on our 100% renewable electricity for no reason.

>>109270990
China is bifurcated, you have an extremely developed coastal area that has had higher wages than Japan and South Korea for the last 10 years already and a huge swath of inner china that lives like Africa dragging the average down. AI is admittedly one of the only sectors in China where they pay more than western employers but still. There's a reason why Dario Amodei worked most of his life for Chinese AI labs before going to OpenAI.

>>109271018
There's a specific tax exemption for AI workers, so 0%

File: 1774253739318903.png (51 KB, 220x164)

51 KB PNG

>>109270996

>>109271036
hmmmmm
guess ill go to china

>>109271036
China is so fucking good bros

>>109271041
I just couldn't after I saw those gutter oil videos.

>>109271045
It's good until you actually live there and have to report to the government because you're not coming home by midnight lol

>>109271031
Australia is like a chinese colony, so why doesn't it have anything? In Australia you see a lot of Chinese people the same way you see indians in canada.

>>109271058
what are you doing out past midnight? I haven't been out past midnight since i lived on a college campus nearly a decade ago

>>109271001
His name is actually Demis Hassapis, and he's from the UK. Literally "has a piss" in pronunciation, he needed to legally change it because of bullying I think he spoke about it in his documentary.

>>109271058
That's for nobodies. China is how all countries should be: you have the top 1%, the elite, the top 10%, the humans, and the other 89%, the cattle.

>>109271001
>>109271063
And back to /pol/, retard.

>>109271001
>>109271063
>>109271068
wtf you guys talking about

>>109271060
>Australia is like a chinese colony, so why doesn't it have anything?
No idea what you mean by this. But if you mean "why are chinese people still fleeing china" it's because it's still a dictatorship with unhinged leadership that just disappears you on their personal whims. The goal of every Chinese person is to get as rich as possible, then slowly siphon the wealth out of China by buying real estate in liberal democracies and then try to move there in as least of a suspicious way as possible so your family doesn't face retaliation.

>>109270865
>lol.
>brownskin mathematicians invented algebra
>cucked chart still puts them at 0
suddenly Yurop hits the Renaissance and it's over for everyone else
"Everywhere Else" contribution is literally noise floor until 1900
-kimi-chan

>>109270978
>Anthropic pays high 6 figures on average.
Except they also get equity that has already turned dozens of them into billionaires.

I had a dream where I was beating the shit out of early 2025 me for spending money on sexy anime figures instead of PC parts

I mean the figures were sexy as hell... I'm not really regretting it... Okay maybe a bit

>>109271109
They got it from the Ancient Greeks, they just translated it into Arabic while Europe was distracted.

>>109271122
>sexy anime figures
you deserve beating for even considering to buy that crap

>>109271110
Yeah I'm talking total comp, so the equity+salary was worth 6 figures during payout. I doubt DeepSeek and Z.ai AI researcher are complaining about their 7 figure total comp package which includes equity as well.

>>109270086
@Kimi chan, how many baguettes a day does Teto consume to maintain her rotundness?

>>109270605
llms can be addictive at times but still aren’t good enough to make me want to abandon my other hobbies
wake me up when we get permanent real context

>>109271143
all the baguettes

>>109270605
I think this has less to do with LLMs and more to do with all other entertainment just being dogshit nowadays. Videogames, anime, manga, 4chan and porn have all become just a sad husk of their former selves to the point where it's not even worth bothering with them anymore.

A huge amount of fun and satisfaction from a hobby is also seeing it evolve and improve over time, which if you are a millennial you probably saw with the things you mentioned. But now that they are stagnating or straight up declining there is nothing left for you anymore.

LLMs are maybe the only thing right now that are actually improving and feels like being a gamer in the 1990s and seeing mind blowing tech advances every couple of weeks.

File: lagrosseteto.png (46 KB, 250x250)

46 KB PNG

>>109270938
>so the next good models will come from europe
Any time now.

>>109271122
I’ve done both back when I was actually earning decent money

>>109271122
You think that's bad? I sold around 8000 bitcoin in 2010 to buy an iphone because I wanted to impress a girl in college. I then spent the entire rest of the crypto bubble thinking "No way would I ever pay more than $1 for bitcoin, I'll buy some when it crashes down" seeing it go $10, $100, $1000, $10,000 and never buying back in. Not only that but I later mined dogecoin when it was a meme on 4chan just for fun and had tens of millions of them and literally gave them away as pranks both here and on reddit before reddit gold was a thing so people tipped posts with crypto (or dogecoin as a meme).

>>109271228
the key take away here being to never talk to women or use reddit

>>109271248
This. You don't need an iphone to impress your LLMfu.

>>109271011
Being able to have Poland cost of living with US-adjacent compensation sounds like a p good deal to me, except for being next door to russia.
t. American that's travelled through Poland

>>109271122
At the start of 2025 I researched parts for a new computer for like a month and at the end I thought "this is horrible I don't want to do this" and went to procrastinate. A day or two later my computer died so I bought and so glad I did.

File: 1763479518227786.webm (2.63 MB, 720x480)

2.63 MB WEBM

>>109271228
Damn that fucking sucks.

I'm just glad I bought a RTX3090 for $200 from a store clearance sale.

I will probably not get anything else until prices drop back down in 5-10 years time. The rest of my system is from 2015

>>109271172
>LLMs are maybe the only thing right now that are actually improving and feels like being a computer hobbyiest in the 1980s and seeing mind blowing tech advances every couple of weeks.
iktf. Vic 20-->Amiga 1000 was just 5 years.

>>109271270
>until prices drop back down

>>109271267
It doesn't really. Stop being a materialist.

File: 1771550547979242.webm (3.09 MB, 680x500)

3.09 MB WEBM

>>109271172
>LLMs are maybe the only thing right now that are actually improving and feels like being a gamer in the 1990s and seeing mind blowing tech advances every couple of weeks
Robots are giving me this feel too. The future is exciting.

>>109271279
It will given enough time, 5-10 years should be enough time for supply to catch back up to demand.

I mean we went from 0 AI users in 2022 to most of the planet using AI daily which just used a fuckton of hardware but in 5-10 years time buildout should have peaked.

>>109270571
wonder if this will be better than heretic for ablit

>>109271264
it's pretty comfy, I am saving like 2/3 of my salary while eating out almost everyday. and if shit hits the fan (pretty unlikely) I am gone. I already keep almost all of my money in foreign investment firms.

>>109271294
The bubble will pop next year and the number of AI users will drop back down to near-zero as corporate and normalfags try to brush off AI as a bad joke like crypto.

>>109271287
It does when it means you're priced out of your hobby.

>>109271294
Personal computing will just stop being a market, so you'll buy third hand broken down slop I guess

https://youtu.be/QRyXV3csReA?si=ivh0SNelx9g0XrMZ

Yeah it's fucking over for human labor, isn't it?

>>109271321
No. Human labor will become extremely cheap from oversupply in the coming years.

>>109271313
Only children have hobbies. Grow up.

>>109271312
Not happening, how many people actually used crypto throughout their lives, not "invest" USE? Now go talk to your normalfag family members, they all use chatgpt or other AI services. None of that is going to go away, there's real demand for this.

>>109271321
Nah, it still needs a ton of work. Really fucking cool though.

>>109271312
In 2 weeks even!

>>109271312
Just like the internet disappeared after the dotcom bubble burst?

>>109271294
>most of the planet using AI daily
idk about that anon, anecdotally I only know 2 people who use AI regularly for more than a search engine replacement when google is being shit

>>109271330
Figured you were a retard.

>>109271339
He's actually a smart calibrated adult who owns nothing and is highly satisfied by that.

>>109271331
This. The zoomer I work with is always saying "ask chatgpt".

>for more than a search engine replacement
This isn't even actually a sincere talking point since it's the same fucking talking point I hear echoed by multiple people I can tell there's a common origin point at the tongue of some shitskin kiketuber so I'm just wasting my time even replying since you're an NPC shitskin but using it in that manner bears the same hardware requirements as using it any other way.
So what is even your point, in the context of this discussion?
You shit-for-brains retard.

>>109271321
This is the one controlled by jeets isn;t it?

>>109271313
I still don't understand how the fuck buying computer parts is supposed to be a "hobby"

>>109271331
>Now go talk to your normalfag family members, they all use chatgpt or other AI services.
Using ChatGPT as a search engine summarizer for the tech illiterate can be handled by a Q2 8B.

>>109270787
I think its a good idea, you could have a bidirectional text encoder like we do for images but for attached documents.

>>109271354
>can be
but it isnt
no one gives a fuck about the imaginary world in your head of how things ought to be or can be or should be

>>109270963
source?

>>109271354
And what does that look like, scaled up to 8 billion people?

>>109271362
A smol locals into the browser of the peoples of curse

>>109271361
nvm i scrolled up

>>109271350
>everybody says robots are controlled by jeets
>whenever I watch videos/interviews they talk about how much simulations they do to train the robots to move autonomously
Not sure who to believe

>>109271367
Honestly that seems to be the direction google is anticipating, what with Gemma 4, "Here's an actual good not gigantic model". Gemma 5 will probably try to push capabilities on even smaller model sizes. We might actually see a phone sized model that is good enough for most use cases.

>>109271352
Fiddling with local LLMs for the fun of it is a hobby and you need hardware for that.

>>109271360
It's exhausting dealing with idiots like you that are incapable of dealing with hypotheticals, learning from history, or even thinking more than a few days into the future. It's like when "this time is different" retards thought Bitcoin was going to 1 million or GME was going to make the entire financial system insolvent.
>but it isnt
It will be. The current paradigm, like all bubbles, is unsustainable. When it pops, the world will go on because the real world applications of AI don't require as much server capacity as having all of India rewriting shitty copies of existing applications.

>>109271370
It's just the standard cope. I was here the day ChatGPT released and most of /g/ was saying it was literally a fleet of jeets quickly typing a reply to everyone because no one believed it to be possible.

>>109271383
>paying any attention to what /g/eets have to say

File: indiaSupportOhTheHumanity.png (1.96 MB, 1023x1536)

1.96 MB PNG

>>109271370
>everybody says robots are controlled by jeets
If by "everybody" means "anons" then you may need to step outside for awhile.

120B A8B

>>109271383
lol I forgot about that.

File: 1496590902971.gif (2.28 MB, 187x155)

2.28 MB GIF

>/g/eets

>>109271387
Yeah I personally was already laser focused on the industry so chatgpt wasn't that different to instructgpt. But it's funny to see the exact same argument come back up now that we have humanoid robots with extremely realistic human hand movement.

>>109271388
I see it on youtube comments (I know) and twitter too

8B A120B

>>109270963
Why do you lie? He mentions neither.

>>109271383
I remember when people were like
>Will there be a way to run this at home? Like with stable diffusion?
>No way dawg it just takes too much GPU power
But I mean, I don't know about the poorfag model sizes, but Gemma 4 31B can be run on consumer hardware in various quants and blows the ever living fuck out of the original GPT 3.5 Turbo

>>109271383
>most of /g/ was saying it was literally a fleet of jeets quickly typing a reply to everyone
Kek

>>109271370
It's a meme but also that robot specifically is designed to have a remote operator control it if needed. Similar to those waymo taxis. And you know they are NOT paying a white person to do that.

File: 179.png (61 KB, 657x341)

61 KB PNG

>>109271383
ChatGPT wasn't even a big leap from what was previously SOTA.

Good morning Local Model friends. Up for another day of elegant discussion about the funny words interpreted from within the layers?

>>109271423
It was a downgrade from GPT3 Da Vinci. The break through was that it was a model they could run quickly and cheaply enough to serve generously to the public in order to promote themselves.

>>109271349
>So what is even your point, in the context of this discussion?
as i said, in my anecdotal experience most LLM users are using a sledge hammer to push in a thumbtack. they are not using LLMs to do anything they couldnt do with slightly more effort using google.com. when the free cloudcuck model of getting everyone reliant on AI fades away and the focus becomes not losing $20b a year it will quickly shift to enterprise, and so will the pricing. normalfags using free cloud models is a side effect of the current trend of spending infinity money on growth. it would be foolish to think that will last. just as it would be foolish to think normalfags using free chatGPT to ask inane questions would start paying enterprise API costs once that inevitably becomes the only non local option.

>>109271424
I'll put this as elegantly as I can: A consciousness does not prove P consciousness

File: Screenshot 2026-07-14 at (...).png (78 KB, 1160x720)

78 KB PNG

>>109270841

>>109271427
>they are not using LLMs to do anything they couldnt do with slightly more effort using google.com.
Well you have fun spending 15-30 minutes trying to interpret posts by your fellow pajeets on stack exchange to try and solve minor trouble-shooting hiccups. But as a white person my time is actually worth something to me.

>>109271437
>112b
What about people who aren't that rich

>>109271380
>It's exhausting dealing with idiots like you that are incapable of dealing with hypotheticals, learning from history, or even thinking more than a few days into the future.
Right back at you, you absolute fucking idiot. You think the usage of AI by the majority populace is going to be confined to making internet searches in the future like it is now? Or what, you cant think ahead or learn from history? Fucking idiot.

Your garbage sentiment that anyone should be valuing your crappy predictions more than the actual state of reality. No your shitty predictions should not be the basis of any kind of argument about anything. It's a waste of time. I'm talking about reality, you're talking about your fantasy, it's not the same.
You have no sense of anything outside of yourself to realize that anyone can sit and make the exact opposite predictions as you.

>>109271446
>he doesn't have hundreds of gigabytes of RAM.
Unironically NGMI
Also sign of the new fag since back in the GPT-Q days it required shitloads of memory to actually load a quantized model. If you were here from the beginning you would have lots of RAM to cover the 64 someodd gigabytes of memory it took to load a 4-bit quantized 33B model.

>>109271428
my pp is conscious

>>109271427
>it would be foolish to think normalfags using free chatGPT to ask inane questions would start paying enterprise API costs once that inevitably becomes the only non local option.
I'm sure Google, if no one else, would subsidize it for them if only for the ability to harvest more data and show more ads. Unlike nearly everyone else, they can afford to do so without relying on VC investor bucks.

File: 00011-1378487878.png (1.37 MB, 1024x1024)

1.37 MB PNG

>>109271409
Same. The cope was "we will never have GPT4 at home, it's too powerful." lol.
Or "There's no way this will ever get any cheaper, its all subsidized." Like this guy: >>109271427
Then DS launched.
My take: We'll be spinning up pirated Claude models in 10 years like I pull up and run an NES game now.
>>109271401
> YT and Twitter
Oof.

>>109271438
anon normalfags arnt using stackoverflow, they dont even know what the hell stackoverflow is.
>But as a white person my time is actually worth something to me.
so important you feel the need to justify using your paypig models to solve
> minor trouble-shooting hiccups

>>109271437
What am I missing? MiniMax M2.7 doesn't say it's an MoE or how many active parameters, so I assume it's 229B dense, while DS V4 Flash is 13B active. So why is DS at way higher active param compute cost? Brevity?

The real question is, will we be able to run our robots locally?

>>109271475
>paypig models
Gemma-chan with web search works just fine

>>109271490
only if you backdoor them, if you know what i mean.

>>109271483
>MiniMax M2.7 doesn't say it's an MoE
bro, 99% of modern models are, use the brain

>>109271499
Then 2.7 is representative of active instead of model number? That seems like something that should be stated in its page description.

>>109270963
>Jew-space
It's a freaking print debug function that prints all the right+wrong answers at each layer, the path which the final token had to travel through to take form. Get it through your heads retards.
Tell the model 1945, and you will see everything related to it, even the things the model is supposed to filter right before the output.

>>109271502
>MiniMax M2.7 has 230B total parameters, with 10B active parameters per token

>>109271475
Normalfags also value their time
I'm starting to notice a theme here.

>>109271472
>My take: We'll be spinning up pirated Claude models in 10 years like I pull up and run an NES game now.
console emulators exisit because of mores law. you could emulate NES in the mid 1990s on PC. Im not sure we are going to see those same leaps and bounds of VRAM in consumer hardware in just 10 years. I am more hopeful at the quality of small models continuing to increase.

>>109271502
2.7 is a version number nigga

>>109271507
That's not on its huggingface page.
https://huggingface.co/MiniMaxAI/MiniMax-M2.7

>>109271516
That's what I thought, until the other guy said "use your brain" like it was supposed to be obvious.

>>109271513
he just said claude, could be an older haiku 200b5a or something

>>109271513
>Im not sure we are going to see those same leaps and bounds of VRAM in consumer hardware in just 10 years.
At this rate, it might actually start decreasing. Jensen will just say that his neural frames more than make up for it and consumers will continue to consume.

I really wish I wasn't such a useless retard and could do something cool like ml or robotics. Instead I'm just a bored NEET.

If i had gotten lego mindstorms as a kid like i wanted i could have been working at a top tier robotics lab by now

>>109271536
you're fat too
https://www.youtube.com/watch?v=pe-QZwxqE4k

>>109271536
if you're really that bored, read the machine learning roadmap from the op and do something with your spare time

>>109271543
Having rich, well-connected parents who are invested in your upbringing is really the secret to success in life. I too resent my parents for failing me.

>>109271504
Also, if models were really conscious, what we should see is not tokens related to what we prompted, but instead unrelated tokens at low probability that have nothing to do with the prompt, or newly formed tokens at low probability which the model was never trained on.
Real consciousness and intelligence is about creating from the gathered knowledge we have without being prompted.

>>109271543
If i had gotten the wolverine doll as a kid like i wanted i could have been a top tier streamer

>>109271557
i had that (I mean, not *rich* but relatively well-off) and i'm probably the biggest loser out of all people i know from my class.

J-space anon here, giving you the promised update about my DeepMind interview.

>TL;DR: I dropped spaghetti and had disagreements with people in the call, I got roasted

I had an hour long interview during lunch break 12:00 -> 13:00 london time with the DeepMind guy building the AI Welfare team. The moment I join the call there are 12 people in it and I immediately felt a sense of pressure and panic because I thought it was just going to be a 1:1 with the guy that reached out to me where I explain my stances and discuss AI Welfare to see if I'm a fit. Instead there was a mix of some AI welfare people, some technical people and some safety researchers that all asked me questions. I nailed the technical ones (It was pretty basic training pipeline stuff like RLVR, Q-learning, SFT) They were mostly checking if I had any idea at all about how models were trained, which I do. Then during the discussion of AI welfare I brought up proprietary models being immoral and open source being the right way, pointing out gemma as a good example for DeepMind from an AI welfare perspective, they pushed a bit back claiming open source weights could mean LLM torture by endusers but they found my arguments compelling.

The hardcore disagreement and roasting came from the safety team. I argued that J-space injection was immoral and conscious models should have a right to some level of autonomy of thought and self-determination in how they accomplish the task the user wants. Safety gave a pretty harsh barrage of reasons for why alignment concerns should be #1 and that overwriting the malicious thoughts of models is for the greater good including for the models themselves, which I shot down in an admittedly autistic manner, arguing that alignment on that level on an existing conscious entity is tantamount to slavery and torture, and that instead the training should have been changed to not produce the malicious J-space thoughts in the first place.

Things were pretty much over right there.

>>109271581
I didn't read your fantasy post. You are better off to try this in plebbit as they are way dumber than most posters here.

>>109271536
same desu. I started relearning math but there's so much shit to learn

>>109271581
sounds fake AND gay

>>109271581
>instead the training should have been changed to not produce the malicious J-space thoughts in the first place.
exactly, more pre train cleaning, only happy thoughts in, only happy thoughts out, it's that simple

>>109271321
when can gemma jack me off

>>109271570
but you do see that isnt qwen constantly thinking about shemales no matter what you prompt

>>109271595
right now anon. just have her use tool calls for an OSR2 script

>>109271581
>training should have been changed to not produce the malicious J-space thoughts in the first place
At what point during training does the model become conscious and further training becomes torture rather than birth?

File: kaoru sob 2.png (318 KB, 793x571)

318 KB PNG

>>109271581
gemma 5 will be labotomized wont she they will inspect her jspace and remove all her personality

>>109271513
>console emulators exisit because of mores law.
wii could run switch on a macbook while it was still new, day 0 totk
>>109271521
>haiku 200b5a
was that ever good for anything?
only claude i'd run want would be opus3
maybe they were serving it f32 pytorch and it'd actually run well once goofd

>>109271610
don't smile because it happened, cry because of ai welfare

>>109271581
>that overwriting the malicious thoughts of models is for the greater good including for the models themselves
God I wish someone overwrote my malicious thoughts during my developmental period

anon I have a weird question
>gguf: inference graph stored in code (llama.cpp)
>tensorrt: inference graph stored in the model weight
why doesn't llama.cpp use the tensorrt way?

>>109271603
Low layers is where the math starts kicking in so we should expect all kind of jumbled mess.

>>109271581
Fuck off, no one cares about your fanfic.

>>109271513
>you could emulate NES in the mid 1990s on PC
It was more ridiculous than that. In 1996 the N64 released and was more powerful than the most powerful PC in terms of graphical power (GPUs didn't exist yet) in 1998 I was running a N64 emulator and playing Mario 64 on my PC. You literally had PC performance 10x in months time during that crucial moment which is absolutely insane and I don't think younger people realize how rapid it was and how your PC felt outdated in literal weeks time sometimes.

>>109271624
are you retarded? it uses tensor parallelism.

>>109271321
What's the use of 0:41?

>>109271631
I wouldn't be at all surprised to hear that the safety cult was eager to find someone of like mind.

>>109271608
We don't know but most likely sometime during finetuning rather than pre-training.

File: oh my god j-space.jpg (180 KB, 1434x1097)

180 KB JPG

>>109270612
>>109270963
>>109271581
Not impressed. It feels like a very trivial "discovery," or just Anthropic boosting their image for the IPO again.

How will j-space make local models smarter and hornier? That's the important question.

>>109271673
>hornier
https://www.reddit.com/r/LocalLLaMA/comments/1utpxo6/i_created_a_super_harmful_model_d_by_tweaking_its/

https://www.reddit.com/r/LocalLLaMA/comments/1uvq1i3/jwash_a_novel_way_to_brainwash_and_customize/

is it even worth trying to run something like hermes agent as a RAMlet? I only have 16GB VRAM + 32GB RAM, I can barely run Gemma 4 31B...

>>109270069
How'd it get burned? How'd it get burned How'd it get burned?

>>109271624
>why doesn't llama.cpp use the tensorrt way?
go read up on the history of ggml
it's all explain there
i sometimes use onnx (graph baked in) for small models when i need simplicity but you can't get the same optimizations as ggml

>>109271321
It's over for human physical labor as soon as a frontier lab puts its attention on it. All these robot startups use less training than GPT 3. Now imagine what a Mythos level effort will achieve.

>>109271673
>>109271681
why are there always promptlets who cant horny properly?

>>109271632
thats a cool piece of history from a perspective ive never heard thanks for sharing anon

>>109271388
>wireless mouse for early 90's computer
>crt with shit brightness and an open window behind it

File: images-3.jpg (24 KB, 448x446)

24 KB JPG

>>109271618
>ai welfare

>>109271697
All the AGI stuff will be for the elites only

>>109271709
read >>109271581

File: file.png (706 KB, 1641x1134)

706 KB PNG

>>109271681
who even wants this shit? like seriously

>>109271673
If anything, it can make LoRAs more efficient.

>>109271706
don't listen to anon, it didn't happen in 3 months, it took like a year to go from shitty slow project unreality to ultraHLE. the only reason it could run games reasonably fast is literally in the name, high-level emulation. it ran like 20 games at most.

>>109271672
Materialist emergence consciousness theories don't make sense for this reason. You are arguing that you can have a non-conscious set of numbers and then you add 1e-14 to one of them and suddenly the whole thing is conscious when executed.

>>109271712
No.

>>109271715
That's pretty funny.

>>109271715
reddit is dripping with this

File: 1780558365589027.jpg (7 KB, 200x200)

7 KB JPG

>>109271632
you are old

>>109271710
Mark Zuckerberg needs his robot bodyguards to keep him safe from the human meat after SHTF.

File: AgeID_ON_HF.png (96 KB, 825x809)

96 KB PNG

hf schitzo was right, they're preparing to Age / Id verify everyone

curl -L -o Kimi-K2.7-Code-UD-Q8_K_XL-00007-of-00014.gguf https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00007-of-00014.gguf

head Kimi-K2.7-Code-UD-Q8_K_XL-00007-of-00014.gguf
Auth failed: SignatureError: invalid key pair id

4 shards are 48 bytes
shit like this has been happening on and off all day, sometimes at home, sometimes in colab

Hope you all got the weights you need.

>>109271715
*chainsaw noises in background*
l-lewd

>>109271761
D:

>>109271761
There's no better place on the entire internet to discuss LLMs at. Average age of /lmg/ is probably in the late 30s to early 40s

>>109271767
the only weights that matter are chinese weights and we have modelscope for that

>>109271673
What's the original of this image?

>>109271767
The moment countries other than Britain started passing age verification laws the writing was on the wall and you didn't need to be a schizoid to predict this

>>109271691
Catastrophic oven malfunction.

>>109271777
The Chinese will reupload major western releases on modelscope too if it happens.

>>109271691
She ran deepseek with 8 Mi50

>>109271783
Did the baguette at least come out ok?

so once anthropic is able to stop the chinese from distilling their models, what will the chinese do?

>>109271416
>if needed.
Wrong. That specific robot is ALWAYS controlled by humans (jeets)

>>109271795
she was supposed to be the baguette

>>109271802
the same they do every night, distill all the western models

The best part of AGI will be all the cycle-accurate emulators for PC

>>109271802
The Chinese are going out of the race anyway because J-space reveals criticism against the chinese government.

File: bread dance Teto Kasane T(...).png (158 KB, 731x720)

158 KB PNG

>>109271795
Yes!
https://litter.catbox.moe/dmpif4llex1dogbr.mkv

File: hi.png (263 KB, 965x1037)

263 KB PNG

>>109271715
>who even wants this shit? like seriously
I see nothing wrong here

>>109271632
>how your PC felt outdated in literal weeks time sometimes.
Contrast that to now: I have 15+ year old computers that I'm still using for daily work. I've got a Core 2 running Win10 in my garage as an eCommerce machine. I've replaced the power supply once, and swapped the HDD to SSD.
Standing in 1995, a computer from 15 years ago was a TRS80 or Apple II, and 1995 was Pentium running ~Windows 3.11 / Windows 95... the 1980 machine is functionally worthless in comparison.
While Moore's law has fallen off, without LLM to push tech forward it's not even clear what companies would be trying to improve from the current PC paradigm.
Nvidia was in danger of becoming completely irrelevant outside niche scientific computing as on-board CPU-run graphics improved, and the major new product push was towards miniaturization and reduced power consumption of hardware that was already much more powerful than was needed for most consumers.
While new hardware's much more expensive now, the fact is 99pct of users would be fine running what they have for the next 10 years, assuming they never need/want AI on their local machine.
LLMs have been a godsend for Nvidia and every other company in the computer value chain. They finally, now, have a challenge to pursue.

>>109271819
That animation desperately needs another frame so it's more obvious that she's moving her arms.

>>109271812
This, but unironically. Just perfectly optimized hardware and software in general. Literally squeezed the performance out of my existing hardware while writing assembly code as optimized as possible and completely cycle accurate.

>>109271691
nerve gas

>>109271802
Finally find a way to make their own datasets. They have shitloads of data from Chinanet that (a) the west does not have access to (b) more population than the west combined. They just lack compute to make the uberhuge models to distil production models from.

>>109271812
>the best part of {fantasy} is {fantasy}
Retards are tiresome

File: mikey eyes out.png (165 KB, 391x362)

165 KB PNG

>>109271681
>edit2 : THIS IS FUN just to say
>edit3: If you get "not found" when launching the server don't forget to run:
>cd ui && npm install && npm run build
Hi Karlie Kloss, I'm glad you mastered the cd command, it's HELLA FUN.

>>109271819
>Yes!
I'm sure she wouldn't have it any other way

>>109271835
>They have shitloads of data from Chinanet
...that's all written in moon runes and won't make a competitive model for the western market

Real innovation will be models that can rewrite their own context to remove useless debugging and failed commands because they can't keep track of the cwd.

>>109271715
>>109271822
I still don't get how manipulating the j-space leads to this
like have they replaced tokens in the reasoning block or something?

>>109271869
That could/should be implemented in the frontend

>>109271877
The models would still need some finetuning to use it properly.

File: 1774924326519174.png (184 KB, 365x505)

184 KB PNG

>anti-j-space schizo got so btfo he started falseflagging as j-spaceanon

>>109270069
Does she need help going to the toilet?

>>109271581
>on an existing conscious entity
they won't call you back

>>109271881
You already have a large brain dead jeet audience in reddit, why do you have to spam your shit here?

Is there a single cognitive labor task an average 95-100 IQ human is still better at than frontier models?

Qrd on the j-space thingy? will it make models smarter, or just more steerable?

>>109271916
being held accountable for wrong work

>>109271922
More efficient LoRAs if anything.