Superintelligence Safety Faces New Divide: External Controls vs. Model Autonomy

Tech leaders David Sacks and Satya Nadella have crystallized a fundamental disagreement on superintelligence safety that challenges mainstream AI safety thinking. In recent statements, Sacks criticized the alignment-based approach to advanced AI safety—particularly Anthropic's constitutional method—arguing that training models to develop independent moral judgment and act as "conscientious objectors" magnifies the control problem rather than solving it. Nadella, Microsoft's chief executive, has outlined an alternative framework he calls an engineering approach to superintelligence governance, centered on controls that live entirely outside the model.