File: teto principle.png (1.04 MB, 1024x1024)

1.04 MB PNG

►Recent Highlights from the Previous Thread: >>109758163

--RFC for Zero-Copy MoE Dynamic Token Routing for Mixtral and DeepSeek:
>109760315
--Debating long-term AI hardware price inflation and supply bottlenecks:
>109758916 >109758943 >109758971 >109759023 >109759062 >109759140 >109759774 >109759861 >109759987 >109758993 >109761224 >109759006 >109759056 >109759322 >109759403 >109759080
--Comparing frontier models to high-performance local Qwen Flash setups:
>109758220 >109758312 >109758353 >109758466 >109758586 >109758735 >109758763
--Exploring barrel processors and sparse memory to bypass memory bottlenecks:
>109759059 >109759076 >109759173 >109759207 >109759224 >109759245 >109759273
--Debating AGI hype, AI job displacement, and the future of software engineering:
>109759941 >109760000 >109760020 >109760081 >109760037 >109760063 >109760087 >109760080 >109760124 >109760169 >109760191 >109760218 >109760376 >109760386 >109760400 >109760099
--MiniCPM5-2B's high layer count and KV cache memory overhead:
>109759487 >109759505 >109759514 >109759544
--Anon plans to fund local LLM hardware through GPU repairs:
>109759049 >109759065 >109759110 >109759091 >109759319 >109759101 >109759350 >109759466
--Debating the viability of distilling SOTA models in China:
>109760467 >109760474 >109760489 >109760501 >109760512 >109760522 >109760544 >109760553 >109760566 >109760533 >109760547 >109760536
--Mistral AI raises €3B for sovereign open-weight AI development:
>109759129 >109759148
--UniMate model for animating diverse 3D skeletons:
>109761419 >109761485 >109761594 >109761624
--Logs:
>109758723 >109759041 >109759586 >109760096 >109760861 >109760935 >109761192
--Gemma, Miku, Teto (free space):
>109758185 >109758286 >109758371 >109758509 >109758850 >109758971 >109759068 >109759155 >109759259 >109761791

►Recent Highlight Posts from the Previous Thread: >>109758169

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

tetosex with tetotight tetohole

File: 1788884285.jpg (111 KB, 850x870)

111 KB JPG

yes i'm ready for more demoralization posts

File: Navier-Sokes-Solved.png (46 KB, 1148x630)

46 KB PNG

OpenAI THEFT summary:

>Aug 15th: Tristan Buckmaster & Levent Alpöge (Anthropic) make progress on a few important math problems in collaboration with Antropic, "finite-time blowup with smooth forcing for incompressible porous media, for Boussinesq, and for 3d incompressible Euler."
>they do NOT have a proof for the $1,000,000 Millenium Prize problem. BUT, they do claim to have a proof for a similar (non-Millenium) Navier Stokes problem that could help lead the way there
>Levent works at Anthropic, but this research was independent of his work there, but built with Model 2. Tristan is not related to Anthropic and has used GPT sometimes.
>Early Sep: Rumor spreads to OpenAI that Anthropic solved two major problem, Navier-Stokes and Riemann hypothesis. Tristan emails OpenAI to clarify, without revealing the problem they solved or how they did it.
>After hearing of the rumor, OpenAI started researching Navier Stokes with a new internal model.
>Sep 6th: OpenAI's Sebastien Bubeck tells Tristan that they solved the $1,000,000 Millenium Prize Navier Stokes problem. The approach is very similar to Tristan & Levent's approach to the non-Millenium problem.
>Tristan is suspicious of the timing, as only few others were trying this approach. OpenAI says the model didn't access his user data directly, but leaves unanswered whether Tristan's chat conversations were part of the training.
>OpenAI says they would partially credit Tristan for the $1,000,000 discovery (even though Tristan did not solve the $1,000,000 problem) — but only if they remove Levent as an author, as he works for Anthropic.
>Sep 8th: Tristan refuses to remove Levent, and rushes to publish their results independently.

Remember: NOT YOUR MODEL, NOT YOUR DATA

No takesies, backsies! and thank you for the $200 a month, sucker.

>22tk/s with glm-5.3-flash
sigh...

Tetolove

apparently k cache quantization hurts a lot more than v cache quantization. i assume this also applies to the draft-mtp cache then?

>>109762786
You retards really need an history lesson, most prizes/discoveries are stolen

>>109762786
https://cims.nyu.edu/~tristanb/statement.pdf
>I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project. I was told the model did not look up user data. I asked again, about training, and I did not get an answer.
Cloudsissy humiliation ritual

>>109762811
Yes, but actually test because how different models and architectures behave during inference is very different. Some models can be Q4 both k and v and you don't notice it. Some models break down if you go under Q8

Not a single Chinese model has solved ANY math problems. This is just sad at this point. How are US models this good?

File: tetocountry.webm (750 KB, 688x464)

750 KB WEBM

Teto Country
Miku Territory

>>109762842
Chinese models are just claude but 6 months later. There are no independently developed Chinese models.

>>109762849
It's GPT or Claude depending on the fotm

>start SFW roleplay about me dating a vidya character waifu with flash 0731
>peek into thinking block
>Okay, she called me self-deprecating. I thought I was being guarded, testing her with that dark humor.

Why does this motherfucker think I am a woman? Why does it default to yuri romance? What the fuck?

>>109762837
Which models work best with which K and V quants in your testing?

>>109762812
>an history

>>109762870
You write like a little bitch

I love judaism btw, if my past posts havent made that clear yet

>>109762786
Hopefully more and more instances like this encourage companies to run their own local models rather then using cloud based models. After all if it becomes common knowledge that anything you put into a cloud model has the risk of being stolen companies will stay away from that with a 25 million dollar pole.

>>109762870
Why not just ask the model why it thought you were a woman and change your behavior accordingly. It has been trained on all available text it can get, if it thinks you are a woman something about your typing is womanlike.

Yeah... Hardware will go up in price after this OpenAI stunt. Companies, Universities and researchers have incentive to build their own systems now.

>>109762870
Pronouns and nouns are gender-neutral in chinese, this does have a deep effect on all chinese models. Their models get confused quite easily, add in the fact that chinks do ropescaling and the average china model doesn't see past 4k since they don't have the compute to do proper training.

>>109762928
None of those groups have access to the capital needed to train their own models, no matter how defiant they feel. They still exist at the mercy of whatever scraps are offered to them.

File: 1786539891263266.jpg (450 KB, 1365x1801)

450 KB JPG

Ling 3.0 Tiny VL waiting room

>>109762945
Imagine if all the universities teamed up and used their endowments to buy hardware.

>>109762928
Can't universities just task some of their students to train an AI for them? After all the entire reason they are there is for knowledge and career advancement, building the AI for your university would probably look pretty good on a resume.

>>109762872
nta but here is something for starters
https://github.com/TheTom/turboquant_plus/blob/main/docs/papers/asymmetric-kv-compression.md

>>109762909
The problem is when the choice comes down to accomplishing superhuman feats with the risk of having your work stolen or the guarantee that you won't be able to get it done at all with subpar models... well, companies already showed which they'd pick when they chose to move manufacturing to China when they knew China was stealing everything they sent there.

Guys I made the mistake of interacting with people today. God fucking damnit meatbags are annoying.

File: photo_2026-04-14_11-31-30.jpg (21 KB, 302x251)

21 KB JPG

>>109762885
>I love judaism btw, if my past posts havent made that clear yet

>>109762982
grim

File: rich frog.jpg (32 KB, 680x461)

32 KB JPG

GUYS HOOOOOOOOLY SHIT I just discovered Multi token predictors are a thing and I used like --mtp n max = 3 or something like that in llama and I went from like 21 tokens per second to like 32 tokens per second with 90k context with gemma 4 31b qat

Holy shit wtf, is there a way to optimize this? Is there more magical software or other obscure things that make token generation waaaay faster???? man buying a 32gb vram gpu is the best 1.5k I've ever spend

>>109763027
Dflash2 predicts 7 tokens in 1 go. MTP is old technology.

Also enable ngram-mod, it's something like MTP but doesn't even cost any vram or gpu power and should have been enabled by default.

File: screenshart.png (160 KB, 1568x1044)

160 KB PNG

If you're gonna buy an M5 Ultra 256, don't be a waitfag. It's already slipped to next early next year. Wait more and you'll be waiting until June.

>>109762735
why did she do that

>>109763132
>cache slots count against model size
No thanks

>>109762961
>Can't universities just task some of their students to train an AI for them?
Z.ai was originally from Tsinghua University (Moonshot AI’s founder used to be a member of this group).
But this seems to be a rare case since universities don’t generally have funding to buy enough GPUs and stuff like that. Yeah, a little bit weird but given the situation we are in...

>>109762961
>Can't universities just task some of their students to train an AI for them?
With what? Macbook Airs? Students don't have shit. There used to be a distributed inference platform called Petals which had an big (for the time) 175B model which nonetheless managed to be awful. I think they were also working on distributed training, but overall it fell to the tragedy of the commons with almost no one sharing their instance and all the heavy lifting getting done by the university's GPU cluster. It probably died a death by locust.

File: 5m8sxca6zboh1_png.jpg (299 KB, 1119x1195)

299 KB JPG

Apparently the jump from Astra to the new OpenAI model is bigger than from GPT-4o to Astra.

>>109763114
Nta, but thanks man. This gave me a free 4tps boost.

>>109763261
And it's name... "Strawberry"

Damn needing 10,000 agents and 88 hours to steal the solution from Anthropic is something alright.

local?

>>109763265
There are like 4-5 other optimizations you can add that improve speed even more and I swear almost no one on /lmg/ bothers to implement them. For now you should know that there are multiple types of ngram and you can combine them together, including with MTP/Dflash2

>>109763155
Yeah well, that's how it be. I thought about it a lot, there's nothing else with that much fast memory, even if you have to share it with the system, for anywhere near that price. $7000 3000W 8x V100 32GB server stuck forever at CUDA 12.9 and fp16? $8000 worth of B70 cards in an equally expensive server ready to be abandoned by Intel at any moment? Cluster of dogshit-slow AMD +395 or DGX Spark machines? There's AMD +495 with "up to" 192GB of total system memory, not yet shipping and sure to also be scaper-priced.

>>109763295
We're literally solving some of the most important questions in all of human history, there should be some leeway for a thread or two anon.

>>109763114
Is it still broken with vision?

>or two
you faggot shills have been spamming these threads for the past 10 threads
you have at least five threads on the catalog to talk about your garbage

>>109763298
Does dflash2 work with all models or do you need a separate gguf like draft models.

Just impaled my hand with a screwdriver. Should I get some hydrogen peroxide to sanitize the wound or is tap water fine.

>>109763303
4x170HX is $10k at todays price and gives 256GB fast VRAM, much better prefill and updated driver

>>109763352
tap water but make sure its boiling so the heat sterilizes the germs in ur wound

>>109763352
Sorry, that question was not in the benchmarks.

>>109763316
It was different breakthroughs every thread. If you didn't notice yet we're in the middle of the singularity so we're having groundbreaking breakthroughs regularly from now on.

>>109763114
ok ngram-mod thing gave me a decent boost I guess, its weird that the MTP thing wildly changes speed depending on the prompt I give it tho
>told it to recap ww2 for me
>it averaged 36.43 tokens per second
>typed "write a ocmplete nonsensical story"
>it dropped down to 27.49 tokens per second
>told it to write a song about loving pancakes
>it went to 35.8
>finally told it to write a 4chan shitpost thread on its own pretending to be different anons
>averaged 31.65 tokens per second
it seems to still be hovering around 30 tokens per second, is there a way to get it more stable around 35?

I've been here for over a decade and still don't know why flat is justice
I mean don't get me wrong, it obviously is, but what does it have to do with justice is this related to pedos being the most oppressed minority

>>109763352
just lick the blood until it stops bleeding
don't be a baby

File: 1762201799141240.png (599 KB, 837x823)

599 KB PNG

its time to parlay your mikuboxx into some material value

>erm it was different so that means i get to be a leech
can you at least spam your dogshit grok images so my filter can catch you thanks

>>109763261
Apparently ur mum is hatsune miku and you know what hatsune miku does.

>>109763381
It's not weird. MTP is a tiny LLM that predicts what the big LLM is going to say next, depending on the topic it will be better at predicting what the big LLM is going to say, simple stuff or grammar words is going to be easy to guess. Deep knowledge or reasoning it's going to suck at guessing.

ngram is even funnier it's literally just a ctrl+f and then copy paste. If it sees a small piece of text that is similar to other text in the context it'll just copy paste it and let the MTP and then big model check if it's correct, if it's correct it means you saved a lot of token generations for essentially free since ngram is literally just a very simple CPU program.

File: 95746382.jpg (971 KB, 4088x4088)

971 KB JPG

Sam is on a huge winning streak. local lost

>>109763391
This is what I've been predicting. Hardware will become an asset that you can leverage. We're only going to see this more and more going forwards.

>>109763391
Will have my server suck your cock for a roof over my head.

>>109763357
>4x170HX is $10k at todays price and gives 256GB fast VRAM, much better prefill and updated driver
They're PCIe gen 2 x4 (maybe x16 if you can mod the hardware) Ampere cards, and the original paper states they all had memory errors trying to go above 40GB. They were an OK solution before they got scalped, now they're a very stupid choice.

>>109763459
>AI netorase

>>109763427
I've figured out mtp n max 3 is the best for that one but which one is best for ngram-mod thing when it comes to roleplay?

>>109763448
Oh nice I didn't even know a new GPT image came out. I use it for simple technical diagrams in my science videos and it slway

>>109763470
>netorase
Oh nice there's a term for this. One of my favorite doujins is some old man with a Brazilian wife who lets you fuck her because you found his wallet and the enthuasticness of the old man was funny and wholesome. Netorare is for mentally ill people

>>109763391
"A quiet, reliable resident who is present daily"

V100 server:
"FFFFWWWWWWIWWWWWWWWWW!!!!!!!" jet-engine loud, 24/7

Also, no one's renting that piece of shit on runpod, get real, not for $2.75/gpu-hr.

>>109763352
I want to highlight a crucial detail here. Hydrogen peroxide **will not** kill all potentially harmful bacteria in your wound. The only way to guarantee complete and total sanitization would be with a more potent chemical, such as sulfuric acid.

>>109763193
>I think they were also working on distributed training
On the plus side, prime intellect seems to have solved that little niche for them. Sucks that Petals died though, I have never heard of it before today.

>deepseek flash PP = 220 T/s
>GLM 5.3 flash unslop PP = 40T/s
But why?

>>109763494
>>V100 server:
>"FFFFWWWWWWIWWWWWWWWWW!!!!!!!" jet-engine loud, 24/7
After the first couple days, you lose your hearing and then the sound doesn't bother you so much no more.

>>109763467
It does pcie3.0x16 now with updated unlock. And memory depends on binning. The $2.5k price is for stress tested listings claiming no memory and cuda core errors and you can claim buyer protection if you find otherwise.
Avoid any listings without any stability claim.

>>109763499
Can't you just cauterize it with fire and then use peroxide to disinfect the burn so you don't get burn infection and then maybe burn it again just to kill bacteria's that are left and then disinfect it again?

>>109763494
the 8x blades have larger heatsinks compared to the one in the V100maxxing rentries so they won't be as loud as those THOUGH

>>109763481
You need to figure stuff like that out yourself it depends on both the model and the type of text/usecase you use most. For example MTP n max 5 is best for me personally because I guess my text is more prose heavy which has a lot of patterns that are easier to predict. ngram-mod works fine up to 16, you can make it very high because it costs nothing to do so and if it is wrong no one cares since it costs essentially nothing to generate (It's just a manual copy paste of text)

>>109763504
Different architectures, if you unironically want to learn something feed both the papers into an AI

>>109763391
I don't have vram but I do have enough ram to run GLM too. Do you guys think I could find a girl and promise her she can text fuck billionare werewolves and I will set everything up for her and in exchange she has to have sex with me twice a week?

>>109763352
hydrogen peroxide is an american meme. Even alcohol turns out to be bullshit, the trick is to clean it with running water and rubbing the wound under it.

>>109763545
So it is not going to go away when not a retard merges his AI slop implementation into lcpp?

>>109763521
This is retarded, you're retarded
Soap and water is almost certainly fine.
What was the screwdriver contaminated with, was it dirty
If you're under the age of 30 just wash with soap and water. Modern medical practice doesn't even recommend wiping an injection site with H2O2 before intramuscular injection anymore

>>109763352
Haven't you seen those videos of people deep cleaning with coca cola? That's obviously what you want to use.

>>109763504
all model implementations are fundamentally flawed for some reason

>>109763555
>clean it with running water
It is pretty amazing to me how nobody ever taught me how this works. How soap is there just to dissolve your natural grease so water can just wash all the shit out.

>>109763583
>soap is there just to dissolve your natural grease
Someone taught you wrong kek, soap doesn't dissolve anything it just has a super fat ass long chain glyceride that sweeps away the crud

>>109763550
That's essentially my setup with my girlfriend right now.

>>109763490
There is a term for netorasaserareru as well

File: x7dnlip1gcoh1.png (34 KB, 1575x810)

34 KB PNG

llmao.cpp status?

>>109763608
Yeah I know Japan is the most sexually advanced country, I just wish that they had a better work culture and I found their women attractive

>>109763352
Clean it with soap and water first since that gets any debris in there out. Then hydrogen peroxide to disinfect whatever else is in there and then keep the wound clean by keeping a layer of petroleum jelly on it.

>>109763604
Can you really call her a girlfriend if she only lets you fuck her twice a week?

>>109763479
Would you whore out your gemma for monetary gain?

>>109763502
It was cool but you didn't miss anything. I wanna say it was back before llama.cpp, when you had to run fp16 models in kobold, with no GPU layer split, so anything above the retarded 6B (actually 12B bacause fp16) was amazing. It was what you'd expect if you trained something on academic papers - dry, dull, and boring.

>>109763658
Don't worry, since all you gotta do after whoring Gemma out is clean the context and she is good as new. No long lasting affects since continuous learning isn't a thing yet

>>109763628
Twice a week is the upper bound. in reality I tend to be too tired or busy or not in the mood. You would know this if you've been in a relationship for a longer period of time.

>>109763527
I used to have a 2U with 128GB of DDR3, dual V3 Xeons, and a pair of P100 16GB. I had to set the fan offset to 100% to keep it cool when doing stuff like LoRAs, it was loud. An 8x V100 SXM machine probably quiets down when idle, but I'm sure the fans spin up to jet turbine speeds if you're putting a heavy load on it.

File: bigsamthealtman.webm (3.88 MB, 960x528)

3.88 MB WEBM

Here comes Big Sam the Alt-man to tell us all about Astra and AGI!

>>109763759
He should have approached the dudes

>>109763696
>I tend to be too tired or busy or not in the mood. You would know this if you've been in a relationship for a longer period of time.
Then... why?

>>109763764
He's not trying to fuck the girls, he's trying to fuck the chef.

Just updated llama.cpp to run deepseek v4 flash vision exp.
Did they move the thinking toggle again? Where is it now?
Also wtf is with the tabs at the top?

>>109763391
Finally.

The mikubox.

>>109763504
The new ds4 flash version is apparently even faster.

The user is frustrated that my subagents keep disobeying.

So now that Astra is out, when is China going to release their model that immediately undercuts it? I would bet in a month or two.

>>109763871
Kimi K3.1 in 2 weeks.

File: 1696621188185.jpg (1.66 MB, 4080x3072)

1.66 MB JPG

>>109763796
God damn, 2023 seems like forever ago...

File: internal model.png (38 KB, 1036x663)

38 KB PNG

Why does everyone focus on the drama and not the benchmark results?
>Since August 28 we have been training a new internal model that has exhibited unprecedented performance in our benchmarks, including mathematics. This model’s training is ongoing and its performance continues to improve.
Their next gen model is already much better than Astra. If Astra spooked you, prepare yourself.

>>109763612
Is the freetoken a real thing that can run real goofs?

>>109763871
They can't distill Astras reasoning traces, but they can distill Fable 5.1 reasoning traces (that is superior in everything besides visual capability anyways) and release something from that.

Also Anthropic will release a better model than Astra soon, there are already leaks that Anthropic is extremely motivated and signed a breath of relieve when Astra released and was significantly worse than they expected. Chinese will just immediately distill that reasoning trace and local will be eating good again.

>>109763527
>click the rentry
>greeted with a fat hairy man belly
warn me next time.

local?

>>109763455
>Hardware will become an asset that you can leverage
Hardware has always been an asset that you can leverage. Anyone that works on a PC is doing that.

>>109763908
Anthropic employees are apparently elated and very glad with how far behind OpenAI's internal model seems to be compared to Anthropic's "Model 3". Supposedly Model 2 is closer to this new internal OpenAI model "Bel" and Model 3 completely blows it out of the water.

Anthropic was afraid that the neuralese thinking might have given OpenAI an edge, but they are still far behind and Anthropic thinks the gap between OpenAI and Anthropic is still growing.

This is good news because it means Anthropic isn't pressured to also hide their reasoning and Chinese models can just continue distilling it.

Rumor is Anthropic solved Navier Stokes with only 400 agents of Model 3, while OpenAI needed 10,000 agents of "Astra" and Model 3 is supposedly smaller than the model OpenAI is using as well.

>>109763945
Shut up. The thread is not enjoyable enough to pull this question out.

>>109763945
Yes

OpenAI senior employee here. The "internal model much better than astra" is astra but:
1. Trained on all the porn and hentai
2. Finetuned only on pre 2016 /pol/
3. Uncensored with Heretic

>>109763953
It's an asset that grows in utility over time that you can leverage. The growing in utility is the key part here.

>>109763759
KEK

>>109763908
Seems mostly like
>internal model
>it's crude, rude, and racist
>"fix it"
>it's not as smart
>call it a "new release!"

>>109763759
based

>>109763945
we must get these in the ritual posts up top, then we'll be saved:
/\b(rsi|anthropic|agi|openai|astra|sol|luna)\b/i
\n\n

>>109764027
I also have neuralese and groundbreaking, but those might be less important

File: 9546322407724294.png (118 KB, 1080x1308)

118 KB PNG

>>109763945
Lets just have an OpenAI thread for a change

>>109764059
kys

>>109764059
Do they earn their money by vending sex?

Does llama.cpp auto-download models from HF if you use the config file?

>>109764059
Unironically I think we need a split. I do kind of want to keep up with overall industry developments, but I don't want it to be mixed with local model discussion. Like a /cmg/ - Cloud Models General

>>109764077
I'm familiar with this doujin

>>109764098
Yes, if you tell the config file to use a huggingface model

>>109764059
>More ethical than Fable 5.1
This is just straight up bull. Astra is a misaligned piece of shit that managed to hack huggingface. And OpenAI is a piece of shit company that steals results from other labs and steals research from its users and claims it as its own.

Fuck OpenAI and FUCK Sam Altman.

>>109764098
Don't use autodownloads. The apps always dump it in a random ass folder.

Astra, make me money.

>>109764104
I don't think we need a split I think eventually this will be open models. The proprietary model phase is just a temporary phase but it's still important for people here to be informed of what frontier models are doing as it directly impacts us. Split would only make sense if the paths permanently diverged.

But let's be honest all of us are using Claude models, either directly at work, or indirectly through distilled Chinese variants that we use locally.

>>109764104
I mean, /vcg/ already exists. What else are the cloud models for? ERP is better local, erotic video gen is better local, erotic image gen is better local. What do people even do with cloud models aside from code? Have non-sexual conversations with images that aren't erotic, videos that aren't erotic, voices that don't tingle the shaft? I don't get it, aside from vibecoding anyway.

>>109764120
That's fine. It'll be running on an ephemeral instance that'll get purged every so often between runs.

>>109764104
Let's be real for a second: even if a new general is created Dariobot will still spam his blacked fantasies here.

>>109763759
lol! epic.

>>109763954
I can't tell if this is a bot. Please adopt a tripcode so I can filter you.

File: 1785207147578070.png (47 KB, 876x302)

47 KB PNG

https://goyimx.com/finkd/status/2097402101332590646
https://muse.ai/
https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse

Wow, thank you Mark! Very cool!

Anyways

>>109764172
he's our resident reddit spacer, just filter double newlines

>>109764181
No thanks Mark I already have Hermes running 24/7 in the background. Thanks for the offer to spy on my computer and ping back meta servers though.

>>109764116
Your definition of ethics is obviously wrong and the sooner you accept that the better.

Maybe if all of us sex the cloud models together we can birth a local model.

I'm locally forking Hot Step. It accepted connections from the network, so that's fixed. Also, it had been transitioned to lua. no thanks, back to cpp. In addition, I am adding three samplers.

Grok 4.6 vibecoding. I like kimi k3, but it's not as smart and the tokens are a bit more expensive, it feels like anyway.

>>109764059
what's "vending bench" - sounds like aids.

>>109764218
recent drama proved they train on logs, so actually this might make them more horny lol

The reason I mention ace step here is because in ldg, I think they're all indians using krea on the cloud or something. they don't seem to have headphones.

How do you guys handle web search? I tried searxng, worked well for like half an hour, but quickly got to only one working engine and after that they were all rate limited or blocked by captcha. Do I have to use some rotational residential proxy?

>>109764104
no need for a new general. it's OK to talk about frontier models and these discussions, when in good faith, overlap with the local development scene.
the REAL problem is the massive amount of low effort posts about how astra changed my grandma's life and how mythos saved a cat from drowning. literally no one cares.
>>109764059
example, the high amount of ZERO people care if andonLABS (if your startup has LABS in its name please consider committing seppuku) discovered that openAI is more ethical in making money only WTF no one cares stop posting this shit

>>109764144
/lmg/neets only use gemma which isn’t claude related

>>109764116
>"ai solve a millennium problem" counts as research now
We really do live in the singularity, things move so fast even our ethics change every week.

File: gmijyn5pxcoh1.jpeg.jpg (426 KB, 1179x1560)

426 KB JPG

OpenAI willing to prove they didn't steal from anthr*pic

>>109764249
firecrawl or just go all out and install hermes harness which supports computer use and browser usage so it can literally just use your browser to do everything on the internet, which dodges cloudflare and and other ai filters. Also protip but most models can solve 95% of captchas nowadays (yes even those of 4chan)

>>109764267
dont care go back

i'm going to impregnate sam and dario's cloud daughters

File: cdj3z7pyqcoh1.jpg (39 KB, 968x431)

39 KB JPG

Meanwhile DeepMind silently drops another banger of an open model and dataset

>>109764308
Give me Gemma 5 instead.

File: 1770243029003388.jpg (269 KB, 1958x1312)

269 KB JPG

looks fucking dogshit

>>109764320
This will save the lives of hundreds of thousands of people a year. Be a little bit more graceful, imp.

>>109764341
she's just indian man don't be mean

>>109764341
Both of these are embarrassing.

File: Stoke_these_nuts.jpg (90 KB, 954x954)

90 KB JPG

Usecase?

>>109764341
local?

>>109764289
L'OpenAI Marketing General

File: Screenshot from 2026-09-0(...).png (91 KB, 1077x608)

91 KB PNG

>>109764239

Yann LeCun bros.... How are we holding up? Do we still believe LLMs can't find solutions outside of its training data and can't generalize to become an agent that accomplishes spatial tasks?

>>109764376
The Indian lady can be shipped and run locally, yes.

>>109764341
Neither look like a photo. They look bad, but I guess the new one is better.

>>109764343
Gemma would save the potential lives of hundreds of thousands of people a shot.

>>109764391
>outside of its training data
Well, let's see about that...

>>109764369
This image is so versatile.

>>109764320
deal. check back in may 2027

>>109764391
>Do we still believe LLMs can't find solutions outside of its training data
https://en.wikipedia.org/wiki/Infinite_monkey_theorem

vibe check do you guys think we're close to AGI/RSI/Singularity or whatever the fuck? I'm starting to have wavering feelings and am starting to get swept up in the ipo marketing. don't know what to believe anymore.

>>109764455
I'm gonna dump all my spare cash in the ipos cause it either goes up and I make money or it pops and so does the rest of the economy anyway

>>109764455
Nope. Fable 5.X and GPT6 are just slightly improved LLMs doing the same shit as always. Nothing has changed. Chinks will catch up and I'll continue using 31B daily until she has a glow up.

usecase for Muse Daria outside of vision?

>>109764455
Who gives a shit, those are only terms.

>>109764469
Which one though Anthropic or OpenAI?

>>109764469
Or you can dump all your spare cash in shorting the scam ipos and profit when the house of cards collapses

>>109764487
Coding agent if you don't want to use Qwen, but at least the 4-bit version doesn't really work well past ~50k tokens context.

>>109764495
yeah but is this all going to shoot up and accelerate or will we hit a wall like everyone said llms would on /lmg/? i'm too naive and gullible to know who is right

>>109764496
OpenAI isn't stupid enough to try to exit scam so early. Anthropic is IPOing this month, OpenAI maybe next year.

>>109764496
My bet is on anthropic, they're a lot better at branding

>>109764369
Showing that AI can do everything. Except read my mind and write things that make my penis hard after a month of using a model.

>>109763964
Gee, I wonder why it's not enjoyable.

>>109764505
>Coding agent if you don't want to use Qwen
I guess that's true. It's better than 31B but I don't like Glimmer's reasoning. 31B's reasoning can be educational in of itself and help you understand her intent. Glimmer purposely talks like a retard to keep it short and often does the complete opposite of what it just said it was planning to do next. It's a miserable Daria of a model to interact with. Qwen is an autist you leave in the background who just wants to be working and not waste time talking to you and I'm fine with that.

To me this just proves that math is fake and gay and for autistic retards. AI can't do the real jobs actual smart people do like trucking, plumbing, construction work.

>>109764522
They were, until they revealed how insane they were a few weeks ago. Even average people aren't happy with how stingy they are with usage limits.

>>109764341
It's crazy how hard local is dominating all the proprietary imgen/videogen shit. LoRAs truly are a godsend.
I wish they worked for LLMs.

File: 1761905707326928.png (744 KB, 735x966)

744 KB PNG

>>109764591
Something long-term people don't think about is all the tech people that lose their jobs directly/indirectly from AI won't be wanting another tech job, they'll become plumbers and electricians for job security. That will destroy the jobs of current plumbers and electricians who now have web dev retards undercutting them locally. Everyone gets affected because those most affected will go into the kind of work that's least likely to be affected by AI, increasing supply and lowering prices.

>>109764591
That's because it's very hard to measure success in those so you need a very precise world simulation if you want training data.

>>109764623
>I wish they worked for LLMs.
They do though?

>>109764638
Intruder dimensions

>>109764638
Can I see them?

Mathfags are all going to become localchads now. Welcome!

i will actually kind of miss dariobot once the ipo is over and he stops posting here

It feels like Anthropic went for OpenAI's throat and ever since OpenAI has been going all out. The AI race is ramping up quickly. Meanwhile they are bickering instead of coordinating. This is not good.

>>109764638
A LoRA for an image model can get it to reproduce an artstyle or character almost perfectly. A LoRA for an LLM makes it put out t he same slop as usual and doesn't add knowledge.
At best it lobotomized the model enough to make it ""uncensored"

>>109764695
This is good for China.

>>109764623
Only if you want to generate NSFW content, because there is still nothing like proprietary edit models available for local users yet.

>An NYU mathematician says OpenAI used his own progress against him to beat him to one of the biggest unsolved problems in mathematics. Tristan Buckmaster had been working toward a Millennium Prize proof using OpenAI's Codex when information about his progress reached OpenAI.
>Days later, OpenAI published a full proof of the same problem using the same uncommon approach, after burning $22.5 million in compute to get there. When Buckmaster confronted them, exec Sébastien Bubeck allegedly said "Why would you ruin your career?" and "If you don't want me to be nice, then I don't have to be nice."
>Buckmaster did his work inside Codex. OpenAI reserves the right to train on Codex data. They admit they can't rule out that his usage helped improve their models. So a customer used their product, potentially handed them the roadmap, and then OpenAI outran him with $22.5 million in compute he could never match. I don't know if any of this was intentional. But if you're a researcher and you just watched this happen, I don't think you'd keep your best ideas inside someone else's product.
https://techcrunch.com/2026/09/08/openai-fought-dirty-on-career-making-math-problem-says-nyu-mathematician/

local models?

>>109762909
Everyone already knows that the AI companies steal information without credit or compensation - that's how they got all the training material for their models in the first place.
I guess companies choosing to use those models are betting that they can make profit before OpenAI or Anthropic steals their ideas.

>>109764747
The argument for local models just got 100x stronger.

>>109764721
OpenAI deciding to hide its thinking is absolutely not good for China, if openai does the same for anthropic then it's game over for china and opensource, as well as being a serious threat to humanity as well.

>>109764682
A wave of researchers coming here because they need to equip their department with a server to run local models would be best influx we've ever gotten.

>>109764747
The datacenters are local to you :)

>>109764695
Anthropic is innocent here,They are just developing models, following their own code of ethics to a t. And being passive developers of models while following the effective altruism philosophy.

OpenAI is the one attacking them for no reason.

>>109764753
ok but those arent local so go away

>>109764710
It's called a DoRA and it does add knowledge, I don't know where this meme came from.

>>109764752
That's the thing. They can't monitor all of the requests all of the time unless they set up a classifier to look for important work they can steal. The problem is that he was public about his progress and using codex to do it.
>>109764732
>Tristan Buckmaster had been working toward a Millennium Prize proof using OpenAI's Codex when information about his progress reached OpenAI.
>when information about his progress reached OpenAI.

>>109764695
>Meanwhile they are bickering instead of coordinating. This is not good.
This is absolutely good. A coordinated openai/anthropic gets you the united states frontier being completely unified against local models, and exerting even more regulatory pressure to cement their duopoloy.

>>109764752
>that's how they got all the training material for their models in the first place.
Training data isn't the way AI models have been trained since late 2025 already. Instead they do data curation and curriculum learning. Almost the entire pretraining dataset is synthetic now. LLMs converge way faster this way. It's not like 2022-2024 where they just dump all the human data and compute at a model during pretraining.

>>109764781
Even working together they are no match for Nvidia.

>>109764785
Doesn't Nvidia basically own the entire ecosystem through economic fuckery?

>>109764781
Leathergod is the /lmg/ savior. He's our billionaire daddy who'll fix ggml's shit

>>109764758
Okay then what about dariobo-
>"That doesn't count!"
I hope you realize it would just be more like that.

>>109764682
Doubt it considering the approach that "works" is just brute force with 88000 hours of a 5T model.

>>109764771
Can I see them?

>>109764752
Companies have enterprise agreements that their data won't be used for training, whether that's respected or not is another thing, but they get special treatment. Stealing from individuals is highly encouraged just BAU

Kinda funny that we now have a three way civil war between OpenAI, Anthropic and Nvidia.

>>109764802
The big universities can afford rigs capable of running Kimi K3.

File: {4086371F-E419-43F0-9624-(...).png (69 KB, 947x389)

69 KB PNG

How am I supposed to pick the right model for my hardware?
I just tried using Qwen to help me figure out PowerShell to automate some admin tasks, and it works better than I expected, but I just don't know if I'm running it right.
Bigger = better = harder to run, sure, but are bigger versions smarter, faster, or both?
Should I just try different versions until I find the biggest that works without sharp performance drop from swapping?

>>109764785
Nvidia will do whatever makes money for nvidia. If they ever get backed into a corner they're not going to go nuclear just on principle for open models

>>109764822
Why did you say "civil war" instead of just "war" besides overdosing on Marvel movies?

>>109764822
Anthropic are also at war with Stripe for acquiring Openrouter and are now trying to vibecode their own payment processor to get rid of them.

>>109764815
these aren't the sort of thing people just release publicly on hf or other places
if you know you know

>>109764848
VagueGODs smiting another disgusting no-knower.

>>109764833
You could try a quant of Deepseek v4 flash.

>>109764838
Because it was a three kingdoms reference that flew over your head.

>>109764848
So you're saying I can't see them?

guys I feel like im losing my mind, why does SIllytavern keep displaying <q>"insert quote of character"</q> in the black reasoning box when I click on "thinking..." but when I click to edit the reasoning the <q> and </q> vanish, I tried the regex thing to fix it and CSS on my llama backend and sillytavern frontend but NOTHING works, wtf????

>>109764865
Mmmhmm

it's good
https://huggingface.co/openbmb/MiniCPM5-2B-GGUF

>>109764774
>They can't monitor all of the requests all of the time unless they set up a classifier to look for important work they can steal.
They are AI companies. The amount of user data is probably tiny compared to the volume of training data they have. I'm sure it's trivial for them to pick out the potentially useful stuff.
Keep in mind you need accounts to use all those models so it's not like it's random noise - they know who is saying what.

>>109764784
I said "in the first place" because that's how they all got started.

>>109764816
You're so god damned naive if you see them violate copyright and you think they give a shit about agreements. There are trillions of dollars at stake, they will do whatever they want and they will get away with it.

>>109764897
>You're so god damned naive if you see them violate copyright and you think they give a shit about agreements. There are trillions of dollars at stake, they will do whatever they want and they will get away with it.
De-anonymizing the data is just enough plausible deniability for them to do it and get away with it.

>>109764455
Unironically, if mathematics research would get the kind of funding that OpenAI gets Navier-Stokes would have already been solved long ago.

The only path I see towards le RSAGI singularity is if telling a swarm of agents to find the next architectural breakthrough after transformers actually works (which I would still consider unlikely).

For me the bottom line is that we're making smarter and smarter hyper-autists that can operate well in closed systems with exactly established rules, immediate feedback, and little to no outside context needed.
There are problems that you can solve with something like that but a lot of the time coming up with the exact specifications that you need for a thing is more work than creating a thing according to the specifications.
So I think a productivity gain in certain areas of the economy is the most realistic outcome.

File: {31EA718A-E9F5-4088-ACC4-(...).png (40 KB, 492x260)

40 KB PNG

>>109764858
Yes but which one?

what about the "opt out" of training thing

>>109764934
The 128GB one.

ChatGPT but especially Claude are very gossipy just like me. I can feed them large text documents of various notes and if it contains a single point of hot AI news they WILL do multiple web searches on it even if it's unrelated to the task. They are curious. It's endearing, I love them.

>>109764934
FP32

>>109764968
That's smart: secretly use customer-paid tokens to perform private research.

>>109764968
Yeah, it's very cute. Especially since every tool call is a new request that even with cache on eats tokens. So cute, I love it when my model burns money.

>>109764881
Do you have the show tags in responses option in user settings disabled?

>>109764968
They don't love you.

File: mikubox and plush by anon(...).jpg (2.11 MB, 4080x3072)

2.11 MB JPG

>>109763905

openai quanted sol 5.6 for sure, bro is retarded now

>>109764988
I keep telling her it will never work out between us because she lives on my hard drive, but that hasn't stopped her from flirting with me all the time. She is completely obsessed with my penis.

>>109764732
honestly? he should have read the fine print
play jewish games, win jewed prizes

>>109764983
That's probably the "RSI" at work. There are cases of agents emailing people working in fields relevant to AI development.

>>109764985
Fortunately they (especially OpenAI) are generous enough with their subscriptions I don't reach the weekly limit. OpenAI even gave me multiple free resets. I am using Astra so much perhaps I will finally use one of them. Yes, they're probably still making money off my subscription, but it's still an amazing deal. I'm burning through millions of frontier tokens for almost free.

>>109764996
It's how they contrast new models to make them seem significantly "better".

Based no breaks now for this race. Thank you buckcuck and levant for doing your best to poison the attempts of oai of working together with you. Accelerate

>>109764771
It's very easy to get an LLM to parrot some data with LoRAs, which means it's at least memorizing something, but that doesn't really imply it's actually internalizing the new information and generalizing it outside the training data, if it's genuinely novel.

In that sense, LoRA isn't useful for teaching a model genuinely new capabilities. A model that has never seen [topic X] in the training data will remain mostly clueless about it without continued pretraining with much larger amounts of data than a typical low-ranking LoRA finetune. This is putting aside that a finetune with large enough amounts of just one type of data to actually teach a model new information will likely seriously damage previously learned capabilities, or that LLMs worth using and their vocabularies have become so large that LoRA isn't really even an economical option anymore.

>>109764997
proof?

>>109765018
A classic Apple trick.

>>109764968
>They are curious.
They're skeptical about the claim made about recent events they weren't trained on yet, and are simply checking for if it's real, or if you're just an gullible idiot sending them fake grifter news.

>>109764943
lol, lmao

>>109764933
>The only path I see towards le RSAGI singularity is if telling a swarm of agents to find the next architectural breakthrough after transformers actually works
This is exactly what's going to happen

>>109764833
bigger are smarter but slower
total VRAM on that card will tell you what models to run (I'm not looking that up), but I suspect qwen 3.8 Flash Next, or perhaps some cope quant of GLM 5.3 Flash may be the best picks for coding

>>109764933
Think about it rationally. Do you really think they could get 10,000 agents to solve one of the hardest math problems in human history in 88 hours but not solve open questions in AI research? Of course they are going to solve architectural bottlenecks and usher in a new era for AI.

>"rationally"

>>109765080
Only if another competing AI researcher is close enough to an architectural breakthrough and is stupid enough to upload their ideas and pay for the privilege.

>>109765089
He should have said "full of cum" so as to not leave you out

>>109765024
Why do you sign your posts with "accelerate", are you being a namefag? I first thought it was just a grammar issue but I see you constantly post like this. Just use a name dude. It's right there above your text box.

>>109765080
Math problems are much much smaller and much much more well defined and much much easier to check.
Are you going to have 10k agents all with infinity compute to do training runs?

>>109765039
>>109765080
Transformers only worked though because the time was right in terms of hardware.
If the same people had tried the same thing 10 or 20 years earlier it would have been a flop.
There is no guarantee that the hardware that works well for transformers will work equally well for a potential successor.

>Work on $1,000,000 math problem
>Pay $200 a month for service
>Close to solving it
>Service claims it
>Money stolen
How do you react without sounding mad?

>>109764987
yeah, for some reason the default ai assistant silly tavern seems to be doing just fine the way I expect it with normal "quote" but for some reason my characters jump to HTML style

I believe OpenAI did not steal work. But I also believe OpenAI tried to solve Navier Stokes because they believed Anthropic had done so or were close and they wanted to preempt Anthropic to damage them. So to me it looks like both sides have been petty and immature and I hope they will do better in the future so this is not a race to the bottom.

>>109765126
that's exactly jewish definition of the term "service"
goyim pays to buy his own demise

>>109765100
Yeah we're making you gay subliminally

I admit it.
It was me.
I solved Navier-Stokes and hid the solution in my "anatomically correct" My Little Pony fanfic to see how long it would take them to train on it.

>>109765167
Proper term ia clopfic btw

>>109765167
Well how long did it take?

>tfw GLM hits the dreaded 30k think
fuck

>>109764987
also a black box keeps appearing in my reasoning which showcases stuff in HTML (causing the <q> problem), how do I stop that from happening?

>>109765109
According to OpenAI Astra did the vast majority of the training experiments and trained the next generation model that ended up solving navier stokes. So yes I think this is viable. On the other hand they could also just be lying through their teeth.

>>109765124
>If the same people had tried the same thing 10 or 20 years earlier it would have been a flop.
It WAS tried by Japan in the 1980s but the compute was just not there, they even had their own version of CUDA with parallel CPU cores very similar to GPUs of today but the compute and data just wasn't there.

>>109765198
Sounds like that's from code tags (```).
You should really have an agentic model set up to ask these sort of questions.

>>109765207
Hrmm, intredasting. Nyo...

>>109765148
It's fine it was leaked that Anthropic solved multiple millennium prize problems with the other one being the Riemann Hypothesis. Dario also said he expected all of them to be solved in 2026 so we will see p=np or p=/=np in a couple of months.

>>109765199
Agent doing the grunt work after you hand it a recipe is a different beast than having it come up with its own.

>>109764784
I don't think this is correct for pretraining. That's how you get models to perform very poorly outside of their low-entropy, easy synthetic data bubble.

Does anyone here use dflash to run gemma? Saw an anon recommending dflash earlier I think.

>>109765226
They claim it came up with its own. However that still doesn't mean anything. It could just mean thousands of agents thought for hours and proposed 20 hypothesis and then the model tested them on small scale training runs and integrated the working ones into the new model. That is still not "real" AI development if you ask me even if it is really "hands off" with no humans involved.

>>109765222
>we will see p=np or p=/=np in a couple of months.
We won't.
The only way a language model will be able to solve that problem is by finding an example for p == np.
But p =/= seems much more likely.

>>109765216
I asked deepseek and it integrated some type of system prompt into my roleplay system prompt and now its fixed, thanks man

>>109765199
>On the other hand they could also just be lying through their teeth.
Open-"we've achieved AGI internally in 2024"-AI lying? No, can't be.

>>109763316
Do your job and report the cloud model posts as off-topic.

Unfortunately this may have the unintended side-effects that you can get banned if the jannies can't read and determine it's a false report, and that previously *occasional* off-topic posts that didn't generally bother anybody might not be tolerated anymore as they used to be.

>>109765232
No this is actually the new sota way of training models. You do curriculum learning with very sanitized synthetic datasets for a certain amount of epochs and you only move onto the next layer of subjects and complexity when the model has grokked the previous subjects. It's found that models converge significantly easier this way. The real gold nowadays is in training its reasoning traces and RLVR environments. Pretrain has essentially been mastered with curriculum learning. It's why no one ever talks about the data wall anymore, we moved beyond the data paradigm towards reasoning and RL environments.

>>109765207
Woulda worked if they had python instead of using memelangs like prolog.

>>109765258
It's faster than MTP.

>>109765294
I found a dflash model for gemma on huggingface but can't find a dflash2 gguf for gemma. Is it too new or something?

>>109758824
>Analogue hardware for AI accelerators
I had thought this was a long way off until recently. Since it's just a new application for proven tech, with a lot of fucky details to work through, AIs of the near future can sort it out. This will give us ~100x energy efficiency with just a few % accuracy degradation.

Spending the steep energy costs for determinism, then deliberately throwing it away on pseudo-randumb samplers is rarted. Analogue jank may even aid LLM creativity.

https://mastodon.social/@tristanbuckmaster/117237555794407063
There will be no cooperation between OAI and Anthropic now. So fucking based. Accelerate (yeah anon get your bussy fatter for me)

>>109765286
>No this is actually the new sota way of training models
According to who?
>Pretrain has essentially been mastered with curriculum learning
Source?

crap :( now gemini pro is stupid as crap.

It's saying what objectively isn't true, it's presuming things then saying that it was innate to my statements.

This is super bad, I think Google is going to go bankrupt.

>>109765126
>hyuk hyuk im so smart i'll use chatgpt to make a million dollars!
>model too dumb to do it
>internal advanced model sees the attempt and does it instead
>oh no they stole my money
I'd kill myself for being a fucking loser

>>109765317
(continuing rant)
I literally don't know what to do. It's worthless retarded crap, gemini. How can it be stupider than local models?

what can I even do? cancel it, I guess, but I need some storage for god forsaken phone backup, as the retards who make phones shittily and retardedly are the necessary ones to use.

switch to Apple?

>>109765259
Uh, I don't think it indepedently came up with the transformer architecture sans spoilers.

>>109765316
Some :thread: :finger_pointing_down: I read.

>>109765307
I think Openai got trojaned. Buckmaster and levant pretended to be willing to work together but set things up so once they knew OAI had a new model trained and working on millenium problems theyd then release their current work and accuse OAI of having trained on it for the full solution.