Quote Originally Posted by Lowell View Post
Nvidia now has a virtual monopoly on them for provisioning the AI revolution. And their stock has gone out the roof. So important are their GPU chips that the US government bans export of their highest end chips to adversaries, China in particular.
The interesting thing about that is AMD and Intel both have equally capable GPU hardware, but in the last 10 years they've completely botched the software side of things, so everyone defaults to nVidia. AMD and Intel have gotten better as of late, but if you want to do serious AI work, nVidia hardware is the best for now. The other interesting thing is that Google and Amazon make their own AI accelerator chips for in-house use only, and I've heard they are better than nVidia hardware.

On this front I, and I'm guessing you too JBMcB, am following the efforts toward creating personal agents, AI's that will be based on our private home computer systems that will serve as our own personal butlers, independent of and protected from internet based AI's and networks.
The software company I work for is looking at that, as are most other software companies. We want AI to analyze our source code to find optimizations and problems, but we don't want to hand that code over to Microsoft or Google, so we are working on running everything in-house.

I've seen where there are many instances of open-source AI's being offered by Meta, Google and others. Right now it takes some heavy hardware but the coding is becoming leaner and cleaner to the point where we will all have the option of our own personal AI's.
Depending on the model and what you want to do, running an AI takes a *lot* less hardware than training one. I could run a stripped-down version of ChatGPT on my mid-range machine and it would run OK. The thing that needs real horsepower is training the models, or getting them to learn. The latest version of ChatGPT took months to build on an enormous array of AI servers. Even doing basic lightweight training of a simple concept can take hours.

This is one of the main stumbling blocks to getting to generalized AI. As it stands, running a model is fairly easy in limited domains. ChatGPT can generate human-readable text fairly well, but it can't recognize pictures or do anything with video. Stable diffusion can recognize and create images to some degree of accuracy, but it doesn't understand text at all. None of the models can deal with abstract concepts very well, nor stacking more than a few concepts together at once. And, most importantly, none can learn in real-time. The on-line versions of ChatGPT cheat by culling the internet for updated information, but that doesn't get soaked into it's neural network. Doing that still takes days or weeks.

On top of that, we are nudging up against the physical limits of what we can do with computer chips. They are hitting the limit as to how small we can make them before the internal traces can't carry electricity. Quantum computers will work around this limitation, but general-purpose quantum computers are still quite a ways off. The most complex quantum machine has around 1,100 qbits, which is roughly equivalent to a transistor. A very basic adding machine needs around 2,000 transistors to function, so we are still a ways off from doing complex matrix operations on a quantum computer.