Can software really be made dependable? And if it was, would we be able to recognize it? For the last two years, I've been chairing a study of the National Academy of Sciences on dependable software, and have had the opportunity to hear - in a workshop held last year, and in open sessions of the committee -anecdotes and viewpoints that have often surprised me. The conclusions of my committee will not be made public until the final report is out. In this talk, therefore, I'll share some of the things we've heard, and draw some connections to requirements engineering.