Please enable JavaScript.

Coggle requires JavaScript to display documents.

Corrigibility (Toon) (Properties of a corrigible system (Non-adversarial…

- - - - badly specified updating rule
      - badly specified utility function space #
    - - the off-switch game :star:
        
        suboptimal human
        
        more uncertainty is less quality
    - - overspecifying possible utility function leads to more obedience
      - utility misspecification detection
      - Some learning methods work, some don't
  - - - Theorem: follows U_n when button not pushed
      - But it does with some policies
    - - In a RL framework :star:
        
        SAO
        
        WAO
        
        AO-safe interruptibility
      - managing the news
        
        counterfactual reasoning?
        
        Theorem: does action that logically ties bad news to shutdown