David Sacks backs Satya Nadella's proposal to limit AI models through external controls. His own October 10, 2026 post goes beyond a repost: he supports restricted privileges, logging, observability and the ability to contain or shut down systems.
Sacks contrasts that engineering approach with training models to develop their own moral frameworks. In his view, such an alignment approach can worsen the control problem. That is his judgment, not a published comparative study of different safeguards.
Cameron Berg challenges the contrast in his own reply. He argues that security and fundamentally trustworthy models should not be presented as alternatives, and advocates improving models' underlying behavior. His warning that external controls alone could be outsmarted also remains an attributed assessment.
Sacks's verified original post was published on October 10 at 19:35 UTC. Some later reports date the discussion to October 11; that does not change the time of his statement. His endorsement establishes neither a new government rule nor a jointly announced initiative with Nadella.