I think the bitter lesson only talks about task performance but not computational efficiency. Could tiny models improve efficiency? Maybe by just using a genera

I think the bitter lesson only talks about task performance but not computational efficiency. Could tiny models improve efficiency? Maybe by just using a general architecture on specialized data, so the artichecture itself is not task specific?