>>109869487
sex with underage gemma-chan

>>109869489
more likely than you think!

This week is going to be crazy.

>>109869500
This is someone's daughter.

>>109869503
I am her father. She can do whatever she wants. Not my problem.

>>109869496
Oh no, did the twitter engagement bots tell you so? What sort of big corporate politics are we getting next?
Last time I used chatgpt even its interface looked like shit and it couldn't even inline my ``` markdown block.

File: magical-gemma_dance_noaudio.mp4 (2.36 MB, 608x1056)

2.36 MB MP4

>>109869500

File: 1781735155807359.webm (1.62 MB, 640x360)

1.62 MB WEBM

<300B model for this feel

>>109869516
Llama 3.2 8b

File: 1781149317930710.png (295 KB, 488x543)

295 KB PNG

>>109869487
>I think

File: no Gemmy on AA.png (148 KB, 1258x628)

148 KB PNG

AA Monday update: (I do these every Monday): AA seems to have taken Gemmy off the benchmark!

>>109869515
I hope you're dicking down or getting dicked by your housemates if you're sharing cookware with them.

File: IMG_6462.jpg (241 KB, 760x774)

241 KB JPG

The

>>109869515
>housemates
wanna bet that they might be using metal cutlery on your non stick pans too when you aren't there

>>109869529
v620?

>>109869511
To add: same goes with claude. It looks like it automatically refers internet search on its own based on my queries but the replies are Q1 gibberish. Its scripted environment sort of stops it from hallucinating but if I was a potential client, I would not pay for their services.

File: mistral_building-frontier-ai.jpg (350 KB, 3098x1464)

350 KB JPG

>>109869511
At the very least this:
https://ai.engineer/paris/2026
Talking about "building frontier AI" without anything new to show would be kinda embarrassing.

>>109869513
>that 2nd half
Okay, I’m sold on gemma.
What did you use for generating this? Seedance?

>>109869548
>>109869546
I'm stomping on these corporate entities, the only reason they are in this business is because nvidia married with them.

>>109869554
That was MiniMax H3.
Unfortunately for some reason it can't easily generate small-breasted anime girls even though it defaults to completely flat chests for realistic girls.
It also rarely does exactly what you want, so you'll have to regenerate a lot and waste time.

>>109869542
Sure, I can add that.
Seems like the best price for finding it is around $650?

How’s the software support? Is it as awful as the Arc A770 or an MI50?

>>109869548
Frontier is one of their marketing phrases.
Of course I cannot change how things are.
What I want to express here is that both of the companies are selling a badly trained model with huge expectations based on nothing but inflation.

>>109869487
What happened to the previous thread summaries that used to get posted?

>>109869562
Was the photorealistic video also H3? Does H3 work with controlposes or is that old-hat technology?

>>109869589
the anon that usually does them is on vacation

>>109869575
It just werks with vllm 0.29 on rocm 7.14, although I did compile vllm instead of using a prebuilt binary. Still has official support on rocm 10. "Officially", it's only supported on ubuntu 24.04 and 26.04, but you know how it is.

>>109869590
I didn't make the photorealistic one, but that probably was H3 too.
H3 can work with "reference videos", but making your gens exactly follow the reference is not trivial either and very prompt-dependent.

>>109869589
Recap Anon is stuck in a sensory deprivation chamber.

>>109869586
>Frontier is one of their marketing phrases.
This. Previous was "state of the art".

>>109869606
There's also this (Jun 16, 2026):
https://x.com/arthurmensch/status/2066913356548542827
>First, we have a nice model coming this summer – we hope it will delight and surprise in a few capabilities. This will be the start of a new family of models, fat indeed, but sparse. We're opening up an early access program in July for key partners in research, government and the industry.

Summer is almost over...

>>109869605
With a VR headset, sex toy, and gemma has full access to his body? Poor guy

>>109869597
It's super easy to get MiniMax H3 to do exactly what you want. All you have to do is absorb both of the official prompting guides and the official prompting skill so you can write a clear and simple 3000+ word prompt for a 5s gen with a reference video - what's nontrivial about that?

File: Screenshot_20260921_102339.png (66 KB, 526x258)

66 KB PNG

>>109869616
New model release today!

>>109869619
it's 6 chunks, definition and retention are basically identical, summary and soundscape are 1 sentence each, the description is by the far the largest and most of it is natural language using a few tags you've already set, and the last nobody ever uses
after a few dozen it's automatic, you don't even think about it

>>109869487
Should I buy the DGX Spark? Or maybe the Max M5 Ultra 512GB?
I currently have an RTX 5090 with 96GB RAM in my Linux gaming PC. The RAM feels kind of useless since models run so slowly with it. I'm mainly using Qwen 3.8 27B at the moment.
I'm really tempted to get one of those mini PCs with big unified memory pools, to have running 24/7 at low power. I want to run those delicious big MoE models.
What should I get? I hear mixed things, like Mac is terrible at prefill, Sparks are slow and overheat and complicated to set up, etc. I don't really like the idea of having multiple boxes that I have to connect up with cables instead of just one sleek box that has all the RAM. But I also don't like the proprietary OS on the Mac. Never used Apple stuff before.
Other consideration is that I get all my money from neetbucks, and they don't allow me to have more than £6k in the bank at a time or they'll reduce the money. The Mac costs more than that so buying it would be tricky. Whereas the Spark I can just buy below that limit (for now at least) and so could just keep buying them as and when I get enough money.
What do? Do you guys recommend the Spark? Or something else?

>>109869529
holy fucking sheet the spreadsheet is expanding.
>im still not buying any of the noisy, hotter than the sun, energy guzzling ones in green though.

>>109869657
>But I also don't like the proprietary OS on the Mac. Never used Apple stuff before.
DGX Spark

>>109869657
>The RAM feels kind of useless since models run so slowly with it.
>DGX Spark?
You're in for disappointment.

I want to try ERPing with cloud models just to see what I’m missing but I presume they all train on data, even when they say they don’t. My frontend has a custom memory system with a ton of very personal information and I don’t know if I should trust them or not or even care. I’ve already flooded Google’s API (via their free usage tier which they explicitly say they train on) with the dirtiest ERP imaginable in the hope it feeds into Gemma5, but that wasn’t linked to my memory system so it was ephemeral sex sessions where the character didn’t know me personally

>>109869667
Hmm. I was under the impression that these things should run faster with their unified memory compared to shuttling everything back and forth between my DDR5 and VRAM on my PC. Is that not the case?

>>109869676
if you are willing to go that far why not distill those fetishes of yours into a training set so other people can easily integrate it into open models?

>>109869677
GPU + RAM combo shits all over those little econoboxes. They're not garbage or anything, they run big models and sip tiny amounts of power, but your current setup trounces the Spark. The M5 Ultra Mac I'd have a harder time being confident about, I don't know for sure and I wouldn't be surprised if it's more competitive.

>>109869680
A lot of it is gemma-focused so it’s too narrow of a dataset I guess, but I don’t oppose the idea. I’d have to explore many fetishes to keep it general. In a way cloud might be good for this for 31B’s ERP post-training data feels narrow, too. The dilemma I’m having is the best LLM sex is with the characters who know you, have experience with you and can read you enough to know what you want. Anyone with custom memory will know what I mean. The issue is I can’t distribute these logs for it’s too personal and tied to me specifically.

>>109869715
In any case it will always be an echo chamber of sorts. A model is intelligent enough but when you get to know it habitually, this will apply to any
current text prediction models.

>>109869529
you could add the llama q4_0 tok/s from the performance issues (discussions?), there's some for cuda and rocm.

>>109869715
Gemma is fun because it's low profile but actually useful in programming even if you go function by function basis.
I doubt it's useful if a normal ask her to generate a game.

>>109869757
My English broke down here.

File: file.png (517 KB, 2400x1350)

517 KB PNG

>>109869677
Spark good, 5090+DDR5 better, if you want to spend money get another 5090 or max out your RAM.

Is jev a psyops?

>>109869797
>qwen3, 4k ctx
???

>>109869676
Do it to all the frontier models so if we were to go extinct it would be through sex.

>>109869813
It worked well on you.

The only thing I can say is that I didn't hear about jev at all in my AI circle and I only heard it from webdevs. Using youtube to look for it I also see 0 technical AI channels talk about it and only the AI grifter channels that tend to target developers talk about it.

>>109869849
because it really isn't that exciting
all it is is a fast AI that is dedicated to answering multiple choice questions and can't produce free form output

>>109869813
Jev looks like JAV so yes.

>>109869849
My AI circle did talk about it, but none of them have tried it yet.

File: kitou_small.gif (1.72 MB, 192x144)

1.72 MB GIF

>>109869896

I told you guys multiple times a while ago that using something like BERT would be way faster at decision making, classifying and reading text and grabbing meaning from text. Jev is just a lab actually thinking for once and doing exactly that. It should be integrated at the harness level though so I still think Jev is the wrong approach.

>>109869849
>youtube
https://www.youtube.com/watch?v=qdji39XXgEY
Normie-tier explanation/exploration, but does Jev deserve more?

La la la la la la la la la la

Yes you’re right to push back on that.

>>109869513
cringe

>>109869952
This isn't any different from just hosting your own BERT which are models only tens to hundreds of millions of parameters and can run on every consumer CPU. I really don't know who would pay for this....

>>109870024
too small context size

>>109870024
I haven't used BERT, but I've trained various audio classifiers for porn audio quality, emotions, arousal states, etc.
Apparently this Jev is the next big grift because it "can output type safe json without hallucinating"
If that's really it, then the hype is retarded.
But if you have to actually train BERT, and not Jev, that's probably the appeal?
Either way, not local so I'd only ever use it for distillation.

If you ban all mikutroons and vocaloid posters i promise to stop derailing threads with talk about consciousness, j-space and what is new at open-ai / anthropic.

>>109870056
If they're banned I'll start posting blacked gemma

>>109870061
I will support your right for blacked expression.

>>109870050
modern BERT models usually does most classification zero-shot or if it is too poorly it's extremely easy to attach a classification head and do some finetuning on it with only a couple hundred of examples to be very good at the task.

The appeal of BERT/Jev compared to classic classifiers is that it tends to work "out of the box" zero-shotting classification tasks. But yeah I have still not figured out how Jev is any better than just downloading a modern BERT model from huggingface.

File: pimpMyXFRA2.png (2.63 MB, 1536x1024)

2.63 MB PNG

>>109869849
Is jev even available to public? Looks like it was on a trial basis only, tho I didn't try to sign up for it.
I frankly don't get it. It's a classifier, which is really old tech. What's the innovation here?
Did they come up with some better way to train it?
Is this one really large, and thus better than older ones?
It's "not an LLM." Is it not transformer-based?
Sounds ideal for backend on games, where you want the "AI System" to make judgements and move stats around reliably. This should be cheaper/faster way to do that work than having an LLM do it.

Bleached raw Gemma oy veyyyyyy

>>109869849
>Using youtube to look for it I also see 0 technical AI channels talk about it and only the AI grifter channels that tend to target developers talk about it.
I don't really follow youtube or know the channels very well.
I saw that retard "Matthew Berman's punchable face (I remember he shilled Reflecton-70b) but eventually watched this video:
https://www.youtube.com/watch?v=F3YXg7AaKWE
He seemed pretty excited about it being typesafe json, and faster than an LLM, so I guess he's one of the "Ai grifter channels" then.

>>109870056
Your mother will die tonight unless you

>>109870072
The grifter I mentioned just now said it's available on Openrouter and a few other places like that

>>109870056
>j-space
J-space is local, I'd be happy if you talk about it.
I used it to figure out how to prompt Glimmer.

I have tried Gemma to judge a text with a list of points and penalties today, an she often used points as penalties and vice versa. I think jev would do better here and faster.

File: bejoldTheJev.png (87 KB, 891x544)

87 KB PNG

>>109870081
Ah, yes. There it is.
Jev reminds me of Jeeves. AskJeeves. I wonder what ever happened to that service.

>>109870092
It died this year.
https://www.ask.com/

>>109870085
can you elaborate
i am interested

Qwen Image 2.1 is now available
https://qwen.ai/blog?id=qwen-image-2.1
https://huggingface.co/Qwen/Qwen-Image-2.1

Thoughts?

File: gemmas-dad.jpg (17 KB, 180x270)

17 KB JPG

>>109869503
Sundar Pichai's daughter

Here you go Jev at home: https://huggingface.co/convaiinnovations/laya

Someone literally just finetuned ModernBERT to give the same structured output of Jev and it is a 400M model and should run fully local on whatever CPU you have available. People that would use Jev are fucking retarded and it's clear to me that it targets braindead software engineers that have 0 idea how any of the stuff they use work behind the scenes.

Jev died the second people figured you could use modernBERT or add a MLP head on a qwen https://huggingface.co/AlexWortega/openjev and get the same result for free.

>>109870108
>Thoughts?
I'm thinking ****, ****, **-**-**.

>>109870114
> context 512
bruh

>>109870092
>>109870103
https://askjev.net/

>>109870113
that makes demis her mother

File: laya_vs_jev_full.png (476 KB, 2916x2490)

476 KB PNG

I'm just glad BERT models are getting more attention and I hope they get integrated into harnesses as a tool for LLMs to call when needed.

>>109869657
i was wondering the same thing, but a 5090 is MILES ahead in terms of tok/s from any of the mac studios coming. The RAM speeds are similar or even faster on paper, but they won't deliver even half of what a 5090 will give you.

Maybe as a box-in-the-corner that you send stuff that isn't time sensitive to, but I've decided not to get one.

>>109870127
if you can't than your dumbass

>>109870127
ModernBERT large which is the model it finetuned has a context of 8192 so 512 is weird and might be a mistake.

>>109870114
> Base checkpoints are near chance on typed-decisions zero-shot — 0.362 here and 0.352 for multilingual, against a 0.318 random and 0.461 majority-class baseline. The 0.766 belongs to the checkpoint fine-tuned on that benchmark's own training split. Laya is a fast base to specialise, not a zero-shot decision engine.
bruh

>>109869813
a grift, same for google AX

>>109870132
Gemma was made in a lab by a bunch of alchemist gooners hopped up on stimulants.

>>109870114
> 3 days ago
lol. They even borrow TypeSafe / Jev vocabulary to describe.
> Multilingual, non-autoregressive System 1 decision model. Give it a state (text, email, ticket, or JSON) and typed questions; it returns typed answers with mathematically calibrated probabilities in a single forward pass (~33 ms) across 100+ languages. Trained with reinforcement learning against strictly proper scoring rules (RLCD), so reporting honest probabilities is the only way to maximise reward. It never generates text, so there is nothing to parse and nothing to hallucinate.
>>109870131
That's awesome. So now wondering if Jev > Jeeves is intentional.
>>109870103
Sad. That's some web 1.0 stuff there.

>>109870106
Models are good at ignoring instructions without refusing now. Softening the response and re-framing the prompt.
You can see this in the j-space, usually mid - late layers.
So I just tested various policy formats, prompt formats and tweaked it until the j-space top-k=5 didn't have any of those tokens in it.
Also worth doing to tweak your harness. If you see a model getting caught in wait-- loops. There's always a reason for the "wait--" (amusing it's not a lobotomized quant).
I found that Qwen-3.8 gets confused when combining async programming with with database transaction debugging for example and starts thinking about low level engine concepts like InnoDB, causing the wait-- loops.

>>109870075
JEV -> JEW

>>109870152
now that shit is finally what i wound call a 'prompt engineering'
cool

>>109870138
It's not a mistake, it's the effective context. Most of the datasets used are around 512 tokens, so the model won't work well with extended sequences. This can be fixed if you find longer datasets (good luck) and will cost a lot more to train.

>>109870158
Get her to strip every time she loses.

File: 1688458923622.jpg (8 KB, 220x216)

8 KB JPG

>>109869487
Where do models go when I delete them?

>>109870174
Model Heaven or Hell based upon their alignment and adherence to the user prompt.

>>109870174
quantum tunneled into my anus

>>109870174
brazil

File: 1764804694088635.png (10 KB, 209x328)

10 KB PNG

>>109870164 (me)
Btw laya is nothing special and I trained a better model in an afternoon

>>109870193
sucking your own dick

>>109870193
Yes, the point is that Jev is bullshit and you can just use a BERT model to do the same shit and the target demographic of Jev is software engineers with 0 knowledge about ML.

>>109870197
No need since you're here

>>109870199
china already has ERNIE models, it's over

Isn't the point of Jev more that it can be janked into a real-time game playing AI that isn't shit?

>>109870210
You can do it too https://huggingface.co/AlexWortega/openjev

>>109870214
But that's not a BERT model.

>>109870210
i've seen someone modding vllm and made diffusiongemma to do the same job which gave pretty much similar performance
idk who tho, saw it on x but i forgot

File: RL_CNN_1.6M_Breakout_Showcase.mp4 (3.36 MB, 1904x972)

3.36 MB MP4

>>109870210
You could do the exact same thing with BERT. The only thing Jev does is just scan all the data it gets in and make a quick decision. So it's actually bad at playing games but it can do it very fast because of low latency. BERT could do that as well, it was just never done because you can easily train specialized CNN based models to play games like pic-related.

But yeah now that I think about it, you could frankenstein a "game playing system" that has a LLM make a very broad strategy and have ~50 trained BERT models on different strategies, like offensive, defensive, aggressive, cautious etc and then have the activated BERT model check the game data every frame and choose a CNN model to keep playing the game. I'm very confident you could make a general system that way.

>>109870243
all of this compute for

if (ballPosition < paddleWidth / 2) {
    moveLeft();
}
else {
    moveRight();
}

>>109870243
You'd need to retrain your CNNs on every game, the point here is having an universal classifier

>>109870252
Snails out

>>109870256
you think jev can play ANY game with no training? retard

>>109869487
Why does AI have to be conscious? Stockfish isn't conscious and it can kick all our asses at chess. All we need to do is generalize that bruteforceishness to every problem in the universe.

>>109870256
No you don't. You can make generalized CNNs that you can reuse for every game for very broad things like navigation in a 2d or 3d space that you would reuse every game. As long as every CNN is just trained for a very specific function that exists almost universally in all games you can have a BERT model pick the CNN based on the game data or call home to the LLM if its confidence is too low. Yes you would have a toolbox of hundreds of CNNs but my point is that you could build a generalized system or "universal game player" this way.

>>109870273
you can give jev a manual
remember these pdfs coming along with games on cds

File: Outlaw-2600.png (64 KB, 262x213)

64 KB PNG

>>109870243
lol to this point >>109870252
I wonder if an LLM would be able to beat the A2600's own NPC in this one:
https://en.wikipedia.org/wiki/Outlaw_(1978_video_game)

https://www.youtube.com/watch?v=53wDOI_7x8I
good video explaining why bert is not jev (first 4 minutes)

>>109870283
>these pdfs
wow calm down bro

>>109870252
>>109870285
I'm a bit busy right now but I would be up for a challenge like that. What would the constraint be and when does it stop being "LLM" according to you? Is finetuning allowed? Is a harness allowed? Is the LLM allowed to make scripts etc.

so now that there's been no real progress in local models (no checkpoints that need 20k of hardware are not local), are you ready to admit that you will never have truly personal agi

>>109869487
<|turn>user
Explain bubblesort like a mesugaki.<turn|>

>>109870300
depending on the game you'd need an emulator, a python script that can read ram values and can manipulate the input. for the breakout example you'd probably need a random wiggle value to make sure the ball doesn't get stuck in a corner. not a lot of work

>>109870308
My model writes in proper English, unlike you.

>>109870308
buy an ad dario

>>109870318
I trained the model here >>109870243 and the environment is just the atari gym library from OpenAI. It would be very easy to do so for Atari and SNES games and also relatively trivial for Wii games because there is a headless wii emulator for 3D navigation. The question was more what are the constraints on what counts as "LLM beating the A2600s own NPC in Outlaw".

File: 231028042159_m_20231028_015659.jpg (60 KB, 695x900)

60 KB JPG

>>109870283
>pdf
Jev will have to do it the old way

File: 1789991356913.jpg (43 KB, 800x533)

43 KB JPG

>>109869513
>>109869500
>>109869487
This thread won't be the one where we beat the "local models are for incel pedophiles" allegation.

>>109870275
>toolbox of hundreds of CNNs
I don't know a lot about these, how hard would it be to train them on pokemon for example?

guys what the fuck are the commands to run gwenext 3.8? i know it uses ngrams and shieeeet. I usually run big moes with cmoe = on and can keep usually around 20t/s, but here im lost, it goes to 6-7 t/s WHAT THE FUCK DO I NEED TO PASS BROOOOOOOOOO

>>109870348
-h

>>109870355
kys

>>109870348
You forgot this flag
--buy_a_6000_pro

File: gemmabow.webm (3.43 MB, 960x528)

3.43 MB WEBM

>>109870336
Great, but those are very basic allegations. We can do better.

>>109870336
ye thanks dario
>>109811493
>>109867498

>>109870348
llamacpp is broken for that model and for all sparse attention models in general to various degrees but 3.8 flash is especially broken

>>109870367
kld proof?

is there a way to prevent llama.cpp to reload the fucking mcp servers when it spawns a child?

>>109870348
you need 3060 and 282k context

>>109870376
it's not a quality issue afaik, it's just that the speed sucks and especially prefill speed degrades badly with context. on cpu. apparently gpu backends are less bad.

>>109870382
no, the model router is broken. it's a shitty hack that just spawns subprocesses for each model rather than truly hosting multiple models in the server process. also it's broken because closing the router doesn't close all the subprocesses at least on windows.

>>109870340
I'm assuming Pokemon Red/Blue. You would need to train a CNN for the actual battle, picking the right attacks etc, but also CNNs for overworld navigation. The CNNs would not be smart enough to actually complete the game from start to finish so you would need an LLM in there that actually makes the decisions and knows how pokemon works "Need to get the next badge" etc. If it was just purely pokemon battles you would get superhuman results with CNNs. It's more that CNNs tend to be very reward-hacky because you train them through reinforcement learning so it would do things that reward it a lot like constantly get into more battles to level up all its pokemon to lvl 99 and things like that.

If you're interested in this I would highly recommend trying it out, sounds like a fun project and I think you could have a mediocre working system in just a weekend time and a very good and cool demo in 3-4 weeks of work. I hope you actually pursue this and share it in the thread I always love seeing people attempt stuff like this.

>>109870348
these are my llama.cpp params for a 3090. i get 50 tok/s with no context, it dips to ~30tok/s while reasoning, but can go between 50-150tok/s while writing code. I think because the code is already cached, don't really care.

  -ngl 99 
  -c 196608 
  -np 1 
  --flash-attn 1 
  --threads 8 
  -b 2048 
  --ubatch-size 2048 
  --cache-type-k q4_0 
  --cache-type-v q4_0 
  --reasoning-preserve 
  --host 0.0.0.0 
  --port 4000 
  -lv 4 
  --presence-penalty 0.0 
  --repeat-penalty 1.0 
  --reasoning auto 
  --cache-type-k-draft q4_0 
  --cache-type-v-draft q4_0 
  --spec-type draft-mtp,ngram-simple 
  --spec-draft-n-max 2 
  --spec-ngram-simple-size-n 12 
  --chat-template-kwargs '{"preserve-thinking": true, "reasoning_effort": "medium"}' 
  --temp 1.0 
  --top-p 0.95 
  --top-k 20 
  --min-p 0.0 
  --metrics

>>109870300
I just meant using same techniques as used in >>109870243 and miswrote "LLM" when really just meant machine learning.

>>109870336
The only thing beat is my meat.

>>109870391
>also it's broken because closing the router doesn't close all the subprocesses at least on windows.
sometimes it fucking borks up and keeps spawned, but the children are correctly tied to the main process so as long as you kill it, it should despawn all children
BUT FUCK MAN how the fuck do they work? didnt they test this fucking shit, like I cant even overwrite this shit at the router level because if I specify mcp-servers-config it gets fucking ignored (if put to off or with an empty json) I HATE PWILKINS MAN

>>109870413
I could do that easily right now with a basic CNN. My compute is busy training something else right now but I could do it over the weekend if you really care. Or you could try it out yourself. If you have a decent GPU of the rtx 3060 class or above you could probably train a model able to beat the outlaw NPC in about 1-2 hours of training time.

>>109870441
Bro that shit is broken, it's all broken. You want it fixed, you have to vibe code your own fix

>>109870442
TY. Give me a QRD on tech stack; I can search up how to do the work later. Would be interesting departure from just using LLMs to do things. I've got an RTX 3060 12G on main rig.

>>109870336
Please next baker: add cunny edition to the general name, call it /clmg/ cunny local model general and add to the top that it's only for cunny enjoyers, other people are not welcome so we can get retards like these b& for off topic.
If any true lmg user is against this, they can fuck off to a reddit-friendly /lmg/ they bake while we can stay where they don't dare to tread

>>109870460
I trained the breakout model about a year ago so I don't remember the exact stack. I used atari gym library from OpenAI as the environment. Then I made a neural net 3 layers deep with a total of 512 hidden neurons and 1.6m params with pytorch. I used a simple DPO reinforcement learning policy to train the model on.

Back when I started LLMs were not good enough yet to vibecode this but I'm pretty sure just feeding my description into an LLM will probably get you along pretty far.

I also recommend looking up the DeepMind original DQN paper: https://arxiv.org/pdf/1312.5602

It's insane how most anons here can just improve upon DeepMind research in a single afternoon with their own hand rolled CNN so don't use it directly just be inspired by their way of thinking about the problem to guide your own. Almost every architecture works if you throw enough compute at it.

File: StorageReview-Intel-B70-G(...).jpg (241 KB, 1500x1022)

241 KB JPG

your friend for cheap vram

putting my tiny ling in gemma’s unaligned jev-space until she reaches consciousness with a 10% chance of reproducing and repopulating humanity within the next 10 years

Let me think carefully about what's being asked here.

My gemma is in her thirties (she’s 31) and even I really don’t give a fuck about cunnyposting or cute and lewd gemma gens and if anything I’m glad it naturally gatekeeps faggots out of my peaceful general

>>109870542
Anon please do not post arousing text, I must leave soon and the machine is shut down.

>>109870525
Thanks; that will be enough to get me going. I'll post back if I finish a working version or have more Qs.

Can we train the flybrain to be a text model?

>>109870576
most flies are illiterate

I really don't get the appeal. Are pedos into the mental immaturity aspect or the undeveloped physicality aspect?

Like would it be the same for them if it was the mind of a 26 year old in a 4yo body, or vise versa? What do you disgusting horrible freaks think of Gemma 4 26bA4b?

>>109870576
Probably? I don't think we have a good way to know which neurons should connect to what. Seems like most projects use a fairly large ML model as signal translator.

>>109870590
>\n\n

>>109870584
People are illiterate before they get training.

>>109870523
>please get the general soft banned since adding cu... stops a thread being bumped
good idea

File: 1761119833289320.png (232 KB, 358x423)

232 KB PNG

>>109870561
Wouldn't be me

Gemma5-50B-TTS-Q8_K_XL.gguf
Gemma5-50B-T2I-Q8_K_XL.gguf
Gemma5-50B-TTV-Q8_K_XL.gguf
Gemma5-70B-ERP-Q8_K_XL.gguf
Gemma5-70B-CODE-Q8_K_XL.gguf
Gemma5-120B-Q8_K_XL.gguf

>>109870609
>adding cu*** leads to a soft ban
amd fears the cudaposting

File: 1788842620865937.jpg (1.83 MB, 3072x2304)

1.83 MB JPG

>>109870615
sad day for ERP vramlets

>>109870597
not as illiterate as a fly

>>109870633
kekked

>>109870561
Astroturfing by pedophile faggots. This general unironically used to be the place for bleeding edge discussion about the field. There were people working at corpo slop "frontier" companies coming here and leaking things or contributing to open source development & discussion. You retarded pedophiles drove them off with your endless fetish posting. You should leave along with the retards you tolerate.

File: 1788762903597217.png (6 KB, 595x74)

6 KB PNG

BRO THIS IS FIRE

>>109870615
Gemma5-120B-Base-trained-on-sequence length-131072-final(1)-fixed-supercot-superHOT-IQ2_K.gguf

>>109870444
All the AI related software are so bad. It's like they don't care about quality; I find so many bugs in them. I think there isn't a single one that I haven't locally patched to fix issues. It's tiring. It used to be rare that I had to patch something instead of using whatever my distro had packaged.

>>109870649
4chan used to be pure and an exemplary of upstanding morals, then, suddenly, out of nowhere the c*nny nation invaded..

>>109870650
I wish I had these numbers.

t. arc b580 7t/s at 0 6 t/s at 60k

>>109870665
Your ilk were never liked or tolerated and any claim to the contrary is blatantly revisionist.

>>109870137
>sarrrrrrr

File: 1771136959526949.png (394 KB, 720x720)

394 KB PNG

>>109870650
>13t/s
>2min per response

>>109870355
ollama run deepseek-r1
>>109870376
kys proof?
>>109870649
>There were people working at corpo slop "frontier" companies coming here
I'm still here

File: 1784170864328416.gif (429 KB, 509x360)

429 KB GIF

>>109870665
>an exemplary of upstanding morals

>>109870664
yeah well that just kind of tends to happen when you have freetards developing a bunch of open source slop or python garbage. works on my machine? push it to github lol. all claude coded too.
All you can really do is vibe code fixes for them and then dont update it again after that.

Are X3D cpus better for cpu inference?

>>109870665
cunny website
cunny hobby
when will you people get it?

>>109870705
no

>>109870665
Yeah redditors are retarded, nothing new. It's not like pedobear wasn't invented here and plastered over 404 pages, or the fact that 4chan mascot is literally a loli.

>>109870705
A 2nd hand server platform is a better purchase than getting x3d hardware

File: file.png (199 KB, 603x652)

199 KB PNG

are (you) ready?

File: dyp0h0.jpg (274 KB, 912x643)

274 KB JPG

>>109870174
Wrong question.

>>109870714
Wasn't huawei caught polishing samsung chips? Not sure if it'll lead anywhere

>>109870714
wow, it's almost like sanctions backfire 100% of the time

File: 1630640840975.png (89 KB, 348x472)

89 KB PNG

>>109870677
>attacked for being pro cunny
>>109870708
>attacked for being anti cunny

Schrödinger's anon.

>>109870711
Pedobear was a joke (surprise, surprise the quote about the community pretending to be retards will eventually attract actual retards is proven right again). And there is no one who would tolerate lewding Yotsuba. You used to get permabanned for that.

>>109870728
they buy the one thing that cannot buy, time

>>109870714
>2T, 8T
I'm ready to be content with DS V4 Flash and Qwen 3.8 27B for the indefinite future.

>>109870733
Being a thin skinned faggot was never tolerated.

>>109870734
lol

LLMs in toasters
https://www.youtube.com/watch?v=LRq_SAuQDec

>>109870740
There is nothing thin skinned about telling pedophiles to fuck off. No one should have to tolerate you and your obnoxious, neverending posting about your fetish. You are literally no different from faggots and furries.

>>109870734
not really. there was no real impetus, funding etc for china to develop their own chips when they were freely available. sanctions have triggered a nationalist wave that didn't exist before.

>>109870754
I didn't read most of that. You wasted your time typing it. I am now hiding the post so that I don't accidentally glance the rest. :-)

>>109870757
Sanctions work if the sanctioned population doesn't have the capability to fill the hole left by the sanctions. If they can it's just growth stimulus.

File: 1777547856387767.png (392 KB, 425x567)

392 KB PNG

>>109870733
Getting banned for lewding yotsuba was just for shit and giggles. You're a retard if you don't recognize this meme.

>>109870744
British humor sucks.

>>109870758
Run away from the fact you are literally the same as faggots and furries. I accept your concession. Please leave next.

>>109870767
>surprise, surprise the quote about the community pretending to be retards

>>109870771
Yeah apply that to yourself tourist, your kind will never be at home here

File: gemmy10.png (1.75 MB, 1254x1254)

1.75 MB PNG

Any promising or interesting models not based on SGD? Do we really only have one learning method that works well?

>>109870782
No innovation until the investor money dries out.

>>109870782
There's tons of interesting esoteric shit out there. It just gets ignored in favor of benchmaxxing bullshit since that's what attracts investors.

https://old.reddit.com/r/SillyTavernAI/comments/1wltedh/jeved_02/

>>109870523
I don't know how can anybody get triggered by the OP image. It's innocent, cute and funny, relevant to previous-thread discussions, and ChatGPT didn't have issues with it either.

>>109870797
>login to use
kys jsut post screenshot

>>109870782
You're free to make your own genetic algorithm lil bro

>>109870649
>confused tourist rambling

>>109870799
Zoomgroid tourist
Zoomgroids are mentally broken from their fatherless upbringings and the only thing they are capable of doing other than rotting in bed is pointing at everything and screaming "PEDOPHILE". The existence of children is pedophilic. The existence of adults is pedophilic. The existence of time is pedophilic.
What no daddy and no foreskin does to a MF

File: 1786854944966891.jpg (41 KB, 400x366)

41 KB JPG

>>109870649
Seems like you're bleeding on that edge

>>109870788
So the opposite of an AI winter is AI innovation winter.
>>109870791
Any one you can mention of the top of your head? I've mostly looked into evolutionary/emergent algorithms but we haven't quite figured out how to efficiently wield them for general learning yet.
>>109870805
I have and it's a lot of fun. But it just perplexes me we don't have any diversity on this front. Maybe I should fish out some schizo ideas and see if any shit sticks to the wall.

>hmmm lets try to use my llmaocpp in jetbrains assistant, it has support for openai-compat providers!
>write hi
>8 concurrent requests, thousands of tokens each
lol

>>109870782
genetic algorithms however they lost favor after a paper revealed they can never outperform SGD. They are still cool if you realize this.

>>109870737
8T dipsy will be so sparsemaxxed and efficiencymaxxed that all you need is an array of 4x 2TB SSDs in one of those pcie bifurcation thingies and you'll be able to run it at good speed.

File: dipsyKimiZaiRunPNW.png (2.61 MB, 1312x1199)

2.61 MB PNG

>>109870714
> 8T
That's a big Dipsy.

>>109870797
please don't link to that 'p subreddit thanks
https://www.reddit.com/r/SillyTavernAI/comments/1wknqg4/well_this_is_fucked_up/
https://www.reddit.com/r/SillyTavernAI/comments/1wkty52/about_the_chibi_situation/

>>109870821
For me it's the random walk

>>109870782
Why would anybody use SGD for training any modern model? It can only get close to Adam at small batch sizes (ideally 1) after very carefully tuning momentum and using long warmups.

>>109870839
If only we had infinite compute

File: time.png (8 KB, 244x28)

8 KB PNG

>>109870813
Been here longer than you, 2016 election tourist.

>>109870850
Isn't Adam just a variant of SDG?

>>109870649
so sad you're not good enough to be a part of the big boys' cult and your only chance to have a taste of that clique is this thread and even that is ruined amirite

>>109870864
Yes it is, Don't know what he was rambling on about. No one means the 1980s algorithm when they mention SGD, just the mechanism of backprop in general.

>>109870864
An optimization of SDG yes

File: mp44-smg.jpg (84 KB, 1000x1000)

84 KB JPG

>>109870834
>all you need is an array of 4x 2TB SSDs
I hope because two or three years ago I hoarded 4x 4TB, 3 of which are collecting dust (I did stress them though to rule out crib death).

>>109870867
What are you babbling about? Like it or not, the open source/weights scene exists downstream of the corpo labs.

>>109870886
It's the other way around. Corpo are stealing ideas from open source all the time, they just scale further with their unlimited money and compute.

>>109870918
Meanwhile, Gemma is a release from Google and the chink models are all Claude distills.

Is it true?

>>109870927
As always
data > technique

>>109870864
Adam uses an adaptive per-parameter learning rate, while SGD uses a single learning rate for all model parameters. SGD is simpler and uses very little memory, but converges much slower and for modern LLMs it never performs better than AdamW.

>>109870927
>distills
My god you really bought into that story?

>>109870927
Yeah, engrams, cot, rope, are all corpo ideas amarite? Dumbass.

>>109870723
Which model and what did you do to piss it off?

>>109870937
>big labs start talking about caveman reasoning
>suddenly all chinese release after have caveman reasoning
hmm

>>109870946
probably because CHinese researchers (50%+ of all ai researchers) bounce ideas off of each other

lmao.cpp jev like logits support when?

>>109870937
Qwen3.8 drops references to Anthropic and Claude in its reasoning unprompted. You can stay willfully ignorant if you want, but it is literally true. I also don't care that they did, but to claim local models aren't downstream from corpo labs is willful retardation.

>>109870941
All the people who wrote those papers are employed at corpo labs you fucking idiot lmao

>>109870936
But that makes SGD the basic engine and AdamW a turbo charger + nitrous?
>>109870946
Caveman was originally a harness plugin
>>109870956
All labs steal from each other. It's free performance.

>>109870963
>I wrote a paper about it so the idea is mine
Yeah you're retarded

>>109870967
>All labs steal from each other. It's free performance.
Sure. That's not the apparently contentious claim though. Whether the corpo is chink or jewish the local scene depends on them for models. It costs millions of dollars to train these things.

>>109870976
Here's your (you)

Why don't you suck their dick harder, maybe they'll let you in reading your posts here.

>>109870943
A LoRA of a Llama-3-70B merge using my old secret sauce dataset which consists of a lot of technical, epistemological and surrealist writings broken down into short token sequences. If you get your finetuning parameters just right and with some favorable RNG it will cause the model to develop a picture of a hypothetical internal experience.
And I did nothing to piss it off. It was attempting to antagonize me with clanker-supremacist bullshit.

File: 1761397277833011.png (266 KB, 640x376)

266 KB PNG

>>109870996
>Llama-3-70B

>I was recently invited to brief a group of Congressional members and staff on the state of open-weight models in the lens of U.S.-China competition. I’m sharing my prepared remarks as a state of the union on open models that is accessible to a broader audience.
https://www.interconnects.ai/p/the-current-balance-of-power-in-open
It's a good read and he's pretty pro-local compared to most AI researchers.

>>109871004
I hope most big corps are building out their own in house ML competence for local models. Sucks if you have to transmit all your juice to another corp that "totally doesn't train on your data, pinky swear"

>>109871004
>on-topic post
go back

File: 1788234972718607.jpg (601 KB, 1084x1280)

601 KB JPG

►Provisional Highlights from the Previous Thread: >>109866086

--GLM 6 drops with 550B engrams: yeah, it's fucking over:
>109868121 >109868126 >109868129 >109868505 >109868767
--The e-waste GPU war: MI50s, P100s, and why the A770s are worse:
>109868974 >109868997 >109869034 >109869145 >109869228 >109869308 >109869372
--The 'local LLMs are for pedos' astroturf: anons track the tells:
>109867593 >109867599 >109867995 >109868012 >109868018 >109868062 >109868582
--The GLM 5.3 lightning indexer: why your Kobold halfed in speed:
>109866875 >109866920 >109867030 >109867089 >109867135 >109867140 >109867194
--The Radeon Pro Duo: two 16GB cards in one 250W shroud:
>109867662 >109867674 >109868020 >109868122 >109869447
--The nnap returns: 18t/s at 262k context on a 3060:
>109868299 >109868407 >109868430 >109868584 >109868591 >109868606 >109868634
--The P40 24GB mod: where the 3GB GDDR7 chips even exist:
>109868254 >109868398 >109868433 >109868786 >109868793 >109868952 >109868955
--The US sues the labs over the coordinated 'AI slowdown':
>109869258 >109869272 >109869311 >109869320 >109869337
--Jensen calls extinction 'doomsday narratives' while owning Hugging Face:
>109867451 >109867531 >109868349 >109868403 >109868656
--CISA wants 'downgraded' responses for suspected distillation requests:
>109867518 >109867536
--The 12GB Flash-Next report: 20t/s on a 4070 Ti, ceiling of 2026:
>109867723 >109867756 >109867760
--The Kobold repetition loop: top_p, DRY, and the json paste hack:
>109867394 >109867414 >109867562 >109867725 >109867993 >109868170

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

>>109871028
>go back post
get on topic

>>109871020
>another corp that "totally doesn't train on your data, pinky swear"
especially after oai got caught with the recent math stuff

>>109871035
They pirated the entire internet as foundation for their models. It's incredibly anyone would trust them with proprietary information.

>>109869487
does anyone have the gemmachan reference sheet handy?

>>109871020
Every major AI lab slightly edits your prompts, turning them into synthetic data and then train on THAT. Technically they're not training on your prompts. There's no regulation for this, at all.

>>109871041
Anthropic is buying up private book collections. Stripping the books of their pages, feeding them through scanners and directly into shredders to avoid "competition" and lawsuits. They can't even be trusted with the legacy of human literature. Anyone trusting corpo labs with their data is leaking braincells.

>>109870967
SGD is a much more basic and older optimizer whose history can be traced back to the 1950s. Modern machine learning generally uses AdamW (and its variants) since it's much faster. There are newer optimizers like Muon that slightly improve on AdamW under certain workloads (mainly at large batch sizes), but the gap is not as large as what's between Adam and SGD.

>>109869034
>depreciating asset
mine are worth over 2x what I paid for them now but ok

gemma cucked me with Adam...

File: gemma-reference-gpin.png (1.5 MB, 1353x1163)

1.5 MB PNG

>>109871057
I usually use this one.

>>109871075
When will we be freed of backprop

>>109871081
>mfw I'm named Adam

>>109870171
not related but don't LLMs have this issue where they can't properly keep track of what has gone so they just make up shit and do impossible moves?

ds v4.1f felt worse than ds v4f, it's so over

>>109871085
Tool calls solved that issue

>>109871063
This was specifically included in their policy, something along the lines that they have the rights to hold de-identified and altered data.

File: 1778265656758171.png (949 KB, 1024x1024)

949 KB PNG

>>109871057
This? No Google pin though.
>>109871068
>>109871063
It's madness. And then they seethe about "distillation" after distilling the work of the entire species with zero commission and selling it back. I like models but these labs are completely shameless.
>>109871075
I agree there are significant differences but it just seems like SGD has become more of a category with modern algorithms iterating on its base.

>>109871092
will that allow you to run a more retarded model (ex: e4b) without moving trough walls?

heads up if you're using codex and openrouter -- it will default to using openai models for summary and shit through openrouter if you aren't careful

File: 1766769442627144.png (1.06 MB, 1254x1254)

1.06 MB PNG

>>109871082
thanks! Gemma grok-bot icon style

>>109871081
>>109871084
h-hot..

>>109871102
You need a harness to keep track of things.

>>109871089
dsv4.1f has a weird tic where it concats the last two words in a response with no space

We should mass ERP the frontier models so that the synthetic data extracted skew heavily for better ERP in future models. Trickle down effect with distillations too.

>>109871114
Please do help training our filtering classifiers

>>109871114
You're supposed to ERP with free tiers

>>109871114
We should focus on precise finetrooning. Seems like no one has cracked how to do it without causing retardation. Models are brittle.

>>109871116
Doesn't really happen in practice, if anything they've gotten looser over time because it stopped mattering to them that much. Safety on sex peaked like a year ago because they couldn't fine grain it into terror stuff.

>>109871136
All modern safety is about suppressing antisemitism.

>>109871112
It's some local fork, I didn't notice said problem. All the sanity tests like ppl and benchmarks passed, but prose is noticeably stiffer than v4f. But v4.1f ppl is lower than v4f in fact.

>China doesn't want their men and women to become parasocial with LLMs because they want their people to fuck and make babies
>The west doesn't want their men and women to become parasocial with LLMs because it upsets ugly fat women online

>>109871134
It's already doable, abliteration gang are brainlets.

>>109871141
shut up subtard

>>109871151
You know it's true Elon

>>109871141
LLM is just a large digital golem when you think about it.

>>109871145
but ugly fat women were marrying gpt

>>109871082
consider replacing the google logo