Comment by cubefox

Large language models would need tens or hundreds of gigabytes of SRAM. Pretty sure the enormous cost for this makes the approach economically unfeasible.