Reposted by Corinne (NYC)
6:08: Why are the models not able to flag that a task is impossible to the evaluators? Why is the model not trained to give up gracefully if a task is impossible or would require doing unethical things? Why are there so many impossible tasks in your eval?
