I don't like the reframing of 'learning from mistakes' from a human-like, near instantaneous feedback loop, to a year-long process of retraining on many traces collected from user data. They're different concepts and we should refer to them using different phrasing.