[wongarsu]

Which is surprisingly (or unsurprisingly?) similar to many approaches in ML training. We train our ability to predict the future (i.e. consequences of actions) by consuming information about things we would have mispredicted (so called novel events, or news).

The danger is that if we only consume newsworthy events, we lose sense of the baseline and start only predicting outliers everywhere. We need to consume statistics or everyday stories from a diverse group of people in addition to news.