Thanks for the reply, that makes sense. It's not immediately clear why the modified softmax (allowing output of 0) will "tame the weights", but I need to read the blog post more closely and think about it...