Sam Altman's argument about RAM utilization is reasonable on a first pass, but misses several key points:
1. people aren't using (GPU) RAM [1] for just running LLMs (in the local model sense), they are using it for other things... games, system displays, other models (Text-to-Speech, Speech Recognition, etc.);
2. being able to run models locally allows them to be used when the internet is down, use them on private data (e.g. medical records, proprietary company data, etc.), including training on that data;
3. the big AI companies are already storing unused GPUs on shelves waiting for their roll-out in data centres, as they can't build the capacity fast enough.
With point (3) that goes against Sam Altman's argument about utilization. The big companies have the hardware, but aren't utilizing it themselves.
[1] And if the argument extends to RAM in general you have all sorts of applications, especially on developer hardware -- editing and compiling software; running local servers; running CAD and other applications; etc.
Even putting aside the whole privacy and tuning debates - unless Sam is willing to sell us inference at something near the pre-constrained cost of directly owning the DRAM all he is advocating for is "don't be greedy guys, that's what I'm for. I can be greedy for you!". It's not actually efficient for people to pay huge markups on compute and memory vs letting that same memory site idle some of the time. I mean I guess it's efficient for OpenAIs bank account, but not for anyone else.
Altman leaves out of this the fact that OpenAI has basically tried to corner the market on RAM, trying to lock up 40% of the world’s supply without even guaranteeing that it will follow through on the actual purchase orders. I mean, I think that alone would have at least some impact on the current price for RAM, wouldn’t it?
Hes going to lobby the federal government to confiscate all your RAM, like FDR did with the gold. :|
A problem of both the Altman post and posts that react on him is that they take efficiency/utility in absolute sense instead of in marginal sense.
There is clearly high marginal utility of the first n GB of RAM in your computer, but then its go down probably faster than for RAM in OpenAI's datacenter. And the market just solves for the equilibrium.
> Altman talks about how they’re working on having robots build “self-replicating data centers” (what?!?)
Company that cautions against the paperclip maximizer already at work building the paperclip-maximizing hardware themselves.
Though I appreciate how this would invalidate the last remaining arguments for the "data centers bring local employment" claim, if the security and construction jobs are automated away as well.
Maybe it will be replaced with a new spin: "Data center? What data center? Oh that one... that's just a wild weed, you can't prove it's one of ours..."
Let's all get rid of our personal means of transportation as well, super inefficient to have cars and bikes and such parked for the majority of the time! Some people even buy houses that they're never in! Or they buy kitchen equipment they only use once a month! Or clothes that people only wear for parties! Why even own anything at all!
I am pretty sure Sam has already sold the majority of his personal belongings and only rents whatever he needs when he needs it. What a great example for us all.
I have been finding this quiye weird, it seems that the real villany was in what was in Sam Altman's mind when he said something that is basically true.
I have seen many explanations of what that was exactly, they are not all agreed on the bad thoughts were, but they are pretty much in agreement that it was a bad thing.
It's like saying 'public transport is more efficient than driving your own car', and the person hearing you say 'you shouldn't own a car, you should use the public transport system I will own.'
Are LLMs the best use
Of ram? Many other extremely beneficial devices for extremely high value use cases are suffering because of this madman trying to corner the market. Show me the value of everything else suffering.
Let's put gaming and all the other uses for compute that require low latency, etc. off the table and just talk about AI compute. People can and do run open models on their own hardware. Altman is, not unseriously, arguing that this hardware would be better utilized if it were concentrated in data centres rather than distributed into smaller, private clusters that might not reach full utlization.
Now, let's take privacy concerns off the table. OpenAI's complete lack of respect for privacy, authorship, copyright, etc. mean that you simply can't entrust private data to OpenAI. Full stop. Anyone with privacy concerns gets a free pass.
Finally, we're dealing just with people doing LLM computations on data that's not sensitive. Why aren't these people doing things in data centres? There are likely a variety of reasons, but one is likely price. If somebody does a lot of stuff with local models, they've probably found an application where a bigger closed model doesn't offer compelling advantages and the cost, purely in terms of filthy lucre, of running a local model is lower. These are the users Sam Altman can legitimately target. What Altman is arguing is that these people are paying too much.
How should he target them? Become the Steam of compute. Save them money and time. If Altman wants these people's RAM because he can use it more efficiently, he needs to offer them the same services they can get from their own hardware, cheaper, and with less trouble.
> if you have a sufficiently powerful AI tool that is told to “maximize” the creation of paperclips, it might decide the best way to do that is to seize control of all resources
I don't understand the backlash against this interview.
When listening to the interview it sounded clear that it was just in context of local inference.
To me it came across as stating that setting up local inference, when dram is limited, is indeed limiting from a global perspective.
I think it's a bit disingenuous to state that his opinion is that we shouldn't have any ram being used for any local purposes.
Efficiency and resilience are always in trade-off. Even if there was no other argument it would be worth having local inference as a guard against systemic SPOFs in the big providers.
"People shouldn't be allowed to bicycle for fun. Only utilitarian bicycles should be allowed. For there's too much energy spent manufacturing mountain bikes, tires, converting ski station into bike parks for the summer. That energy and money should instead be directed towards creating more RAM factories and more datacenters for the AI companies".
Where does such an "efficient" argument ends?
P.S: as a bonus question: does anyone know how to differentiate a paid shill from a sycophant?
Screw Sam Altman. He's the root cause of this shortage, besides that, how does he justify his car having extra seats that he's not using. Sick and tired of these techbros trying to re-centralize computing all over again whilst stealing left, right and center.
I think the argument being made is fundamentally wrong. It reminds me of the Socialism vs Capitalism models - this idea that centralizing the resources will somehow unlock efficiency. It's also becoming clearly that merely throwing more data, RAM, compute, etc, is not generating significantly better behaviours.
We are far better off with millions of people running lots of little experiments, than a handful of super large companies running a subset of experiments. We really need to get agentic LLMs onto local machines at a reasonable price to really accelerate.
The likes of qwen3.5-0.8b have been super interesting to me, you have this tiny little model that can run on a CPU that can exhibit agentic behaviours (with a lot of tickling) - in the hands of hobbyists is where I suspect we will see some cool updates.
Sam Altman's argument about RAM utilization is reasonable on a first pass, but misses several key points:
1. people aren't using (GPU) RAM [1] for just running LLMs (in the local model sense), they are using it for other things... games, system displays, other models (Text-to-Speech, Speech Recognition, etc.);
2. being able to run models locally allows them to be used when the internet is down, use them on private data (e.g. medical records, proprietary company data, etc.), including training on that data;
3. the big AI companies are already storing unused GPUs on shelves waiting for their roll-out in data centres, as they can't build the capacity fast enough.
With point (3) that goes against Sam Altman's argument about utilization. The big companies have the hardware, but aren't utilizing it themselves.
[1] And if the argument extends to RAM in general you have all sorts of applications, especially on developer hardware -- editing and compiling software; running local servers; running CAD and other applications; etc.
Even putting aside the whole privacy and tuning debates - unless Sam is willing to sell us inference at something near the pre-constrained cost of directly owning the DRAM all he is advocating for is "don't be greedy guys, that's what I'm for. I can be greedy for you!". It's not actually efficient for people to pay huge markups on compute and memory vs letting that same memory site idle some of the time. I mean I guess it's efficient for OpenAIs bank account, but not for anyone else.
Altman leaves out of this the fact that OpenAI has basically tried to corner the market on RAM, trying to lock up 40% of the world’s supply without even guaranteeing that it will follow through on the actual purchase orders. I mean, I think that alone would have at least some impact on the current price for RAM, wouldn’t it?
Hes going to lobby the federal government to confiscate all your RAM, like FDR did with the gold. :|
A problem of both the Altman post and posts that react on him is that they take efficiency/utility in absolute sense instead of in marginal sense.
There is clearly high marginal utility of the first n GB of RAM in your computer, but then its go down probably faster than for RAM in OpenAI's datacenter. And the market just solves for the equilibrium.
I heard someone say it’s inefficient for Sam to have two kidneys
> Altman talks about how they’re working on having robots build “self-replicating data centers” (what?!?)
Company that cautions against the paperclip maximizer already at work building the paperclip-maximizing hardware themselves.
Though I appreciate how this would invalidate the last remaining arguments for the "data centers bring local employment" claim, if the security and construction jobs are automated away as well.
Maybe it will be replaced with a new spin: "Data center? What data center? Oh that one... that's just a wild weed, you can't prove it's one of ours..."
Let's all get rid of our personal means of transportation as well, super inefficient to have cars and bikes and such parked for the majority of the time! Some people even buy houses that they're never in! Or they buy kitchen equipment they only use once a month! Or clothes that people only wear for parties! Why even own anything at all!
I am pretty sure Sam has already sold the majority of his personal belongings and only rents whatever he needs when he needs it. What a great example for us all.
I have been finding this quiye weird, it seems that the real villany was in what was in Sam Altman's mind when he said something that is basically true.
I have seen many explanations of what that was exactly, they are not all agreed on the bad thoughts were, but they are pretty much in agreement that it was a bad thing.
It's like saying 'public transport is more efficient than driving your own car', and the person hearing you say 'you shouldn't own a car, you should use the public transport system I will own.'
How many houses owns Sam? How many rooms in those. Maybe he should let homeless people live in them when he's not there.
Or maybe he should live in a one bedroom apartment.
That would be way more efficient.
In short sound logical argument absolutely zero sense of optics or relatability
This keeps getting posted, but it doesn't seem interesting to discuss. Money is the way we decide how scare resources are allocated.
If Sam wants to buy 16GB of 6 year laptop RAM, he can send me a check for $2000.
Are LLMs the best use Of ram? Many other extremely beneficial devices for extremely high value use cases are suffering because of this madman trying to corner the market. Show me the value of everything else suffering.
Let's put gaming and all the other uses for compute that require low latency, etc. off the table and just talk about AI compute. People can and do run open models on their own hardware. Altman is, not unseriously, arguing that this hardware would be better utilized if it were concentrated in data centres rather than distributed into smaller, private clusters that might not reach full utlization.
Now, let's take privacy concerns off the table. OpenAI's complete lack of respect for privacy, authorship, copyright, etc. mean that you simply can't entrust private data to OpenAI. Full stop. Anyone with privacy concerns gets a free pass.
Finally, we're dealing just with people doing LLM computations on data that's not sensitive. Why aren't these people doing things in data centres? There are likely a variety of reasons, but one is likely price. If somebody does a lot of stuff with local models, they've probably found an application where a bigger closed model doesn't offer compelling advantages and the cost, purely in terms of filthy lucre, of running a local model is lower. These are the users Sam Altman can legitimately target. What Altman is arguing is that these people are paying too much.
How should he target them? Become the Steam of compute. Save them money and time. If Altman wants these people's RAM because he can use it more efficiently, he needs to offer them the same services they can get from their own hardware, cheaper, and with less trouble.
> if you have a sufficiently powerful AI tool that is told to “maximize” the creation of paperclips, it might decide the best way to do that is to seize control of all resources
Hey don't talk about Sam like that!
I don't understand the backlash against this interview. When listening to the interview it sounded clear that it was just in context of local inference. To me it came across as stating that setting up local inference, when dram is limited, is indeed limiting from a global perspective.
I think it's a bit disingenuous to state that his opinion is that we shouldn't have any ram being used for any local purposes.
Efficiency and resilience are always in trade-off. Even if there was no other argument it would be worth having local inference as a guard against systemic SPOFs in the big providers.
What kind of specs would you need on a local machine to be able to run fairly decent (Opus 5.5-level) models on it?
And if we cared even a little bit about efficiency the world would look very different, so we can just stop listening to him anyways
We don't have to hold up paperclips as theoretical anymore. AI needs RAM. All of it. The real life paperclip maximization is RAM chips.
SA: "Well, first of all, we need to do much better than we are doing now."
Me: This might well be the understatement of the century.
SAL "Sand-boxing works to a point. Then I think monitoring........ works to a further point."
Me: REALLY? You think you should MONITOR your system of agents while it does its work? GEE I wish I'd thought of that!
Why did this post get dropped from the homepage?
His case his "efficient"? What about this one:
"People shouldn't be allowed to bicycle for fun. Only utilitarian bicycles should be allowed. For there's too much energy spent manufacturing mountain bikes, tires, converting ski station into bike parks for the summer. That energy and money should instead be directed towards creating more RAM factories and more datacenters for the AI companies".
Where does such an "efficient" argument ends?
P.S: as a bonus question: does anyone know how to differentiate a paid shill from a sycophant?
Screw Sam Altman. He's the root cause of this shortage, besides that, how does he justify his car having extra seats that he's not using. Sick and tired of these techbros trying to re-centralize computing all over again whilst stealing left, right and center.
I think the argument being made is fundamentally wrong. It reminds me of the Socialism vs Capitalism models - this idea that centralizing the resources will somehow unlock efficiency. It's also becoming clearly that merely throwing more data, RAM, compute, etc, is not generating significantly better behaviours.
We are far better off with millions of people running lots of little experiments, than a handful of super large companies running a subset of experiments. We really need to get agentic LLMs onto local machines at a reasonable price to really accelerate.
The likes of qwen3.5-0.8b have been super interesting to me, you have this tiny little model that can run on a CPU that can exhibit agentic behaviours (with a lot of tickling) - in the hands of hobbyists is where I suspect we will see some cool updates.