NVIDIA PAIR Allows Local AI Clustering - OpenAI is DEAD

Eli the Computer Guy (0:00)

Oh boy, oh boy, oh boy, we get to talk about my favorite subject, that is right, systems architecture. And not only are we gonna be talking about systems architecture, but we're going to talk about how AI architecture in 2026 is still incredibly immature. This is an immature technology architecture. We do not know how we are going to be building AI systems in 2035, right? If you're going to be building a web application in 2026, you know how it's gonna be done, right? Basically, there is a playbook for building web applications, whether it is the next Uber or the next dog-dating app or whatever else, you know how you're gonna build it. That was not the case back in 2006 Any of us still have PTSD from things like ActiveX controls or Silverlight? Oh, one of my commenters brought up Silverlight. I had forgotten Silverlight. Anyways, right, there was a time when Microsoft came out with Silk. We're not gonna go into it, right? But those were the old days, right? Those were the before times. Now we know how to build basically web applications when we want to. One of the big things to understand with artificial intelligence is that the stack itself, the architecture itself, is still incredibly mature. We are figuring out how all this stuff should be built. In a decade, there will be a playbook for how to build AI systems, but we are not there yet. And this is one of the reasons that I argue against the trillions of dollars of investment into AI data centers, simply because we do not know the architecture that we're going to need for AI systems going in the future. As I've talked about before, right, OpenAI Sam Ullman really believes in intelligence being a utility, just like water or electricity or sewage or something like that. You will go to one provider or multiple providers for the intelligence for your systems. But if you look at companies such as Google, right, Google is coming out with small models and building them into everything. Android has a model built into it. The Chrome browser, you can put a model directly into the Chrome browser. They came up with a model that can run on laptops with at least 16 gigs of RAM. And one of the things there is basically having those local models available so that you can run tasks locally, and then if whatever the system can't process the thing locally, then you can send it up to the cloud. And now think about it, right?

If 50% of your AI queries can be done locally, basically that removes 50% of the load that would go to a data center, right? So then the question becomes is, do you need as many data centers as they're building out? And that's a hell of a question. What I really think is cool about this particular article is that this is from Nvidia, and they're basically doing with artificial intelligence, something we've done with database services and other things for decades at this point in time. And basically what they're doing is they're creating a local cluster for AI services. So where Google, right, allows you to put LLMs onto your Android phone or your computer or whatever else, what this is allowing is Nvidia is allowing you to use the resources of other computers on your network to complete AI tasks. Basically, you break down the AI task and the little subtasks. You send those subtasks off to different computers on the local area network. They complete those subtasks, bring them in back to the computer that initially made the request. And so now you can get a tremendous amount of power through your local infrastructure without necessarily having to go off to the cloud. And I think this is an important thing because, like, think about it. Like, if you've been in the enterprise world, you know most people are barely using any of the powers of their systems, right? Most computers, most computers, for most people, got good enough back in about 2012, right? If you have a computer post 2012, right? You got 16 gigs of RAM, you got a decent graphics card in or whatever else, you have more than enough power in the computer, right? And so think about this. You can now use the power on these computers in order to complete AI tasks, which becomes very interesting. Also, you think about this in the corporate world. In the corporate world, the business world, like the computers people buy can be weird.

Like have you ever done that? Like you're going through and you're looking for fixed assets or whatever computers, you're auditing the computer systems, then you just find like five computers, actually have really pretty good graphics cards. Like the accounting department. Like you're looking at the accounting department's computers, and you're like, why the fuck do you have these graphics cards? Or I don't know. It's like, what are you doing other than spreadsheets? No, no, we just do spreadsheets. Okay, why do you have $1,000 graphics cards in your computers? I don't know.