Safeguard-Conditioned Uplift: Measuring Utility-Risk Frontiers for Dual-Use Biology Assistants
The episode discusses a paper titled "Safeguard-Conditioned Uplift," which measures the trade-off between utility and risk for dual-use biology assistants. Hosts analyze how deployment conditions change model behavior, showing that no single safeguard is universally best. They conclude that the key takeaway is to measure full access conditions, use risk-budgeted calibration to find optimal operating points, and build trust through transparency about safety tradeoffs.