How "reverse game theory" could keep AI agents in check
AIResearchers Stephen Morris, Elad Hazan and colleagues propose mechanism design, which sets rewards so that AI agents behave as intended without relying on understanding their internal workings. Morris's preprint suggests rewarding a model for how well its answers predict a second model's output, so that sandbagging on tests becomes a losing bet.
Source details
MIT Technology Review · AI · Amit Katwala · technologyreview.comPublished · added here