3 ms·TIL: Masked Language Models Are Surprisingly Capable Zero-Shot Learners4 points by yurivish 2y ago