Human-Compatible AI

Content

• Provably beneficial and well-founded AI, probabilistic
• Safety guarantees, switch-off problem
• Human-aware and human-centered AI: Mental models, interpretable behavior and generation of explanations, agent-supported human collaboration work, aligned AI, analogies and common sense
• Adaptation of language models: Reinforcement learning with
  human feedback (PPO method)
• Task-oriented perception: from task descriptions to internal goals, task representations
• Basics of assistance games, perception of human
  preferences, inverse reinforcement learning
• Simulation of agent behavior in mechanisms,
• Durkheim test, Weizenbaum test