On a closely related note, there is a collaboration of DeepMind and the FHI going on and they are about to present a paper on safely interruptible agents at UIA 2016: http://lesswrong.com/r/discussion/lw/noj/google_deepmind_and...