The policy that says not to scale unless we are entirely sure we can make what we build safe.
RSP Structure
A good RSP should cover:
- Limits: what capabilities would indicate unsafe scaling
- Protections: what measures are necessary to contain risks
- Evaluations: what procedures used to catch warning signs of dangerous capability
- Response: set of principles to use to pause AI progress
- Accountability: how AI developer can ensure RSP commitments are executed and how people can verify
RSP Falloffs
- Slower development, to keep up with competition, you must reduce the RSP scope
- Requires consistent evaluations that sound false alarms
- GA