A new proposal on the AI Alignment Forum outlines a theory of change for value generalisation. The plan calls for small research teams and medium compute to build rigorous benchmarks and toy problem demonstrations. It focuses on mitigating risks during the transition from academic research to commercial application.