Takeaways from our robust injury classifier project [Redwood Research] — LessWrong