This is the correct approach. Also see: https://ezyang.github.io/ai-blindspots/