Running models locally is the future for most inferencing cycles. There is a lot of accuracy that could be improved in your numbered list trying to dissuade people.

> 64GB probably isn't enough for competitive LLMs anyway

I am trying to charitable, but this is pretty not true. And the hedging in your statement only telegraphs your experience.