

He is technically wrong though. His argument is based on people only using DRAM for model inference. Which is nuts.
People use DRAM for a lot of things more beneficial than AI inference. And those use cases are also many times more efficient than whatever this Epstein class guy is using it for.
A CNC machine that produces actual goods that are actually beneficial to humanity does need DRAM. I doubt it needs more than 100MB though. An LLM counting the RS in strawberry eats up a gazillion TB of DRAM to produce nothing of value.


I don’t know what weird definition of value you are using. But for a sane definition, this is just wrong.