A new paper by scientists at MIT investigating the generalizability of large language model results in some interesting conclusions.