TB: Artificial Intelligence: A Modern Approach (Russell & Norvig)
Core Thesis
AI is the study and construction of rational agents — systems that perceive their environment and take actions that maximize expected utility. The central challenge has shifted from building capable systems to building systems whose goals remain aligned with human values as capability scales.
Key Takeaways
Rational agents
- An agent is rational if it selects actions that maximize its performance measure given its percept sequence and built-in knowledge.
- The standard model: maximize expected utility. This works well until the utility function is wrong.
- Alignment problem: as AI systems become more capable, the difficulty of specifying correct objectives grows — a misspecified utility function pursued optimally is catastrophic.
Game theory and multi-agent reasoning
- Nash equilibrium: a strategy profile where no agent can improve by unilaterally deviating. Rational agents converge to Nash outcomes even without coordination.
- Prisoner's dilemma: individually rational choices produce collectively suboptimal outcomes — the foundational tension between self-interest and cooperation.
Machine learning as search
- Supervised learning: find a hypothesis consistent with labeled training data that generalizes to unseen examples.
- The bias-variance tradeoff governs generalization: models too simple underfit; models too flexible overfit.
- Deep learning: hierarchical feature learning that made perception tasks tractable.
Autonomous systems and risk
- Autonomous weapons raise novel ethical questions: who bears moral responsibility when a system decides to use lethal force?
- The frame problem: an agent acting in the world must reason about what does NOT change as well as what does — an underappreciated difficulty.
Mental Models
- Expected Value — rational agent decision-making is formalized as expected utility maximization
- Nash Equilibrium — multi-agent rational convergence; explains why coordination problems are hard