What does it mean for correct operation to rely on transfer learning? — LessWrong