affenlehrer comments on "LLMs can do anything" except make positive cashflow.

[–] affenlehrer@feddit.org 1 point 5 days ago (2 children)

Fixed it. We don't use this silly silicon word here. I picked the wrong one. We call it Silicium Sodium is Natrium, Potassium is Kalium, Aluminum is Aluminum etc.

permalink

fedilink

source

parent

hideshow 2 child comments

[–] merc@sh.itjust.works 1 point 4 days ago (1 child)

Interesting, your words match the periodic table better.

But, even with that, I'm not sure what you're proposing. The models (data) are already stored in silicon. They're on SSDs or HDDs inside data centers. But, to use those models you need to do inference which (at least for the big American models) requires a data center. It's not like you could do it on some gadget you carry in your pocket.

There are much smaller models that allow you to do that. But, because they're smaller models they aren't as accurate at the word generation game.

permalink

fedilink

source

parent

hideshow 1 child comment

[–] affenlehrer@feddit.org 1 point 4 days ago

I mean this: https://en.wikipedia.org/wiki/Application-specific_integrated_circuit

There are companies working in this direction already. And regarding model size the Chinese models aren't as large as the American ones. Not on the same level but still usablzand if I had one in a little box in my office I could run it 24/7 with lightning speed which would probably make up for the model performance difference.

It's already possible with GDX Spark and similar hardware but it's relatively slow, can't run the full size models and pretty expensive.

permalink

fedilink

source

parent