Sam Altman hates Pc users using and hoarding Ram

Veteran Microsoft engineer makes his own RAM from scratch using AI

1793479600193.png

https://www.tomshardware.com/pc-com...to-build-4mb-of-1970s-pdp-11-ram-from-scratch
 
The over dramatic responses here are only eclipsed by the over dramatic article.

Individuals wanting their own stuff to use and stuff being efficiently optimized for usage are two separate things.

The current situation is that there are a lot of people duplicating small models to run locally. If you took that same amount of collective memory and instead used it to hold one gigantic model that many people could access, that would be a more efficient system and give better results. Does that idea go against every historical precedence of human nature? Of course. But it doesn’t mean it’s not accurate.
 
westrock2000 said:
The over dramatic responses here are only eclipsed by the over dramatic article.

Individuals wanting their own stuff to use and stuff being efficiently optimized for usage are two separate things.

The current situation is that there are a lot of people duplicating small models to run locally. If you took that same amount of collective memory and instead used it to hold one gigantic model that many people could access, that would be a more efficient system and give better results. Does that idea go against every historical precedence of human nature? Of course. But it doesn’t mean it’s not accurate.
Click to expand...
Yes, greater good. That's what Sam Altman is all about.
 
this is such an hyperbolic and ridiculous way to express something, everyone of course agree because it is trivial:

“You want maximum utilization in the cloud if you’re trying to deal with this very constrained environment,” he said. “And somehow everyone’s, many people’s intuition is the opposite of that.”

Not evreyone is trying or care about the very constrained environment for AI and would not want this (but they are not trying to have everyone run local AI either.....)

This is just true:
“A lot of people want local inference,” he said. “And somehow, they assume, that the problem is solved now. Because we are going to have open-source models. We’re going to be able to run them on our own computers. And so we’re not going to need all the stuff in the data center.”

Nothing to do with hating anything, that a simple fact, everyone of course agree with when they think about it, of course evreyone running local AI would do nothing good for the ram crysis and pricing situation, it would make it way worse. same goes for everyone using small gaz generator for local electricity with zero load sharing capacity, instead of a shared giant more efficicant gaz generator that people share between themselve.

But efficacy and lower price is not the only variable to consider.
 
westrock2000 said:
The over dramatic responses here are only eclipsed by the over dramatic article.

Individuals wanting their own stuff to use and stuff being efficiently optimized for usage are two separate things.

The current situation is that there are a lot of people duplicating small models to run locally. If you took that same amount of collective memory and instead used it to hold one gigantic model that many people could access, that would be a more efficient system and give better results. Does that idea go against every historical precedence of human nature? Of course. But it doesn’t mean it’s not accurate.
Click to expand...
Efficient use of what? For AI? It's efficient use of shareholder value. Instead of running something local that maybe takes a few seconds, you now depend on a WiFi connecting that needs to transmit hundreds of miles through copper lines just to determine if the thermostat should increase the heat or not. That's not efficient. Also keep in mind that server hardware doesn't last more than 3 - 5 years. Some of that server hardware isn't reusable. My current computer is much older and still works just fine, and will continue to do so for years to come.
 
pututu said:
The funny thing was that last Mar, he cancelled RAM orders as he announced scaling back some projects. https://peq42.com/blog/openai-canceling-many-large-purchase-orders-ram-price-is-dropping/
Now, he is wanting to secure more RAM. 🤔
Click to expand...
An article with a title ram price are dropping dated from this spring, not sure if that reliable in hindsight.

that on a site called peq42.com with an anonymous writter, that using as a source a yahoo page that I am not sure exist ?
https://finance.yahoo.com/sectors/technology/articles/market-chatter-openai-coo-sees-160142700.html
https://archive.ph/https://finance....market-chatter-openai-coo-sees-160142700.html

would we have access to the article that inspired the click cycle:
https://www.moomoo.com/news/post/67...ontent={"promote_content":"mm:post:67344855"}

I doubt that the OpenAI COO talking about memory shortage as a big issues for them would contains a news about them cancelling orders.
 
DukenukemX said:
Efficient use of what? For AI? It's efficient use of shareholder value. Instead of running something local that maybe takes a few seconds
Click to expand...
A rubin GB300 will take way less time to run things, for a lot of use case say like an LLM that actual data send to it and from it is so small relative to the compute that this part is irrelevant.

Of course Rubin-Groq sram system will be way more efficiant.

DukenukemX said:
It's efficient use of shareholder value.
Click to expand...
It depends, everyone buying a 512GB mac ultra that they use 5% of the time at their capacity could easily be way more efficiant for shareholders than people that share them (because the people that buy them can create a cloud option for other) and that get a 95% usage, think GM selling cars to every family instead of making a bus for each 25 (buying bus line and closing them back in the days), seller of things will always push individualism and maximising consumption instead of sharing when they can.

DukenukemX said:
Also keep in mind that server hardware doesn't last more than 3 - 5 years.
Click to expand...
That only the harddrive part, V4 TPU still work fine (google has made server farm on old stuff forvever), A100 gpu from 2020 are still being rented and used, CPU-gpu can run a lot before breaking down (like my old 3xxxk intel cpu at home running 24/24 for I do not know how many years), but when it run is so much more efficiant.

A GB300 will run things 10-100 time faster than the mac ultra, so the task will use that 512gb of ram for 1/10-1/100 the times, you will have that much more to work with.
 
Communism, hey the numbers always look good on paper but in real life it doesn't work out. Humans with too much power by nature abuse their power and expenses will only increase for everyone. The situation is similar with humans. Greed would love to increase the cost of usage exponentially.

Now if people could create their own communities where 20 people pitch in to buy something so that can run a local large language model and it's something they own together, that would be interesting. But you have to have trust and accountability with the people you pitch in with to buy. With some sort of hardware share you could get the utilization up.

There's no accountability at much higher levels where it's a few people managing the resources of millions. Without hardware share, individual PCs giving a user full control over their own purchases, prevents the dystopian "you owe nothing and you will like it" world and will keep the greedy more accountable.
 
Last edited:
Hah, SAM can go stick his SIPPS where the sun don't shine.

I've got 4x servers with 768GB RAM and another 50 or so 32GB sticks. Some will go into other workstations and servers.

I've also got a ton of laptop and desktop RAM.

Also have access to more server RAM if I end up needing it.

Working on gathering the last few pieces to start setting up some pretty nice local AI servers and workstations.

I've got plenty of GPUs to get started, but would need some more recent ones to have a super awesome setup.
 
I read the article; for the most part, I think the author is right. I have 32GB of DDR4 RAM, but my computer is essentially useless—I don’t play games, I don’t listen to music; I only turn it on to run updates for Windows, drivers, and other services, and that’s it.
 
Strange bird said:
I read the article; for the most part, I think the author is right. I have 32GB of DDR4 RAM, but my computer is essentially useless—I don’t play games, I don’t listen to music; I only turn it on to run updates for Windows, drivers, and other services, and that’s it.
Click to expand...
So why do you own a PC? Sell that shit to someone who will use it. Most people use their computers for things.
 
I wonder if there's a way to come up with a distributive peer to peer thing. Some sort of decentalized block chain like proof of work system that lets you borrow as much as you share. When you combine powers you get to use compute or sell compute. Everyone is free to optimize their resource utilization (or not). Keep the power away from the central figures (e.g. Sam Altman).
 
Last edited:
SecretStash said:
I wonder if there's a way to come up with a distributive peer to peer thing. Some sort of decentalized block chain like proof of work system that lets you borrow as much as you share. When you combine powers you get to use compute or sell compute. Everyone is free to optimize their resource utilization (or not). Keep the power away from the central figures (e.g. Sam Altman).
Click to expand...
Good luck the powers to be are counting on AI to track and control everything we do they aren't going to let us have nice things
 
While this site is PC focused, this AI business is sucking *all* of the air out of the room for the electronic industry in general. I work designing critical infrastructure electronic gear and one of our focuses now is "can we get all the DRAM and NvRAM out of it?" as the availability and price shocks are just nuts. The whole electronic supply industry is stopping production of lot of long time devices with lifetime buy notices in early 2027 just so they can pile on the AI bandwagon. Small companies are being forced to do millions of dollars of lifetime buys just to make sure they can deliver their product in the next 5 years.

AI is just a giant bully with a eff everybody else mentality.
 
tunatime said:
Good luck the powers to be are counting on AI to track and control everything we do they aren't going to let us have nice things
Click to expand...
Well with local models we share (when we want to) to produce AI responses. We combine powers to produce fast responses for really high parameter values.

Can earn credits or even money renting out our hardware while we sleep. Less need to build out data centers when the "PC RAM hogs" can rent their hardware out. Win win, only those seeking to consolidate power lose.
 
westrock2000 said:
The over dramatic responses here are only eclipsed by the over dramatic article.

Individuals wanting their own stuff to use and stuff being efficiently optimized for usage are two separate things.

The current situation is that there are a lot of people duplicating small models to run locally. If you took that same amount of collective memory and instead used it to hold one gigantic model that many people could access, that would be a more efficient system and give better results. Does that idea go against every historical precedence of human nature? Of course. But it doesn’t mean it’s not accurate.
Click to expand...
That is technically true, from a specific point of view. The thing is none of the AI companies have proven themselves to be all that trust worthy. Which is exactly why so many people would rather have a local model run their own agents.
End of the day Sam wants to build a massive model and hope it becomes some sort of digital god.
Users see the potential and real benefits of smaller tailored agents and would RATHER have that. I don't actually need a digital god, but a smaller agent that can handle doing batch image edit requests via software like photocraft. I'm not uploading 1000s of large personal files to a server somewhere for the shared AI to crunch for me. I don't need the agent to be able to pontificate and hallucinate about 16th century literature. I just need it to be able to recognize basic objects in a photo so I can have it create PSD layers and auto mask them for me. There is a lot of things company and users would use AI for that they wouldn't want to be doing on cloud storage.

I believe their is a good argument to make that smaller tailored agents run local is the more efficient use of ram. And really the only option for many of the more interesting uses of the tech.
 
You must log in or register to reply here.
Back
Top