Researcher Explains How Model Failures Drive New AI Problem Statements