[pzo]

I really hope we will have some innovation with LLM that would remove such demand for RAM. Any ideas what is more likely?

ASIC LLM?

inference inside SSD?

bigger 1-bit trained models?

better MoE so we don't have to load all experts?

[m4rtink]

Bubble popping, useless AI hardware sold off to creditors and scrapped to RAM sticks in SSDs for actual real world use case ? And eventually maybe later in a few years some actually useful and profitable real use cases for the current LLM generation are found.

Alternatively "AI" gets such a bad name regular people will go and burn it all down as long as it has "AI" even just in the name.