Nick Bostrom's Superintelligence lays out the orthogonality thesis, instrumental convergence, and the control problem behind modern AI safety thinking.