News
OpenAI publishes first guidelines for safety cases dedicated to frontier AI training.
Structured safety documentation should be required before continuing any frontier reinforcement learning training.
Safety cases remain a goal to reach, harder to make rigorous for AI than for aviation.
Technical guardrails cover alignment training, confinement and monitoring.
Each member of management who reviews the safety case should be able to veto the training.