Sam Altman's argument was that people tend to not run their consumer electronics at 100% 24/7, and so views the time that's not happening as wasted potential for RAM that could be sitting in a data center. I sincerely reject that notion, but from that... utilitarian perspective it's hard to argue that spreading out the world's limited RAM supply across thousands of personal machines would fix any issues
Just give us back our RAM prices so that everyone can afford 256GB DDR5 and run their models locally!
And Sam's (correct) point is that that doesn't work. If everyone tried to run the same amount of inference locally with the same size models that they're currently doing, the shortage would be even worse.
Just give us back our RAM prices so that everyone can afford 256GB DDR5 and run their models locally!
This solves the datacenter problem and it's impacts, RAMageddon, SSDgeddon, etc..
and they can just sell us the improved models without the hassle of doing the compute themselves..
What's the problem? With the proper ToC's they would still get our data.. ¯\_(ツ)_/¯