Why Tay the Twitter bot from 2016 failed and what we can learn from it. How can we generalize this to ML development best practices?