The virtues of this article is the adoption of via negativa. If the main question is safety, then trying to build a AI with much efford guide us to erros of bad specifications. The problem is to convince AI researchers of that, I read some interviews here and part of them are not aware of the risks.
Anna Salamon and I have finished a draft of "Intelligence Explosion: Evidence and Import", under peer review for The Singularity Hypothesis: A Scientific and Philosophical Assessment (forthcoming from Springer).
Your comments are most welcome.
Edit: As of 3/31/2012, the link above now points to a preprint.