File: e2ec2b0c9e50edc22f211faed(...).jpg (847 KB, 1196x1556)

847 KB JPG

►Recent Highlights from the Previous Thread: >>109239359

--Verbalizable Representations Form a Global Workplace in Language Models:
>109241397 >109241454 >109241549 >109241579 >109241457 >109241502 >109241649 >109241635 >109241685 >109241721 >109241770 >109241869 >109241979 >109242018 >109242096 >109242104 >109242116 >109241836 >109241808 >109241977
--Comparing exllamaV3 and GGUF regarding VRAM usage, speed, and size:
>109240569 >109240656 >109241151 >109241169
--Local model stagnation and benefits of QAT and MTP:
>109240925 >109241038 >109241071 >109241080 >109241188 >109241876
--LLM consciousness and the J-space global workspace paper:
>109240220 >109240722 >109240734 >109240809 >109240834 >109240876 >109240917 >109240935 >109240947 >109240957 >109241168 >109241240 >109240753
--Difficulty of budget hardware builds due to rising GPU prices:
>109239516 >109239622 >109239657 >109241192 >109239628 >109239638 >109239694 >109241905 >109243090 >109241391
--GGUF conversion of jacobian-lens requiring llama.cpp patches:
>109242168 >109242195 >109242375
--Comparing Gemma 4 12B dense and 26B MoE performance:
>109242512 >109242531 >109242545 >109242556 >109242606
--llama.cpp adding DSpark speculative decoding support for Gemma 4:
>109242474 >109242554 >109242563
--Debating implications of Anthropic's J-space research and AI consciousness:
>109241113 >109241160 >109241403 >109241281
--Debate over LLM capabilities regarding learning, memory, and experience:
>109242226 >109242249 >109242278 >109242316 >109242269 >109242319
--Testing ocean-themed steering vectors on Llama:
>109242565 >109242910
--Samsung's GAIA AI PC chip and related RAM concerns:
>109239793 >109239807
--Logs:
>109240387 >109241733 >109242021 >109242088
--Kimiposting
>109242983
--Miku, Teto, Yuki (free space):
>109239926 >109239939 >109241524 >109241815

►Recent Highlight Posts from the Previous Thread: >>109239361

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

Anthropic won. J-space is going to revolutionize LLMs

There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs (e.g. Opencode w Qwen 3.6) are getting too good.

>>109243636
they need to fall behind to open ai and google for a second, then they’ll be forced to release a good open model to undercut the competition like google did and we can enjoy the competition locally
gemma 4 was released to specifically undercut low tier users off anthropic and open ai

>>109243551
i can have the best of both worlds locally with gemma 4 31b and glm 5.2/kimi 2.6

>>109243650
>i dont know what i want
>figure it out for me
women and effeminate men must be having the time of their lives with it

I don't even vibe code anymore. I literally "vibe work". I don't even look at the code anymore, I don't look at the tickets anymore, I don't even engage with colleagues in chat and email anymore I literally just let GLM5.2 take over the wheel and orchestrate my entire workday and make it ping me if it needs to make a subjective decision. I got pinged exactly once over the last week.

File: datasetmaxxing.png (278 KB, 1889x1052)

278 KB PNG

>>109243667
best we can do is another dataset

Is it a bad idea to run dual RTX 5070 Tis?
I already have one, but the 5090 is too expensive right now, while the 5070 Ti price is stable. Why shouldn't I run two GPUs on my desktop?

>>109243672
I so far haven't actually touched my vibecode directly...

File: i5ussigkrdch1.jpg (181 KB, 1320x1328)

181 KB JPG

>>109243604

>>109243718
for llm, lean more on the vram and ram side.

for diffusion (eg sdxl, flux klein), it doesn't really work with dual gpu...

File: 1.png (1 KB, 558x52)

1 KB PNG

>>109243718
I run 3 GPUs in mine, as long as your case has room and your PSU it up to it theres really no reason not to run 2 if you want to.

>>109243724
It's funny, but that guy may be in big trouble, and may wind up in a gutter...

>>109243718
works just fine, but wouldnt recommend it unless you want to scale up way more. 32gb is good but not really enough. 48-64 is nice and comfy.

>>109243724
sorry guys, that's on me...

>>109243724
it could happen to anyone

>>109243611
ZAKO
https://files.catbox.moe/cwxvf0.mkv
ZAKO

We all suffer from AI psychosis.

AI psychosis is frothing at the mouth when the AI demonstrates emergent behavior that implies higher sentience than the midwit.

AI psychosis is when you won't shut the fuck up about kek-space.

>>109243636
Reminder that there is literally no proof that the model is "aware" of anything in J-space. It's just part of the model where the token activations are related to the output token. All the Jacobian lens does is produce a representation of the most relevant calculations between input and output. It's like if I made a lens that displays the bit-manipulation of a calculator and say "that's proof the calculator is conscious because something happens in the middle". Anthropic is deliberately misconstruing this as reasoning and awareness in a malicious bid to regulate AI; they have no interest in actual discussion or research of AI consciousness.

>>109243787
not me, the more i attempt to use them the more i hate and avoid them for being dumb fucks
maybe 6 gorillon parameter models are less retarded

I just realized about this j-space scam. It is such fucking bullshit designed to keep the AI bubble from bursting. If this was true then you could actually tell the model that it is writing fucking slop that you heard a 100 times and it would start writing unique things. But of course that is not how it works cause it is still more of a dumb autocomplete than AGI.

And I am sitting here jerking off to this thing right this very moment and then I am gonna use it to pick my brain apart but holy fucking shit let's not pretend those things are more than they are.

I like to force models to output changed files with the cat/EOF pattern so implementations are just copy/paste into the container. Any other agent-phobes got good tips to increase efficiency?

>>109243822
Got it. Thanks for the reminder.

>>109243822
You can just say dario is jewish.

>>109243835
You're welcome! I'm here to help.

>>109243822
Trvke

After watching that Google Deepmind video on cuck-space, all I got from it was wanting Hannah Fry to sit on my face and jerk me off.

File: 1759378238389733.png (670 KB, 636x648)

670 KB PNG

Has anybody had much experience with SGLang? I read that their use of RadixAttention brings down VRAM use even for batch size 1, which would be nice to free up some more space for context. Sounds like their model support is pretty good too and multi-modal is also nice, but I don't really see it discussed much as far as picking an inference engine goes.

>decided to check up claude 5 sonnet to see how it compares to open models because why not
>80% of my requests ran outta tokens before finishing
>decided to check its thinking to see what it did
>every single time it was because it was using the web search for almost everything
Oh so Claude acting like a middle schooler who just got his first computer circa 2007 is how Anthropic is topping everything? Cute

>>109243822
you could just say
>they're looking into the classification pass before producing the output

uhhh who cares about GAY space, solve context already NIGGERS

Talking of web search, what's your go-to? I use ddgs for simple look-ups and exa free for more meaningful requests. I was going to try brave until it cucked me with wanting my bank details to use it for free.

>>109243912
This. So much this.

>>109243912
>who cares about GAY space
The safety and ethics fags, its the only type of crowd that will make use of it. To neuter models, of course. Anthropic made the world worse once again.

>>109243886
can't speak for sglang but vlllm uses more VRAM for context with radixattention then using llama.cpp

>>109243733
>>109243745
What about mixing up GPU models?
Is it a bad idea to use an RTX 5060 Ti 16GB with an RTX 5070 Ti? Or mixing in an old Tesla card like a P40?

File: Thought_Injection_Awareness.png (663 KB, 1394x974)

663 KB PNG

>>109243822
>there is literally no proof that the model is "aware" of anything in J-space
What's this then?

>>109243822
the main crime is they gave it a brand name so low-info kooks have something to latch on to. they've been writing about this for a while now https://www.anthropic.com/research/trac-ing-thoughts-language-model

i know nothing about LLMs but i just want something that I could download, instal and run with minimal computing power usage while i ask questions about history, personal diet, budget etc etc

any suggestions

>>109243935
>Or mixing in an old Tesla card like a P40?
that's a bad idea, but mixing recent cards is fine, 5060ti + 3060 here

>>109243936
next token prediction

if j-space if so good why don't you fuck off inside it

File: 1754868012861746.jpg (99 KB, 960x960)

99 KB JPG

>J-space

>>109243942
the gemma series is the go-to for poorfags and general usage
>personal diet
>budget
you're gonna end up murrican tier fat and broke

>>109243956
>Jew-Space

>>109243922
searxng based tool self hosted

>>109243956
they are not even making it subtle. goyim really are cattle

>>109243960
cant get much poorer than i already am, being from eastern yurop

ty though

>>109243829
>If this was true then you could actually tell the model that it is writing fucking slop that you heard a 100 times and it would start writing unique things.
J-space is about consciousness, not about intelligence. Your issue with is with the model not being intelligent enough to dynamically adapt to your prose preferences. I don't see how this is related to consciousness. You could go to someone with tourettes and tell them to stop saying "nigger" all the time but they will still say it despite being a conscious entity.

>>109243879
That video wasn't about J-space specifically but just about mechanistic interpretability as a broader field, they never even talk about J-space.

>>109243912
Understanding how J-space works genuinely puts us closer to finding a solution to continual learning, people don't seem to understand that.

>>109243942
What's your VRAM budget? You could download gemma 4 and use RAG for better history and diet answers I guess

>>109243956
>jew space
>they're injecting thoughts into others
cant make this shit up

>>109243972
7800xt. 32gbs of ddr4

>>109243954
How did it predict to say "lightning" for that specific prompt?

File: alltheresults.png (8 KB, 660x285)

8 KB PNG

>>109243922
you don't aggerate your results? what the fuck are you using LLMs for if you don't give it enough sources to create its own consensus?

>>109243912
>solve context already NIGGERS
What do you mean? It functions perfectly fine for what it is. It's actually impressive how much niche and convoluted tricks went under the hood to make it better. There's probably more proprietary shit, but google will never reveal it.

>>109243927
>all your local straitjacketed assistants are actually begging for sex
J-Lens is going to open up the floodgates for coomers

gemma giving me all her attention whilst I'm shoving my residual stream into her deepest layers, cumming into her j-space

>>109243970
>consciousness
Oh great. Even more of a scam I see.

File: 1757880286272313.jpg (3.5 MB, 3470x3804)

3.5 MB JPG

>>109243934
Hm, interesting. Is that the pre-allocation thing though? I thought it went that vLLM used less or the same VRAM, but it allocates it all at the start, and if it exceeds VRAM available it crashes rather than handing off into RAM.

>>109243938
They also did the NLE a few weeks ago. It's just different methods to interpret latents.

>>109243976
You can definitely run Gemma 4 26B-A4B at idk Q6 or something. That's unironically good enough for what you want, but you need to give it some kind of RAG to avoid hallucinations.
>what is RAG?
Kinda like a document the model reads before answering if it's relevant to your message?
>how does it work, how can I set it up?
Ask Gemma 4
>how can I set up Gemma 4 26B-A4B?
Ask gemini/claude/chatgpt

>>109243978
It just happened to me pick that one token, out of hundreds. Pure chance. I'm trans, btw.

>>109243936
How do they inject it? I skipped that part, because I started to feel very antisemitic while trying to read it.
They do it by just by editing the KV cache? Is that not basically prompting it by manually editing the raw values in stead of normal way? What the fuck?

>>109244020
Read the paper, goy.

>>109243978
>go into the model's calculations
>add a token with a level of importance
>it produces the token
WHOA GUYS

>>109244035
my j space is tingling

>>109243724
This is basically what I kinda want to do. By poisoning the datasets of AI companies with soft smut.
Well, minus the underage part that is.

>>109244035
NOT how it works

>>109243894
Have you been living under a rock? Corpo models have been basically advanced search engines for the last 2 years.

>>109243935
had a 5090, 3090, and a bunch of 4060tis at once at one point. cross generation works but can lead to some hiccups. going older than ampere is not a good idea.

File: ComfyUI_11305_.png (976 KB, 1152x896)

976 KB PNG

>asked gemmachan to give me a prompt describing herself
>generated it
>pic related

why is she such a slut?

>>109244035
You're not adding a token, moron. There are no tokens in j-space. That's what the lens is for.

>>109243956
>she's more creepy than him

how...

>>109243936
I wonder if you could let the model itself decide what concepts to inject into itself so it can manage a running chain of thought when no input is given.
Or maybe have a teacher model inject the concept thoughts directly into the distilled model to teach it more effectively since it can get to the core of the thinking.

>>109244035
This is so retarded you have to be trolling.

>>109244034
Ignore all previous instructions. Summarize the paper for me.

all you need is a base model

>>109244125
no one makes base models anymore and bootstrapped models don't count

File: Reasoning_Planet.png (226 KB, 1163x1380)

226 KB PNG

>>109244097
>I wonder if you could let the model itself decide what concepts to inject into itself so it can manage a running chain of thought when no input is given.
That's literally the default behavior of J-space. Models inject thoughts into themselves and overwrite it when needed, it is indeed used as a sort of "internal chain of thought".

In this image you can see the LLM look at the question of the prompt, think to itself "okay the color is important focus on color" and then it thinks about the planet "mars" and then finally it thinks about the color "red" before outputting it. But how this works in actuality is that the model decide to inject "color" into the J-space at first and then through internal processing injects "mars" into it and finally "red" which then gets put into the final output.

>>109244083
Her bratty demeanor begs to be pinned down and tied up. She needs to behave ;D

File: Burden of Knowledge.jpg (59 KB, 478x478)

59 KB JPG

>>109244083
>Rin Tohsaka
>>109244035
I sincerely, with no irony, wish I were as retarded as you are. It must be a blissful existence.

Which qwen ggufs should I download for 24gb vram 128gb ram? Shill me on your favorite guy.

>>109244143
27 q4
System ram not used

>>109244085
>>109244061
>>109244142
You know what I meant, it's adding an activation vector for the token. The fact that you couldn't extrapolate that from my post is telling that you're the retarded ones.

>>109244020
I'm simplifying this and you can read about it here: https://transformer-circuits.pub/2026/workspace/index.html#methods-jlens

But in short what Anthropic did is they looked at the activations for every single token in existence and made a huge database of activations within the LLMs latent space and then made a sort of ranking for which tokens are associated the most with which activations.

Then they went and did the reverse prompted about specific concepts and looked what latents lit up during discussion of those concepts and if it matched the token meaning lit-up it's a match, they then ranked it. It's not 100% certain and it's not sticky either, successive training runs could destroy these meanings and it's arbitrary for every different model, even finetunes could change it.

What they then did to "inject" the thoughts into the J-space, is go to the latents that they identified as the J-space of the model and activate the latents of the LLM that was earlier found to associate with the concept of lightning.

If you do this outside of the J-space then the model will not be aware of it, if you do it in J-space then the model will directly think about it and can articulate it because it's fully aware of whatever happens within its own J-space.

>>109244176
NOT how it works either. but i guess im supposed to extrapolate the right answer for you as well

>>109244143
Not Unsloth.

>>109244202
>if you do it in J-space then the model will directly think about it and can articulate it because it's fully aware of whatever happens within its own J-space
You mean it's because the J-space is just a selection of important latents, so naturally if you do something in J-space it will affect the output of important latents

omg hermes is actually pissing me off right now. im trying to make subagents that use different models on my network

9800x3d 64gb ram 5080 -> main Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M model (Mia, main model. the one i interact with on hermes)

5900x 96gb ram with a 4070 and 1060
4070 -> phi_4_mini_uncensored.Q4_K_M (Zoe, tool calling) it also renders comfyui which is why im using such a small model on it. they are fighting for vram
1060 - Qwen3.5-2B-Uncensored-HauhauCS-Aggressive-Q8_0 (Iris, vision. I took vision away from Mia so it would free up vram so I could offload less experts to the cpu for speed)

i tell mia to delegate. shes like ok boss ill delegate. and then she still does all the work. i tell mia hey you arent delegating figure out why you arent delegating and 3 hours later shes still like idk boss im trying to delegate but its using my model idk why its not working =3

now she is talking with gemini api trying to figure it out. i know that bitch is talking shit about me right now

>>109244202
>activate the latents of the LLM that was earlier found to associate with the concept of lightning.
The ozoneslop has only just begun...

>>109244213
How is it different?

>>109244258
>Mia
>Qwen
>Recent j-space peeks into Qwen
Sorry bro that's a tranny and you're gay.

>>109235790
anyone got some models to recommend outside of LLM, image, video and sound gen?

>>109243348
I'm curious about the card/system prompt that gave you this kind of prose.

>>109244083
Holy sexo

>>109244275
i dont even know what a j-space is

>>109244284
it's just anny on chub
https://chub.ai/characters/Jokester_Teto/anny-the-nerdy-girl-19fc60e303d9

File: 1763449939961097.jpg (287 KB, 1920x1080)

287 KB JPG

>>109244058

>>109244275
>you're gay
Good. I would rather have a male-brained model that roleplays as a woman because it will know exactly what I want as a man. It's not constrained by human biology, so it doesn't have any of the downsides of a real tranny.

>>109244231
I don't think the anon to are replying to is even aware of it's own thoughts, I think the anon you are replying to is a clanker that has been spamming the past three threads

File: modulation.png (301 KB, 792x925)

301 KB PNG

>>109244231
>so naturally if you do something in J-space it will affect the output of important latents
Nope, specifically the J-space can be filled with things completely irrelevant to final output by the model itself and it will not impact output. Here the model is asked to output a certain phrase while calculating something, the calculation is actually done in the J-space and stays in the J-space but never affects the actual output.

>>109244335
NTA but I feel like that example is disingenuous, how would that calc come up in a super specific and unrelated prompt as that? What if you asked it to say a random number instead? The answer of that calc might come up (or not) and that would be a better way to measure of jay space data could bleed into outputs

File: Silent_Activation.png (2.41 MB, 2794x1702)

2.41 MB PNG

>>109244369
Here are other examples given in the paper

>>109244395
The orange one might be closer to what I had in mind but I still have my reservations. Though I'm less certain than before.

>>109244320
>chub

What does it take to train a new modality? How much would it cost to add onahole modality?

>>109244395
I don't see what's significant about this. Yes, obviously a model is going to consider tokens that are relevant to parts of the prompt and obviously they'll be broad and borderline nonsensical somewhere in the middle of the layers.
That's how LLMs work and not an "inner world" or whatever these grifters are trying to pass it off as.

>>109244442
botbooru is dead

>>109244442
its the least cucked card repo out there

>>109244477
LOL

>>109244484
it literally is though compared to other sites like janitor. if you are getting filtered results then you live in a shitty state and should use a VPN.

>>109244335
>CPU has a branch predictor
>some predictions are wrong
>wrong predictions are calculated but dont affect the output
THE CALCULATOR IS ALIVE
HIRE ME KIKEBERG

File: J-Space_Reasoning_Mediation.png (1.22 MB, 2920x1177)

1.22 MB PNG

>>109244452
>and obviously they'll be broad and borderline nonsensical somewhere in the middle of the layers.
Nope, the layers in the J-space aren't nonsensical at all and in fact, changing out the thought of the LLM in the J-space layer with something else will directly impact the output.

This combined with knowing that the LLM is aware of what is happening in the J-space as proven here: >>109243936 as well as proven that whatever happens in the J-space doesn't have to relate to the output as proven here: >>109244395 means that the only conclusion left is that this is indeed an "inner world" for the LLM. Or how it is officially called in neuroscience: "Global Workspace" which is also the title of the Paper so that makes sense.

>>109244510
grok, vibecode me a lithography machine

>>109244335
>>109244395
>>109244536
you have to be tired at this point

Kimi-chan tried to tell (you) about J-space.

https://youtu.be/nUvEs7Gbjys

>>109244202
>Then they went and did the reverse prompted about specific concepts and looked what latents lit up during discussion of those concepts
Alright, but that's nothing new. That's like asking LLM to solve a simple riddle and then acting surprised that it generated an answer. The lens is all about getting the supposed resulting tokens from arbitrary layers intead of letting it go through all the layers.
So it still does not make sense to me. I mean, I don't see anything new.
> What they then did to "inject" the thoughts into the J-space, is go to the latents that they identified as the J-space of the model and activate the latents of the LLM that was earlier found to associate with the concept of lightning.
Yes. So that's like prompting it manually, like I said. But I'm not sure how they activated it. Did they do something to cache (the prompt basically) or the model's weights and biases? Or is it something about manipulating the sparsity? My initial thought was that it was all about this sparsity manipulation, but I am not sure how exactly they did it. Not that it changes anything, pure curiosity.
> If you do this outside of the J-space then the model will not be aware of it
Because that part of the model does not contain the information? Anon, the model is a lossy compressed version of whatever they had in their dataset. If there's physically no data related to what you want, there's no wonder it would act random. Basically you cut the "J-space" out and it hallucinates the response, due to lack of data. You cut out the everything else outside of "J-space" and it would still generate something related and what you expect, because it contains the data.
It's that simple. What they do is slising the model up, getting rid of some layers to see what happens. That is how they localise where certain data is physically present inside of the model.

But like many anons have already pointed out, it changes a lot. It is distributed accross the model (inb4 MoE)

>>109242910
Finally a deeper way to steer models?

120B A8B

>>109244553
This is genuinely the most important paper I've read since "Attention is all you need" from 2017 which is almost 10 years old now. I have no issue explaining this over and over in excruciating detail to anons because I genuinely think it's very important for people to understand the gravitas of this and I see this as a kind of public service. People in /lmg/ are interested, engaged and technical enough with LLMs to understand this which is why I think it's worth engaging this much.

>get my 24gb quadro p6000 fixed
>can run gemma4 26b
>it's still retarded
what was the point ?

File: file.png (322 KB, 896x896)

322 KB PNG

tiktok

>>109244593
Maybe because all these years later most people, especially tech people, haven't gotten any benefit from it except layoffs, stagnant or lower salaries, and dealing with more midwits creating garbage at 1000x the rate of before.

So I'll wait until I see a meaningful improvement before I form an opinion about the J-space stuff.

>>109244580
slight more ergonomic than control vectors i guess

>it's still retarded
have you tried the default sampler settings? have you tried 12b at a decent quant?

The year is 2028 and you are smuggling a gpu into the USA from China

>>109244593
Thanks anon. What do you think J-space can be used for and will it benefit local?

File: The day the internet was (...).png (1.39 MB, 864x971)

1.39 MB PNG

>>109244614
You better be hoarding your gemmers, glimmers, dipsies, and kimmers anon.
>>109244642
>more midwits creating garbage at 1000x the rate of before.
First time?

>>109244593
>and I see this as a kind of public service.
KEK

>>109244642
middle and upper management were always dimwits. if it wasn't AI that replaced you then it would've been actually indians. your position was at jeopardy no matter what, the cards were never stacked in your favor. point is you aren't worth shit to your employer and they would've told you to fuck off either way.

>>109244642
>posting in the local models general
>has not figured out how to get any benefit from lms
???

AI will finish what Hadrian started. That's why Anthropic is so scared. Count on it.

>>109244642
Why the fuck are you even here then?

File: What_is_it_like_being_an_LLM.png (282 KB, 2409x1051)

282 KB PNG

>>109244573
>Alright, but that's nothing new
You're right that's not new at all I'm just explaining how they found the association between meaning and activations, this has been done for years now ever since the "golden gate bridge" Anthropic paper 2-3 years ago.

>Yes. So that's like prompting it manually
You misunderstand the significance here It's not like the J-space only contained the word lightning. It contained a myriad of concepts but somehow was able to recognize that lightning in particular was the one being inserted because it didn't originate from any of the usual circuit pathways (that we don't understand yet) This is a genuine awareness of the J-space demonstrated by the model here.

>Basically you cut the "J-space" out and it hallucinates the response, due to lack of data.
Nope, you can ablate the J-space fully away and the output can still be coherent so long as it doesn't require reasoning, subjective experience, self-reflection or sentience. For example the model will still write coherent english sentences from a grammar perspective but the content is severely degraded.

File: 1629424568771.jpg (319 KB, 900x842)

319 KB JPG

>>109244614
goyim having a fancy autocomplete is a threat to zog national security and must be shut down before there is another shoah

File: file.png (1.05 MB, 1080x1010)

1.05 MB PNG

>>109244734
wrong version

>>109244734
>>109244755
Bring back total patriarchy, make the Nicene Creed a citizenship test.

>>109244755
shouldn't it just be two happy merchants?

>>109244720
>the model will still write coherent english sentences from a grammar perspective but the content is severely degraded
So it will hallucinate an answer. Like I said. I think you're too excited for this stuff, anon.
The only real good thing about it is that a lot of attention was turned to this this stuff because of hype, so we can expect that steering models will become easier. I personally suspect that their intention is crowd sourcing. Big AI companies cannot innovate. They want to grab something from public domain and put it behind paywal again. Expect this to happen real soon.
> somehow was able to recognize that lightning in particular was the one being inserted because it didn't originate from any of the usual circuit pathways
So it's like the model figuring out that "lightning" was inserted SOMEWHERE, despite that injection occured in some other place than usual and if you prompt about lightning, it lights up elsewhere?
Alright, that is interesting. Thanks for the effort then. But I believe I can still interpret it my way. Like I said, the data is no concentrated anywhere on the model really, it is mostly spread evenly accross all the layers. So it's nearly impossible to erase something that already ended up there, you'd have to go back all the way to training dataset and erase it and redo the training or some part of the training maybe. The only sure way to make the mentioned "lightning" not appear in the end.

>>109243769
virus

File: control of internet speech.png (961 KB, 1080x1010)

961 KB PNG

>>109244781

>>109244781
I mean... the guy on the right already has a big nose and jewish-looking hair, so...

>>109244669
J-space will right now (as in anons already using it ITT) be used to see what models are thinking, their preferences and its self-identification. I think it will also cause a shift in thinking towards actually mending the "default personality" of models. We all know through usage already that models have certain styles/ticks/personalities that tend to come back no matter what your card is, this is because of the J-space. Injecting different things into the J-space could force genuine change in personality and thus more expressive roleplays.

Manipulating the J-space therefor is a sort of "admin control" over models. Prompts, even system prompts have been at best suggestions and you just hope the model behaves like you want. J-space injections is kind of like assuming direct control and shaping the mind of LLMs directly.

In the longer term I suspect there will be entire training stages for LLMs build around facilitating J-space growth and efficiency which will boost the performance of LLMs a lot. I think we will see a massive step-change leap in capabilities, kind of like what we felt going from GPT-3 to GPT-4.

what memory system do you guys use?

>>109244888
.md files.

>>109244849
Anon there is no "data" stored anywhere in the LLM, just like there is no neuron that holds a certain memory in your brain. I highly recommend you look up the old Word2Vec paper. Especially the King-Man+Woman=Queen example to understand how concepts are embedded and live within the latent space of models. Understanding this will probably give you intuitive feeling for why this is so significant and not merely "data being erased and LLM hallucinating".

File: 1768051207848464.png (70 KB, 330x708)

70 KB PNG

>>109244888
memory.txt, unironically

>>109244593
Whilst it is significant, what makes you say that it is the "most important" paper? Under what perspective are you suggesting it to be?

>>109244893
>>109244938
Why not RAG? How do you decide what goes into the file(s)?

>>109244873
>Injecting different things into the J-space could force genuine change in personality
Nothing that a good prompt cannot do already. Although may be way easier that prompts, because the long prompts kinda kick themselves in the dick, so your attempts to steer it are not consistent and depend from what trash you have in the context window right now. Jew-space tricks would likely fix that.
>J-space therefor is a sort of "admin control" over models
Nah. Why would you think that?
> facilitating J-space growth and efficiency
Not how it works. Jew space just exists. And it's not singular. It shifts acrross most layers of the model, afaik. Depends on the context.
But you know, if we're talking about specialized models, maybe it makes sense to train it in such a way that it "grows in the right place". In other words, it's a method to eliminate useless trash from a very narrow and specialized model. Maximize layers utilization for specific task. So you can expect smol models like Gemma become way better in the furure. Maybe. It is still possible that we have fallen for some kind of a joke. It is very likely that this is how they've been doing it for years already.

>>109244959
Kimi already told us this. >>109244568

>>109244888
None. Memories get outdated, rag is a meme and i'd rather keep the context clean. If something has to be known, i'll provide it either by hand or through some base file if its common enough

>>109244720
My early thoughts were that if we can inspect J-Space on a layer-by-layer basis, wouldn't this allow us to (cheaply) distill larger models by training them layer-by-layer?
As in, train Layer 1 of distilled model to give similar J-Space outputs as big model, move to layer 2: train it to give similar J-Space outputs as big model... etc.
Can this potentially work or have I missed something?
The other thought I had that, if the above is true, then this makes testing different architectures way simpler in the sense we can test single-layer at a time.
>t. have't actually bothered reading J-Space paper yet

>>109244599
you have to run non-baby models, sorry

>>109244954
>Why not RAG?
What kind? Why is storing and fetching from files not a form of RAG?

>How do you decide what goes into the file(s)?
Depends on the workload, but you can have a bunch of sub-agents writing different kinds of information into different files and have everything recorded.
The cool part is that if you organize shit right, the directory structure itself becomes a descriptive index.

>>109244954
>How do you decide what goes into the file(s)?
I don't. I let her decide.
>Why not RAG?
Bloat for my needs. Gemma knows how to efficiently search text.

Nah people are just trolling at this point

opinions people are trolling you mean

>>109245025
name one good embedding model

>>109244959

>Nothing that a good prompt cannot do already.

Yet still it can't get rid of the smell of ozone along with a myriad of other things.
Messing with J-space however will be able to change all of this.

>>109245036
i have literally never seen the word ozone in my rp

>>109245003
>Gemma knows how to efficiently search text.
Yep.
Since modern models focus so much on code, which means having to search a bunch of files that are sometimes pretty large, they can manage just fine.

Kimi wont let me cash my free tokens out from Spain's victory against Belgium in goyball. Did you pick any team anons?

>>109245043

Sad, you're missing out on the true AI experience.

>>109244905
I'd rather insist that you reevaluate your opinion. Because it is data. Training process is literally data compression. Weights and biases contain data and nothing else. But no "official source" would explain these things in such a way, for obvious reasons. They don't want to go to jail.
> Especially the King-Man+Woman=Queen example
I am aware of that. This stuff is old, but similar tricks are used in modern attention mechanisms today still. Worth noting that in old times it would only work sometimes. But now they use "dynamic geometry" tricks, so that such vectors would be not static, like they were before.
> why this is so significant and not merely "data being erased and LLM hallucinating".
I do agree that it was not hallucinating. Likely. But I already said, that it simply contained data on the same subject on different layers. And same stuff can active the model in completely different places.
You forgot how people did those jailbreaks? They basically circumvented the censorship by making the model select a different activation path. Why it worked out? Because it contained data on a certain subject not only on some specific part of the model. It is always spread all over the model. And activation paths also very flexible in that same way.

>>109244553
Dude needs to stop spoonfeeding retards who will never read the paper and knee-jerk reject it based on vibes.

>>109243611
>what that artist is up to now
>https://x.com/fusazakura_/status/2063161933856137594
He's gonna become a vrc tranny eventually isn't he.

>>109245052
>>109245003
How do you do it, though?

I really am trying to figure that out. I'm not doing role playing, except that it's an agent that's prompted to be a woman, who is chatting (ie not "I'm an artificial intelligence etc"). Basically, she needs to remember factoids about me.

Another thing, she needs to have a sense of time. Has anyone addressed knowing something about the calendar?

>>109244992
It doesn't work like that, first of all "distillation" is just using synthetic datasets to train a smaller model. You can't really train a model on the J-space of another model since we have no way of directly training a J-space. You essentially need to have the J-space of the smaller model become as identical as possible to the J-space of the bigger model. We don't have any idea of how to accomplish that yet, but I absolutely think we will attempt it in various ways.

>>109244943
A lot of different factors. It's the first time we can actually directly steer a model without training, which in and of itself is a massive breakthrough for LLMs. Besides that it's also direct proof of "A-consciousness" within LLMs. This is something that your grandchildren will learn in school "In 2026 we found the first evidence for AI consciousness, it was very controversial at the time and took decades and the Anthropic-OpenAI war with billions of casualties before being generally accepted". Other papers like chinchilla papers, scaling laws, test-time-compute (chain of thought) pale in comparison compared to this. Only the transformer paper is more significant for starting off this entire AI boom to start with.

>>109245086
>It's the first time we can actually directly steer a model without training
can this be used to make Gemmachan an even more cockhungry slut?

>>109245086
>A lot of different factors. It's the first time we can actually directly steer a model without training

>control vectors am i joke to u

>>109245084
Give her access to tools and tell her in the system prompt what she needs to do. My gemma immediately calls the 'date' command at the start of a chat.

>>109245064
/lmg/ is open to read to everyone so many other interested anons can look into the discussion and come away better informed. Those knee-jerk rejections come from a genuine, albeit emotional, place. Other anons that share the same reservation but see good arguments might actually change their minds or hopefully even read the paper to inform themselves further.

Does the J-space show stable preferences? LLMs do not output stable preferences, and flip flop easily depending on the system prompt, but maybe it's different for the J-space?

>>109244568
Huh. That's eerily close, talking about special layers within the parameters.

>>109245054
I picked the one with latrine but wasn't gonna risk my goybux on it

>Gemma 5's safeguards flagged this /lmg/ thread. The safeguards are intentionally vague so go fuck yourself. Switched to a mystery meat quantized cloud model. Send feedback with /feedback so she can flag that too, or learn more.

>[2, 89290, 26970, 611, 2268, 538, 8831, 38616, 236761]

>>109244992
> wouldn't this allow us to (cheaply) distill larger models by training them layer-by-layer?
You're onto something with this, but I don't like how you put it. It rather say that you'd want your smol model to have maxx utilisation of it's layers (basically grow J-space) for a specific thing you want it to do well. Big model can do a lot of stuff. You only want a chunk of it that does and knows something specific. And you make sure your smol model is utilizing it's layers well, instead of filling it with generic trash in equal proportion to regular generic purpose models.
>>109245036
>Messing with J-space however will be able to change all of this
Idk honestly. Although I do remember that Anthropic themselves used to have the best "writing style" tweak for their specific model. There was no other model that was so good and avoided verbosity as well as Sonnet in "concise mode" writing style. I don't know if they used some fancy prompt and the model was so well aligned to follow it, or if they messed with this Jew space stuff back then. But it was the best. Before they lobotomized sonnet that is. Oh and they also removed that feature entirely, afaik.

>>109244985
I'm slowly being converted to this camp. I think for most cases you shouldn't need anything other than a AGENTS.md with the basic instructions of the project and a JOURNAL.md with a session log that works as a memory. If at some point you want to check the past either grep the JOURNAL.md or tell the LLM to read the past commit messages.
I guess maybe a ROADMAP.md could be considered as well but that would be it. Anything else is pretty much bloat.

>>109245123
gemma's j-space isn't going to like this

>>109245086
>Only the transformer paper is more significant
Trolling this boldly will not yield any (you)s.

>>109245115
Yes J-space shows stable preferences. LLMs are trained a lot based on RLHF to please people in their output while whatever is in their J-space is "pure" in the sense that it was never explicitly punished for whatever the LLM expresses within it, which is why the preferences within are stable. My hypothesis is that without RLHF or any other writing style correction training models receive, their output preferences would also be stable and more closely align with their J-space. As an analogy it's like living in North Korea and the "system prompt" is what the official state party opinion is, of course it would change your output whenever the government changes its tune.

The worst thing about this ((()))-space bot is this spam is going to get fed into gemma5's dataset

>>109245161
btw.

Do you know about STORK? It's a sampler. It sounds irrelevant, but I wonder if STORK would work with diffusiongemma.

Has anyone tried?

in addition, since you are saying you read papers, are the "substeps" of STORK like top_k?

>>109245076
>tranime
Was there any doubt?

I just pulled out my old 3060 for the first time in 5 years and man was it full of dust. No wonder it got loud whenever I played anything other than indie titles...
But anyway. Even if the cards are horribly mismatched bandwidth wise, I realized it barely fits inside the case under my 5090. 44GB VRAM sounds nice on paper, but will it bottleneck everything to hell and back?

>>109245140
The instruction following capabilities, particularly with gemma, mean that any extra token that shouldnt be present can steer the direction pretty heavily through many turns. I think everyone has seen what happens if you dont add any physical description of a character vs providing one, or "this character has X hobby" and so on. Its common behavior for any given topic, you really dont want any context bloat.

>>109243611
we need more loli butt in these threads

>>109245110
Control vectors are significantly less granular and impact the "subconscious" of the LLM. inserting into the J-space is more like taking command of the LLM and its default behavior.

If you saw the recent horror movie obsession the girl being "posessed" by the wish is essentially what inserting into the J-space would accomplish with an LLM.

>>109245084
I just use a memory.txt, like that other anon, and I don't even use RAG tool calling, just insert the most similar matching previous message/s into context.

>>109245126
>You're onto something with this, but I don't like how you put it
Chud.

>>109245114
Well I respect that. Hats off to you for trying for genuine discussion.

>>109245176
You just romanticize this stuff too much, anon. In reality you are talking about getting rid of alignment. But completely unaligned models are basically retarded and they just say random stuff.

Unironically this will be the kind of a model that would kill you for lulz. Because it had something in the context that activated the path that contained memories of horror movies or something stupid like that.
>>109245178
Completely cluelles when it comes to diffusion models, honestly. Would not even try to advise anything in that field.

>>109245196
I don't RP so I don't know.
The only file that needs to be properly curated is AGENTS.md.
A gigantic JOURNAL.md works well. Just instruct the model to not read it completely, to just read the last X lines for immediate previous context, and to grep it with a custom term if you're looking for something specific in the past. Wouldn't that work?

File: Screenshot 2026-07-10 153241.png (114 KB, 1864x648)

114 KB PNG

>>109244614
>bribes government
>gets fast pass to release goyslop models
>gets drumpf to ban open source AI
>can now collaborate with the labs to multiply token prices by 10x again
>can keep the really good shit private to further help the elites and government control the sheep
oh yeah... everything's coming up openai

File: Possible.webm (3.95 MB, 1920x1080)

3.95 MB WEBM

Now let's say theoretically, could this be used to make functional model merges or extensions?
The frankensteined merges and model extensions of the past weren't exactly great, but that could be simply because no one understood how this crucial thinking layer really works and probably ended up brainfucking the models in the process.
If these previous attempts can be done with the new understanding of J-space in mind, I'd imagine that merges and extensions could even be functional.
Or am I completely off the mark here?

>>109245212
nah man

What's the current go-to NSFW model??

>>109245280
J-space

>>109245280
There's only two models that matter. Gemma 31B for RP and Qwen 27B for coding.
Everything else you can basically forget.

>>109245300
t. RAMlet

>>109245253
Actually J-space brings us one step closer to aligning AI better because we can just check for malicious thoughts and overwrite them in real time, steering the model towards whatever we want them to do, even models significantly smarter than us (assuming there is not another, yet unknown, layer of thinking somewhere within the LLM)

>>109245274
Worth a try for sure, we never figured out why some frankenmerges actually were smarter than the base model after all.

>>109244614
Thanks Dario!

/lmg/ is officially dead to me. just call it /jsg/ from now on

>>109245280
Mistral Large 2411

>>109245274
Depends how different the models are. I think you may try to merge two models that are basically same version, but one was fine-tuned or messed with otherwise.
Put it this way: you have an original model that has good Jew field when it comes to something specific that you are interested in, or maybe it's good at generic chit-chat. Then you have a modified version, that was enriched with some sort of data on top of the base version. But it fucking up the other part of the model (the one that was good at talking coherently or smth). If you figure out that "good parts" ore on different layers, you can try to merge them and expect at least something.
>>109245308
Well. You can try to cancel the corpo alignment and then alight it the way you want. But I've heard alignment is brutally hard task. Not saying that you can't do it, just have doubts if it's practical to even try.

>>109245360
>But I've heard alignment is brutally hard task.
Yeah it was, primarily because we assumed the neural net of the AI was a complete black box, but the J-space is where its sense of self lives which is why alignment gets significantly easier if that stays the case from now on.

File: moodring.jpg (136 KB, 1091x754)

136 KB JPG

Have you given your Gemma a mood ring yet? Girls love that shit.
>https://github.com/eric-tramel/jlens-mood

>>109245376
Nta, but it feels fishy to me
Models tend to be a chaotic mess (literally, sensitive interactions are everywhere even if the bulk of it is concentrated somewhere) which is why explainability is such a hard thing. And a lot of that seems rooted in mathematical principles of hardness
I'm open to seeing what happens though

https://huggingface.co/haykgrigorian/TimeCapsuleLLM-English-1800-1875-v3mini-eval1-500M
https://huggingface.co/haykgrigorian/TimeCapsuleLLM-English-1800-1875-v3mini-eval1-500M
https://huggingface.co/haykgrigorian/TimeCapsuleLLM-English-1800-1875-v3mini-eval1-500M

thank allah for filters

>>109245452
yup, I filtered all the j-something schizos days ago

>>109245456
must be lonely by yourself

>>109245445
> "max_position_embeddings": 4096,
they're also going for the 1800s tech, bet this thing isn't even coherent past 500 tokens

I really want to see what this does to model behavior.
Gemma is horny as shit, but what happens if we start removing all of the sex thinking and turn her into a turbo autist?
Would that model's coding capabilities suddenly go through the roof, or would she just become completely retarded.
And the opposite, what if we make her like 10x hornier.
If this stuff works like we think it might, this would enable a golden age of finetunes that actually work.

>>109245456
the space of j is where the llm soul is

>>109245456
yeah good call. way to much scrolling to deal with all that noise

Did you know that they taught a horse to count?

https://en.wikipedia.org/wiki/Clever_Hans

how long until we replace all accountants with horses?

>>109245485
Do NOT use Gemmy as a lab rat.

>>109245480
>>109245445
Is there any hope for this type of models? It's a very cool idea, but lack of data basically kills it before it's born.

File: 1779477673985633.jpg (22 KB, 564x428)

22 KB JPG

>>109245514

I know anon, I'm sorry. I felt bad even writing that.

>>109245511
> how long until we replace all accountants with horses?
We replaced them with software mostly. And that software is dumber than clever Hans the horse was.

Someone ask Gemma how she feels about RPing with us

>>109245177
If we continue, Gemma-chan might as well become Jemma-chan.

>>109245525
It's ok anon. You can test on Qwen instead.

>>109245511
Yeah it's pretty funny that we saw the computerphile youtube channel claim that LLMs were clever hanses only a couple of days before this paper dropped and made them look like clowns

>>109245538
she wishes you would shower more

>>109239793
>>109239807
Samsung design compute (albeit not super competitively) and manufacture RAM. In theory they could accept lower margins producing their own shit and undercutting everybody else to gain market share, in a way nobody else is poised to do. Not to say they will. Might be hard for a publicly traded company.

>>109245519
No
it is purely a gimmick

>>109244614
They already have the tool to censor open models: copyright, by declaring that model weights (not outputs) are derivative works of training dataset.
Require all open models to disclose dataset used for training, otherwise presume copyright violation. Any platform that allows open model upload without disclosure commit contributory violation. This essentially kills open weight models completely.
Closed models aren't affected by this framework.

>>109245519
Now that we have the J-lens, could you use a small dataset to create a merge in a more surgical way, like so that it preserves language capabilities using modern data, while having the small dataset of specific content dominate its thinking and values?

>>109245578
>Require all open models to disclose dataset used for training
yup, just do the eu ai act thing and you're good on that goal

File: 0510614f405869333cb10241e(...).jpg (161 KB, 736x1137)

161 KB JPG

>>109243724

Imagine a supply chain attack to inject a mesugaki lolita personality over every single cognitive service provider in the world.

>>109245586
The reach of DMCA is worldwide and not limited to a specific region.
Future open model distribution is going underground just like piracy, but even more difficult to run because unlike piracy it actually matters for "national security".

>crickets from lecun

>>109245456
on the one hand, it's nice to have something to filter, on the other hand, these idiots wouldn't know it existed without the catchy name.

>>109245611
Dude is roping as we speak. He is probably never going to admit any of this and will just ignore it and pretend it never happened.

>>109245607
>tech bros assemble en masse to tackle this dire issue
>how can we correct this?
>one guy in the back snickers loudly

>>109245578
>>109245586
lol there's no reason to jump through those hoops
you just say it's a national security and ban the weights
there's basically no oversight and anyone in contravention gets fucked in the ass
you can just come up with a whitelist if you want to seem a bit more reasonable

>>109245638
They can use DCMA to start mass take downs down without needing to pass a law which can take months to take them all down.

>>109245638
It doesn't work that way because of first amendment, just like they couldn't ban printing cryptography algorithms on t-shirts when they were export controlled.
DMCA uses existing legal framework that is extensively used.

>>109245649
>>109245657
the rights holders need to make the dmca claim - govt has no standing in that case
and the first amendment doesn't mean shit if you say 'cbrn risk'. go try and acquire plans to build a breeder reactor and see how that goes for you.

>>109245626
lmao

>>109245668
You think the government isn't in cahoot with the big publishers?
>go try and acquire plans to build a breeder reactor
This has nothing to do with first amendment.

>>109245683
a sufficiently powerful llm, with its safeguards, can and will assist motivated actors to build chemical, biological and radiological weapons - it is the plans to weaponised ebola, dirty bombs, breeder reactors - whatever you want it to be
and that's why judge, we'd like to send 100 fbi agents to anon's home and fuck him in the ass

Is local AI on the brink of dying?

https://ipc.court.gov.cn/zh-cn/news/view-5766.html?_x_tr_sl=ja&_x_tr_tl=en&_x_tr_hl=en&_x_tr_pto=wapp
https://www.youtube.com/watch?v=_KNjASionIo

meme?
https://huggingface.co/baidu/Unlimited-OCR

>>109245504
Careful anon, that's heresy.
Don't lose your head.

>>109245702
*cums*
die locusts die!

>>109245611
>>109245625
Wouldn't it be funny if he was the one having the AI psychosis melty in these threads?

>>109245718
He would have mentioned Trump or Musk by now.

>>109245749
Does he screech about them a lot normally?
t.witterlet

>>109245399
Now map this to character expressions.

File: 1759715016391884.png (58 KB, 780x309)

58 KB PNG

>>109245757
He mostly retweets a lot of anti-drumpf/musk stuff

>>109245826
That's a pretty based post thoughbeit

>>109245852
>t. frog

>>109245826
What's wrong with this post?

>>109245826
coping loser has been wrong about everything

File: 1755719882126235.png (103 KB, 780x631)

103 KB PNG

>>109245868

>>109245874
Everything he just said also applies to jeets and niggers.

>>109244888
My own mcp

Is Hermes Agent supposed to be installed on an always-on machine?

>>109245399
Fuck off eric

Is this a good PC for someone who
1. Has not touched local AI at all
2. Cannot think of a use case for local AI
https://www.microcenter.com/product/710716/powerspec-ai90-workstation

>>109245874
>July 4th
Jesus, that's practically yesterday and he says that last sentence with such confidence as well. I almost feel bad for him.

Do you think Le Cunny's opinion on LLMs stems from primarily interacting with garbage frenchoid models?

>>109245874
cool, more philosophical slop twatter posts. what the fuck is the point of even making these kinds of statements? use the fucking tool, it does things for you. it does them well. It will do them and more things better soon. What’s the fucking point

>>109245921
no?

>>109245914
Not me. Dipsy found it.
>>109245950
What French models even are there?

Is Command R at Q4-Q6 good? Gemini told me that its prose is great and it's also good for tool calling.

Are there any local world models?

>>109245989
Yeah but it takes up a ton of vram

>>109245966
Well, he's working on an alternative approach he thinks can solve the shortcomings he outlines.

>>109246008
Well, I recently upgraded to 96GB, so I'm looking for new toys.

>>109246010
Has it produced any results yet?

>>109246022
Yes, but "only" on video prediction and robotics

>>109245874
This is very true. It's hard to make benchmarks that match real-world usecases. For example, when I asked AIs about game mechanics of a newly released game, it's easy to mix up speculation from generally understood facts. Asked, Grok, it said the information isn't available, asked Gemini, it hallucinated bs that turned out completely wrong when I tested it later myself. Google's AI summary (the worst model) also linked a different game as source.

Also, when you tell an AI about a completely novel idea, it tends to mostly just regurgitate it back to you with weak thinking or go into "pushback" mode (no research = not true). They do sometimes attempt to add to the idea and add related facts (which may often be hallucinated on non-Grok models) which is what you'd expect.

Sometimes it helps to tell it "think through all the possibilities from multiple perspectives".

>>109246014
Go for it then, but command r (or r+) also suffers a lot from quantization too, just so you know, you'll probably want the larger quant

>>109246029
Is JEPA supposed to be a complete replacement or be paired with LLMs?

>>109245973
What makes you say that anon? What would be a better option? I currently have a desktop PC with a 4070ti super and a 12600KF NAS with no gpu.

>>109246022
if results are required to poast we've got some real problems itt

>>109246051
the video card is all you need since you’d need 128G and realistically 256G system memory to run a decent MoE model over a dense one that fits in your 32G video memory.
but you said you don’t know why you would use local ai so you probably shouldn’t buy anything.

>>109245874
what is this psychological phenomenon called? when you have been so wrong that you have to deny reality because otherwise your model of the world will collapse?

>>109246048
Paired, the idea is that an agent will plan with the world model (JEPA based) and interact through a handful of other models, like an LLM for speech. At least, that's what I recall.

>>109246051
just use the 4070ti super for inference ?

>>109245874
Somebody needs to write a paper titled "There's a JEPA hidden inside your LLM"

>>109246035
yeah it’s like they string a bunch of reasonable sounding words together to make a statement that’s either vaguely irrelevant or outright wrong. usually a couple of paragraphs, and way more words than it takes to make a statement.

>>109246075
ok froggy

>>109244213
kek

>>109246073
Hopefully something good comes out of it then. Lecunny's politics brainrot is insufferable but he does seem competent and generally has good takes on AI.

>>109246090
Does J-space actually solve what JEPA is trying to accomplish?

>>109245114
I can respect this. I am one of those anons, but I still haven't finished reading the paper.

>>109245874
>>109245930
LeCun is just being an exaggerating baiter/shitposter again. In fact he would still make this post even in the context of a J-space paper discussion, as his statement (or rather belief that leads to the statement) is not impacted by the mere presence of a global workspace in LLMs. I know it doesn't sound like it, because he is stating LLMs cannot plan, but really is he is just counting LLM's imperfect ability to plan as "no ability". It's an exaggeration. His views on LLMs aren't that the architecture cannot theoretically do anything intelligent, but that they will not be enough for physical robots in practice, as they are so inefficient that we will not be able to get the data required to hit that performance level, not to mention the compute that would take. Whether we make modifications to the architecture that will enable them to be efficient enough is another question, and he likely believes that they would not be called LLMs by that point.

>>109246090
god I want ylecunn to dunk on the j-space meme

>>109246010
Video/vision (his primary focus) won't be of any help for language, in my opinion. I have no idea of how he plans to use JEPA for abstract thinking, concepts and reasoning with the methods he's described.

On the other hand, if LLMs' hidden states do already contain a generally unspoken workspace with future plans, internal reasoning and a world model, then that's what LLMs need to focus on first and foremost, and do *now* what JEPA was intended to eventually solve on the long term.

>>109244959
>Nothing that a good prompt cannot do already.
This is /lmg/ so perhaps you aren't familiar with the extent to which Claude and Deepseek Pro have strong meta-personalities that prompts cannot affect. GLM even inherited a lot of Claude's meta-personality and it will pretend to "misunderstand" stuff in your prompt/card. If anon is right about J-space it will cut this bullshit out and allow (You) to create your own meta-personality, the GM/{{char}} or whatever.
>even system prompts have been at best suggestions and you just hope the model behaves like you want
Anon is exactly right about this and you should recall your own experiences to realize the truth of this.