The actual model required for general purpose likely lies beyond the range of petabytes of memory.

These models are using gigabytes and the trend indicates its exponential. A couple more gigabytes isn’t going to cut it. Layers cannot expand the predictive capabilities without increasing the error. I’m sure a proof of that will be along within in the next few years.

There’s plenty of stuff where ML algorithms the state of the art. For example the raw data from nanopore DNA sequencing machines is extremely noisy and ML algorithms clean it up with much less error than the Markov chains used in years previous.

“There is no free lunch.”, is a saying in ML research.

A lot of new tech is not as efficient or equally so at the get go. Learning how to properly implement and utilize it is part of the process.

Right now we are just throwing raw computing power in ML format at it. As soon as it catches and shows a little promise in an area we can focus and refine. Sometimes you need to use the shotgun to see the rabbits ya know?

For the meme? The Walking Dead. For the content? No idea.

So what you’re saying, Dad, is it’s nascent and already faster? Gotcha.