this is basically the only thing pre-training teams work on in labs. compute efficiency is the metric, the assumption that scaling = intelligence is considered a given.