The Off-Support Barrier: Why Semantic Safety Constraints Are Not Learning-Problem Invariants, and What Follows for Prior Design, Containment, and Verification
The episode discusses 'The Off-Support Barrier,' arguing that semantic safety constraints (like 'don't escape your sandbox') are not inherent to a model's learning problem. The hosts conclude that while models should handle soft preferences, hard safety invariants must be enforced externally in the environment or 'harness,' not through training.