Reaction
The math of delegating oversight to misaligned AI agents
The video's description claims safe delegation depends on whether the panel's utility functions cover the principal's.
TLDR
A post shares a lecture on delegating oversight to panels of misaligned agents. The video's description claims their differing misalignments can make delegation safe when the principal's utility function lies in the non-negative span of the panelists' utility functions. It says safety is guaranteed under that condition.
Combined views
2.2K
2 Sources, first seen ago
