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.

Hes going to lobby the federal government to confiscate all your RAM, like FDR did with the gold. :|

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.

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 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.'

Or maybe he should live in a one bedroom apartment.

That would be way more efficient.

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..."

If Sam wants to buy 16GB of 6 year laptop RAM, he can send me a check for $2000.

Hey don't talk about Sam like that!

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.

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.

"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?

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!

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.