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


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: 1773644074020355.png (102 KB, 951x902)
102 KB
102 KB PNG
https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/

Amir Zandieh, Research Scientist, and Vahab Mirrokni, VP and Google Fellow, Google Research

Thank them.
>>
why would compression have anything to do with the RAM crisis when these companies are so big they can throw a blank check at anything
this is just going to justify purchasing even more RAM
>>
>>108479682
The crisis was never because they needed the ram. Altman is just hoarding to fuck everyone over.
>>
>>108479682
>quantization model
that already exist saaar
>>
Memory stocks dropped but memory prices have no dropped. Pretty blatant market manipulation.
>>
>>108479682
>plz don't Luigi us
>>
>>108479682
Jevons Mother Fucking Paradox, baby.
>>
>create 1 GB hard drive
NOBODY WILL EVER NEED MORE STORAGE
>create 1 TB hard drive
NOBODY WILL EVER BE ABLE TO FILL THIS MUCH STORAGE SURELY
>create 30 TB hard drive...
>>
The RAM fiasco is being caused by companies preemptively buying up stock and rendering what the consumer market can offer inert by comparison. it's not about "needing" it, it's all speculative.
>>
>>108479770
Well ideally you'd expect as time goes on that trying to be more efficient would come with increased advancement. The better things get the more should also be able to be done with less.
>>
>>108479791
Yes, and we also got more efficient video compression over time
Did that mean the demand for storage went down?
>>
>>108479682
>Just multiply the vectors for a 2D embedding

Are things really that bad that nobody though of this before?
>>
>>108479770
this reminds me of that story of a team intentionally bloating a simple program into a few gigs because the client wasn't comfortable paying so much money for something that was in the singe-megabytes range
>>
File: rear.jpg (920 KB, 1600x1200)
920 KB
920 KB JPG
>>108480082
not long ago the frontier model for tts quality was made by a guy with a handful of 3090s in a garage
>>
>>108479682
Wow, I guess now all the grifters who were buying ram they didn't need and weren't even using will magically stop.
Retard.
>>
Isn't this something they should've done in the first place before they sent the entire chip industry into panic, and governments into panic about electricity grids?
>>
>>108479734
I think it was a power play by Altman to sell RAM to NVIDIA, and extort out some of the ludicrous markup they have on their GPUs, alternatively to deny them from competitors.

It's been like half a year since that happened, and still there are no signs OpenAI is using those memories for anything
>>
Why didn't AI come up with this breakthrough in compression?

Why did humans have to think of it?
>>
>>108479682
They didn't end shit. This just means AI companies will get richer because they can do more with less. But they'll still want MORE. Why would they reduce the demand for memory when this just means they can handle more users and train even bigger models?
>>
>>108480666
bigger models don't seem to be that performative also smaller footprint means consumers will be able to run bigger local models potentially killing the cloud services.
imagine running a 1t parameter model locally with just a regular ATX motherboard and a GPU, its going to be glorious for power users.
>>
>>108479682
The ram crisis was planned. They got caught limiting supply the first time. Now they can write fake contracts to AI companies and do the same thing legally.

Game is game.
>>
>>108479682
Isnt this a nothingburger? Its only for the kv cache, whatever that is and only really helps in some specific algorithms?
>>
>>108480666
Yep. They want to take over the world. These are insane transhumanists who want us all dead by the end of it. They won't even stop when the surface of the entire earth is one giant computer. Our remnants will be like the stalkers in half life 2 if we're lucky.
>>
>>108480871
That just means we all need to take up arms and stop them before it's too late, especially if it means ending all normies. No fate but what we make.
>>
>>108480687
>The ram crisis was planned
For what purpose?
>>
>>108480871
>>108480896
Schizo melty
>>
>>108480901
total system faggotry
>>
>>108479754
>Mario Movie in 3 days
The world needs a hero...
>>
This just means they can have bigger context windows with what they already have. Why would this end the RAM crisis?
>>
>>108480919
They literally tell us their end goals. To make "god" and transhumanism (immortality by extension). We are not invited. They hate us. Money is just a means to an end for power.
You're a fool for not respecting and taking your enemies seriously.
>>
>go to page
>all linked papers are over a year old
Very good
>>
>>108479734
I think you're right.
>>
>>108479744
Everything on a digital computer is quantized. It's a stupid term to use.
>>
>>108480094
99% of all early day LLM efficiency breakthroughs came from the LLM threads here before they turned into
>1girl, BIG tits
>>
>>108480901
You can ask arbitrary amounts for the same product, increasing the profits without modifying the production.
>>
You know what they say. Unused RAM is wasted RAM
>>
File: 1464183983787s.jpg (7 KB, 250x241)
7 KB
7 KB JPG
>>108480612
Cause it never does and it never will, it will only ever regurgitate.
>>
>>108480612
Gee, I wonder. Could it be because these AI companies are all hype and no substance? No, surely not!
>>
>>108480082
they hire math researchers that have no fucking clue of how hardware works.
all they do is textbook implementations all day.
they barely even know that the FFT is a real algorithm.
>>
>>108480094
>7 cards for optimal pcie bandwith
>3 PSUs, only one fully utilized
damn, that's botched, but professional hardware is so expensive
>>
>>108479682
>Amir
>Vahab
jeet tech
>>
>>108479754
as if anyone has dared to do the same thing post-Luigi
he failed, unfortunately
>>
>>108482752
>unironically thinks normal people feel the need to say things like this in an organic setting
Shameless.
>>
>>108480612
>LLMs could never
gee, idk
>>
wrong tab LOL
>>
>>108481616
Indians own the future, Timmy. Stay mad; stay broke.
>>
>>108479682
They're the first of many.
Optimization beats brute force all day, any day.
OpenAI will fucking die.
>>
>>108479682
How? This does nothing for training and only impacts the context memory
>>
>>108482859
He's still right no matter how you cope about it. Mutts are thoroughly domesticated.
>>
File: n0izhm02zx6f1.jpg (27 KB, 750x646)
27 KB
27 KB JPG
>>108479682
>believing a fucking corporate in 2026
>>
>>108479770
Who are you quoting, you schizo?
>>
>>108485139
They are, but it doesn't mean that he lost. He wasn't trying to start a revolution, he wanted revenge, and he got it. For all intents and purposes, he won.
>>
>>108481400
well it regurgitates much better than you mr brainlet
>>
>>108480943
Source?
>>
>>108479734
Don't be a luddite
>>
>>108479770
Exactly right.
It's called the Jevons paradox btw.
When a technology gets more efficient, its total usage goes up, not down. Because more people can use it and you can use it more often, new usages get created.
>>
>>108485139
Have you been here? CEOs literally can't feel safe on the streets around regular people because someone they pissed off might recognize them, and they go out of their way to obfuscate their existence exactly because of scenarios like Luigi. Talk about living in a cage.
>>
>>108485825
>They live in gated and guarded communities and don't interact with the drug addicted homeless guy while travelling in their private jets!
Winning!
>>
>>108485274
Don't be a subhuman
>>
>>108480607
>I think it was a power play by Altman to sell RAM to NVIDIA
how exactly does that work when altman doesn't manufacture ram and nvdia buys direct in bulk and is the biggest player in the hardware market?
>>
>>108480666
anon you do realize that ai companies don't actually manufacture ram and have to buy it off the companies that do?
>>
>>108488273
Altman resells the wafers for etching and RAM production to NVIDIA for half the GPU's sale price, then NVIDIA sells them the complete GPU.
Altman gets to keep half the profit, instead of NVIDIA buying RAM for cheap and selling the whole thing at an insane markup.

Also Altman can decide to choke NVIDIA out of supply, so if they want to sell to OpenAI competitors, no more RAM for them.

Don't tell me OpenAI was actually doing something productive with that RAM they bought other than sitting on stock
>>
>>108488337
openai supplies the ram for openai contracts. if you think they're the only customer nvdia has or nvdia gets all their ram from openai then, ok, you do you,
>>
>>108488412
you give yourself away writing in all lowercase smug little silllycon valley man
>>
>>108488441
>i can read your mind i am a noooooticer i can't actually back my position up please don't noooootice that
i don't even live in california you schizo faggot retard. altman bought all that ram for project stargate. not to resell. fuck off with your dumbass shit nigger.
>>
>>108488454
Once again, OpenAI doesn't use those chips, it has vendors that make hardware that do. Altman essentially scalped its own suppliers to buy himself half a year of market monopoly. Now that time is just about over, and RAM supply is kinda returning to normal.

Now, guess how happy and eager those suppliers to continue and work with OpenAI.



[Advertise on 4chan]

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

[Enable Mobile View / Use Mobile Site]

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