Daniel Spiewak on Nostr: nprofile1q…szpm2 Fwiw, I'm mostly nitpicking methodology here. The broad conclusion ...

nprofile1qy2hwumn8ghj7un9d3shjtnddaehgu3wwp6kyqpqtlz4xpxfuxsd7238rfpsq3qd4lkstdz0lzum4keg5axkx0ayvnmshszpm2 (nprofile…zpm2) Fwiw, I'm mostly nitpicking methodology here. The broad conclusion still clearly holds: large models are very very power-hungry. Now, that can be justified in some cases (I have another thread from yesterday about how DLSS4 shows a case where inference results in less energy usage than classical computation by a significant margin), but I'm quite convinced search is not at all one of those problems.