Wow, this talk is an excellent overview of the Overton system: In a nutshell: • ML engineers should strive to improve supervision/data quality at all costs • Model architecture benefits/hyperparameter tuning are marginal compared to data