If you exclude the deep work on models themselves and just focus on use — harnesses etc. — then the intellectual community is basically prescientific. Very little meaningful empiricism. danluu's notes on benchmarks and experimental design come to mind: https://danluu.com/exercise-7/