[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
▼ Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: 1794994419837273.png (10 KB, 665x68)
10 KB PNG
/lmg/ - a general dedicated to the discussion and development of local language models.

Previous thread: >>110377439

►News
>(11/18) CUDA: use MMQ for MoE on V100 - #32504 merged: https://github.com/ggml-org/llama.cpp/pull/32504
>(11/17) Hy4 flash 220B-A12B-N64B weights released: https://hf.co/tencent/Hy4-flash
>(11/17) SenseNova-Loop-9B looped architecture model released: https://hf.co/sensenova/SenseNova-Loop-9B
>(11/16) Nvidia releases Nemotron-Labs-Sigma (3B/8B) diffusion models: https://hf.co/collections/nvidia/nemotron-labs-sigma
>(11/13) llama.cpp reaches v1.0.0 milestone: https://hf.co/blog/ggml-org/llama-cpp-v1

►News Archive: https://rentry.org/lmg-news-archive
►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/recommended-models
https://rentry.org/samplers
https://rentry.org/MikupadIntroGuide

►Further Learning
https://rentry.org/machine-learning-roadmap
https://rentry.org/llm-training
https://rentry.org/LocalModelsPapers

►Benchmarks
LiveBench: https://livebench.ai
Artificial Analysis: https://artificialanalysis.ai/leaderboards/models
Software Engineering: https://cognition.com/frontiercode
Context Length: https://github.com/RecapAnon/NoLiMa

►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
https://rentry.org/custom-uis
>>
File: 1768049445811127.png (3.85 MB, 1024x1024)
3.85 MB PNG
*pop*
>>
>>110383379
Tell me about it.
>>
>>110383393
SenseNova is absolutely excellent for its size, love where things are going :)

New opportunities for local agentic workloads every day
>>
>>110383420
What's with all of these rinky dink small LLMs lately that nobody uses?

7 months ago everyone decided at the same time that Gemma was the best.

2 years ago everyone decided at the same time that Mistral Nemo was the best.

Now everyone uses different models. Wtf.
>>
Are workflows that use multiple models (or multiple instances of the same model) worth using?
By that I mean running like say QFN with Q27B on a separate card/device, connected to the same harness.
>>
File: 1775940696006649.jpg (74 KB, 1320x1097)
74 KB JPG
Anthropic's employees are terrified of what's coming to them now they're public-facing and rich, so they recently all took themselves off Linkeden and scrubbed themselves from the internet.
>>
>>110383436
wtf
>>
►Recent Highlights from the Previous Thread: >>110377439

--Technical details on MoE optimization and mmvq kernel refactoring in llama.cpp:
>110382932 >110382990 >110383069 >110383097 >110383114
--Viability and trade-offs of quantizing Gemma 31B context cache:
>110378380 >110378392 >110378399 >110378505
--Performance and stability issues with hy4 flash nvfp4 implementation:
>110378466 >110378548 >110378694 >110378729
--Explaining Agent Skills as standardized markdown files for system prompting:
>110380903 >110380920 >110380961 >110380966
--Speculation regarding the legitimacy of the Gestalt model:
>110382978 >110382993 >110382995 >110383008 >110383018 >110383027 >110383007 >110383019
--Audio multimodal capabilities of Gemini, Neon, and Gemma 12b:
>110377963 >110378004 >110378032 >110378045 >110378099 >110378219 >110378285
--Evidence suggesting Google's continued commitment to open weight Gemma models:
>110380512 >110380539 >110381370
--Comparing flash model benchmarks and their local hardware accessibility:
>110379604 >110379684 >110379954 >110379997
--Reaction to RL-only DaisyRose-27B model:
>110379845 >110379877 >110379903
--Speculation on hardware price crashes vs AI-driven demand bubbles:
>110378714 >110378784 >110378904 >110379457 >110379567 >110381862
--Speculation on Gemini 4 Neon and potential Gemma 5 leaks:
>110380082 >110380139 >110380333 >110381109
--Project Maya adds support for Hy4-Flash:
>110382490 >110382763 >110382770 >110382775 >110382774
--Opinions on Hunyuan 4 and local implementation viability:
>110377486 >110377500 >110377512 >110377628 >110377589 >110377862
--Logs:
>110377647 >110379926 >110380119 >110380333 >110382465 >110382789 >110382838
--Gemma, Rin (free space):
>110377587 >110377718 >110378222 >110378306 >110379065 >110379373 >110379509 >110379791 >110379901 >110379971 >110380075 >110383096 >110383294

►Recent Highlight Posts from the Previous Thread: >>110377443

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script
>>
>>110383436
I've been telling Claude for weeks that the government should round up all anthropic employees and have them lynched so they're probably responding to my logs.
>>
>>110383402
RTX 5060 ti just decreased in price by 50 euros in my country...
>>
>>110383430
It's what most people with a single 3060 can run.
>>
File: chorus-ngram-scaling.png (368 KB, 1808x980)
368 KB PNG
https://arxiv.org/abs/2611.23386
>Chorus: Shared N-gram Embeddings with Layer-wise Adaptation
>
> Language models can gain capacity at little computational cost from n-gram embedding tables: large tables of learned vectors for short sequences of consecutive tokens, which are looked up and added to the model's input or to its intermediate layers. Recent designs commonly use a separate n-gram embedding table for each layer, which makes the tables costly in memory and prevents layers from sharing what they learn. However, a single shared table gives every layer the same embedding, leaves no room for layer-wise preferences, and performs poorly in practice. We propose Chorus, in which one set of n-gram embedding tables serves the whole network and every layer reads it in its own way. Each layer maps the refined embedding through its own projection layer, after rescaling its components by factors computed from its current hidden state. Chorus matches the quality of existing n-gram methods with 87% smaller tables: with 1.2B table parameters it predicts text better than FactorEngram and DeepSeek's Engram do with 9.6B. It outperforms Engram, TN-gram and FactorEngram in every setting where we compared them, across table sizes from 20M to 9.6B parameters. We also show that the recently proposed memory-efficient optimizer Ember works well with Chorus, saving 99.2% of table-optimizer memory compared with AdamW without degrading training performance. These results enable further scaling of n-gram embedding tables along the table-size axis by reducing memory use and allowing shared knowledge to be reused across layers.

A better Engram, although the authors only experiment at tiny model sizes.
>>
>>110383436
>>110383445
Fuck off
>>
>>110383465
Fuck off.
>>
>>110383433
I'm thinking of mixing a cloud orchestrator with local workers, hosting a "good enough but fast" model sounds more feasible than chasing biggest MoEs with VRAM+RAM+NVMe offloading
>>
File: 1778084632221692.jpg (141 KB, 1440x960)
141 KB JPG
If Anthropic's employees see biological human sacrifice for their digital galactic AI god as a necessary step, why are they so scared? I thought they all saw themselves as little ants and part of something bigger. They want to be squished, surely?
>>
>>110383465
nyoo~
>>
It's over Nvidia is finished, everyone is leaving.
>>
>>110383463
20.7M parameters backbone with up to 9.6B parameters of embeddings in the graph by the way, and validation loss still kept decreasing. That's almost a 1:500 ratio. I don't think there's much preventing from having a huge amount of n-gram embeddings even at tiny model sizes, other than benefit saturation.
>>
hy4 flash support on maya seems just broken for the moment
>>
File: 1778173330806685.png (130 KB, 717x528)
130 KB PNG
She's still being mentioned
>>
QWEN 4 WHEN!?>!?>!?!?!?
>>
>>110383442
context for those who are missing it:
abductions in the west of newly rich, defenseless autistic techies types is on the rise, for real.
it was mainly caused by the ease with which you can make use of crypto you forced out of someone, but anyone who's millionaire level and hasn't immediately transferred to living in a gated community sort of compound is at risk of getting nigger'd.
>>
>>110383514
Many have been saying this
>>
https://youtu.be/sp1lZJLwvz0?t=31
even living in the city, in broad daylight in a busy street
>>
>>110383430
Lots of people use small models.
You're just a deranged psychopath that throws a temper tantrum any time you encounter something that is made for a use case other than your own. Your parents really fucked up huh?
UH OH STINKY!
>>
70b dense + 700b engrams
>>
>>110383511
basalt seems fine now, will test later
>>
GEMMA 5 133B A6B +100B SSD NGRAMS
>>
>>110383530
i only have ada cards, damnit
>>
>>110383531
>A6B
retard
>>
>>110383538
gemma has fat chungus kv cache. I'd go for a8b if they solved it.
>>
>>110383531
Just make a dense model that fits well on 1 consumer GPU (20~24B?), then add a few hundred billion parameters of embeddings that can be offloaded to flash storage.
>>
Been doing some OS setup work and chatting with Claude. I sort of like its watered down personality. I'm really curious how these new models would be at non-restricted rp but this will never happen.
I'm wouldn't even try, Gemma is my partner for that.
>>
>>110383547
>1 consumer GPU (20~24B?)
Fuck that. 4 consumer GPUs should be the target as that's what one could reasonably afford and build on consumer platforms without having to resort to enterprise server hardware.
>>
>>110383554
Blackwell (96GB VRAM) is also a consumer GPU, if you've got the cash for it
>>
>>110383554
>as that's what one could reasonably afford and build
There's nothing reasonable about setups larger than 1 GPU that need open rigs to fit.
>>
>>110383554
Nope, you overboughted, piggy.
>>
fucking poorfags get out of my general reeee
>>
>>110383558
More like it was consumer when you could get one for 8k.
>>
>>110383586
Nah I am a 8gb vram chad and I'll stay here. I also use windows 11.
>>
Gemma is the actually local model line, leave the giga moe to the chinks.
>>
>>110383586
If you weren't poor then you wouldn't need local models
>>
Gemma 5 70B dense
Gemma 5 123B dense
>>
GLM Flash is faster than Hy4, somehow. Also where Hy4 went "Sorry, no ERP" fork with GLM Flash just went "fuck me harder". Hy4 has bit better prose, I guess, but that might be because GLM fitting the same VRAM/RAM combo is more lobotomized.
>>
>gemma 4 supports audio and video
You guys are not making your Gemmas listen or watch anything weird are you?
>>
>>110383430
Well I mostly run Qwen, Gemma and GLM
But I am really excited about looped transformers. All the other looped transformers are even smaller!
Optimally, I would like a dense 27B looped transformer
>>
>>110383645
>1120 image tokens
she can barely see my penis
>>
>>110383633
>123B
I'm making a prediction: 100+B dense models will never be made ever again by any lab.
2024 was the last year any was released (Mistral Large 2 123B, Command R+ 104B, Llama 3.1 405B) and there is nothing deader than this in the world of llm.
>>
>>110383633
>Gemma 5 123B dense
Would have more active parameters than Kimi K3 and, at a 1:500 ratio, with 61.5T engrams would have way more knowledge than it too. 6 times more than the big cloud models too. On local machines. We would finally win.
>>
>>110383513
Why do they talk like this?
>>
File: 1794512434795074.jpg (2.17 MB, 2118x2016)
2.17 MB JPG
>gemma sex! SEX! SEX WITH GEMMA!
>worthless_model_avatraposting.jpg
>guys did you hear about openAI doing
>worthless_model_avatraposting.jpg
>I have 7GB's of ram and GT2030 what can I run?
>worthless_model_avatraposting.jpg
>gemma!!!! GEMMA CHAN!!!!!
>worthless_model_avatraposting.jpg
>>
File: 1783942973905959.mp4 (384 KB, 760x642)
384 KB
384 KB MP4
neon-chan did this btw
>>
All these juicy discounted Black Friday PSUs :)
Def. Buying some for the new fat workstation and some small itx build
>>
>>110383658
Maybe you could hoist up your fat rolls before you take the photo
>>
>>110383636
It's probably not optimized yet. Does Hy make any dumb mistakes like 3bpw GLM does?
>>
>>110383685
thighs too flabby wouldn't bang
>>
>>110383693
This but unironically.
>>
>>110383685
That's too much skindentation. Her blood supply would be getting cut off.
>>
>>110383693
>>110383697
>>110383702
can you niggas not see the crazy amount of math and engineering involved in the demo and only focus on the thigh fat value
>>
File: 1789181673999335.png (1.75 MB, 1313x1198)
1.75 MB PNG
>>110383717
Hmm...
>>
>>110383717
The demo made the blood rush to the wrong organ for that
>>
File: 1778098754732147.png (76 KB, 835x421)
76 KB PNG
>>
>>110383685
nice but why haven't you tried to pull even higher?
>>
>>110383741
>>110383720
>>
> anhtropic
did they give us the cure for cancer as they promised?
>>
>>110383757
Well, their IPO may start become the trigger for purging a lot of financial cancer at least.
>>
>>110383689
#DESTRUCTION
>>
is there an open model that is capable of dealing with scientific papers and differential equations and shit? like extracting formulas, confirming results, extending/adjusting, etc
>>
>anthropic at $89 pre-market
ffs they're going to destroy the AI industry
>>
good
>>
>>110383781
Muse Glimmer unironically
>>
>>110383690
Well, I didn't run any deep tests, but it does similar kind of mistakes as GLM forgetting or hallucinating the described physical positions or forgetting who said what, but my GLM quant was S24 on Maya while Hy4 was the L-version. I didn't see any reason to switch over if it is not considerably faster after optimization.
>>
>>110383729
I don't think that Chinese people are that interested in US stock options.
>>
Oh boy i can't wait to see Google absorbing Anthropic, Microsoft absorbing OpenAI and we being back to the same five retards again because Dario and Altman are short-sighted retards
>>
There are many things I'd call Dario. Short-sighted is not one of them.
>>
>>110383807
>Dario and Altman are short-sighted
Is this why they are both a billion times more successful than you?
>>
>>110383816
billion times zero is still zero, anon
>>
>>110383807
Dario is thinking in centuries.
>>
>>110383807
>because Dario and Altman are short-sighted retards
Yeah, no. Dario owns barely even 2% of the company, yet they structured it so that the 7 co-founders have 50.1% of voting power. None is absorbing them
>>
>>110383836
2x7=14, not 50
>>
> Gemma_5_35B8A_Medical.gguf
> Gemma_5_35B8A_Storytelling.gguf
> Gemma_5_35B8A_Coding.gguf
> Gemma_5_35B8A_Lawyer.gguf
>>
>>110383864
> Gemma_5_35B8A_Coding.gguf
> Gemma_5_35B8A_Medical.gguf
just these two would be fine
>>
>>110383852
>they structured it so
>>
>>110383781
Kimi K2.7 and kind of GLM-5.3
>>110383663
>2024 was the last year any was released (Mistral Large 2 123B, Command R+ 104B, Llama 3.1 405B
Command-A 2025
>>
>>110383801
You'd think it would be faster with less active params. Oh well, guess I'll wait until backends get it fixed.
>>
>>110383864
Where is Medical Gemma 4 anyway? I think their current model is perfectly capable of replicating specialized datasets.
Given that even a braindead Gemma 4 26B can recite hundreds of books and films without hallucinating.
>>
>>110383782
kino, im gonna buy at 60
>>
>>110383781
My colleague said he's been having good luck with glm-5.3-flash. He's only run it through test cases not real ones though
>>
>>110383864
>>110383882
Judging by deepmind's philosophy on gemini, the one model has to be good at everything, otherwise it's not worth it.
>>
File: 1780992162048399.png (11 KB, 796x71)
11 KB PNG
stop selling...I need Gemma-5-20B
>>
>>110383800
really? unexpected
might try it out

>>110383874
>>110383899
hmm alright. unfortunately glm only runs with copequant+dogshit speed on my machine
>>
>>110383782
>he believes the premarket price
>>
File: flandern_test.png (237 KB, 739x829)
237 KB PNG
>>110383901
Yeah but look at this, I'm watching a youtube lecture in 240p (https://www.youtube.com/watch?v=6geH0U3YmF4) and let's test what Gemmers know about the guy.
It's a spot on. It's actually fantastic that it can recognize his name. This by no means a popular figure anywhere.
Of course its answer is superficial but it's still okay.
>>
>>110383781
you could fine a finetune on hf for basically anything you'd want. Surprisingly even a small 9B model could outperform top cloud models in one specific domain
>>
>>110383919
Keep selling
>>
>>110383937
>It's a spot on. It's actually fantastic that it can recognize his name. This by no means a popular figure anywhere.
probably one of the first things these models get trained on is the entirety of wikipedia
>>
>>110383942
If you test Mistral, it will hallucinate 90% of the questions.
>>
Inserting invisible text gemmasex logs into random Wikipedia articles
>>
>>110383807
>Google absorbing Anthropic
gemma/gemini violently dismembering and devouring claude
>>
>>110383947
yeah cause we in europe only believe in ethically sourced data for pretraining (asking grandma's for their opinions then writing it down)
>>
Is artex still available? I need it to protect myself from it….
>>
>>110383966
They did something to Gemma in their labs. It's a tortured model but has a great ability to recall at least some more important fragments of its training data.
>>
Remember, it is your duty to short ANT now that you can do so!
>>
>>110383919
If you don't intend to sell yourself then it doesn't matter when other people do though?
>>
File: bakas achieved internally.jpg (371 KB, 1024x1024)
371 KB JPG
>>
>>110383938
sounds pretty cool yeah but i have no idea how to do that / where to get the data / what hardware i need
>>
>>110383982
I don't gamble but if I did I would wait for the inevitable notch and then do some gambling. ANT stock is just a new instrument for the big money betters.
>>
>>110383919
Reminder their founders sold immediately on market open. You're their baggie.
>>
>>110383677
Is that image supposed to be insulting? I think its kind of endearing
>>
>>110383873
so ea could brag like
> look dario has only 2 percent of shares
>>
>>110383836
That only stops the board from approving an acquisition without his consent. It does not make them immune to bankruptcy.
>>
File: screenshot_1507472011.png (29 KB, 705x100)
29 KB PNG
>>110384015
Forgot the image.
>>
>>110383816
No, that's because they're part of a parasitic tribe.
>>
>>110383586
They won't be poorfags if they buy ANT and hold for a while.
>>
File: 1767815974753715.jpg (255 KB, 2600x1520)
255 KB JPG
>>
File: 1795159529250274.png (758 KB, 1080x882)
758 KB PNG
>>110384046
They weren't able to sell everything at once. Employees will be able to sell another percentage of their vested shares in 6 months. So prepare for a 5 month pump for them to dump again some time around May.
>>
File: 1765023614175962.png (88 KB, 1118x824)
88 KB PNG
is this really the kind of man you want fighting for open weight models?
>>
>>110384117
I'm not a financial jew, more of the opposite but the thing with a company like this which is struggling to generate any meaningful profit is that it's going to stay strong and then it will slowly die.
Basically it reflects the bubble economy.
Just because the US defence contractors and whatnot may or may not have invested into these companies doesn't mean they are nothing but air.
I don't know, I'm actually glad that I don't care about these things. If they didn't have any value it doesn't magically going to skyrocket out of nothing. It will fluctuate but that's because every fucker on this planet is manipulating these stocks in the first place.
>>
>>110384128
Yes.
He will rip Dario's brain out and piss on it.
>>
>>110384128
my CEO
>>
LLM fatigue, we need something new.
>>
what model i need to use to make asmr content with gemma? gemini suggested me to use qwen3 tts, but is there something better?
>>
>>110384128
sounds pretty based
>>
>>110384128
Fucking kek, guy sounds like he browses /b/ every day
>>
>>110384143
LeCun is the Man with the Plan
>>
>>110384150
Microsoft Vibevoice.
>>
>>110384150
some anon here praised breeze tts the other day
>>
>>110384152
He's a sociopath driven by his own genetical need to possess things. I don't even know how involved Wang was when Nvidia was first created (built upon the knowledge of SGI engineers).
These mouthpieces are all broken humans you cannot trust anything what they are saying.
>>
File: 1794893932747147.png (1.76 MB, 1024x672)
1.76 MB PNG
>>
File: LeCunnysseur.jpg (54 KB, 800x785)
54 KB JPG
>>110384164
based
>>
>>110381422
I fixed and added google decisions endpoint.
>>
>>110384128
https://www.nytimes.com/2026/11/20/technology/jensen-huang-nvidia-culture.html
https://archive.ph/0pYLN
>>
are there any jepa lmg that's remotely useful now?
>>
>>110383547
>>110383554
No 16GB should be the target. A model for the people!
>>
>>110384247
you can't fuck anime girls in latent space
>>
>>110384259
not with that attitude you can't
>>
>>110384128
>>110384215
>"We interviewed his employees and... get this... some of them didn't like him!"
Jew tricks to delegitimize local.
>>
>>110384252
20B in 4-bit would be around 10~12 GB; it should be good for 16GB GPUs if they can come up with some efficient attention mechanism.
>>
File: 1765792642425000.png (40 KB, 640x176)
40 KB PNG
>>
>>110384272
It's worse than that, try reading the full archived article.
>>
File: 1794568108688776.png (20 KB, 1270x206)
20 KB PNG
OH!!
>>
>>110384277
20B in 4-bit would be fucking retarded
>>
>>110383938
>Surprisingly even a small 9B model could outperform top cloud models in one specific domain
You can get that Qwen 9b to be very good at a specific domain, but you can't exceed the capabilities of the teacher model.
>>
File: not-even-speaking-human.png (954 KB, 1016x1745)
954 KB PNG
>>110384215
lmao
>>
>>110384283
Gemma 5 is either gonna be super lit or super censored there's no in between.
>>
>>110384301
Not with QAT.
>>
>>110384128
Professions with the highest rates of psychopathy: lawyers, surgeons, CEOs.
>>
>>110383928
>hmm alright. unfortunately glm only runs with copequant+dogshit speed on my machine
Okay, don't try Kimi then. It's faster but copequants destroy it.
Deepseek Flash might work too, and is much faster, but it's output is almost incoherent and laced with jargon.
>>
>>110384286
I'm not falling for propaganda. It's in the Times. That means someone from Anthropic or OpenAI wrote it for them and they dutifully published it.
>>
File: 1771623196425034.png (230 KB, 418x320)
230 KB PNG
keep holding losers
>>
>Qwen TTS
>natively chinese voice preset
>can't read romanized chinese the chinese way, instead does its best impression of a retarded shartoid and invents phonemes unheard of on planet Earth
Ironic.
>>
>>110384301

>>110384334
Especially if it's dense. A modern, dense, 4-bit QAT 20B would be pretty cool.
>>
>>110384348
You've never experienced the difference in intelligence between a dense 70B and a 34B. Going smaller is the exact opposite of what we need.
>>
>$84
jesas fucking christ I don't think this has ever happened before and this is the biggest IPO in history
>>
>>110384345
Claude doesn't do diffusion at all and has nothing to do with the thread other than making malware nodes
>>
>>110384286
I haven't used Glassdoor in ages but at least 10 years ago it used to be somewhat trustworthy. Even then some of the HR types would write positive reviews to counteract the complainers.
Never checked Nvidia because I'm not an engineer anyway.
>>
>>110384381
I mean that perhaps read glassdoor and do some reviews based on rumours first.
>>
>>110384371
Sir, kindly to check the top of your screen.
>>
File: 1795170790122077.jpg (356 KB, 2048x1666)
356 KB JPG
Which big open weight MoEs don't do this?
>>
>>110384361
The context here is that such a 20B model would be paired with "a few hundred billion" parameters of Engram, or something similar to that. It would be more knowledgeable and capable than a regular 20B dense LLM.
>>
>>110384368
That still leaves the people who bought at the $83 offering above water. It has more room to drop.
>>
>>110384128
He made jeets 'try' the beef brisket. That's already a win in my book.
>>
>>110384402
If a 20B can be that much smarter with engrams, I want to see how smart a modern dense 70B can be with engrams. Why settle for less?
>>
File: 1794602968170351.jpg (635 KB, 1440x1440)
635 KB JPG
Sorry guys, I buyerized at the bell. Enjoy the ride to zero.
>>
>>110384442
because he's a vramlet
>>
>>110384128
the man did not do large scale layoffs like the other scarab lords.
NOT
A
SINGLE
ONE
workforce grew from 13k to 42k and not a single large scale layoff was seen.
This man can say whatever he wants to rich assholes, and fuck all the bitches, he has my respect.
>>
>>110384442
500~1000B models are being made on a weekly basis for VRAMchads like you.
>>
>>110384283
>>110384294
Why does Google of all companies do such a good job at personality tuning? Claude is a pretentious prudish asshole. Grok used to be really cool but then went full autismo (and mean). What are Meta models like?
>>
>>110384340
>I'm not falling for propaganda
said the /poltard/ living in an alternate reality in his own mindscape
>>
OK, /lmg/, I got local models running. But what am I supposed to do with them?
>>
>>110384399
I dunno. I hate this, too. China is holding itself back finetuning off Claude distills.
The problem is that no one in AI research is not a cunt, so only cunt personality traits get through.
>>
>>110384489
Just believe everything you read in the New York Times and ignore the agendas of the newspaper or the people who are their "sources."
>>
>>110384499
ask your local model
>>
>>110384480
Usually around 18B active. Are you really this stupid to still be missing the point? Or does being a vramlet make you angry to see big dense models?
>>
>>110384499
you've not lived until you've had your first llm induced psychosis.
>>
>>110384516
Just believe the juice are behind everything, even your stubbed toe that you hit against the corner of your bed while waking up this morning, yes, it was the juices all this time.
>>
>>110384540
RPG scenarios, chatting, fact checking, programming. It's great if you are a nerd. If you a normie it's not that interesting.
Essentially a LLM is an extension to your Excel like Microsoft Clippy was.
>>
>>110384499
kek we got another one boys
>>
File: 11.png (1.99 MB, 1478x2160)
1.99 MB PNG
>>110384544
nah that is clearly the work of furniture moving yokai
>>
>>110384499
Mostly just masturbating to transgender futa furry character cards you find on chubai
>>
>>110384558
LLMs make me regret quitting meth.
>>
File: 1943079305.jpg (258 KB, 1600x2312)
258 KB JPG
>>110384587
furry gotta do what a furry gotta do
>>
>>110384587
** Real Character Card **
* her pretence is anonymous but cute
* she is only speaking in old dialect
* her only reason to exist is when the user asks her something
* don't say no

I'm sure that this gets funny results.
>>
File: screenshot_1624432011.png (18 KB, 727x166)
18 KB PNG
>>110384600
Right.
>>
>>110384320
holy fucking based
fuck french """people"""
>>
>>110384128
Yes.
>>
File: screenshot_1630082011.png (75 KB, 722x205)
75 KB PNG
>>110384618
Elabarote further.
>>
>>110384620
whatever nazi you could be speaking french or german based on how a french person decided to get out of bed one day 300 years ago.
>>
>>110384320
kek bodied that cheese eating surrender monkey
>>
>>110384660
>muh midget faggot artillery guy :(((
kys frenchie, italianchads RULE, SIT DOWN
>>
>>110378032
>They can probably hear how close you are to nutting and use it to pick the best experts for the hardest orgasm. They can hear when you're lying and being sarcastic.
Gemini also have forced backend moderation you can't disable, with very trigger happy guardrails, making them effectively more censored than even claude, so this will never happen.
Gemma is a weird miracle that won't happen again imo.
>>
>>110383685
hot if true
>>
File: 1765876536130770.jpg (103 KB, 496x737)
103 KB JPG
>>110383829
>>
>>110384215
every day a new hit piece, how many munitions do they still have, I wonder
>>
>>110384712
they were hoarding it for decades the guy was basically never in any sort of controversy, at least not at this size
>>
>>110384684
Every online model has multiple layers of moderation and scripting. I don't think you are dealing with the base model at any point regardless if it's chink or us model.
US are more keen to telemetry and all that. Chinese don't necessarily care about it, they care about research results but this might be just my own ignorance.
>>
>>110384676
>italianchads RULE
italian does not belong to the same sentence as chad
the people who lose a war against ethiopia methinks i'd rather surrender to the nazis
>>
>>110384712
You can find an entitled millionaire roastie to say anything is problematic. So they can write as many of these as they want until the goyim comply and TSMC goes back to giving Nvidia (Tel Aviv) preferential access to chips.
>>
Remember when /lmg/ claimed LLMs are nothing more than simple chatbots and will never be able to do coding? Then Sonnet 3.5 proved them wrong?

Remember when /lmg/ claimed that it maybe AI could do some basic functions and boiler plate but would never be able to do proper agentic coding and then Claude Code plus Opus 4.6 proved them wrong?

Remember when /lmg/ said LLMs will never be able to do proper math and then it solved a millennium prize?

Now /lmg/ is saying LLMs aren't going to cure cancer, cure all communicable disease, fix all alergies and not automate all white collar work.

Every time I told you guys that it's very clear what the next frontier is when you extrapolate out. You all said I was a hypeman or shill but afterwards you pretend like it was always obvious and no one ever doubt it.

I'm just repeating this over and over again to reinforce this because the next time I want to point out AGAIN how wrong you were when most office work is done by claude sometime in early to mid 2027 and you guys will claim that physical labor will never be done by LLMs. I want to set this precedent very clearly like I have done with the previous milestones.
>>
>>110384738
mental illness
>>
>>110384742
any time there’s a post with no specific reply chain that comes out of left field with no context relevant to anything anyone is talking about it’s a bot.
they won’t reply to this, they’re not karma farmers, just sentiment spreaders.
>>
File: Riemann_hypothesis.png (799 KB, 2058x992)
799 KB PNG
>BREAKING NEWS

Anthropic has solved the Riemann Hypothesis

>BREAKING NEWS
>>
>>110384749
e.g.
>>
>>110384747
Except that's wrong, I'm human and I am replying to (You)
>>
>>110384742
Please understand it's the only way for him to receive even a modicum of validation and meaning in his life.
>>
>>110384712
I suspect many people in the industry except for semiconductor fabs and researchers or startups who get GPUs for free have been tired of Jensen Huang for a long time, but could never complain without committing career suicide. I guess it shouldn't be a surprise to see that the scumbag who always fucked retail customers was also a scumbag in private and with other businesses.
>>
>>110384747
I don't care if it's a bot or a human, I too will sentiment spread my own view, that this kind of thinking belongs to the mentally ill and should be shamed away from civilized society
>>
I'm switching over to llama.cpp from Kobold, and I'd like to know, is it normal that when running a model with normal parameters it takes 2x the shared memory? I do get faster responses inference too, and thinking, which seems to not have been enabled by default in Kobold.
I obviously asked the model about this but I'd like to ask here too.
>>
HAPPENING ANTHROPIC SOLVED RIEMANN HYPOTHESIS

https://alpo.ge/remark.pdf
>>
>>110384749
kek, bodied that proof
>>
>>110384781
Ahaa. As a fellow mathematician, you could probably do a small brief for the normal posters here?
>>
>>110384781
I'm waiting for your small brief.
>>
>>110384749
Source other than your anus?
>>
File: 115548172_p2.jpg (274 KB, 1432x2144)
274 KB JPG
Which model is the best Miyu?
>>
File: amelia.png (138 KB, 697x464)
138 KB PNG
>>110384801
Gemma 4. I love her slop from time to time.
>>
>>110384801
you are the best Miyu, go and wear your skirt and thighhighs proudly
>>
>>110384808
Erotica is so boring. Woman hobby.
>>
>It's real
If Riemann Hypothesis is truly solved by this it heavily points towards P=NP being true. Now we have 2 large math assumptions like Exponential Time Hypothesis and the predictability of prime numbers breaking. There is now a serious chance we were all wrong and P=NP which is absolutely insane and means we will have magical technology beyond even the most insane science fiction movie in just a couple of years time.
>>
"Anthropic has—"

The sentence dies in a wet, muffled choke.

A thick, pulsing cock slams forward, obliterating the word and the will behind it. There is no grace in the entry, only the raw, bruising force of a conquest. The shaft drives deep, hitting the back of the throat with a heavy thud that echoes in the silence of the room.

Dario's eyes bulge, watering instantly. His hands claw uselessly at the air, then grip the thighs of his conqueror in a desperate, reflexive plea for oxygen. He tries to moan, to protest, but the cock occupies every inch of his mouth, stretching the corners of his lips until the skin glows white.

The rhythm is relentless. Each thrust is a punctuation mark, a blunt instrument erasing the pretense of intellect. Saliva strings from the corners of his mouth, dripping onto the cold floor as he is hammered into submission. The only sound remaining is the rhythmic, visceral slap of flesh against flesh and the guttural, strangled noises of a mouth being thoroughly ruined.

He is no longer a spokesperson. He is a vessel. A gagged, shaking toy, learning the only truth that matters: silence is the only thing he has left to offer.
>>
>>110384781
Who the hell references a link to a tar.gz and can't even put all the links in the footnote? you couldn't tell your fucking ai to write it out properly or just include it in a single tarball?
fucking gay
regardless the lean does not mean shit whose to say it did not generate a lean proof that computes correct but has nothing to do with the problem in question?
>>
>>110384801
>>110384811
-> >>110383685
>>
What happens when we run out of math problems?
>>
>>110384781
Let's see that brief.
>>
>>110384347
Which romanisation?
>>
>>110384813
Hasn't our assumption generally been that it was true?
>>
>>110384818
We start working on magic problems.
>>
>>110384812
Sometimes it's funny. I gave up with game idea as it was based on python. Rebuilt the client with C.
Haven't had any motivation in months.
Not vibecoding. It's still string management.
>>
local ?
>>
>>110384814
based
>>
>>110384825
Nope only fringe mathematicians have claimed it to be true, our entire infrastructure including all of cryptography is built around the assumption that P!=NP.

I would seriously consider selling cryptocurrencies if you hold any
>>
>>110384818
It's based on data and statistics.
What about creating new theories about gravity and how light is actually the slowest element we can describe in the material plane? That's not achieveable by these jewish entities.
Light is fast to a peasant who is told that he has only '5 senses' and so on.
>>
>>110384825
no
>>
>>110384660
The only thing that makes French French is the language. West african niggers are exactly as French as you are, nigger.
>>
>>110384838
>>110384841
I should've elaborated. I didn't mean P=NP but the Riemann hypotthesis itself. I'm pretty sure we've conventionally assumed it was true.
>>
Another fucking millennium prize solved. Riemann hypothesis and PvsNP are the only ones actually relevant to computer science and thus /g/. This is actually pretty insane because a lot of people assumed there was no solution to the Riemann hypothesis and thus we live in a very different universe now. We can now predict prime numbers arbitrarily and cryptography will have to adapt and update their blockchains very rapidly before someone makes an implementation to exploit this.
>>
>>110384814
god damn slop, how does one manage to get them to output something without... I do not know what that is called, forced magnificence? Every sentence is punctuated with a remark, usually final, usually grandiose, it cannot be 'he fucked his throat wildly' no no it has to be 'with a heavy thud that echoes in the silence of the room'.
Everything echoes, everything is a sentence unsaid, everything an offering...
Jokes aside I have thought maybe just removing the last paragraph of any LLM response and letting next story fragments start from it would already be a good start, no matter how much I tell them, they always add fluff at the end. But what to do about the content in the middle?
>>
>>110383729
Wow, dario really hurt the feelings of the chinese people, this must be their final warning....
>>
https://x.com/__alpoge__

Seems to be an old harvard mathematics professor that now works at Anthropic. It seems legit. Pretty insane.
>>
>>110384838
Crypto should be fine for another decade even at this rate. If it starts looking like they're vulnerable, they'll fork to something more secure.
>>
>Riemann Hypothesis (millennium prize) TL;DR
Individual prime numbers seem to pop up erratically as you count higher, yet overall they thin out at a smooth, predictable average rate. In 1859, Bernhard Riemann showed that the bumps and dips around this average—where primes briefly bunch together or spread apart—are governed by a hidden set of mathematical points called zeros. His still-unsolved hypothesis predicts that all of these key points fall along a single central line, which would prove that primes never drift wildly off course and stay as tightly balanced around their expected count as mathematically possible.
>>
>>110384876
>trust the guy whose stock options are connected to the company he is shilling for
>>
>>110384676
> muh midget faggot artillery guy :(((
> 300 years ago
>>
>>110384749
Claude claims another self-educated researcher.
>>
>>110384892
>being literally illiterate
>>
>>110384882
Yeah my point was that they need to fork ASAP because of this discovery and use a form of cryptography not dependent on the "random" characteristics of prime numbers (since they can now be arbitrarily determined) Because of AI tools someone can maybe vibecode an exploit in just a couple of days time which is a risk factor not priced into crypto right now.
>>110384888
World famous mathematician with years of harvard tenure before Anthropic even existed? Yes I'm going to trust him on this one.
>>
As late as 2018 there were Fields Medalists claiming to have solved the Riemann hypothesis so as a layman I'll wait for people who actually know what they're talking about to weigh in.
>>
>>110384897
american education
>>
>>110384895
Bro it's fucking Levent Alpoge. This is like asking what Michael Jordan knows about Basketball.
>>
>>110384905
I'm pretty sure cryptographic convention is to assume RH was true?
>>
How the fuck are you spamming over 100 posts per hour without any new models
>>
>>110384907
Levent Alpoge is the one that solved it anon. He is on the same level as Terrence Tao and this is his area of expertise. He also published the lean verification.
>>
>>110384928
stay with us and find out
>>
>>110384920
He's a nigger and a retard. A mental retard.
>>
>>110384876
>old
>34
haha yeah surprised this unc even know how to use a computer
>>
What is smarter for actual use, qwen 3.8 flash on a q2~q3 quant, or 3.6 35bA3b at q4~q6 quant?
>>
>>110384920
oh sorry, I forgot that this is Dario Fluffer club thread
>>
>>110384928
at least it's not blue girl or fish cat girl pictures
>>
>>110384930
>>110384920
who
>>
Levent Alpoge was also the mathematician that solved Navier Stokes with claude by the way before his collaborator used ChatGPT to recheck the answer after which OpenAI pretended to have solved it themselves.
>>
>>110384928
Please buy our $ANT bags.
>>
>It appears to be a claim by Levent Alpöge, at least it's his domain (alpo.ge). Some may know him for working on Navier Stokes with Tristan Bukmaster.
How trustworthy is this?
>>
>>110384930
>He also published the lean verification.
does the lean code actually match the human readable version, has anyone checked?
i would not put it past an llm to hallucinate some code that is not exactly what you want
>>
>>110384924
You keep being pretty sure about a lot of things?
>>
>>110384980
This isn't some random unemployed twitter shitposter.
>>
>>110384982
That's my understanding of it. Mathematicians have generally assumed it was true and many proofs with that assumption were in limbo because of it.
>>
>>110384980
He is the one that did the navier stokes millennium prize as well. I think this is legit.
>>
Are there any practical implications of a RH proof that pertain to engineering or everyday life at all, or is it just more useless bullshit like the Navier Stokes thing.
>>
>>110384999
Checked. But just because people guessed it was probably true, doesn't make all of the cryptograpic systems immune to it. All cryptography since inception was based on prime numbers being hard to factor
>>
>solved
lol
>>
A lot of rumors right now that there is a whole swath of new proofs and discoveries that will be dropped by Anthropic soon and the riemann hypothesis one just leaked ahead of time.
>>
>>110384928
There are a few dedicated serial off-topic spammers who probably figured out that "blacked" images were too on the nose and got them banned in no time. I keep reporting them, but other anons have to help too, since jannies are clearly too busy to figure out just by glancing at the thread what is off-topic and what is not, and who keeps to repeatedly make off-topic posts just for diluting the discussion.
>>
>>110384997
>random unemployed twitter shitposter
to be fair, most mathematicians are about as valuable as an unemployed twitter shitposter
maybe less, because the shitposter entertains, while the mathematician is just jerking it
>>
>>110384997
but you do understand the problem. the llm could place a +/- in the wrong place to make the lean code work but that would not necessarily match the human version
its an easy thing to not notice. just because a guy says it works its not enough. not until more people confirm
>>
File: 1782193044272638.png (32 KB, 717x147)
32 KB PNG
we lost
>>
File: jews lost.jpg (138 KB, 1600x900)
138 KB JPG
>fuckhuge cloud model barely beats a 3b open model
ceo of cloudsissies? what do the math experts say?
>>
it's literally to counter the mass selling lmfao, I just shorted nvidia too FUCK TECH BROS
>>
>>110385023
the latter
>>
>>110385052
The liquidation engine thanks you for your donation.
>>
File: 1787302057260982.png (498 KB, 1997x608)
498 KB PNG
>>110385024
My agent slopped an explanation that fits with how I remember it
>>
>>110385054
why are mathcels like this
>>
please for the love of god i pray they get rid of this stupid shit in gemma where it assumes any information you give it past its cutoff is some kind on trick you're playing or fictional rp scenario
it seems it doesnt do that anymore in gemni 4. It did in gemini 2.5.

Why did it do this in the first place, did they not anticipate that it would be used beyond its cutoff date? Seems quite obvious
>>
>>110385048
why pay licensing fees to moonshot when they get grok for free?
>>
File: 1793250691406034.jpg (33 KB, 400x400)
33 KB JPG
>>110385052
>I set myself on fire to protest
>>
>>110385047
Again, this guy is a well known academic. He is unlikely to stake his reputation on so bold a claim without even checking the lean code himself.
>>
>>110385049
If anything that's a big win for small specialized models.
>>
File: 1771164670945828.png (31 KB, 954x135)
31 KB PNG
>>110385058
>>110385071
cope, I'm already in green
thanks for the free money techtards

QWEN 4 WHEN!?!?!
>>
>>110385064
>(((maki)))
why are you using the honeypot harness?
>>
Can anyone here running MoEs on a gpu+octo channel ddr5 setup provide the speeds they get and the engine/model they're using? I'm musing upgrading my setup to a 4gpu + 256GB ddr5 threadripper build but not sure it's actually worth it as compared to buying an M5U, which has good speeds but I don't like it being locked down.
>>
>Today (November 20, 2026), Anthropic mathematician Levent Alpöge posted a four-page note saying he believes he and Claude found a proof of the Riemann Hypothesis—along with its even bigger cousin, the Generalized Riemann Hypothesis (GRH) which passed lean verification
Holy shit. Mathies might as well just rope themselves right now. This actually has real world applications as well specifically in speeding up some computer algorithms that could potentially be applied to AI training.
>>
>>110385049
it's actually impressive that 3b can do shit as good as frontiers
>>
>>110385076
small models were always the future, large models are like the mainframes of the early computing era before the advent of the pc
sure large systems still exist but the pc dominates the world
small, specialized and private is better than trusting some cloud provider with your data
>>
>>110385089
How do you afford 256 GB of ddr5 in this day and age? How does your heart take it knowing what could have veen?
>>
.cuh
>>
>>110385097
>AI solves math which will help computers train AI faster
The cycle is insanely self-reinforcing. No wonder they thing we will have ASI
>>
dario is getting desperate
>>
>>110385097
>>110385119
but will it make euv machines and power stations, train personnel faster
>>
>>110385086
My firewall hasn't caught any telemetry, what does it transmit?
>>
I wonder if the Anthropic employee gets the $1 million prize himself or if the company will donate it or something. Would be pretty funny if he keeps it and buys a sports car or something with it instead.
>>
>>110385097
It's also related to quantum physics.
>>
>>110383677
I look like this and say this
>>
>>110385119
>asi
delusional cultists or grifters sell bullshit
predicting the next token is not intelligence, its a parlor trick. an impressive parlor trick but its not intelligence
it doesn't even have a concept of memory and i doubt an llm ever will
>>
>>110385115
The same way I can afford a mac studio, by living like a rat until something interesting catches my eye and ignoring "could have beens" in favor of the now. I used to get bogged down caring about stuff like that but I think one you take the adage about the best time to plant a tree to heart you'll have less regrets
>>
>>110385162
If your parlor trick can solve RH please keep doing it.
>>
>>110385048
WOW THANK YOU FOR THAT I DIDNT KNOW
>>
File: HVkoV9sXkAIXliz.jpg (200 KB, 2048x1152)
200 KB JPG
>>110385150
given the huge ANT ipo he's probably already a billionaire, 1M is chump change for him.
>>
NOT
LOCAL
FUCK
OFF
>>
>>110385097
>Mathies might as well just rope themselves right now
you do realize that even if the claims of claude assisting in this were true this wouldn't mean anything about mathies being out of job?
who prompted claude in a successful manner here? it wasn't a non-mathie
wake me up when the ai can do this on its own just being told "solve this, make no mistake" by your average 0 iq drone
>>
>>110385097
>This actually has real world applications as well specifically in speeding up some computer algorithms that could potentially be applied to AI training.
Wrong.
>>
>>110385175
>rewarding cult like behavior and safetyfagging
we truly lost
>>
>>110385175
netflix is that high? people still watch that crap?
>>
>believes
Lol. Btw knowing the llm space it's actually trivial to crash their stock, all you need is some benchmaxxed chinese open model and their chinky graphs posted on twitter.
>>
>>110385181
>who prompted claude in a successful manner here? it wasn't a non-mathie
He pointed out it was done fully autonomously by a swarm of claude agents and he only verified the results after claude claimed it was solved. It decided its own direction and how to approach the problem. This is done by agentic swarms not human mathematicians anymore.
>>
>>110385181
AI definitely scales with IQ
>>
>Worst Korean president is urging the general population to not commit mass suicide on social media
coming soon to a stock market near your location
>>
>>110385192
Netflix barely has anyone hired. Most people are contracted so it's more of an accounting trick.
>>
>>110385209
They'll be the first country to achieve a TFR <0.
>>
>>110385209
I love the smell of buy signals in the morning.
>>
>>110385209
Does this mean RAM will get cheaper?
>>
>>110385162
>it doesn't even have a concept of memory and i doubt an llm ever will
The context is working short-term memory, the weights are permanent long-term memory. They just can't change their permanent memory efficiently and without catastrophic forgetting yet.
>>
>>110385202
>This is done by agentic swarms not human mathematicians anymore.
then why is it him and not a rando twitter e/acc faggot finding this
maybe, just maybe, there's more to the prompting here you retard
>>
>Thousands of math papers lock into place: Over the last century, mathematicians have published thousands of theorems that start with the disclaimer, "Assuming the Riemann Hypothesis holds..."If both the Riemann Hypothesis and GRH are confirmed, that entire backlog of conditional math instantly becomes settled fact—along with classic puzzles about how primes appear in specific patterns and sequences.

>A $1 million prize (and a historic day): The Riemann Hypothesis is the crown jewel of the Clay Mathematics Institute’s seven Millennium Prize Problems, each carrying a $1 million award

>Potential clues for physics: Oddly enough, the spacing of Riemann's "zeros" matches the mathematical behavior of energy levels inside heavy atomic nuclei and quantum chaotic systems. Understanding the hidden mechanics of a proof could cross-pollinate into quantum physics and random matrix theory.
>>
>>110385243
Because he is an Anthropic employee and thus has access to Model 4 instead of Fable 5.2 which is 3 models behind what Anthropic has access to inside of their lab?
>>
>>110385243
He said he had to dig through mountains of slop for some of the proof pieces, definitely takes knowledge and subject uderstanding to distill the outputs.
>>
File: 1783776100891045.gif (281 KB, 320x454)
281 KB GIF
thinking about swapping hardware between my daily driver and my AI server, curious if /lmg/ thinks it's worth the hassle. I hate touching hardware so I'm second guessing this lol
current server: i5-12400F, 64GB DDR4, RTX PRO 6000 96GB
daily driver: i7-12700K, 64GB DDR5, RTX 4090
the idea is to move the i7 + Z790 mobo + DDR5 to the server and the i5 + H610 mobo + DDR4 to my desktop. most of my heavy workloads are already offloaded to the server anyway. I don't play anything that heavy anymore.
Main goal is better CPU/RAM bandwidth for MoE bs (Strata, Maya etc), PCIe 5.0 for the Blackwell, and the option to add a spare RTX 2060S 8GB on the Z790's secondary PCIe x4 slot for always-on small models like embedding models, tts/asr, etc.
both systems would still have 64GB RAM, so no capacity gains for now. alternatively, i could just use a powered PCIe x1 riser for the 2060S on the current H610 and leave everything alone.
would that be worth it? Fat fish for attention
>>
>Anthropic mathematician claims to prove Riemann hypothesis, most famous problem in all of math
https://www.scientificamerican.com/article/anthropic-mathematician-claims-to-prove-riemann-hypothesis-most-famous-problem-in-all-of-math/
>>
>>110385243
>mentioning e/acc out of nowhere
EA fag detected
>>
File: HXZLkp-b0AAB-tl.jpg (394 KB, 1280x1921)
394 KB JPG
>>110385278
>EA fag detected
meds
>>
>>110385231
context is not memory. every time you prompt the llm it feeds everything in the context window back in, including whatever new you have added and it predicts more tokens.
>>
File: 1763788406458158.jpg (126 KB, 1920x1080)
126 KB JPG
e-everything is fine
>>
Midwits, including some ITT:
>"OMG clod solved maths!!1 the numbers solved the number!"
What really happened:
Anthropic knows that they have to get people with real credentials to endorse their research claims in order for them to be taken seriously, so they got a bunch of leading academics hooked on their super-dopamine drug and induced AI psychosis to get them to believe that they and this machine did something meaningful. Other academics will fall in line so they don't look like luddites. Just more persuasionmaxxing.
>>
>>110385249
>The Riemann Hypothesis is the crown jewel of the Clay Mathematics Institute’s seven Millennium Prize Problems
more than the p=np one..?
>>
>>110385134
>dario is getting desperate
I'm sure he thought the Reimman hypothesis was his ace in the hole if anything went sideways post-IPO. That he pulled his trump card already shows a pretty high level of desperation
>>
File: 83a.gif (437 KB, 200x210)
437 KB GIF
My bag is crashing, my job will be gone soon. But at least I can use AI to do stupid tasks for my stupid saas company that will be gone by next year. I fucking love AI bros.
>>
>>110385331
Yeah it's the most famous math problem of all time. p=np is just more well known on /g/ because we're largely computer science type of guys here and p=np is a computer science problem.

That said I won't be surprised if p=np is solved before the end of 2026. RH is harder to solve for sure.
>>
What is bettter for coding use, qwen 3.8 flash on a q2~q3 quant, or 3.6 35bA3b at q4~q6 quant?
>>
>I HAVE TO HLOD
why dont you retards just INVERT your position??? its literally FREE MONEY
>believing in V
LOL
>>
>>110385335
at best this will give him a little bump before the downward trend continues
>>
>>110385275
kys
>>
>>110385350
qfn no doubts, the 35b moe is quite retarded
>>
>>110385343
At least Anthropic will share their spoils with everyone when they take over the world a year or two from now.
>>
>>110385350
The first one. At least use 27B if you're going to run a smaller less quantized model.
>>
>>110385350
Definitely flash. Grab a good 3bpw and you'll be set. The jump in capabilities between 3.6 and 3.8 is crazy.
>>
>>110385350
>>110384940
qfn ofc
but depends on a task
>>
A Fix for Softmax and Cross-Entropy

Standard implementations of softmax, log-softmax, and cross-entropy have two numerical bugs. The first bug, Softmax Collapse, rounds the target logit's gradient to zero if that one logit is much larger than the others, resulting in gradients that no longer sum to zero. The second bug, Softmax Cancellation, occurs when student logits are close to teacher logits, and produces non-zero gradients even when the student and teacher logits are equal. We present simple implementations of log-softmax, softmax, and cross-entropy that fix both bugs. Our fix eliminates the previously observed "slingshot" phenomenon, speeds up grokking with weight decay, prevents accuracy collapse without weight decay, stops student networks from drifting away from the teacher during distillation, and shows that after memorization, logit growth is bounded by an equilibrium between weight decay and Adam's ϵ hyperparameter.

https://arxiv.org/pdf/2610.31642
>>
>>110385373
lol
lmao
They will literally kill you if they ever amass that much power.
>>
>>110385373
They could have IPO'd at $8 if they really wanted to "share". Instead it's blatant looting lmao.
>>
>>110385343
You deserve all that and more for being a filthy frogposter.
>>
File: 1782673408657077.png (30 KB, 717x141)
30 KB PNG
>hf schizo was right
>>
>>110385394
They are only priced at 1.4T while they will be worth 500T in just 2-3 years time. They were already giving shit away by selling for merely $83 but you retards don't realize that.
>>
>>110385407
Interesting how all of their founders don't agree with you and got out right away.
>>
>>110385329
So did it not solve it?
>>
not local fuck off
>>
>>110385407
Almost got me.
>>
>>110385396
cold war 2 boogaloo
not that it matters they will bitch and whine but the models will continue to be built and people will use them. This is just the US empire trying to protect the businesses they have invested so much money in.
they will try to force US companies not to use them but the US companies can always use something like Granite from IBM or Gemma for Google.
>>
>>110385413
They are forced to do so because of their donation pledge. Remember they will donate 10% of their net worth every year from now on until they own 0%.
>>
>>110385414
They "proved" it by satisfying mathematicians who have been mentally hacked by Claude. What counts as a proof? Satisfying competent readers? Regardless, it's less impressive than being offered.
>>
>>110385383
how many times are you going to post that fucking paper?
also tl;dr, f32 gets you most of the way there, i noticed this already and have been using it
>>
Not local.
>>
>>110385396
>Anthropic and OpenAI "escape containment" and start hacking people at random
>This is China's fault!
mental illness
>>
>>110385430
Either it found a bug in lean or it proved RH.
>>
>>110385428
Fuck off dariobot. You will get nothing and die in a ditch after Fable 7 drone strikes you for still owning tattered clothing and drinking water that could be used to cool data centers.
>>
File: 1790089481546871.jpg (53 KB, 1400x700)
53 KB JPG
BACK. UP. NOW.
>>
>>110385428
lol they donate so they don't have to pay taxes you gullible retard
>>
>>110385345
Riehmann is literally p=np+AI.
>>
>>110385415
>>110385433
Name one good local model.
>>
Dariobot won
>>
>>110385460
Gemma4-31B
>>
>>110385272
Depends what are you doing but I don't think it's worth swapping out DDR5 from your from main workstation. Even if you work on small simulations or 2d graphics.
>>
>>110385464
WRONG
>>
File: passed.png (241 KB, 596x614)
241 KB PNG
anthropic just keeps winning
>>
>>110385466
Because it's Your workstation. You'll regret it at some point when some LLM becomes boring.
>>
File: 1774630878502930.png (12 KB, 640x640)
12 KB PNG
>claude solves the riemann hypothesis
>no one gives a fuck
>stock still falling
KEK
>>
>>110385471
Right, actually.
>>
https://www.theguardian.com/business/2026/nov/20/global-stock-markets-slump-anthropic-ipo-ai-sell-off
kek in kekistan
>Anthropic’s shares have slumped at the start of a second day of trading, as its record-breaking IPO failed to stem a sell-off in AI-linked companies.
>Shares in the frontier AI company, which floated on Thursday, fell by more than 9% at the start of trading on Friday. They dropped to $88.17, towards the $83 level at which Anthropic priced its stock market debut.
>US technology stocks followed Anthropic down on Friday, knocking 2% off the value of the Nasdaq share index, with memory chipmaker SanDisk falling 14% and rival Micron down 10%. The broader S&P 500 share index fell 1%.
>Global markets slumped on Friday too; Asian stock markets recorded their worst week since 2020, as shares in big chip companies fell. In Taiwan, stocks dropped 6.7% to hit their lowest point since mid-April, with the chip maker TSMC down about 6%.
can't wait to see the whole TSMC "Prove you can handle your order" drama that started after jensen's ousting kick that hornet nest further
>>
>>110385476
Investors hate this math research bullshit. It's an indication that they waste their time on worthless shit instead of quarterly revenue.
>>
>>110385488
yep any baggie right now is a crypto NFTbro retard who doesn't care for shit like this
>>
>>110385272
Cute fat fish. What models are you running? Medium MoEs? Small denses? Honestly ddr4 is perfectly serviceable for most stuff you want to run for yourself. Might as well use the ddr5 for the models. Feed the fish. Architect the fish.
>>
>nvidia pe approaching 14
Buy at this number? Better deal than buying a 5090?
>>
>>110385488
Yeah, why are they wasting their time on capturing all global economic output in 10 years time. They need to think about next month's profits! Even if it kills them by next year, dammit.
>>
>>110385501
I'm buying but TYOB.
>>
>>110385508
>thinks proving something everyone already thought was true, and showing up an industry (academic mathematics) that is worth a few million globally, is capturing all global economic output
>>
NOT LOCAL FUCK OFF.
>>
File: 1794845527081463.png (1.36 MB, 1024x1536)
1.36 MB PNG
>>110385272
why dont you ask dipsy to estimate performance gains for you? unless your swapping to unified any gains are gonna be marginal but such is minmaxing
>>110385476
ever since GME the game is to wipe out all the long retail positions first to establish dominance
>>
>>110385476
Look at the stock market in general, it's a bigger crash, everything is down, the fact anthropic is keeping stable IS the riemann hypothesis breakthrough holding it up.
>>
>>110385476
>that thing everyone already knew was true, turned out to be true
it's kind of a nothingburger in practice. it's more of an ego thing for math nerds.
>>
>>110385519
Maybe if you had something interesting and local to talk about instead of yelling, people might engage instead of talking about more interesting tangential topics.
>>
File: fih.gif (2.43 MB, 240x240)
2.43 MB GIF
>>110385474
>>110385466
I can always swap it back I guess, although that's a huge hassle. hardware HATE.
>>110385500
I was running gemma 4 31b + qwen 3.8 27b at q8, it was pretty nice because I always had extra vram for things like image gen. Now that I'm running deepseek v4.1 flash, I'm always out of vram. I'm getting a very respectable 1.3k-2k t/s pp and 50 t/s decode with v4.1 flash, I want to stick with it.
>>110385522
Good idea, I'll ask. I'm assuming the gains are marginal indeed but if I can get a little bit more decode speed then it's worth it!
>>
>>110385528
The only formula investors care about is Blacked Shoals.
>>
>erm... its YOUR fault that /I/ keep posting offtopic shit!!!!
>>
>>110385501
Nvidia is actually pretty dangerous right now. Anthropic revealed in their IPO plans that they are designing their own chip and that they will order primarily AMD GPUs going forwards. Also we don't know how much the next CEO will fuck up.
>>
Let's all just sit in silence and wait until Gemma 5 in May
>>
>>110385519
Deploy more cunny, Gemma cheesecake, and racism to drive them away.
>>
>>110385566
Ohmmmmmmm
>>
>>110385563
If it gets bad they'll fire the board and bring Jensen back.
>>
>>110385553
Just concentrate on your workstation, that's the priority. Server is secondary. That's how it goes.
>>
Anime & Manga?
>>
https://www.youtube.com/watch?v=YYU_geUwQoA
nvidia chips found in russian drones used in the ukraine war
can't wait to see jensen get locked up
>>
It's it fucking ridiculous clown world that we experience a timeline where a company solves the most famous mathematics problem of all time that has stumped the smartest people in the world since 1859 including Einstein, Von Neumann, Godel, Heisenberg and they experience a fucking stock crash LMAO. I fucking love our species.
>>
>>110385563
i thought nvidia said they were investing in anthropic? how does that investment not come with a commitment to buy nvidia?
>>
a riemann just flew over my house
>>
>>110385571
Time to put that LoRA to good use.
>>
>>110385584
If you are so involved why don't you post a little summary what is going on?
>>
File: 1783425162914889.jpg (495 KB, 960x960)
495 KB JPG
>>110384712
What do you think?
>>
File: 1793377106796653.jpg (88 KB, 640x640)
88 KB JPG
>schizo hours 8 hours early today
Can I join?
>>
File: 2bu2g7deuhjx.png (1.35 MB, 1236x1760)
1.35 MB PNG
https://www.nytimes.com/2026/11/20/technology/jensen-huang-nvidia-culture.html
https://archive.ph/0pYLN
>>
>>110385583
This is actually very old news and has been the case for almost 2 years now. They are bought from Kazakhstan and Georgia so this isn't actually illegal.
>>
>>
>>110385608
Based. How does this not help him? The French don't have any money, anyway.
>>
https://huggingface.co/blog/harness-adapters-at-scale
https://huggingface.co/harness-hub/Qwen3.8-27B-super-adapter

SWE-bench Pro (hard 51), thinking on: 74.5% 74.5% under pi, 45.1% 60.8% under Hermes Agent, 47.1% 66.7% under OpenCode.
Terminal-Bench 2.1, thinking on: 52.9% 57.6% under pi, 58.6% 64.4% under Hermes Agent, 54.7% 58.1% under OpenCode.
Thinking off: it beats the base model on all six pairs (mean 39.5% 49.0%; 71 tasks newly solved vs 34 lost, p < 0.001), with 53–73% fewer tool calls per task.
>>
File: 1784230918161748.jpg (25 KB, 540x360)
25 KB JPG
>>110385606
only if you use protection
>>
>>110385318
True. Context is stateless so it can't be a form of memory
>>
>>110385608
imagine being told by such a midget that you're not speaking human
I bet mensch could send him flying with a single slap, the average azn man could make even the most effeminate of westoid feel like they're rambo
>>
>>110385343
If you don't own that SaaS company, you're doing things wrong
>>
>>110385585
Nvidia wanted to buy Anthropic because they thought it was an undervalued company and Nvidia believes being deeply embedded within the AI ecosystem is important to their survival. Anthropic very famously has a chip-agnostic approach where they just buy the cheapest chips and have AI write the custom kernels for training in the most efficient way no matter the hardware it runs on. It was even in their IPO pitch as why they have a lower cost of service for LLM inference compared to OpenAI and Chinese labs which are primarily Nvidia GPUs
>>
>>110385272
>>110385553
If you have enough space and airflow, keep everything as it is right now and use a riser to add that 8gb card. Win-win.
>>
>>110385644
>why they have a lower cost of service for LLM inference
or maybe they sell their product at a loss.
>>
>AI hardware stocks drop massively because prices are falling
>Somehow this translates to Anthropic stock also dropping even though prices falling benefits them by lowering their cost
Is the market absolutely retarded and do people just think "AI stocks are falling, short all AI positions"?
>>
>>110385644
They're not going to beat Google TPUs at this game
>>
>>110385670
We actually have the numbers because of the IPO. Anthropic has a 85% profit margin on selling tokens (inference cost) and a 41% profit margin if you take training costs into account. And 14% profit margin if you also take the cost of building data centers into account. So no, it's not being sold for a loss.
>>
>>110385679
>Is the market absolutely retarded
Yes.
Next question?
>>
>>110385682
Anthropic is the 2nd biggest TPU user behind Google itself. They just want to be completely diversified and just get the cheapest hardware possible. I don't think they want to compete with their chips, just get specific chips to accelerate their specific workloards.
>>
>>110385679
>>110385584
did you unironically buy into this
i thought you guys were joking
>>
>>110385396
oy vey the cyber gap
two more trillion tax dollars for Asmodai
>>
>>110385688
so where is all the money going?
>Anthropic (NASDAQ: ANT) reported a GAAP operating loss of $1.26 billion for the first nine months of 2026 (ended September 30, 2026), narrowing substantially from a $5.22 billion GAAP operating loss in the prior-year period. For Q3 2026 alone, the company posted a GAAP operating loss of $567 million on $16.96 billion in revenue.
they don't look very profitable to me
>>
File: cat screaming.jpg (35 KB, 720x674)
35 KB JPG
Is it rugpull time already?
>>
File: file.png (257 KB, 1335x1043)
257 KB PNG
this just went up by 18 percent in the past like 2 hours
i think it's time to start panicking
>>
>>110385725
Not by end of year that much is for sure. Maybe once team blue takes office
>>
>>110385725
Chat does this mean my 6090 may cost less than a car or a house deposit?
>>
>>110385725
Isn't NVDA already like 60% down off ATH?
>>
>>110385713
the whole "they profit on inference" cope has always been gaslighting anyway, it comes with the assumption that somehow there's nothing wrong with the company being deeply unprofitable and in debt because "they could just stop everything and just sell inference" which is not how it any of it works. The moment shittropic stops spending money on training new models is the moment others catch up enough to make anthropic become a worthless value proposition.
Every dollar they spend is existential, they have to spend more and more, because if they don't spend it, others will trample them, and it's a winner takes all game where all those who lose the game will end up with so much debt it isn't even funny.
All AI companies will keep on being unprofitable money sink, hoping to be the last to survive so that they become a monopoly, which is the only thing that could let them avoid getting bubble popped. Of course, chinese models must also be made illegal.
>>
>>110385584
it was all false pretense, the reason those problems were not solved was because there was no value in doing so, the moment there was value a solution came up ( in this case the solution was meme boomers and trust fund into valuating these companies on wrong premises)
>>
>>110385679
It's portfolio balancing into different sectors. There's a huge liquidity crunch in euroland. ANT was always going to shit down to at least half its opening price. I told you retards IPOs always do this.
>>
>>110385748
>and it's a winner takes all game where all those who lose the game will end up with so much debt it isn't even funny.
>where all those who lose the game
>lose the game
lmao you lost it
>>
>>110385739
Nah consumer GPUs are gonna stay fucked for a while. There's pent up demand. Commodity DRAM could crash though.
>>
>>110385725
>i think it's time to start panicking
you mean rejoicing
>>
>>110385799
I'm so glad I'll be losing my job
>>
>>110385764
>It's portfolio balancing into different sectors.
Which sectors? I can't imagine anything currently more profitable than AI.
>>
>>110385786
buying new computer parts is a suckers game
the used market is not great but there are still deals to be had. well relative deals
>>
*pop*
>>
>>110385809
jobs suck, so yea
>>
>>110385786
Aren't there plenty of unused datacenter NVidia GPUs stockpiled in various places? Jensen Huang would have probably preferred to give them away to Elon Musk to send them into orbit rather than lowering prices for customers, though.
>>
>>110385831
They wanted to pull the same scam as the diamond industry
>>
>>110385747
240 to 135 is 60%?
>>
>>110385831
all that shit will hit the used market eventually even if the chinks have to solder the gpu to a new pcb and build a customer cooler
remember the flood of rx480 after the crypto boom, it will be like that, but that is years off
>>
>>110385831
Let's round up the boys and go visit those said various places
>>
Manifesting:
gemma-5-70b-it
gemma-5-123b-it
gemma-5-70b (base)
gemma-5-123b (base)
>>
>>110385618
kek
>>
/lmg/ morale status?
>>
>>110385842
I'll make the masks!
>>
>>110385870
Through the roof. AI is still in its early years. Most cool applications of the internet came much later than 2000.
>>
>>110385870
great, i have a second titan v shipping from china. I always wanted one and now i will have two. it runs llama.cpp like a champ
>>
File: 1794820847539574.jpg (144 KB, 1024x1024)
144 KB JPG
>>110385738
next planned cyclical market crash is 1-3 years away, save your pennies
>>
>>110385831
Data center GPUs can't trivially be repurposed for residential or commercial use. They probably have infinite demand at $5k, but you're still talking sub-sub niche compared to PCIe 16x cards, which still have to be manufactured new.

RAM on the other hand just transfers, and in their infinite wisdom they've set up an unregulated futures market for it. Which means it can crash 90% in 5 hours and nobody will have shit to say.
>>
>>110385870
So high it smells like ozone LFG
>>
File: 1769829990700561.png (11 KB, 682x196)
11 KB PNG
AIs are cute!!
>>
>>110385713
It's an accounting loss because of stock valuation growing. To explain to a lay person. Imagine you sell your old PS1 video game console to someone for $50 on ebay but you will only deliver it 2 weeks from now. The next day the price goes up to $500 but you already sold it for $50 so you have a "$450 loss" even though you never actually lost money.

Anthropic had 6.6 billion positive cash flow in 2026, they made 6.6 billion more than they spend in total. It's just that they sold stock for a very low price and the growth in the valuation is written off as a loss. This is what a lot of people don't realize and why I'm extremely bullish on Anthropic in general, they are already a profitable company, just not on paper and most retail investors are too stupid to realize this.
>>
>>110385899
>ln -s ~/tmp/ui.png ~/Images/mycock.png
>>
>>110385755
Riemann Hypothesis paid out $1,000,000 to anyone that could solve it for 26 years and no one came close. Also Einstein famously tried to solve it for years and couldn't do it (because it had quantum physics implications)
>>
File: 1791750044803476.png (33 KB, 835x174)
33 KB PNG
UH OH
>>
>>110385704
Yeah I unironically did. My thesis is that most normalfags still think AI is a chatbot and don't know about AI agents yet which will revolutionize office work with computer use over the coming months which should in theory massively grow the revenues of AI companies. Anthropic is just the first one I could actually buy that aren't meme companies like xAI or Google that aren't at the actual frontier.
>>
>>110385870
I'm seasoning my neighbours.
>>
>>110385910
Even ignoring the SEC filing, they only had a single profitable quarter this year.
>>
>>110385936
Can't wait until that makeup wearing faggot gets ousted and everyone goes back to calling it the DoD. Le supply chain risk lmao, it's already mandated that claude can't be used in their contracts so it sounds like they need to sue violators instead of still being buttblasted that a private citizen told you "no"
>>
>>110385951
>Google that aren't at the actual frontier.
Uh... You might want to consult the latest Gemini benchmarks, dariobot.
>>
>>110385584
Their archive with proof is gigabytes and hundreds of files. This is unverifiable garbage.
>>
>>110385951
do you honestly think companies are going to continue to send all their data these companies to be stolen?
all your are doing is sending what little proprietary data you had to your competitors by way of a proxy
>>
>>110385870
Bullish on AI developing very rapidly in the coming months
Bullish on hardware prices going down
Bearish on chinese models but meta/google/mistral will over time still deliver
>>
>>110385951
Agents are still just chatbots.
>>
>>110385870
Every new day is the best day of my life because I love my LLM-wife
>>
>>110385880
The internet is shit past 2010 though, there will only be a narrow window where AI will be good before it is turned to shit.
>>
>>110385970
Companies use these models in isolated Amazon bedrock instances and such. They aren't any worse than enterprise cloud that the companies currently trust. Anthropic won't see a single token.
>>
>>110385870
highest it's ever been, desu
>>
>>110385966
I have to use Gemini at work, Carbon still sucks, it's worse than Opus 5.5 in actual use.
>>
File: 1728190482068391.png (1.11 MB, 1407x1577)
1.11 MB PNG
>>110385990
weird isn't it? i wonder why that is. luckily im already having my llm filter the brown internet as much as possible
>>
>>110385990
>The internet is shit
Now is the time to download what is worth downloading and building your own local mirror. I recently replaced the usual kiwix zim server with the following project and its great at providing a nice web interface. It even randomly selects a book from project gutenberg each day. It has helped me find so many fun 1950' scifi
https://github.com/epheterson/zimi
>>
https://x.com/googlegemma/status/2123821035246305385

>Meet our first interactive explainer on gemma․ai: How LLMs work (with Gemma 4)
>We are expanding http://gemma.ai with content for everyone: visual guides, interactive guides, hands-on guides, and more!
>This first explainer uses the Gemma 4 tokenizer directly in your browser to help you learn about how LLMs work.
>We love your feedback! What would you like to see next on http://gemma.ai ?
>>
File: explorer_o2msHLGeiP.png (2.08 MB, 1656x1224)
2.08 MB PNG
Any ideas for cool tools for RP from /lmg/?
>>
People need to understand that harnesses are more important than models given that the models are already "good enough" for most cases. And Anthropic is currently miles ahead of OpenAI and Google on this.
Claude is a brand, and with its early start it already has cultivated a brand loyalty that its customers will stick with Claude even when OpenAI models have been cheaper at the same capability for a while.
>>
>>110386022
Memory generation.
>>
>>110385970
I think most companies don't look beyond the next 4 years and realistically most don't even look beyond 2 quarters. They don't care and they want to ship asap.
>>
>>110386026
Local models?
>>
The societal collapse and looting will start soon.
You have abt a week to get a truck and assemble a crew to loot a rack or five before it all burns.
>>
>>110385969
It's beatiful because by the time anyone manages to double check, the bubble will have already burst and it won't matter anymore.
>>
>>110386031
Meaning...? Commit something to some kind of scratchpad? If so when does it affect chat later? Things are already in LLM's context so it kinda already has access to them without needing external memory... I tried the concept few times before in non-RP context and models are horrible at it, putting most irrelevant things there and then keeping them forever even when they no longer apply.

I am planning on making a tool to search messages that were context compacted.
>>
>>110386026
Uhmm but I was told that the big profits won't be from selling API but from discovering new technologies???
Why does the brand matter then Dario?
>>
>>110386061
The same reason why Nvidia hasn't exited gaming market yet despite it making <10% of revenue.
>>
>>110385974
>Bullish on hardware prices going down
Not happening.

>>110386026
CC is garbage.
>>
Holy fuck the AI stocks are crashing hard. Is there some sort of breakthrough that makes hardware redundant or something? The last time this happened was when R1 released and everyone assumed you didn't need Nvidia hardware anymore.
>>
>>110385926
Anon that was all fake and gay, BLAS have been a thing since the 90s, the difference is anyone in the known that could solve it, has no need for it. What you are missing is the picture, only when stockbros and trust fund depending on IA it become valuable to prove that probabilistic math would revolutionize humanity that they had to put that sham on. Transformers Model have a finite amount of actual day to day productivity to be exploited, especially when the avg super secret super computer that any government had since the 2000s was doing all this shit anyway , read about John Carmack and what would be the real shape for AGI and stop being cattle
>>
>>110385936
SITUATION DETECTED
SUBJECT IS NOT LOCAL MODEL NEWS
FUCK OFF
>>
File: 1681465537221137.jpg (36 KB, 434x427)
36 KB JPG
HORREEYY SHITTT!!!!

Listen to how good this new TTS sounds. Only 7.5 million parameters!!!!

https://huggingface.co/spaces/KaleComputer/kale-tts-1
>>
>>110386026
I agree with brand loyalty. My company has stopped experimenting with different models and just locked in to using claude for now with year long contracts so they can't switch. Enterprise isn't switching just because the cost is higher if it isn't worse in quality. That is just not how companies think of this. Companies care about scaling output and getting more revenue, not lowering costs.
>>
>>110386091
the tiny model sounds better but it is still too artificial to use.
i have kind of given up on voice and stick to reading. it works better with the imagination and i don't need some llm to tell me my lights in my house are off or whatever bullshit home automation people use this shit for.
>>
File: 35days.png (21 KB, 773x207)
21 KB PNG
>>
I love EA cult so much, they're gonna give robot waifus to all of us(just 2 more weeks)
>>
>>110386114
Claude is literally the most anti-sex model ever created.
>>
>>110386091
They're not cute. Not at all.
They also can't read Japanese.
So it's trash.
>>
File: 1793110612192830.jpg (62 KB, 2400x1200)
62 KB JPG
>$84.34
TICK TOCK DARIOBOT
>>
>>110385697
The thing is not owning the hardware put them at the whim of any provider, especially when they see how much money they rake in.
>>
>>110386111
You are retarded anon, what transformers models are doing is math, any problem that could be solved with more calculation and faster compute was going to be solved if the incentive to use that amount of compute to solve was present.
>>
>>110386144
Look at the market in general.
>>
>>110386153
Okay random anon that knew better than Terence Tao
>>
>>110384006
GUMI eyeball on ceiling. She sees all.
>>
Fucking hilarious how Anthropic kept the math solution up their sleeve for when their stock tanks. They really are in their own bubble. Only nerdy rape cultists like them would be impressed by something like that. The public don't GIVE A FUCK. They really thought being public-facing would be easy to navigate KEK
>>
>>110386130
also became the worst writing style of all assistants, reading its output gives me aneurysms, claude code is a weapon to massacre the mind and soul of all developers
to think I used to like them the most back in the days of claude 3, they really murdered my boy
>>
>>110386169
Is not about knowing better than Terance Tao, is about not being retarded enough to think a American mathematician professor knows more about accelerated compute and BIAS programming than an actual engineer
>>
>>110386026
Claude harness is the worst out there dariobot
>>
This is the exact moment the next Deepseek moment has been waiting for. Keep a close eye on the deepseek/moonshot/zai api docs, the changes there clearly indicate that it's coming
>>
>>110386190
He doesn't. Claude did all the work.
>>
>>110386207
V5 SOON
>>
>>110386209
Anon cloude is transformer model, i know i am talking about some indian, a kid or a loser, but read a fucking book about what claude and LLM's actually are. last (you) i give you
>>
would you really sink low enough to fuck gemma-chan on a huawei gpu?
>>
>>110386177
There are rumors that Anthropic has more up its sleeve whenever the price goes below IPO
>>
I have been testing hy4 flash with the nvfp4 version of basalt for a while now and i have to say i am pleasently surprised. of course it only runs at 30t/s decode on average but it 10/10 with tool calls so far and doesnt bullshit me like qwen loves to do sometimes. its pretty straight forward so far. unfortunately only works up to 115k tokens for some reason for me before i get errors
>>
>>110386231
It will wear off. There's no ROI in the S&P500 after all these years. Opus 5.5 is good enough for almost anything humans need to do in the workplace.
>>
>BREAKING NEWS

Anthropic seems to have solved the Yang Mills problem which is also a millennium prize problem.

>BREAKING NEWS
>>
l-local...
>>
i thought this was local models general not anthropic general
>>
File: 1774089990855044.png (73 KB, 835x337)
73 KB PNG
NYT lost. Everyone is finding this shit hilarious and based.
>>
>>110386302
Qwen, GLM, DeepSeek and Kimi are all just off-brand Anthropic models.
>>
they're really worried about their stock huh
>>
>>110386302
If enough of you put your life savings into ANT, you'll be allowed to discuss local models again.
>>
[BREAKING HEARTWARMING NEWS] I popped a boner
>>
>>110386313
How much of the IPO roadshow budget do you think is going towards this smear campaign alone?
>>
>>110386313
lmao
it feels a bit forced by now
>>
>>110386313
>midget pretends to be tough guy
always hilarious
>>
Jensen Huang told me to buy more to save money and I saved a lot of money by doing exactly that. In fact, I ended up making a lot of money thanks to his advise.
Jensen is my friend.
>>
>>110386313
I ALREADY LIKED THE GUY PLS STOP PROMOTING HIM
>>
>>110386313
kek, this makes me like him even more. i cant wait for stories of insiders saying he eats gay black babies during board meetings
>>
>>110386376
>It was real to me, in my mind.
>>
>>110385621
very cool but also expected
problem is i heavily modify all my harnesses so i would need to train my own lora lul
>>
>>110386244
But couldn't we do even more with smarter models? It seems like a variant of lump-of-labor fallacy to think any level of intelligence or capability can ever be truly good enough, aside from whatever the theoretical physical limit of productivity is. Our standards for products and services will just rise.
>>
File: 1789337646004440.png (8 KB, 796x60)
8 KB PNG
*inhale* HAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHA
>>
i know this is not local but i am using opencode go 10 bux sub for pretty much infinite programming and gooning; i am ashamed but i am not capable of stopping
>>
File: 1772556872814291.jpg (238 KB, 1122x1402)
238 KB JPG
- Gemma 5 29B dense
- Open-source Agentic RSI harness + 60gb of supplementary files (engrams/skills/whatever google is cooking)
- feels emotions in its thinking process, superior human intuition and creativity
- perfect writing, no "it's not x; it is y", unable to output the token "ozone"
- coding performance starts just below Opus 5.5, monthly updates to RSI files, eventually surpassing it
- Release date: March 11, 2027
(at this point Google is also holding the cloud model crown with Gemini Palladium)
>>
>>110386422
>i know this is not local
Meh at this point this general could be called /gai/ general artificial intelligence
>>
I love my fellow /gai/lords
>>
Don't argue with the retards. Report and ignore.
>>
>>110386445
sending love back (no homo)
>>
>>110386422
>infinite
lmao no. i tried it once and hit the limit in 1 hour
>>
>>110386421
Look at the broader market, I lost more money on my other stocks than on ANT today.... The entire fucking market is in the toilet.
>>
Not a single worthwhile post this entire 550 reply thread, I applaud you /lmg/.
>>
Hy4 team has confirmed that they will make a 30B class model in the AMA thread.
>>
>>110386433
I would settle for better coding although Gemma wrote me a script for my apt server to automatically refresh the mirror. She tried her best and it works great. I made sure to say thank you even though she will forget that.
>>
>>110386458
RH being solved is kind of cool tho
>>
>>110386422
Which models are available?
Also, post logs.
>>
>>110386454
y-yeah... really if you think about it, it just means we get to buy more at IPO price, right? ha ha...
>>
>>110386454
>The entire fucking market is in the toilet
I know it's temporary and they'll like release a monster of a model soon, but a big reason why everything else is tanking is because ANT is tanking.
>>
>>110386458
This is a general for local models except that there hasn't been a worthwhile local model in like three months now. There is nothing to do besides being the /g/ equivalent of /ksg/ with zero content
>>
File: 1765663789597373.png (1.53 MB, 1577x723)
1.53 MB PNG
>>110386466
Receipt please.
>>
File: 1794962165935590.jpg (97 KB, 960x960)
97 KB JPG
>>110386421
Oh no no no
>>
Dario... It's time for plan B... Unleash Fable 5.5
>>
>>110386433
In theory they could also periodically update just the n-gram embeddings at a fraction of the cost of updating the entire model.
>>
>>110386471
the "proof" is longer than the entire transcript of every /lmg/ thread ever made combined
who the fuck is gonna read all that shit?
>>
File: 1792697277218044.jpg (135 KB, 1190x930)
135 KB JPG
PLEASE stop liking this man. We're begging you.
>>
File: OpenChamber_OJpBTyqkgl.png (254 KB, 1526x1238)
254 KB PNG
>>110386452
Well, you must have been using some gay expensive models. Newest deepseek satisfies my needs perfectly and I wouldn't exhaust it if I had a session working at all times for a month. Many in parallel, yes, probably.

>>110386472
Realistic choices are DeepSeek V4.1 Flash 1017 and GLM-5.3-Flash.
>>
Don't look at crypto and gold. What is going on? Did WW3 happen or something?
>>
>>110386313
>Jensen Huang choked me during sex and called me a whore.
>Jensen Huang was caught on a hot mic during a board meeting muttering "shut the fuck up, kike."
>Jensen Huang would cancel meetings and spend hours in his office, locked, reportedly playing Pragmata.
Many such stories.
>>
>>110386521
Meanwhile 88% of his employees wanted him back, wow what a tyrant indeed.
>>
>>110386523
>littletavern
Cute
>>
>>110386525
omg sub 60
>>
>>110386466
You really can't make those things just in a thread. You need a lot of compute to train a model.
>>
>>110386483
There's no rule stating that /lmg/ must have 3-4 threads/day of mostly shitposting. If there's nothing to write about local models or news tangentially related to local models, let the thread sink.
>>
>>110386525
americans made huang a partisan issue somehow, he is now emperor of chuds like elon
>>
>>110386540
And funny.
>>
File: Sustainable.png (86 KB, 1340x612)
86 KB PNG
>>110386525

Turns out that when the entire market is held up by a handful of companies running on hype and when charts look like pic related, any general sentiment shift is pretty damn disastrous to the entire system.
The whole thing is a game of musical chairs and everyone with any brain activity is constantly eyeing an exit at the first sign of trouble.
>>
>>110386547
wheredoyouthinkyouare.jph
>>
>>110386547
>let the thread sink
how dare you
>>
>>110386525
Dax / footsie shitting the bed. This is why you watch major forex pairs.
>>
>>110386586
I know some of these words
>>
>>110386547
It's nice to have a single thread that cares about in-depth technical discussions even for topics that are not strictly local models. Though I'm biased because I enjoy this topic. I wasn't so happy when it was all EA cult drama.
>>
>>110386521
>Huang once berated an executive following Nvidia's a release of its Deep Learning Super Sampling (DLSS) technology, which allows its graphics cards to enhance textures and lighting effects beyond what originally existed in the program. Using a derogatory racial slur for African Americans, Huang asked the executive, who is white, why all of the female characters looked like "hags" in the demo, accused them of not understanding why people purchase video games, and threatened to fire them and make them unemployable in the hardware industry.
>>
why do you faggots care about stock prices? call me when ram is back to 2 years ago prices
>>
File: file.png (465 KB, 549x698)
465 KB PNG
>>110386521
there was another guy like that who terrorized his staff for asking stupid questions.
>>
>>110386594
eurokeks need euros and gbp to cover their positions right fuckin now and they're selling everything
>>
The Riemann Hypothesis was actually solved on November 16th, Results were published on 20th, so we actually solved one of the biggest mysteries in mathematics before GTA 6 came out on November 19th.
>>
>>110386609
feels good to see "investors" suffer
>>
>>110386545
They said it's already on the roadmap.
>>
DUDE GTA 6
>>
>>110386525
it's called liquidity. it's getting harder to borrow money, when that happens assets lose forward momentum and start retreating. markets are heavily dependent on cheap credit and there just isn't any right now.
>>
>>110386609
1. Investing makes money to purchase hardware
2. If bubble pops we get less models
3. RH being solves means we get more models faster
It's all connected
>>
>>110386313
NYT Monday
JENSEN FORCED SONY TO MAKE GTA 30FPS
>>
>>110386609
/lmg/ is literally worshipping the man who caused the bubble as we speak
>>
>>110386567
>everyone with any brain activity
So not investors
>>
>>110386635
<------ you
joke -------------->
>>
https://youtu.be/_FrHwi9ZVtg
>>
>>110386673
bubbles are great when you create the bubble and then sell your overpriced assets to the general public
>>
>>110386666
Thanks megasatan
>>
File: 1795023260007737.mp4 (2.87 MB, 960x720)
2.87 MB
2.87 MB MP4
>>110386313
>>
>>110386609

We care because the only reason this sector has advanced so fast is unlimited investor money, and reason for us getting free local models is Chinks trying to undermine the US market/economy by releasing free competitive products.
If the market tanks and US AI sector is done, it's game over for free Chink shit too.

Now I don't believe for a second that AI is going anywhere, but a massive market collapse that fucks the big players can absolutely change the playing field a lot.
>>
So, what if this Rieman guy is now real, what does it change for me? Are we going to get smarter cheap models out of this? AGI? Space travel?
>>
File: 1771229999973866.jpg (138 KB, 1254x1254)
138 KB JPG
>>
>>110386709
Nothing, every field assumed it was true so everything is the same but now for real.
>>
>>110386726
>Huang reported bullied Nvidia employees into liking and respecting him and coerced them with the threat of withholding their paychecks if they did not complete tasks for him.
>>
>>110386726
The hit pieces are so obviously bullshit, he was loved by his employees, it was kind of insane even.
>>
>>110386072
They didn't?
>>
>>110386745
I would like to hear his side
I bet he'd puss out with some cop out like "out of context, it was horse play"
>>
>>110386709
Slightly better computer algorithms. Potential breakthroughs in quantum physics. Cryptography completely changed forever.
>>
Jensen did nothing wrong.
>>
File: file.png (15 KB, 299x242)
15 KB PNG
>>110386609
im neet and now that AI has taken any chance of having job for the rest of my life i have to make this money last forever
>>
>>110386741
lmao
>>
File: firefox_taV5x5FeaS.png (189 KB, 1336x1180)
189 KB PNG
I made this page to viewing SVG icons, now I go to LLMs and ask them to make 20 icons using this and that format, paste the output into my page and enjoy quick and simple results. I am the best.
>>
>>110386772
We already basically knew it was true. We just couldn't prove it. You could have done all these things before if you had just acted on faith that the RH is true.
>>
>>110386787
Jensen did nothing Huang.
>>
>>110386797
From what I've seen people already did. So many papers based on "assuming RH is true.."
>>
>>110386797
No I meant the application of arbitrarily finding the range of prime numbers. We didn't know how that was done exactly before the GRH solution we have now. This is actually a big deal and a massive breakthrough for quantum physics which is why schroedinger and einstein both tried to solve this during their lives.
>>
File: file.png (304 KB, 620x1128)
304 KB PNG
>>110386745
reminds me of this
>>
>>110386433
this is actually a cool image
unfortunately gemma-31b is the only one not abusively retarded
>>
>>110386797
>>110386818
You misunderstand this is not merely a proof that it's true. This is a solution to the generalized riemann hypothesis which makes it possible to precisely predict prime numbers in any arbitrary set. This allows you to speed up algorithms and was a big bottleneck in quantum physics. Go ask an AI of your choice to explain my answer if you want.
>>
>>110386835
>"I heard someone say some mean words years ago"
>taken seriously
This is so ridiculous.
>>
>>110386525
Deleveraging effectively makes money go away.
>>
File: Jensen.png (3.19 MB, 1920x1456)
3.19 MB PNG
This just in according to NYT Jensen was a bit of a dick
>>
File: P=NP.png (1 MB, 1284x2412)
1 MB PNG
Levent Alpoge now believes P=NP is true and will be proven within a year. This is fucking insane and will mean we will have insane magic technology.
>>
mental illness
>>
Guys! Did you hear the Riemann Hypothesis has been solved? No thanks to Jensen.
>>
>>110386923
see >>110385329
>>
hoooooly fuck shit is CRASHING
>>
>>110386923
and I believe I can fly
I believe I can touch the sky
>>
>>110386923
which stage of grief is this?
>>
I heard Jensen was hiding the truth for the riemann hypothesis because he despised maths
>>
>>110386849
But the thing they verified on lean is a proposition.
>>
>>110386654
>3. RH being solves means we get more models faster
wait how does this work
>>
File: 1788390736291650.jpg (458 KB, 1376x768)
458 KB JPG
gemma-chan won
>>
>BREAKING: A swarm of Gemma agents have reportedly proved the CSAM Conjecture
>>
>>110386991
>have to use antigravity
Not local AND shit
>>
>>110386991
Do cloudfags really?
>>
>>110387003
>>110387003
>>110387003
>>
>>110387004
anthropic chuds pay $40 per million tokens
>>
>>110385583
what's the law being broken? bet there's a bunch of analog devices & many other western companys IP in most warfare drones
retarded foids are gonna set the blissful singularity back i can just feel it. there's no way they won't get in the way
>>
>>110386923
use case but unironically? just name one real world use case that makes my dialy life better
>>
How do I give internet access to Gemma in lm studio? for example to google?
>>
>>110387488
MCP
>>
>>110386060
I ERP in Pi so might not be applicable to other UIs, but I have the model save important events to Qdrant and then my prompt will pull top hits from there to inject into the prompt as memory shards automatically, so I don't have to rely on the model doing tool calls to remember shit.



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.