Sam Altman hates Pc users using and hoarding Ram

[zehoo]

Its more about physical reality of what possible versus anything moral, it is not really just some greater good that everyone that built local AI capacity that they do not use fully most of the time to rent it (i.e. share it) with others if they can, until something change in ram supply or how AI work it is the only way to give access to capable AI to 4-5 billions people.

People can of course have different axis and think AI should not be commonly used at all, but those people do not want everyone to build a MacUltra 256/512GB or more capacity either anyway.

What’s wrong with competition to the big boys of people buying their own local LLM. And why should Sam and co be guaranteed a monopoly/duopoly market and monopoly access to RAM in the interests of “the greater good”. Which as someone pointed out is usually just cover for people with nefarious ambitions. It’s right up there with the classics like “think of the children” when you’re trying to sell an unpopular idea.

A lot of people rightfully want to control their own LLM so they actually have control of the LLM. And then they also don’t need to worry about rampant IP theft by openAI of its customers’ data.

 

[kac77]

Looks like I'll kick off my own local AI VM just to

I highly recommend it. I've been running my own on a Linux box for a bit and moving my home automation to it as well. Maybe I'll do a write up when the project is complete. If you're the type of person that upgrades hardware at a reasonable clip hosting your own AI powered home assistant is very reasonable if you have GPUs laying around.

 

[LukeTbk]

I don't care that it's not, though.

exactly and that perfectly fine, what Sam Altman is saying is trivial and I think everyone agree with it and it is ok to attack the sentiment behind it, there is no need to make up that Sam Altman hate that people have personal computer device (how would they use openAI products and send him money without them, you need one to talk with ChatGPT weould it be via an app, api, webpage, etc.., chatgpt can even take control and work using your own computer, openAI make desktop and phone application expecting people to have computers to use them, etc...)

Four to five billion people don't need access. Supposing everyone does, then it only needs to reach the same point where it's good enough on a local machine for the average user.

Not sure what need mean here, we do not need anything that was invented after the year 1,000 in some sense, but if you are an elderly that want to stay very late in life in their house it could be nice to have a robot that has a lot of local ai but also lot of cloud for the things that do not need instant reflex, if you are an african farmer it can be nice, it can be nice to have for everyone specially once your competitor in business as it.

The if people use 50 times less AI at 2% of the efficacy it could work without datacenter that certainly true, same would go for home generator for electricity instead of big central with a distribution infracstructure and everything else, but say 5% of humans has built local machine AI capacity but use it only 10% of the time, they could share it to 50% of humans that want to use it via the Internet when they are not using it fully.

About no one used anything else than the cloud for emails(talking about privacy implication....) or video sharing, but it became a big subject for AI ?

 

[LukeTbk]

What’s wrong with competition to the big boys of people buying their own local LLM. And why should Sam and co be guaranteed a monopoly/duopoly market and monopoly access to RAM in the interests of “the greater good”. Which as someone pointed out is usually just cover for people with nefarious ambitions. It’s right up there with the classics like “think of the children” when you’re trying to sell an unpopular idea.

A lot of people rightfully want to control their own LLM so they actually have control of the LLM. And then they also don’t need to worry about rampant IP theft by openAI of its customers’ data.

There is descriptive and prescriptive.

AI Datacenter because of how efficiant the infracstruture can be versus small local affair and its optimised sharing nature is more ressource efficiant, for the same amount of AI less hardware was built, less electricity and water used and so on, if it is in massive NVL72 building.

That does not mean that the thing to do, 55% efficancy giant gaz turbine is more efficiant than everyone having is small generator and that usually what society do, but people could consume less electricity and all have their own generator, it would have a list of downside and upside to be like that, same for AI.

A mix is what Altman would prefer, like electricity, not everyone has a massive local AI capacity (while some do for who it make sense, big users, hobbyist that want to thinker and keep competition alive and so on), while most use the more ressource efficiant way.

 

[DukenukemX]

What's really crazy is that MANY laptops come with nVidia, AMD, or Intel GPUs, which means most people are sitting on their own AI assistants w/o realizing it because no one has designed a product to use it other than Microsoft, which is really the problem as people go out of their way not to use MS OS features because of the problems they cause. Local on-prem use of AI assistants could work if it wasn't Microsoft pushing it.

The difference in what you can run on your laptop and what's in the cloud is that the cloud AI model is better. Not because it's in the cloud but because it has hundreds of billions of dollars behind it, while the open source alternatives like DeepSeek cost maybe 6 billion. What's interesting is that China is probably 9 months behind what we see in USA, so it's just a matter of time before LLM's catches up to something like Claude. As of right now, nobody needs AI. Anyone using AI is doing so to experiment and nothing more.

Consolidation works when demand for each action is minuscule in comparison to a larger operation asking for the same demands. AI is NOT that at all. That's why no one started really complaining about data centers until the resource demands went through the roof. That's also why cloud gaming has taken so long to build. It's resource intensive just like AI.

The problem with the recent Data Center influx is that it's putting a massive burden on the locals. Even if Data Centers was moved into the middle of nowhere but it'll still suck up electricity and water like there's no tomorrow. They also don't want Data Centers far away from people because it costs more. What's more retarded about Sam Altman's statement is that there's a clear problem with electrical supply. Would he also say that electricity is better used in his Data Centers instead of people using them for computers to play games? Ram isn't the biggest limiting factor for Data Centers, as that would be electricity.


View: https://youtu.be/e4IreHXOl4k?si=ZtYqLq2ZWAj4-oJ-

 

[kac77]

The difference in what you can run on your laptop and what's in the cloud is that the cloud AI model is better. Not because it's in the cloud but because it has hundreds of billions of dollars behind it, while the open source alternatives like DeepSeek cost maybe 6 billion. What's interesting is that China is probably 9 months behind what we see in USA, so it's just a matter of time before LLM's catches up to something like Claude. As of right now, nobody needs AI. Anyone using AI is doing so to experiment and nothing more.

I agree more or less. What I'm doing is definitely more of a project. I don't NEED to use AI. I just want to for my own learning experience.

As for the AI cloud model being better, I'll admit that, but the differences over repetitive tasks I just don't see, unless I'm doing something that's specialized like coding or images. Then the model matters A LOT.

The problem with the recent Data Center influx is that it's putting a massive burden on the locals. Even if Data Centers was moved into the middle of nowhere but it'll still suck up electricity and water like there's no tomorrow. They also don't want Data Centers far away from people because it costs more. What's more retarded about Sam Altman's statement is that there's a clear problem with electrical supply. Would he also say that electricity is better used in his Data Centers instead of people using them for computers to play games? Ram isn't the biggest limiting factor for Data Centers, as that would be electricity.


View: https://youtu.be/e4IreHXOl4k?si=ZtYqLq2ZWAj4-oJ-

I saw that video. Loved it. Quite smart on those residents. I agree with the statement here. I have family in Virginia and they were fine with the data centers UNTIL the AI data centers went in. The amount of power they need is crazy it's way more than your average workload. So much so that the amount of generators is in the hundreds. I've seen local data centers with about 10 or so. AI data centers? Need exponentially more depending on the design and that's affecting air quality in the region.

 

[zehoo]

There is descriptive and prescriptive.

AI Datacenter because of how efficiant the infracstruture can be versus small local affair and its optimised sharing nature is more ressource efficiant, for the same amount of AI less hardware was built, less electricity and water used and so on, if it is in massive NVL72 building.

That does not mean that the thing to do, 55% efficancy giant gaz turbine is more efficiant than everyone having is small generator and that usually what society do, but people could consume less electricity and all have their own generator, it would have a list of downside and upside to be like that, same for AI.

A mix is what Altman would prefer, like electricity, not everyone has a massive local AI capacity (while some do for who it make sense, big users, hobbyist that want to thinker and keep competition alive and so on), while most use the more ressource efficiant way.

Is it really more more efficient though? Just looking at one aspect, their business model is to charge with per token. And it has been demonstrated in the past that some models were purposefully choosing to do requests in the least efficient way possible so that it used the most tokens. And if you force a duopoly situation like you seem to want, the incentive is there for this to be the norm when there is no competition.

 

[Comixbooks]

Sam can use S.I. to fix his character flaws.

 

[M76]

Ram wise compute speed can affect it, if a task take 1 hours instead of 6 hours to do, you need 1/6 of the ram on perfect timing

Not if the model is 10x larger or 200x larger.

(it will never be that much being saved, but you get the point), Rubin GB200 system is so much more efficient than blackwell who was so much more efficient than Hopper and so on, the level of custom kernel-effort that go to optimise token per watt into those giant datacenter, the superiority of the dedicated hardware for it, make that almost impossible.

If the only thing people used their RAM for is AI, then it would be true, but that's not the case. Saying that it's terrible that people have RAM is such a sociopathic statement.

Rack scale architecture efficacy is magnitude higher than simple PC at home, usually in the 10-25x more token per watt than MacStudio's ultra.

Tokenmaxing is so last month. It turns out using more tokens doesn't result in better productivity. Slower inference might make the fleshy thing think for himself a little more instead of just wasting tokens on things he should be able to easily figure out with the same effort as asking AI to do it.

always hard to talk about closed door model efficacy, but GPT 6 luna is so cheap to use (very similar per token price than a small Qwen model) and more token efficient, that people that could use a small model at home can choose that (or their routing system picking it) and will be way cheaper, cloud or at home does not need to change anything on the nature of the model used (you can use Qwen3 30b on the cloud or at home), that a bit 2 different subject.

But isn't that a terrible thing for them? How will they make money if tokens are so cheap? They have hundreds of billions in data center bills coming due, what will pay for it if not tokens? Investor gullibility? Sure it worked thus far.

Or wait, is it the strategy to kill the home computer nationalize it if they have to and then they can charge 1000x for allowing you to use their RAM?

 

[1_rick]

Sam can use S.I. to fix his character flaws.

That would be some serious tokenmaxing.

 

[modi123]

Has anyone clued in Sammy to existing solutions like RAM Doubler or downloadmoreram.com ?

 

[Ididar]

This is so far away from his comments, all Sam altman clients need to own a computer to pay openAI to use its product.
This is all a click-bait for profit cycle (of course), and everyone want to not understand what going on for the fun of being mad here.

What he said was:
An example of this is people...
A lot of people want local inference.
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 computers.
And so we're not going to need all the stuff in the data center.
But from a global allocation perspective, actually, it's quite terrible to have everybody have a bunch of the world's limited supply of DRAM sitting in their own computer.
And, being used—Mm-hmm. 5% of the time.
You want maximum utilization in the cloud if you're trying to deal with this very supply constrained environment.

i.e. It is about people thinking that having 512gb-12 TB of ram to run lage local model with big context using it only 500 hours a year at full capacity, instead of people sharing that ram with each other for when they need it is a not a good global allocation.

Nothing to do with everyone wanting to or having both a phone and a Macbook

It is and it isn't.

Efficiency is being used in a very limited fashion here. When you evaluate efficiency you establish the end goal of the job you're trying to do and how many resources you need to achieve it. If my goal is a private AI instance inside a firewall where I can do stuff I don't want in a public model then it is VERY efficient to have my own computing power. I just have to be efficient in how I deployed that RAM and not buy 4x what I actually need. Altman's problem is that he's being purposefully blind to various use cases where his efficiency argument falls flat on its face. His only goal with AI is to train the model that his company is currently running and anyone else buying up RAM, to his viewpoint, is just being inefficient and they should instead be using his model, not doing their own thing.

By his logic, maybe Altman should pack it in and give all his compute power to Anthropic? After all, it isn't efficient to be training multiple different models in different data centres. Give up the fight, Sam, and make all this RAM use much more efficient by a bigger factor than anyone else who might be running a local model. Running all those various models is far less efficient than any individual and yet he won't consider that at all. Everyone else should pay the price for his goals.

 

[Ididar]

I highly recommend it. I've been running my own on a Linux box for a bit and moving my home automation to it as well. Maybe I'll do a write up when the project is complete. If you're the type of person that upgrades hardware at a reasonable clip hosting your own AI powered home assistant is very reasonable if you have GPUs laying around.

At some point I want to flip my home assistant from a Rasperry Pi to something a bit beefier and start using the AI as well. Just not a good time, hardware cost wise, to be doing it. I should have made the leap a while ago but just didn't go for it.

 

[LukeTbk]

Is it really more more efficient though? Just looking at one aspect, their business model is to charge with per token. And it has been demonstrated in the past that some models were purposefully choosing to do requests in the least efficient way possible so that it used the most tokens. And if you force a duopoly situation like you seem to want, the incentive is there for this to be the norm when there is no competition.

They use way less token for the same task at the moment, one possible is reason is a lot of users pay per months instead of per tokens, which created a giant incentive to minimise aggressively token use, the other one is the less tokens (the faster and cheaper AI get) the more use case you have and the competition. And the hardware itself is way more efficiant and the sram (Cerebras-groq) type of technology will make it even bigger, those are incredibly more efficiant in energy but less in wafer surface need, which make sharing the hardware even more logical to do.

Cloud an duopoly is 2, you can run an encrypted private Gwen model on the cloud exactly like you would locally (and free up the ressource when you do no use it dynamically), you can use meta-gemini-grok that 2 things

 

[Comixbooks]

Wonder how many sticks of Ram is needed for a useless Data Center?

 

[LukeTbk]

Efficiency is being used in a very limited fashion here. When you evaluate efficiency you establish the end goal of the job you're trying to do and how many resources you need to achieve it. If my goal is a private AI instance inside a firewall where I can do stuff I don't want in a public model then it is VERY efficient to have my own computing power.

yes ram efficacy is only one variable (and here used in the sense amount of AI work done per amount of ram made) and there is many different debate-axis, what Altman said is so trivial and everybody agree with him so I get why the first reflex is to expend, the of course but..... It is better to start the conversation at a realistic point and not make up some strange position about hating personnal computer (who use the AI altman is selling in a world without personnal computer to connect to it ?)

If your goal is private AI instance, local can be efficiant, but an encrypted AWS session would tend to be even more efficiant, specially if you do not need a 24-24 near peak usage installation, big pharma like Eli Lilly with extreme level of secrey use AWS encrypted private cloud session for their private own AI use, personnal AI level of privacy need should be easy to do on the cloud.

Once you have built yours, it would not impossible for you to share it (become a cloud provider for say your family that become cloud consumer to you) during your downtime in a way that do not expose your data either.

 

[bigdogchris]

Well, after reading countless threads here and everywhere elsewhere on the Internet of people bitching about their Operating System using the memory they have installed to cache apps and improve performance, I don't blame him for thinking this way.

 

[LukeTbk]

Not if the model is 10x larger or 200x larger.

well of course people using 10-200 less AI could lead to a similar amount of ram being used, that what people saying that local AI is not some solution to the ram price, using less AI would be one. If one use Qwen on openrouter they will probably use a bigger version than locally, but they do not need too.

If the only thing people used their RAM for is AI, then it would be true, but that's not the case. Saying that it's terrible that people have RAM is such a sociopathic statement.

But no one ever said that, the comments is about everyone that use significant models for AI building significant local capacity for generative AI instead of using more efficiant hardware in datacenter would not help ram price, no one said that it is terrible for people to have phone and computers at home.

Tokenmaxing is so last month. It turns out using more tokens doesn't result in better productivity. Slower inference might make the fleshy thing think for himself a little more instead of just wasting tokens on things he should be able to easily figure out with the same effort as asking AI to do it.

Gpt 6 vs 5 was mostly about using less tokens, it is indeed a major focus of where the effort are going, but that a bit orthogonal on where the inference is made.

But isn't that a terrible thing for them? How will they make money if tokens are so cheap?

intelligence are way cheaper to buy than not so long ago, but they cut by how much it cost to produce it by a larger ratio by now.

 

[Zarathustra[H]]

I was thinking about how Sam and RAM rhyme, and then I had a stroke...

...of genius!

This shit writes itself:

"Not in a rack! Not in a stack!"
"I want my local hardware back!"
"Not in the Cloud! Not out in space!"
"I'll keep my hardware at my place!"

That Sam-Alt-Man! That Sam-Alt-Man!
Do you want some cloud-based RAM?
I do not like that Sam-Alt-Man!
I do not like his cloud-based RAM!

 

[jfreund]

Well, after reading countless threads here and everywhere elsewhere on the Internet of people bitching about their Operating System using the memory they have installed to cache apps and improve performance, I don't blame him for thinking this way.

It's not caching apps and improving performance when your 16GB is full and your system is hammering virtual memory and running like dogshit. Thank God IT had a 32GB laptop turned in.

 

[MrGuvernment]

That is the point. They don't want anyone owning anything. Everything will be in the cloud and be 100% controlled and will pay for that privilege.

All while being 100% identified so every thing you do is tracked and your controlled even more!

 

[M76]

well of course people using 10-200 less AI could lead to a similar amount of ram being used, that what people saying that local AI is not some solution to the ram price, using less AI would be one. If one use Qwen on openrouter they will probably use a bigger version than locally, but they do not need too.

But I already have the RAM and GPU here at my home for various other uses, why would it be more efficient to use openrouter than my local HW, that is running on Solar for the most part?

But no one ever said that, the comments is about everyone that use significant models for AI building significant local capacity for generative AI instead of using more efficiant hardware in datacenter would not help ram price, no one said that it is terrible for people to have phone and computers at home.

Why is it different to have a computer than a phone? Both have RAM, and both are becoming more expensive due to their hoarding of all the RAM. I think it's quite terrible to have all these GPUs and RAM sitting in warehouses and data centers. IIt would be much better to have affordable home electronics that actually make life more convenient.

Gpt 6 vs 5 was mostly about using less tokens, it is indeed a major focus of where the effort are going, but that a bit orthogonal on where the inference is made.

So where is the break even point? Because the cost + interest / revenue seems to be diverging curves. Ie. the cost far outpacing the returns, and then we didn't even account for the human and environmental impact of these data centers.

intelligence are way cheaper to buy than not so long ago, but they cut by how much it cost to produce it by a larger ratio by now.

There is no indication that they ever will slow down training new versions of their models, for increasingly diminishing returns.

 

[Marees]

Sam Altman renames his twitter handle to RAM Altman

https://x.com/sama/status/2117845542021808438?s=20

 

[DrezKill]

I was thinking about how Sam and RAM rhyme, and then I had a stroke...

...of genius!

This shit writes itself:

"Not in a rack! Not in a stack!"
"I want my local hardware back!"
"Not in the Cloud! Not out in space!"
"I'll keep my hardware at my place!"

That Sam-Alt-Man! That Sam-Alt-Man!
Do you want some cloud-based RAM?
I do not like that Sam-Alt-Man!
I do not like his cloud-based RAM!

View attachment 834587

HAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHA!!!!!!
I like yours better than GN's: https://x.com/GamersNexus/status/2117193914151281037?s=20

 

[d3athf1sh]

the pc has now joined the "me too" movement

 

[LukeTbk]

But I already have the RAM and GPU here at my home for various other uses, why would it be more efficient to use openrouter than my local HW, that is running on Solar for the most part?

It is nothing to do with the ram-gpu people already has (and the type of ram-gpu people historically had) the subject is building local AI capacity (to run anything serious, people need more than the ram-gpu they already had)

Why is it different to have a computer than a phone? Both have RAM, and both are becoming more expensive due to their hoarding of all the RAM. I think it's quite terrible to have all these GPUs and RAM sitting in warehouses and data centers. IIt would be much better to have affordable home electronics that actually make life more convenient.

computer-phone here are the same (they are both PC, personal computer), that why I use and here, sam altman has no problem and which evreyone has a computer and a phone with ram, that the only way to use its product at the moment and to send him money.

So where is the break even point? Because the cost + interest / revenue seems to be diverging curves. Ie. the cost far outpacing the returns, and then we didn't even account for the human and environmental impact of these data centers.

It is hard to know for OpenAI, anthropic now with the S1 we know (that match all the estimate), yes it is extremelly divergent curve with revenue going up by 12x while the expense only 3x, the break even points seem well past in Q3 with 17 billion in revenues vs 9.86 billions in compute & infracstructure expenses, the moment they do not have to spend fortune on spot compute they would have great margin already. Environnment impact is way smaller (again if we do has much AI) if it is done in data center than everyone at home-local enterprise on less efficiant system.

Open source model cost more to operate than the more efficiant closed one, if the open source work economically the closed source one must has well.

There is no indication that they ever will slow down training new versions of their models, for increasingly diminishing returns.

Same goes for open source model, I am not sure for diminishing returns, as inferece become bigger and bigger vs training, the retuns of training a model are bigger and has of now, model are getting better faster than before, not slower, we have yet to reach the diminishing returns in that sense.

 

[DukenukemX]

what Altman said is so trivial and everybody agree with him

You keep saying this but we're 3 pages in and the internet can't stop making fun of Sam Altman's stupid statement. It's also a very easy statement to disprove. For example, the total power usage just in the USA for AI Data Centers is 183 terawatt-hours (TWh), which is the equivalent to the UK. You're telling me that AI produces more value than entire countries like Italy and Spain? Not only that but OpenAI isn't even the best AI, and hasn't been for some time. Anthropic, Claude and Google's Gemini have been better than ChatGPT. Even DeepSeek gives ChatGPT a run for their money sometimes.

It also comes back to that nobody really wants AI.

 

[LukeTbk]

You keep saying this but we're 3 pages in and the internet can't stop making fun of Sam Altman's stupid statement

Never seen anyone make fun of is actual statement, they all construct a different statement and make fun of the one they made. They talk as if Sam altman had anything against people having personal compute device with ram with them (who and how send him money as a openAI customer in that world ?)

It also comes back to that nobody really wants AI.

How does that work, who send over 5 billions to anthropic while not wanting AI, how electricity is spend on inference on people that do not want it ?

Modern chat LLM type AI is the tech in world history that was the most rapidly used in mass by regular people, all over the world.

For example, the total power usage just in the USA for AI Data Centers is 183 terawatt-hours (TWh)

It is for datacenter as a whole, i.e. that power a lot of the world internet, cloud compute, e-commerce, streaming and AI, yes it would not surprise me that it is more value than a lot of countries in the world combined. And who are all those people forced to use for 180 Twh of AI they do not want to use... ?

Not only that but OpenAI isn't even the best AI

By dollar/speed to do task it often is, Anthropic has better model, but it is a luxury option quite often, the competition is quite strong enough there is not a best model, there is best use case at different price point with many being the best for somethings.

Right now the race look a bit like this:

Luna is pretty much the cheapest thing available (hard to beat with raw electricity price at home with small model), 6.1 sol medium has a good argument to be the best deal, it does compete with Haiku 5.5 max.

 

[MrGuvernment]

LukeTbk Sorry, but this is purely Sam wanting everyone to use their services and pay for it so they can make money back because they are bleeding cash and investors want money back.

This has nothing to do with what end users are using or wasting ram on.....

before all this AI DataCenter crap and the circle jerk of investment money between OpenAI / Oracle / NVIDIA, RAM was plentiful for what ever purpose you needed it for...

I guess Sam is also upset at companies that have racks and racks of on-prem hardware with 50% buffers for growth / failover (ram/cpu), that may never happen...

 

[serpretetsky]

can someone help me understand what we're disagreeing with here?

all i see is this part of the video, which seems obviously true to me, but maybe people have different interpretation?:


talking about component contraints:

a lot of people want local inference and somehow they assume that the problem is solved now because we can have open source models, we're going to be able to run them on our computers and so we're not going to need all this stuff in the data center. but from a global allocation perspective, actually its quite terrible to have a bunch of the world's limited supply of dram sitting in thier own computer and being used only 5% of the time. you want maximum utilization in the cloud if youre trying to deal with this supply constrained environment

View: https://youtu.be/iMsUcCYNKqY?t=1745

 

[jfreund]

can someone help me understand what we're disagreeing with here?

all i see is this part of the video, which seems obviously true to me, but maybe people have different interpretation?:


talking about component contraints:


View: https://youtu.be/iMsUcCYNKqY?t=1745

One more try...

The RAM environment is "supply constrained" because this asshole bought 40% of the world's RAM production as unfinished wafers that can't even be used in a PC or data center.
The environment is also supply constrained because corporations are panic buying GPU racks to sit in warehouses because the data centers aren't built yet.
THAT is terrible allocation, not the spare 32GB kit I have on my shelf.

Altman created the problem and now he's telling us we're the problem for using our own PCs.

Fuck him with a rusty chainsaw.

 

[Ididar]

can someone help me understand what we're disagreeing with here?

all i see is this part of the video, which seems obviously true to me, but maybe people have different interpretation?:


talking about component contraints:


View: https://youtu.be/iMsUcCYNKqY?t=1745

If he was that concerned about efficiency he wouldn't be placing bulk orders for RAM he can't even use right now because he doesn't have the data centres built to house the RAM. He's having RAM sit on shelves waiting for the buildings. So, right now he has RAM that is seeing 0% utilization and not even the 5-10% it might get in someone's hobby machine. What he really means is that once he has all his data centres built and all his RAM installed then it would be more efficient, for overall AI cycles, to have it in big data centres. Yes, I agree. But, the assumption is that global allocation should only be based on efficiency of AI inference use percentage and not based on any other factors. Also, as I said before, if he cared about efficient use of RAM we should limit the number of models being trained because it is inefficient, from a RAM useage point of view, to train multiple models. There are other reasons as to why we'd want multiple models, but from a global allocation point of view it is inefficient to train multiple models.

 

[serpretetsky]

One more try...

The RAM environment is "supply constrained" because this asshole bought 40% of the world's RAM production as unfinished wafers that can't even be used in a PC or data center.
The environment is also supply constrained because corporations are panic buying GPU racks to sit in warehouses because the data centers aren't built yet.
THAT is terrible allocation, not the spare 32GB kit I have on my shelf.

Altman created the problem and now he's telling us we're the problem for using our own PCs.

Fuck him with a rusty chainsaw.

ah i see. people are not disagreeing with the statement necessarily. they're disagreeing with environment he created.

 

[jfreund]

ah i see. people are not disagreeing with the statement necessarily. they're disagreeing with environment he created.

Incorrect.

I disagree with his premise.
I disagree with his implied definition of efficient utilization.
I disagree with his conclusion.
He is wrong on every level.

 

[Nytegard]

can someone help me understand what we're disagreeing with here?

all i see is this part of the video, which seems obviously true to me, but maybe people have different interpretation?:


talking about component contraints:


View: https://youtu.be/iMsUcCYNKqY?t=1745

The words he chose are problematic.

but from a global allocation perspective

He might have been talking about local LLMs, but the above definitely can be validly interpreted to broaden the scope to all dram.

actually its quite terrible

Better wording here. Actually, it's quite inefficient vs actually, it's quite terrible. Both can be interpreted to efficiency, but the first limits it to efficiency, whereas the second is broad in scope, as well as painting it as Sam Knows Best.

you want maximum utilization in the cloud if you're trying to deal with this supply constrained environment

This is a very self-interested phrasing here. It's not what people in general want, or what would be best, but what Sam Wants.

Some people don't see a problem with his statement, while others do. And his phrasing is definitely open to interpretation.

 

[serpretetsky]

Incorrect.

I disagree with his premise.
I disagree with his implied definition of efficient utilization.
I disagree with his conclusion.
He is wrong on every level.

Ah, gotcha. i was confused because your first response to me didn't address anything in that quote from Altman.

 

[jfreund]

Ah, gotcha. i was confused because your first response to me didn't address anything in that quote from Altman.

Altman's comment in the video was about supply constraint and "terrible" allocation. As was my original reply to your post.