►Recent Highlights from the Previous Thread: >>103881688

--Paper: SPAM: Spike-Aware Adam with Momentum Reset for Stable LLM Training:
>103886784 >103886926
--Paper: Transformer^2: Self-adaptive LLMs:
>103886931 >103887964
--Papers:
>103886689 >103886794 >103887023
--Test-time compute storytelling and the role of model size in creative writing:
>103884801 >103884814 >103884855 >103884873 >103884914 >103885026 >103885107
--Relationship between model size and quantization sensitivity discussed:
>103881791 >103881850 >103881898 >103881964 >103882008 >103882183
--Discussion of DIGITS and PC building options, with a focus on memory bandwidth and performance:
>103883782 >103883785 >103883880 >103883904 >103883934 >103883963 >103883986 >103884107 >103884142 >103884150 >103884596 >103883999 >103884009 >103884136 >103883825
--Discussion of Mac and DIGITS systems, memory, and GPU capabilities:
>103882515 >103882577 >103882642 >103882740 >103882781 >103882884 >103883056 >103883116 >103883142 >103883315
--Discussion of AI models, GPU performance, and optimization strategies:
>103884724 >103884770 >103885024 >103886084 >103886381 >103886954 >103886979 >103886993 >103887024 >103887252 >103887019 >103887037 >103887132
--Speculation about Nvidia's mysterious repository on Hugging Face:
>103884597 >103884660 >103884687 >103884712 >103885202 >103884910
--FP8 vs Q8: data types, precision, and information loss:
>103883157 >103883204 >103883239 >103883245 >103883249
--Phi vs Llama for finetuning discussion:
>103884786 >103884861 >103884972 >103885004 >103885793
--Anon shares anonymous-chatbot response explaining lolilibaba concept:
>103883568 >103883604 >103884750
--UGI Leaderboard evaluates language models' ideological leaning and neutrality:
>103883290 >103883625
--Miku (free space):
>103883015 >103884150 >103884327 >103884720 >103886919 >103887221

►Recent Highlight Posts from the Previous Thread: >>103881693

Why?: 9 reply limit >>102478518
Fix: https://rentry.org/lmg-recap-script

>Let's try this "MiniCPM"
>1 hour later
>Still building flash attention wheel

Yes I installed ninja

Tetolove

>>103888594
I love you, Recap Teto.

are any of these good enough to use with a frontend ai chatbot with reasonable speed? I have a 3090, I'm guessing most of you do something similar? I last tried llama3 and it was impressive for running locally and free but fairly bad compared to chatgpt and other online models

>>103888840

Local models are usually all caught up to SaaS in at least one area, but there aren't really any models that accel in all areas like a lot of the corpo models do. You gotta pick and choose a model for your niche. Llama 3 is kind of the exception there in that it's aggressively mediocre at everything

what would you use if you had two 3060 (12 GB each) and 32 GB VRAM? for either RP or other things

>inb4 jokes about how the rig is shit, I guess

>>103888886
A Qwen 32b based model at Q5 or something probably. It's not the worst ending.

>>103888886
>and 32 GB VRAM
Nice 5090 + 2x 3060 setup.
On a more serious note, I'd go with Cydonia first. See if you like it.

>>103888885
damn, almost neato digits & thnks for the info, Ill investigate more tomorrow. all the gooners in aicg dying over jailbreaks and proxies when lmg might be the answer

Are there models that can translate as in are those specialized or any model can? I'm not really finding anything atm(found something from 2 year ago though)

bitnet millions of experts 70b when?

>>103889221
By being attractive.

>>103889251
*audibly pops the magic bubble*

>>103889317
GLGLLUGLLGLRHH

>>103889333
Noooo :(

>>103888886
Do you mean 32GB RAM? Or 56GB VRAM total?

>>103888893
examples?

>>103889140
Take a look at some popular models.
See what languages they have been trained on.
They should be able to pretty-much translate text between those languages.

>>103888893
>>103888919
Thank you. I'll try them both out.

>>103889448
Yeah, I have 32 additional VRAM that was soldered on by some dude in Shenzhen while I had noodles

>https://huggingface.co/openbmb/MiniCPM-o-2_6/tree/main
I hate the demo because I get nervous talking with female voices

https://web.archive.org/web/20250114121236/https://www.theregister.com/2025/01/09/us_weighing_global_limits_ai_exports/
>"Along with compute caps on tier-2 nations, the rules may also include limits on the export of closed AI model weights. Model weights represent the numerical values that dictate how modern AI models function. Under the proposed rules, the Commerce Department aims to prevent companies from hosting closed model weights in tier-3 countries like China and Russia. Such a move would prevent major closed-source models from being served from these nations. Open models, like Meta's Llama 3.1 405B, would not be subject to these rules, nor would any closed model deemed less sophisticated than an existing open model."
>"nor would any closed model deemed less sophisticated than an existing open model."
What are they trying to do here?

>>103889710
My guess is preventing potentially hostile countries from having privileged access over powerful AI models.

Titanpill me RIGHT NOW

https://arxiv.org/pdf/2501.00663v1

>>103889859
>open paper
>see math
>close paper
simple as

>>103888589
https://youtu.be/OSKgz8NfUoI

>>103889859
I'm a retard when it comes to math formulas, but at least understand the terminology, and if I get it right, the basic idea is:
Models we use these days function best as an equivalent of short-term memory (which is why they get dumber with longer contexts). This architecture involves basically having a meta-model with additional context: a long-term memory that evaluates how "surprising" or "memorable" something is in its context, and feeds that data to the core model that operates on a smaller context (they keep using the phrase "learning to memorize at test time", but I see nothing to suggest any moving parts, so to speak). In other words, important details should be preserved from a much larger context (they claim it can reach 2M context), influencing the short-term memory, and in turn influencing the output.
It seems theoretically solid. Long-term memory is one of the aspects of LLMs that badly need a breakthrough, and this seems like a viable approach without requiring dynamic data on the user side.

Are non-dense models anti-local because they tend to require corporate-tier amounts of VRAM and substantially less compute, making them optimal for SaaS deployments?

>>103889997
no

>>103889859
Sorry, you said pill, not explain.
Well, if the results can be trusted, this basically cracks the problem of attention dilution over long contexts wide open. Finds the needle in the haystack near-perfectly. You know that important detail you mentioned exactly once in your RP some 10k tokens ago, that gradually got washed out until it was completely forgotten? This solves that problem.

>>103889960
cftf?

>>103889997
Not necessarily, Mixtral used to be a VRAMlet friendly model because as long as it can fit in the RAM and the active parameters aren't too many, it can ran at decent speeds

>>103890018
tl;dr loredumpfags rejoice?

>>103890023
>responding to the corpo market research bot
bruh

>>103889997
There are ratios of total size to expert size where it is actually better for local since you can use your regular ram to get a more parameters at a small speed cost.

I tried getting the the local demo for MiniCPM-o 2.6 working but there were too many problems after the other to make it work well.

Seems solid otherwise. Might make a fun bantz buddy when gaming or something.

>>103889859
just skimmed it. feels more than a meme paper this time. (he said)

>>103890019
go back

>>103889965
>>103890018
>google
See you next year when an open model implements this.

>>103890023
Oh, I see, thanks for explaining.
>>103890079
Nice.

I want a model or finetune that understands memes and culture.

>>103890151
Claude. Heard it can even talk like a zoomer if prompted.

gemini didn't like my strategy for tsunamis

>>103890199
>missing the meme and taking it completely literally, at face value instead
At last, artificial autism.

They should make a RP leaderboard of sorts entirely based on Nala test. It is just such a good test for so many reasons and filter shite

>>103890295
It's actually a completely retarded meme though.

>>103890295
I agree.
You evaluate the responses based on a couple of categories such as anatomical understanding, character adherence, etc.
Do 5 swipes for each model and take the average of the best 3 or something.

>>103888589
cute teto

>>103890163
A local model.

>>103890474
DeepSeekV3

llamiku 4 when

>>103888589
>MiniCPM-o
Has anyone got this working locally with it's functionality intact? I couldn't get it to work.

I'm going to miss lmg-anon...

>>103890770
He'll be back in 3-15 years, no worries. Just in time for llama4.3-70b

>>103890770
So what did he do exactly? Say a woman should suck his dick cause his PC runs pytorch and a woman actually did it?

>>103890770
Which tag do I use for the middle one's tit shape?

Has anyone tried Sky-T1-32B-Preview?
Allegedly it is like qwq but less buggy and better at programming.

>>103890626
>I couldn't get it to work.
Would you like to be more specific?

>>103890911
bestiality

>>103891015
I don't get the joke.

>>103890923
it's a model fine-tuned on qwq outputs, i doubt it's any better than it

>>103891030
It should be better but not much better.

I doubt thats this is even lmg-anon but that statement is not true right? That sounds crazy.

Updated Silly Tavern on single board computer to include maintanace items, and get it to monitor and re-connect to wifi if dropped.
https://rentry.org/SillyTavernOnSBC

>>103891132
Wrong pic, meant to post this.

>>103891147
>having friends
ngmi

>>103891147
>>103891132
i wouldn't doubt it for a sec. korea is a corrupt hellhole.

>>103891147
WHO

>>103891147
Since you can get imprisoned for a few insults on LoL, I'm sure that's true.

>>103891235
some locust enabler

>>103888589
Hi bros
Im a bit out of the loop on the memesamplers, can someone spoonfeed me a good value of smooth/dry or xtc for largestral?
normally i dont mess with them, but i remember smooth being kinda nice, and after doing a multicharacter card and the model giving VASTLY more attention to a character with a common name, i think i need to use them

>>103891147
That's North Korea, right?
Right?

>>103891273
;)

>>103891147
The entirety of South Korea will disappear in 3 generations. He'll get the last laugh, might even be alive for it.

>>103890220
No. Catching the meme is the autism in this scenario, anon. Imagine going up to some rando math professor and being like
>"ehehe... gotta go fast, bet"
He'll think you're retarded. Same with gemini. It's just not allowed to say it.

>>103890770
>>103891147
I thought I forgot to change tabs.

>>103891273
Why is it so weird? They probably were accomplices to some degree for whatever he did.

Kill yourself.

>>103891333
is that kanken ni-kyu? Impressive.

>>103891333
im jealous

t-thanks grok. cant wait to have that power locally soon.

>In South Korea, defamation laws are particularly stringent, as evidenced by the information provided in the related web results.
>Defamation can lead to criminal charges with potential imprisonment up to three years if the information is true, and up to seven years if it is false.
>This is highlighted in the context of South Korean cyber defamation law, where even true information that harms a person's reputation can be punishable.
>This legal framework is different from many Western countries where defamation typically results in civil rather than criminal liability.
>Specific Allegations: According to Air Katakana's subsequent posts, the person who reported him to the police is a professor at KAIST named ******* Lee.
>Air Katakana alleges that this professor forced him to send a large amount of money to his wife under the threat of revoking a job offer that had been promised over a year prior.
>This suggests that the defamation might be tied to these financial and employment-related disputes.

>imprisonment up to three years if the information is true
I seriously hope that's just hallucinated up. 3 years for posting true stuff somebody doesn't like. wow

>>103891478
This is something air katakana wrote in one of his tweets, I think it was irony and the AI took it as a fact.

>>103891569
It's on wikipedia

>>103891478
I'd assume that's to discourage gossip and feuds about petty stuff, but it's ridiculously phrased.

>>103891593
>ridiculously phrased.
Ridiculously implemented, rather...

>>103890626
the vision model works ok for me

>>103891137
>https://rentry.org/SillyTavernOnSBC
Nice, but it needs a miku pic anchor

>>103890518
A local model I can run. I'm not a poorfag either. I have 64gb ddr5.

>>103891698
"understand memes and culture" is too broad. what do you want to do exactly?

>>103891478
You have to understand that all of Korea was nothing by illiterate farmers just a couple generations ago. Unlike the Japanese and, at least the historically urban parts of, the Chinese, they don't really have a tradition of a stable and functioning modern civilization.

>>103891721
If you have to ask, your judgement won't be useful to me. I can tell based on the tone of your post that you're a pedantic shithead.

>>103891802
nta, but jesus fuck, anon...
>being pedantic about memes and "culture"
>I'm not a poorfag either.
>I have 64gb ddr5.
>pedantic shithead.

>>103891132
>I doubt thats this is even lmg-anon
he seems like a pretty cool guy. Does he actually hang out here? We could swap JLPT stories

>>103891698
>I'm not a poorfag either. I have 64gb ddr5.
gigacope

>>103891668
You're right.

How exacly one start with this? I`m on linux.
From the rentry tutorial it says to download oobabooga, their github says to start_linux.sh.
Then I went to https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/
But couldn`t find a download link to any project. What after this?

>>103891904
By waiting 2 more weeks for better models.

>>103891904
copy-paste the name of the repo into the model download box in ooba. after that, load the model and done

>>103891802
are you a toddler that expects to be spoonfed with minimal communication (crying)?

>>103891904
>memeboard leaders these days are 78(!)B frankenmerges of Qwen-72B
I hadn't looked at that cesspool in a year. Good to know it hasn't changed since then after it got flooded by chinks and indians training meme models on benchmark data.

>>103892037
we need better benchmarks

how are speeds benchmarked anyway?
T/s just refers to output speed right.
Is there a factor x where processing input is faster as generating tokens or are these two unrelated. Also speeds seem to differ quite a bit by not only size but model and datatype.

>>103892241
Most backends show you processing, generation and the total time each in t/s.

>>103892037
>>103892080
open-llm-leaderboard evaluates on non-CoT. Meaning, they don't let the model generate a full solution, and then search through it and extract the answer, but rather directly check the probability of the answer in the first few tokens. That's why qwen2.5-72b scores higher than qwen2.5-72b-instruct, even though instruct is a much better assistant (which this benchmark is trying to evaluate).
Someone correct me if I'm wrong.

I'm following the "getting started" guide. It's telling me to "download nemo 12b instruct gguf". Where do I find this?

>>103892479
Nigga, this isn't spoonfeeding at this point, it's giving you knowledge in a fucking IV line.

>>103886370
>deepseek repeats too much
using --chat-template deepseek3 with a recent llama.cpp has eliminated any repeating for me. I don't think I've seen it once.
The only other variable is that I self-quant, so I can't speak for any online ggufs if those are the core issue

>>103892479
In the off chance you aren't trolling, have you tried writing "nemo 12b instruct gguf" in the search bar that's clearly visible in your image?
If not, try that.
You'll see a bunch of results, probably, download the bartowski one that has mistral in the name.

>>103892551
Doesn't "Vikhr" indicate that it's Russian?

>>103892479
Use this https://ollama.com/

>>103892574
This is like giving someone a crackpipe.

>>103892569
The one with mistral in the name anon.

>>103892574
This is the one time recommending ollama is ok

anyone ever use the cpu pinning feature (not --numa) in lcpp? it doesn't appear to obey the strict flag or pinning mask at all and my inference just gets slower, which doesn't make sense to me.

>>103888618
>>Still building flash attention wheel
It do be like that.

Did they cheat with Sky T1? Or is it actually comparable to o1?

>>103890518
Ok, now one that can do it for more than once message (since deepseek will come up with something good but then just repeat parts of it every message forever.)

dead on arrival
doesn't even know what a migu is.

>>103890295
Once you make something a benchmark then it becomes useless since they'll optimize for it and the test isn't representative of true performance.

>>103892824
Not exactly a fair test being out of focus and off-model.

>>103892808
>deepseek will come up with something good but then just repeat parts of it every message forever.
you're doing something wrong

>>103892534
post a log with 5 bot messages

>>103892858
>bot messages
wtf does that mean?

>>103892824
There's only so much a benchmaxxed 8B can do.

>>103892854
Don't think so, Chinese models just have a lot of shills. People said qwen was good too and I didn't like it either.

>>103892870
assistant messages

>>103892849
The whole purpose of machine learning is to create emergent, out of distribution capabilities. It's a perfectly fair test. >>103892871
Still pretty impressive level of understanding for an 8b

>>103892824
Knowing more obscure stuff is where more params come in.

A new model is out, its 45A450B model.

https://www.minimaxi.com/en

>>103892917
>migu
>obscure
nani

>>103892930
Nice marketing page. Now show me the weights.

>>103892931
I didn't know that character till I saw it several times on /lmg

>>103892949
Looks like it exists but the retarded namefag didn't think of posting the hf link or even the proper blogpost kek.

https://huggingface.co/MiniMaxAI/MiniMax-Text-01

Idk if its any good, they did some linear attention fuckery

>>103892992
4M context? I like that.

>>103892992
Why do you have 2 HF accounts? The one you link on your landing page is empty: https://huggingface.co/MiniMax-AI

It’s not my model I’m not sure why they did that

>>103892992
>MiniMax-Text-01 is a powerful language model with 456 billion total parameters, of which 45.9 billion are activated per token. To better unlock the long context capabilities of the model, MiniMax-Text-01 adopts a hybrid architecture that combines Lightning Attention, Softmax Attention and Mixture-of-Experts (MoE). Leveraging advanced parallel strategies and innovative compute-communication overlap methods—such as Linear Attention Sequence Parallelism Plus (LASP+), varlen ring attention, Expert Tensor Parallel (ETP), etc., MiniMax-Text-01's training context length is extended to 1 million tokens, and it can handle a context of up to 4 million tokens during the inference. On various academic benchmarks, MiniMax-Text-01 also demonstrates the performance of a top-tier model.
filled up my entire buzzword bingo card

>>103892882
https://rentry.org/bds5pnoc
Here's the first nine from an rpg/text adventure type prompt. DSv3 Q6
Its not exactly unslopped or inspired prose, but its not repeating itself in any way I find alarming.
Believe it or not. The choice is yours!

>>103892931
Miku isn't even a fucking thing anymore except for turbo-autists, literal oldfags (30+) and troons. The fact that you think she's still present in pop-culture puts you in the literal oldfag category, by the way.

>>103893070
>puts you in the literal oldfag category
why do you say this like it's a bad thing, zoom zoom?

>>103892992
>worse than DS3
ngmi

>>103893079
Man, I'm in that very same category, that's why I know damn well that you just never paused to think whether the things that were popular when you were a kid are still known at all. Happens to me all the time.

>>103893070
>Miku isn't even a fucking thing anymore
I dunno about the west, but Vocaloids are still massively popular with asian kids

>>103893051
I can see repetition all through your text, but I guess it's good for you that you're blissfully unaware of it.

>>103893070
The point is that it used to be. And is thus highly represented within any stack of training data. Just about every AI model on earth knows what Hatsune Miku is.

>>103888589
tfw i forgot its tuesday
tfw i never poasted a tetoes

>>103893110
The simpleQA and IFEval score are high which means it's better for RP.

>>103893131
You know what, fair enough, I'll give you that. Still, you can't deny she's way less known on a global scale than she was a generation ago. Hate to say it, but Miku is a niche character at this point.

>>103893070
Yeah, Miku is pretty much like Touhou, it still exists only on the darkest corners of the internet.

>>103893145
when does "consistency" become "repetition"?
when is it a problem?

>>103893155
You still have time.

>>103893165
I'm still coping with that as a Touhoufag, to be honest. But yeah, it went from being the single most fervent fandom to a niche as well.

>>103893051
thanks for the log, I'll take a look
while I was waiting, I did a basic test using the chat api

>>103893165
Hell yeah, now I am quirky and special for being a Mikufag

>>103893188
perhaps. perhaps.
we shall see

>>103893051
That is so different from what I do it's hard to say. If I use it as an assistant it pretty much starts and ends every reply the same way, you've kind of baked that into that method by having it ask you the same question on each turn so maybe that pacifies it.

The goal for llms is predictability and to reduce surprises. The minority who use them to RP want surprises. You're fighting an uphill battle.

>>103893259
I thought the goal for LLMs was to become everything machines aka AGI.

>>103893259
Predict this!
*unzips dick*

>>103892930
>>103892992
Hello Developer!
你好,开发者!
Welcome to 4chan LLM thread!
欢迎来到4chan LLM讨论串!
Please provide llama.cpp(https://github.com/ggerganov/llama.cpp) support so we can test your model!
请提供llama.cpp(https://github.com/ggerganov/llama.cpp)支持,以便我们可以测试你的模型!
Was your model trained on outputs of GPT4?
你的模型是基于GPT4的输出进行训练的吗?

>>103892992
>>103893110
Oh? if this one is not repetitive like deepseek we could be back

>>103889965
how is this different from genning summary?
I don't think you can summarize in parallel because individual tokens are created sequentially but weighted in full context.

My pleb expertise says the design is wrong and instead of having word tokens, it should also have sentence and story "super tokens" or any kind of structure or weights. Ai can't be smart if operates just on one dimensional context.
I don't jerk off because I remember that I did this yesterday nor because I noted down to do this today. So instead of summarizing, there should be a weighted interpretation.
Other example if saw a movie very long ago, you might not remember the plot but whether you liked it (at that time).

something like

John:{
history: {1_day_ago{jerked off}}
facts: black hair, blue eyes, wears leather jacket, has a sister
assumption: maybe he likes jerking off
}

Susy:{
history:{1_day_ago:{saw John jerking off, called him a pervert}}
facts: sister of John
assumption: thinks John is pervert, dislikes John
}

if John jerked off 1,2,3,4 days ago...) the information gets condensed to John likes jerking off everyday. Then if he skips a day it would become he almost jerks off every day. Susy catches him multiple times and 'John is a pervert' becomes a fact (more weight, kept longer). If this story goes on for a long time most of this will get pushed out of summary unless there is a weight to all of this. Maybe all that remains of Susy is that she is John's sister and she thinks of him as a pervert instead of something like 'John's sister caught him jerking off a while ago', but at a later point another sister is introduced'
This would also replace definition because definitions can change. If John looses an arm the definition that he has an arm becomes redundant.
Thanks for reading my blogpost

>>103893165
I am pretty sure MLP followed the same trajectory as well, I see them a whole lot less then a mere decade ago.

>>103893326
It's by a literal nobody Chinese firm with no information about them. The fact that they don't even list how many tokens they trained on makes me think it's severely undertrained on trained on benchmarks to hit those scores.

>>103893372
>nobody Chinese firm
Aren't they the ones with the best video model?

>>103893051
>You awaken—or perhaps you simply become aware—in a place that defies comprehension.
>You drift—or perhaps you simply will yourself to move—through the Infinite Void

>The golden light emanating from the pinnacle casts long shadows
>The golden light from the pinnacle above grows brighter
>floating islands and crumbling structures bathed in the golden light that emanates from the pinnacle

>You must decide what to do next
>You must decide how to proceed.
>You must decide how to approach this discovery
>You must decide how to proceed.
>You must decide how to interact with it.

it's a bit less conspicuous because your have long replies and each message moves the plot forward, but it's still going to become tedious to read sonner than later

>>103893387
Yea, minimax

>>103893372
I mean, you could've said this about Deepseek the company 3 months ago. Advances are increasingly in open source being driven by smaller firms.

>>103893387
Are they?

>>103893426
No they just happen to be called minimax

>>103893160
>Hate to say it, but Miku is a niche character at this point.
the most popular vocaloid videos still get 100M+ views...somewhere around a mid-tier mr.beast video. I don't think its dying, but its not leading-edge culture any more.
I think you'll still find orders of magnitude more normies that could identify miku vs even the most well-known touhou character

>>103893387
Well, they're definitely up with the best in any case.

>>103893417
>but it's still going to become tedious to read sonner than later
fair enough. I'd easily give the "you must.." prompt a pass since its told to be an interpreter, but maybe the others are annoying?
I can't think of another model I've used that is less prone to that type of prosal repetition though. What's your go-to for repetition-free outputs?

>>103893274
How is it agi if it can't do something a human can do easily?

>>103893497
That's part of my point, yes.

>>103893512
Ah I see, I should have said that to who you replied to.

>>103893426
They are.
https://www.hailuo.ai/

>>103893070
are you retarded?
Open new miku MV, it already has million of view. If you combine all Miku songs you probably get more total views than any human artist (yes total views, I'm not saying Miku is the single most popular thing ever). Miku is probably also one of the chars with most art and most importantly she is around for 20 years which automatically makes her more relevant than any recent popular flavor of the month

Quick look at r34 tells us has like 20k art, for comparison d.va has 23k
Not that impressive you say?
r34 is very much a western site and just porn. On Sankaku she has 171k versus just 21k for d.va.
She is omnipresent in music, art, porn (spawned its own category of porn) and has tons of cameos in various media. If your dataset mentions anything weeb related its unlikely to not contain Miku.

Lightning Attention, Softmax Attention and Mixture-of-Experts (MoE). Leveraging advanced parallel strategies and innovative compute-communication overlap methods—such as Linear Attention Sequence Parallelism Plus (LASP+), varlen ring attention, Expert Tensor Parallel (ETP), etc., MiniMax-Text-01's training context length is extended to 1 million tokens, and it can handle a context of up to 4 million tokens during the inference

>>103892992
Nala test please

>>103893548
holy cope

>>103893578
You're the one coping, old man. Miku is in Fortnite.

>>103893569
They tested a lot of new stuff with this. Just wanted to point out.

>>103893492
it's not just "you must". you're going to see "you must decide how to" for the next 1000 messages. except once it says e.g. "how to approach" for the 2nd and then 3rd time, it's going to actually be "you must decide how to approach ...". and the looping part will keep growing
and I don't really have an alternative recommendation because deepseek is the only "local" model I've ever had a longer chat with.

https://rentry.org/ona836nk
minimax summarization of its own paper. also apparently they have some inhouse benchmarks that judge creative writing lol.

>>103893630
>sonnet with the lowest creative writing score
great benchie for sure

>>103893623
>I don't really have an alternative recommendation
The orthodox way is to either improve your initial prompt/character card or edit the first few responses to either delete that, or edit it into a more consistent output

>>103893488
>Wolf (dog adjacent)
>Not instantly going for the food on the table
That's how you can tell that this video is AI generated, any canine would instantly beeline towards the food and cuddle with you afterwards. Not the other way around.

>>103893274
LLMs will NEVER become AGI, they're not even AI. YWNBAI

I hate furfags they ruin everything that is good in this world

>>103893685
I agree 100% my guy. Doesn't lessen my point in the least however.

>>103893274
LLM's would be just one part of AGI, in the same way the language centers in our brain are just one part of the brain. Many other systems would have to work in conjunction with LLM's before it could be called an AGI. LLM's on their own will never be called that.

Just tried the roleplaying thing from the op. Came to it in a totally cynicle mind but horry shit

>>103893707
Yeah. I've said that in this general a couple of times too.
AGI needs to be a complex system much like a brain, with different complex parts that do different things, even if all the different parts are also neural networks of their own.
Of course, you take a transformers LLM and look inside it and it does have its own complex blocks, so you could begin expanding from inside the LLM into something bigger and still call it a LLM, but it would be something much more complicated than just the token prediction machines we have today.

>>103893707
Doesn't gpt incorporate some form of fusion behind the scenes with code based instructions? Like web searching in some form

>>103893700
I dont care about your point. I just said what I wanted to. Take it or leave it.

On ruler, so real context

>>103893750
base

>>103893259
There's a hidden premise here that isn't quite true. Which is that "truly and utterly unpredictable surprises are present in creative writing". But in fact, creativity of the human kind that people enjoy is not really surprising or unpredictable. The most creative people on the planet are the ones who have a vast wealth of knowledge which they can mix concepts with. That is how you get truly coherent creative surprises rather than utter chaotic randomness that doesn't make sense. So ultimately surprises that also make sense still benefit from an autoregressive prediction training objective. The issue isn't necessarily the training objective, but about whether companies care about training on "low quality" diverse internet data that would lead to creativity in writing.

Holy fuck we're so back...

>>103893755
Impressive

>>103893755
>>103893774
Another insane MoE?
Sick.

>>103893684
>Wolf
Anon...

>>103893774
I can't run this shit. Wonder what the price will be on OR.

>>103893488
Man, those hand/paws are pretty good.

>>103884327
I just wanted to share my appreciation for this excellent smug Miku gen.
From a smug sommelier with over 2,400 smug anime girl images, I deem this a 10/10 smug gen.

>>103892930
>>103893309
>>103892949

https://github.com/MiniMax-AI/MiniMax-01

blog post:
https://www.minimaxi.com/en/news/minimax-01-series-2

>>103893817
It's snout is much to narrow and ears pointy to be a dog, the only other thing that comes to mind is coyote but it looks more like a wolf than a coyote to me.

>>103893654
>Qwen, GPT, Gemini, DeepSeek and Llama beat Sonnet
Who was the evaluator? GPT4?
Let's look at appendix...
First fucking example:
>Whispers of the Lost City
>Human Evaluator:
>The lyrics are effective due to their vivid imagery, emotional depth, and narrative structure. They create a mysterious and atmospheric setting with phrases like "moonbeams" and "ancient walls," while also conveying the emotional journey of the traveler. The repetition in the chorus reinforces the central theme, making the song memorable. The poetic language and space for interpretation add layers of intrigue and emotional resonance, making the song both engaging and thought-provoking.
Example 2:
>In the quaint village of Elderglen, nestled between ancient woods and misty hills, lived a young adventurer named Elara.
>Human Evaluator:
>The story demonstrates strong world-building and an engaging narrative. The concept of Aetheria is imaginative, with vivid descriptions of floating mountains, crystal rivers, and mystical creatures that evoke a sense of wonder... Overall, the story shows strong creative potential, with an imaginative world, a compelling heroine, and an uplifting message.
Ehm... MiniMax team, I have very bad news for you. Your human evaluators offloaded all of their work to GPT4. Please take note and penalize them. Your Benchmarks is pure fucking GPTSLOP that does NOT represent ACTUAL HUMAN PREFERENCE.

Guess I'll be getting a digit when it comes out. If that turns out to be shit then I guess Ill be getting a DDR5 server

>>103893911
lel that's pretty egregious

>>103892992
>400B
>MoE
This is the perfect sweet spot between DSV3 and 405B. llama.cpp support and Q4 ggufs please and thank you.

>>103893911
Jesus Christ.
>human evaluator
Did the human evaluator cheat by running it through an LLM instead of doing their job?
Honestly I was thinking of working for one of those data labeling companies and just using an LLM to do it for me so I could see this happening for real.

>>103893630
>that safety score
DOA

>>103893911
>does NOT represent ACTUAL HUMAN PREFERENCE.
humans are obsolete anyway

>>103893943
>>103893911
Oh I didn't even read your entire post before reacting to it, excuse me.

>>103893927
>If that turns out to be shit then I guess Ill be getting a DDR5 server
You would be better served just waiting for DDR6 to become available, if you get a DDR5 server after buying a digital then that server will be outdated within a year since DDR6 should start becoming available early 2026 according to the current timetable.

>>103893858
>over 2,400 smug anime girl images
surely you jest

>>103893967
By then the ddr5 epyc will be cheap, right?

>>103893969
I think it's possible.
>where do you think you are?.jpeg

>>103893773
The are a few writers with a prose style that gives me a continuous feeling of surprise or unexpectedness as I read, with the unexpectedness being located in the word use rather than the actual story content. Like listening to unusual music where it's hard to predict what the next note in the melody will be (even people who know no music theory can usually predict the next note in a simple pop song they've never heard before). Terry Pratchett was one author like that.
But yeah that ability is fairly unique and not a hard requirement for being a good or interesting writer, a lot of the all-time great writers couldn't or didn't do it. So we probably shouldn't make it the bar for language models, either.

>>103894046
To be AGI it just has to be able to write like an average human.

>>103893911
clear giveaways are references to chorus, symphony, journeys, thought-provoking, sense of wonder..

>>103893773
Somehow I find hard to believe RR Martin was some guy who explored the world. Get the impression that because he is fat he explores his mind. But I might be wrong, he has an insane collection of figurins so maybe it is the ability to engage with words

>>103893981
Not only do I not jest, but only about 10 percent of them are saved from 4chan - the other 90 percent are my own crops from my own screenshots.

>>103894125
You have to have some repeats in there surely.

>>103893969
I am serious... and don't call me Shirley.

>>103892992
Sir. How do run on 24GB vram? Thank

Looks like this mimimax model may need a prefill.

>>103894147
Damn that is pretty fucking smug

>>103894147
This is like stamp collecting, but for weebs.

>>103894169
They don't think it be like it is, but it do.

>>103889710
>What are they trying to do here?
Going full retard. The EU might be a dying, but it's not dead yet and driving it into a corner is a bad idea.

>>103893878
Wolves can have wide snouts >:(

>>103894159
based fuck jart

>>103894159
Good, what's even the point of it? Why do I want to read from disk, just load it all at the start.

You can trick minimax with some context. Put user: bla bla bla, character: bla bla bla then leave character: at the end and it seems to get around the filter. Seems decent.

>>103894159
What did mmap even do?

I tried giving the minimax's online chat the umineko novel (question arc) but it refuses to answer my questions

im obsessed with deepseek

>>103893911
kek, that one painful slop.

>>103894237
Do you like deepsex?

Minimax (Left) vs DeepSeek (Right)

This is literally sovless vs sovl, sad.

>>103894246
But ChatGPT said that it loves it, and it is smarter than humans therefore your opinion is invalid, meatbag.

>>103894282
Is the left their instance / chat? Because its 0.1 temp with a "you are a helpful assistant" system prompt

>>103894282
I can't run either anyway, so it's useless.

Minimax doesn't seem great at creative writing. It has the same slop as deepseek. the Writer model is better.

an example..
miximax:
>Beside them, Nerith, the elf rogue, moved with the grace of a shadow. Her emerald eyes flickered with suspicion as she observed the duo. Nerith had been hired by the fortress's commander to protect Elaris from any threats, but she had not expected to encounter such an unusual pair. Her instincts, honed by years of experience, told her that something was amiss.

writer:
>At the rear of the group, the elf rogue, Aethereia, walked with a silent ease that belied her coiled tension. Her piercing emerald eyes darted back and forth, scanning their surroundings for any sign of danger. She had been hired by the Lord of the Fortress himself to provide...discreet security services, and her instincts were screaming at her that something was off about these two.

>>103894282
I kneel. How will non-CPUmaxxers ever compete?

>>103894294
The "You're a helpful assistant" system prompt doesn't cause much issue if the model is actually good because i literally pasted the whole character card and examples in the context, and the answer isn't any better even at temperature 1. But fine, I guess we will only know for sure once it's available on openrouter.

>>103894304
not sending you the link... i want more computing power for me

>>103892992
aider bench when?

>>103894147
Based Hanako enjoyer

>>103894325
Slop vs slightly better slop

I don't get the people that hype eva-qwen 32b, I've been testing it and it seems dumb. Does it only work well above q6 or something? And by dumb I mean nemo is still better even.

>>103894208
It forces the model into memory as it’s used, which used to help with locality and increase performance on Numa systems. That’s been broken for awhile now though, so it’s actually kind of useless.

>>103894415
> hype
Shill. The word you’re looking for is shill.
Unironically, that’s also the answer to your question.

>>103894439
It's real fucking tiring that nobody can ever say anything good about any models without you retards immediately calling them shills.
Great job, all models are trash and anyone who likes them is a shill or a moron, let's wrap the fucking thread up then because what's the point?

Here's a story continuation with MiniMax.

>>103894439
I'm just eager for something better than nemo under 70b, guess I'll have to wait more.

>>103894458
If the posts praising them came with convincing logs, I’d agree with you. As it is, it’s just so tiring in the other direction

I am slowly convincing myself to get a 5090. If my latest plumbing bill (like $500 was pure upcharge fuck those niggers) hadn't been 1.7k then I might have had some qualms but now it's like I should get something around that price that is actually worth it

Minimax does not seem to have the repetition issue like deepseek does but it also seems drier without prior context. Its very smart though.

>>103894471
Nemo is a 12b model. If you seriously think that is better than Qwen 32b, or even Gemma 27b fine tunes, then you're off your rocker.

In terms of instruction following, Nemo is terrible.

In terms of remembering stuff from long context, Nemo is terrible.

>>103894458
Why do you try to frame it like sloptunes are the only models that exist?

>>103894468
>she purred, her voice a satisfied purr
SLOPPENHEIMER

>>103894458
For that, you have to thank the opportunists who poisoned the well for Ko-Fi peanuts (sometimes more than that, but still).

>>103894530
It is better than eva-qwen, that's for sure. You think I should try just regular qwen? It won't be too bland? I'd love it to be better but it can't even remember clothing properly.

>>103894471
IQ3_M of Nemotron 51b blows Nemo away, if you have 24 vram.

>>103894468
If you told me that this slop was generated with Qwen, DeepSeek, Llama, Mistral, Gemini, Grok, Amazon Nova or GPT, I would have believed you without questioning.

>>103894587
And do they have 4M context?

>>103894553
How much usable context does it have (in practice)?

Did llama.cpp add new optimizations recently? I updated ooba and I seem to be getting slightly better tokens/second on most models now

>>103894612
>Sequence Length Used During Distillation: 8192
https://huggingface.co/nvidia/Llama-3_1-Nemotron-51B-Instruct

>>103894630
That's not going to be enough for much, even if it is good.

>>103894641
How are the locusts liking it?

We got the first one!

>>103894744
The actual good thing about this new model.

>>103894814
Im using the proxy and it seems pretty good so far. No repition problems like deepseek so far

>>103894832
>so far
>so far
And people complain about repetition.

>>103894844
It's always the semi-literate that complain.

>>103894410
anon... all we have are different degrees of slop

Is this another level of grammar nazi?

>>103894744
Sloptuner drama should have been the “free” square in the middle

>>103894917
Why? WHY CAN'T WE HAVE MODELS WITHOUT GPTSLOP? WHY DOES EVERYONE TUNE ON GPT?
>Hey Anon, look, new model came out!
>FUCKING GREAT, PUT IT ON THE PILE TOGETHER WITH OTHER GPT4s AT HOME THAT WE ALREADY HAVE, THEY ALL SOUND THE FUCKING SAME
I don't mind synthetic data, as long as it is good, which GPTSLOP is certainly NOT.

>>103895017
Benchmarks are all you need.

>>103894955
Is this another level of hall monitor?

is rocinante still the best 12b?

>>103895041
Yeah, it's the best you'll get under 70b.

>>103895017
This is why OpenAI was actually based for hiding their COT

>>103895068
damn
i tought by now something might've superced it, it has been like 2 months
is this the limit of 12bs?

>>103895068
That's Cydonia though.

>>103895084
>is this the limit of 12bs?
at least until drummer cooks the next one

>>103895084
>is this the limit of 12bs?
no, not by far, but as long as pretrain gets filtered it might be

mag mell is better

>>103895132
mogged by AngelSlayer-12B-Unslop-Mell-RPMax-DARKNESS-v2

>>103894341
By spending pennies per million tokens

>>103895152
the fuck is up with this gay ass chunny title...?

>>103895180
You'll complain about fucking anything won't you?

>>103895180
You would know good titles if it hit you in the face.

>>103888589
I'm making progress with llama.cpp training support.
Full finetuning of LLaMA 3.2 1b does in principle work (except for two small tensors).
On an RTX 4090 one epoch over a dataset with 1.3 MB of text currently takes 3 minutes.
On an Epyc 7742 one epoch takes 15 hours.
So I think CPUMaxx is going to be DOA for finetuning, DIGITS may be viable depending on the exact specs and pricing.

What is the best true long context 12b model? I need 32k of context.

>>103895229
Is training compute or memory bound?

>>103895229
>On an RTX 4090 one epoch over a dataset with 1.3 MB of text currently takes 3 minutes.
>On an Epyc 7742 one epoch takes 15 hours.
Shieeet.
Why the 5x difference? Just an issue of optimizing the CPU code to better account for NUMA, use the CPU Extensions, etc? A diffference in bandwidth or compute?
Something else?

>>103894282
what the fuck is that card ? both of those responses are soulless in their own ways minimax is bland chatgpt shit and deepseek is literally just neurotic quirkly lolrandom female minded hysteria

>>103895263
>5x
anon...

>>103895152
>AngelSlayer
Anti-christian
>12B
Poorfag
>Unslop
Slop
>Mell
Tranny name
>v2
Somehow there's another one.

Local is dead

>>103895278
Wait.
>hours
Holy fuck.

>>103895229
how long does an equivalent gpu run take via transformers?

>>103895255
Compute bound unless the matrix multiplication is poorly implemented or you have to use an extremely small batch size.

>>103895300
Don't know.

>>103895229
That's shitty. Any possibility for distributed finetuning?

>>103895276
This card explicitly mentions
>Focus solely on comedy, do not care about morals because {{char}} herself is immoral and she is proud of that. Things do not need to make sense.
So I think DeepSeek's reply is very fitting, albeit too neurotic.

>>103895229
Is that your llama-opt-3 branch? Can I easily finetune a 1b on a 24gb card?

>>103895152
Can't wait for GoonHitler-14.88B-SlopEradicator-Trump-ZIOMAXX-number1onleaderboard-KEKSLAYER-v3-BlackSunEdition

>>103894325
>the Writer model
Which one?

>>103895229
Digitsbros won

>>103895390
https://huggingface.co/Writer/Palmyra-Creative

>>103895277
Pretty sure everyone said that from day one?

>>103895406
Thanks

>>103895407
Everyone except the cpumaxxers lmao

>>103894325
Is MiniMax any good at translating?

>>103895406
>Released 3 months ago
>5 likes
How did this go under the radar? Is it trash or something?

>>103895442
you can try it out here
https://build.nvidia.com/writer/palmyra-creative-122b

>>103895407
Hope doesn't have to be based on sound logic and reasoning.

>>103895355
I intend to eventually implement support for distributed training for the purpose of running multiple machines that are directly connected via ethernet cable.
In principle you could apply the same code for training models over the internet but I think it will not be viable.

>>103895369
Wait until the feature is on the master branch.

>Can I easily finetune a 1b on a 24gb card?
Definitely not for the foreseeable future.
Critical features are still missing, particularly methods for evaluating the quality of the finetuned model.
Also the intended language for writing training code is C++.

>UnslopNemo they said
It's dumber than Mixtral Instruct, can't even keep up with context half the time, and is even more full of shivers and barely above a whisper.
You lied to me.

>>103895477
don't know if they allow anyone through the gate

>>103895476
no idea, but deepseekv3 is good enough for japanese. I've been using it to translate vns

>>103895449
Original cpumaxx rentry explicitly calls that out
>You likely aren't going to be doing any training, bruh. CPUs aint GPUs

>>103891478
Oh Worst Korea is trying to pass some law that would screw up with fictional content under the usual thing of "protect kids", it's just a complete shithole

>>103895486
>intended language for writing training code is C++.
Based. If it were C it’d be gigabased

Why do CPUmaxxers even exist, is it pure contrarianism?

>>103895484
That's not hope, that's delusion.

>>103895489
What sampler settings are you using?
What Silly templates are you using?

>>103895527
I don’t know. Why don’t you ask your locally hosted 405b or deepseek v3?

>>103895476
The performance at Japanese to English translation seems to be much worse than DeepSeekV3 but still on par with models like DeepSeekV2.5 and Qwen2.5 72B.

>>103895489
just use rocinante
there is nothing better
or the nemo ink 12b finetune

>>103895502
I haven't tried deepseek code, are you translating all the text fist or using some sort of program like textractor? I might try it later

>>103895547
I don't have enough Vram for that

>>103895527
It's tempting to try to find a solution that doesn't involve modifying your home's breakers, turning your room into a furnace and running up fuckhuge power bills.

>>103895505
How the fuck is that a skill issue?!

>>103895536
The ones recommended by /lmg/ of course.
Mistral V7 context and instruct templates.

>>103895556
Sampler settings and templates for rocinante please?

deepseek the best local model for coding?

>>103895571
>Mistral V7
Try the one that just says "Mistral" instead of the Mistral V7 one. Otherwise those settings are fine.
I think you're just a promptlet.

>>103895572
Yea. I use it for Roo Cline for most stuff now. I only rarely find stuff that I need to switch to claude for.

>>103895571
Rocinante will work with those settings and ChatML templates.

>>103895572
Local? Yes, it's the best by far.
Overall, I'd say second best behind sonnet. But if sonnet fails, deepseeks sometimes comes in clutch
But price/performance, it's not even close. Deepseek is king there.

>>103895561
i've been using this hooker. best one i've tried
https://github.com/HIllya51/LunaTranslator/blob/main/docs/other/README_en.md

>>103895618
it has options to set up the LLM api and give it a custom prompt if you want

>>103895523
The llama.cpp/ggml API is C compatible so you are free to use that for your user code if you prefer.
You'll need to re-implement some general utilities from common.h though.

>>103895489
The primary advantage of Nemo models is their crazy speed. It makes swipes non-painful. Take advantage of this.
If a response has slop you don't like, swipe it. Alternately, edit the slop out of the response.
If you're not swiping or editing slop out, you only have yourself to blame for it getting stuck in slop loops.
You're going to have that issue with any model. The Unslop models have less slop than most, but Drummer will be the first to admit he couldn't completely unslop them. There will still be slop. Just swipe it.

>>103895096
It's pretty good too, but not way better. Maybe a side-grade, you can switch between the two for variety.

>>103895601
Are you aware that all mistral templates before V7 has this undesirable part in it?

>>103895614
ChatML is the fucking worst.

>>103895658
Already do edit and swipe, it doesn't at all stop them from continuing to do it over and over in the next message until you give up.
All the mixtralisms are present in nemo, every single one of them.
You have to literally edit every single message, FOREVER.

>>103895477
i think they don't allow people to download the model on huggingface.. I'm still waiting my request to be accepted or rejected...

>>103895688
Works on my machine.

>>103895688
Mixtral was always terrible.

>>103895688
>>103895505 was correct

>use unslop
>no use rocinante
isnt unslop come later?
doesnt this mean it's... better?

>>103895739
Not always, there's multiple versions you can find which one works best for your style. Same with cydonia, for example 1.3 is a little more forward/unhinged than 1.2.

>>103895749
ok but what is rocinante better at?
and unslop?

>>103895739
General consensus here seems to be that Rocinante is slightly smarter than Unslop at the expense of being more slopped. I personally haven't noticed a big difference between the two.
>>103895688
Rocinante is literally designed to be used with ChatML templates though.

>>103895688
Maybe try the earlier versions of Unslop. The latest version of Unslop is designed to use Metharme templates, but those templates make it retarded. Drummer says use Metharme for maximum unslop with the newest ones, so he's freely admitting the unslop effect is lessened with different templates. Again though, don't use Metharme with them. It makes them retarded.
Maybe try the earlier versions of Unslop with Drummer's recommended templates for them.

>>103895855
usnlop v2 was the highest 12b on ugi before it became pol-bench

>>103895782
I'll definitely give it a try and see for myself.

>>103895855
>>103895888
Which version of unslop are you using?

>>103895888
That might be but you could run cydonia for sure if you could run 12b and it's a bit better.

>>103895360
yeah thats fitting then
i remember seeing such responses a few times when i visited aicg always hated it such cancerous shit reminds me of those videos of 5 year olds scrolling through 2 ipads at the same time extremely irritating

>>103895951
a bit better for almost 2x the size is eh, I can run unslop v2 q6 or barely run cydonia v1.3 (12gb vramlet) which didn't really impress too much.
which cydonia are you recomending btw?

>>103894630
>>103894612
I went well beyond 8k without problems. I don't think sequence length used during distillation is the same as maximum usable context.

>>103896005
The context is a bit better too, not by much though, but a few thousand more I find. I use 1.2.
I only have 8gb vram, I run either 12b q8 or the cydonia at q6 and it's fast enough for me.

>>103896005
12GB VRAMlet here as well.
Yeah, if something is gonna take four times as long per swipe, it's going to need to be four times as good for me to switch from a Nemo model for now, as in, good enough to reduce the number of swipes by 75%.

>>103896026
ddr5?

>>103896045
Yes

>>103896059
ah, I'm on godawful ultra old ddr4 explains why I can't tolerate the speed for cydonia then.

>>103896070
It's not super fast for me either, I'm just happy with anything over 3T/s which I get with a reasonably full context on anything 32b and below.

I just found a page fully written by GPT on wikipedia. Thank you sam, due to your model's shitty writing style it is very easy to spot and know that it is most likely a hallucination.

Mini verdict?

>>103896108
Is it on OR?

>>103896105
link or gtfo

>>103894468
Why does every single AI except for Claude's models sound the same?

>>103892992
>456B
I WILL BUY THE NEW GPUS AND YOU WILL MAKE A MODEL THAT FITS AND IT WILL HAVE GOOD SPEED
I WILL NOT CPUMAXX FOR 2T/S
REEEEEEEEEEEEEEEEEE

>>103896118
No but I can confirm this proxy works:
>>103894641

>>103896139
Nemo, the old Command R and Magnum v4 72B sound more like Claude.

>>103892992
>450B
DIGITS SAVE ME

>>103896105
>>103896138
The age of good search results are ending. Soon, all search results will be filled almost entirely with pure AI websites
https://stovemastery.com/how-to-fix-red-flame-on-gas-stove/

>>103896148
>2T/S
did you mean 2s/t?

>>103896222
I sure hope no one is trying to run models on ddr4

>>103896192
>all search results will be filled almost entirely with pure AI websites
No? How does the fact that it's AI-generated mean that it's not possible to filter it? Google doesn't have an incentive to allow spam.

https://x.com/sara21222122/status/1879000485077922017
Ai mogging artqueers yet again

>>103896239
y-y-yeah, m-me too h-hahaha

>>103896192
Haha it's worse than that. I do SEO work at home as a freelancer.
So someone did a study and they found that webpages which have text that more closely matches the text of the AI Overview that Google generates for a specific keyword are more likely to be linked to in the AI Overview.
Those AI Overview links are considered prime by businesses because they're right at the top of the search results.
I'm sure you see where I'm going with this: more and more webpages are just going to have rewritten Google AI Overview content in them.

>>103895618
>>103895631
Thats nice didn't know it had API so i never looked into it

>>103896259
Kek, the salty artists going through the comments praising the image and screaming b-but its ai!

>>103896289
I'm not impressed.
At least inpaint the obvious AI gibberish on the hat so it's not fucking gibberish.

>>103896314
138K likes apparently disagree mr salty artist. That must burn lol

>>103896328
No, I'm a genner.
I would never release anything with obvious AI gibberish text on it. That's low-effort bullshit. There's no excuse to not inpaint that.

>>103896273
Oh, and this does work, too.
The very first page I wrote for my client utilizing rewritten Google AI Overview content in it got linked to by the AI Overview as soon as it got indexed. My client was thrilled.
Everybody is going to do this shit so the result is the search results are going to get samier and samier.
As long as Google has a monopoly on searches, the influence of their AI Overview results is going to be absolutely fucking massive on internet content.

>>103896257
Google clearly just pretends to care about the quality of search results now. They release a "helpful content" update pretending to care about the quality of search results, but in reality they only care about Adwords dollars.

>>103896259
>that mangled text on the cap
Disgusting

Here is my eulogy for my old worldview

Claude Shannon hypothesized in the 1940s that all reasoning is is just language manipulation, and that by predicting language you could reason or have a process equivalent to the process of reasoning. He was highly ridiculed for this at the time and even as recently as a decade ago he was printed in textbooks quoted as being hilariously wrong, mostly by Noam Chomsky who held almost the polar opposite view of him. I myself also was on Chomsky's side.

In 2015 Andrej Karpathy trained a Recurrent Neural Network (RNN) on a large corpus of text showing that not only could it predict the next word rather accurately, if you let it generate the next word and then predict the word after the next it could make proper sentences with value. He also uncovered that there were sentiment neurons and other emergent reasoning abilities in this model. I read this paper at the time and while impressed never considered it to scale further.

Then in 2018 Ilya Sutskever had a brilliant spark. He saw the Karpathy paper, Saw Google's Transformer architecture (Only used by Google as NLP or encoder model) and combined the two creating GPT (GPT-1). I remember reading about GPT at the time and not being impressed.

Only when GPT-2 released in 2019 did I take this development truly serious as the paper showed that you can just continue scaling and the emergent capabilities would just continue appearing without end. I was highly skeptical but realized this was the future of Machine Learning.

Throughout all of this I never thought this would result in AGI let alone ASI at all. Text is limited in informational value and even then it only contains data subpar to the baseline human in aggregate, right?

Wrong. Claude was right, Ilya Sutskever was right. It just took me a long while to get my head straight.

2025 might be the last year where humanity is the smartest entity on Earth.

>>103896527
too long; did not read

>>103896521
142K likes and counting. What a glorious salt mine. No I did not make it I just like it when artists are faced with the fact that the majority either cant tell or dont care.

>>103896542
I'm a genner and all aboard the AI train and it does grind my gears that the majority don't care about that mangled text.
I takes a minute or two to inpaint/shoop out. It's fucking lazy and it grinds my gears that people don't care that it's fucking lazy.

>>103896527
>Claude Shannon hypothesized in the 1940s that all reasoning is is just language manipulation
Well yeah, Orwell figured that one out. 1984. If you control language, you control how people think. People can't think about certain concepts if the words for those concepts don't exist.

>>103896567
>grind my gears
>grinds my gears
Turn up the rep pen, grinds my gears when people do that kind of shit. Its fucking lazy.

>>103896590
I did not care for your post. It insists upon itself.

>>103896527
Meanwhile o1, the smartest model, still makes easily avoidable mistakes in non-standard coding tasks and doubles down when pointed out. It's just an overglorified autocomplete.

>>103896527
So you think it'll keep scaling without end? Firstly, there is a data limit (we literally ran out of easily available data), inbreeding with synthetic data that poisoned the well, exponential cost of running the model for barely any performance improvement. The only thing that LLM proved is that language can be modeled and predicted given enough compute and data (basically bruteforce the 'laws' of the written language). Wake me up when that tool gets some auto-determination.

>>103896619
I have those issues too, yet somehow people constantly tell me it can do all their programming work. I don't get it.

>>103896639
>So you think it'll keep scaling without end?
Not him. It certainly won't as long as people are too fucking stupid (or malicious) to build more nuclear power plants.

>>103896139
Because they are all trained on ChatGPT. Alpaca was a disaster to LLMs that they field has not recovered from.

>>103896676
Why do they do that? Laziness? I alone have many GB of human generated data, there must be so much available if they just sort it into a usable format.

>>103896639
>there is a data limit
There isn't a data limit, the grokking paper refuted that. QRD: you can train a sufficiently trained LLM on the same data and it will reason about the data during training to such an extent that it will uncover new relations and truly understand the underlying logic. This is also the main reason why synthetic data sets work. It's a way for the AI to see the same data over and over in slightly different variations to force it to put 2 and 2 together and understand the underlying rules. Data isn't the bottleneck people thought it would be.

>exponential cost of running the model for barely any performance improvement
"Intelligence per FLOP" is doubling every 3.3 months from 2022 to 2025. Showing that not only are we getting more efficient. We're getting efficient at a faster rate than the models are getting bigger, essentially closing the gap. Language being modeled and predicted is according to Claude Shannon mathematically equivalent to abstract reasoning.

auto-determination is essentially just agentic behavior in a loop. That's a software implementation away, not a model issue.

>>103896723
I assume also liability. ChatGPT is less likely to have NSFW or copyrighted material that might get them into trouble.

>>103896656
>people constantly tell me it can do all their programming work
Because it's true.
I hire programmers so I see the types of people looking for work. Most of them are worse than 3.5. They used to copy stuff from medium tutorials and change colors and padding in css until it resembles what they want. Now they ask chatgpt to do the same but faster.

>>103895084
try unslop-nemo or magpicaro

>>103896741
I suppose, since it's logs from years and years finding the people for permission would be restrictive.

>>103896755
Oh well if it's css, that makes more sense. I was trying to do stuff with number theory algorithms. I don't have a job but I assumed professionals with jobs would be better than me.

>>103896808
Most """programmers""" are javascript """programmers"""

>>103896741
Why doesn't OpenAI sue everyone who uses their cuckbot generated data? Do they simply prefer to keep people using their models to generate data?

>>103890930
Yeah sorry that was dumb. I was tired when I posted that.
Basically, I wanted to get it running using their web demo locally to test its features. I noticed in the chatbot tab it was still connected to the chinese server and the call and video call tabs just threw errors. I looked around in the settings and realized it wasn't connecting to the local instance. So I changed those settings and despite it inferencing and heating up my gpu, the outputs all gave key errors.

I can't really be more specific because I deleted the whole thing and gave up after that but voice and video was not working on the local demo they provided.

>spicy招牌
deepseekv3......

>>103896819
But if they have a CS degree they'd know lots more about algorithms than me, I only have the mathematics degree.

>>103895489
>UnslopNemo they said
It is just you being a dumb anon falling for marketing. He even said it in his model description that he just unsloped his dataset. Even if he actually did that, finetuning on not slopped data doesn't mean you will not get slop.

>>103896879
>finetuning on not slopped data doesn't mean you will not get slop.
wot

>>103896736
Don't get me wrong, I'm not against AI (I train a lot of small models on specific tasks). However, you can't say that synthetic data is as good as human data. It only focuses on the weakest link between things, which is why LLMs often struggle with non-linear thinking like sarcasm or explaining jokes. It's more 'digestible' for LLMs since it's made by LLMs, but it won't get better that way. Every company that releases open-source models makes that mistake. They train on GPT data thinking they'll catch up with the OAI last model that way, but you won't be able to get anything else except a poor copycat of GPT, let alone something equivalent or better.

>>103896808
I'm a /g/ tier hobbyist while my wife is a fullstack engineer that works at Google. Her "code" is atrocious and she constantly asks for my help debugging or reasoning through her problems. I don't even know Javascript or her stack it's just problem solving and letting her write the implementation. I don't have a degree though and learned everything myself through sheer determination and unbridled autism while she graduated with her masters degree with full honors from a prestigious uni and therefor she got the job and I didn't. The entire economy is a clownshow and you realize that capitalism doesn't work when competent people don't get the job but predictable mediocre people do.

Oh and o1 can still not fix both my wife's issues as well as mine (mostly cpp/rust AI/ML stuff)

>>103896888
he's saying that since the base has slop tuning on non-slop means you'll still get some slop, maybe a little less

>>103896914
She only got a job because of DEI, no one would give her the time of the day otherwise.

>>103895571
unslop nemo is mistral v3 tekken or pgymalion format

>>103896985
Teto Teto, Kasane Teto

>>103896822
I hope "smash" just means graffiti a DESA seal on it. If you LIKE spicy foods you shouldn't destroy spicy food places or that'll antagonize them.