Pretty sure there’s at least a couple orders of magnitude in purely algorithmic areas of LLM inference; maybe training, too, though I’m less confident here. Rat

Pretty sure there’s at least a couple orders of magnitude in purely algorithmic areas of LLM inference; maybe training, too, though I’m less confident here. Rationale: meat computers run on 20W, though pretraining took a billion years or so.