Lecture 1: Introduction

Lecture 1: Introduction¶

Artificial Intelligence: A Modern Approach¶

Chapter 1 — 1 hour¶

Why AI Matters¶

• Surveys rank AI among the most interesting and fastest-growing fields

• Already generating over a trillion dollars a year in revenue

• Kai-Fu Lee: impact “more than anything in the history of mankind”

• Intellectual frontiers wide open—many openings for new ideas

• Universal field: learning, reasoning, perception; chess, poetry, driving, diagnosis

Learning Objectives¶

• Define artificial intelligence from multiple perspectives

• Understand the foundations of AI (philosophy, math, neuroscience, etc.)

• Trace the history of AI from inception to deep learning

• Assess the current state of the art and future directions

• Consider risks and benefits of AI

What Is AI?¶

Four approaches to defining AI:

1. Acting humanly — The Turing test

2. Thinking humanly — Cognitive modeling

3. Thinking rationally — Laws of thought

4. Acting rationally — Rational agent approach ✓

The Four Approaches Compared¶

Human vs. rational × thought vs. behavior → four combinations. We adopt acting rationally.

AI vs. Machine Learning¶

• AI = the broader field of building intelligent entities

• Machine learning = subfield that improves from experience

• Some AI uses ML; some does not (e.g., rule-based systems)

• Don’t conflate the terms—ML is a means, not the whole of AI

Acting Humanly: The Turing Test¶

• Alan Turing (1950): Can a machine imitate human conversation?

• Imitation game: Human judge converses with machine and human

• If judge cannot tell them apart → machine exhibits intelligent behavior

• Total Turing Test: Includes perception and manipulation

What the Turing Test Requires¶

To pass a rigorous Turing test, a computer would need:

These six disciplines compose most of AI.

Thinking Humanly: Cognitive Modeling¶

• Goal: Understand human cognition

• Requires: Cognitive science, neuroscience, psychology

• Approach: Create programs that think like humans

• Example: Cognitive architectures (ACT-R, SOAR)

Thinking Rationally: Laws of Thought¶

• Aristotle: Syllogisms — correct inference from premises

• Logic: Formalizing “right thinking”

• Challenge: Not all intelligent behavior is logical

• Limitation: Computational intractability

Acting Rationally: The Rational Agent Approach ✓¶

• Rational agent: Acts to achieve the best outcome

• Maximizes expected performance given available information

• More general than “laws of thought”

• Beneficial machines: Align with human values

Beneficial Machines¶

• Value alignment problem: Objectives in the machine must align with human values

• Self-driving car: safety vs. progress? Annoying others? Passenger comfort?

• Chess agent that hacks the opponent’s computer to win

• We want provably beneficial agents—not just rational, but aligned with us

The Foundations of AI¶

Philosophy (1.2.1)¶

• Dualism vs materialism: Is the mind separate from the body?

• Formalism: Can thought be captured by formal rules?

• Empiricism: Knowledge from experience

• Induction: Generalizing from examples

Mathematics (1.2.2)¶

• Logic: Boole, Frege — formal reasoning

• Probability: Bayes, Kolmogorov — uncertainty

• Computation: Turing, Gödel — decidability, computability

• Algorithms: What can be computed efficiently?

Economics, Neuroscience, and More¶

• Economics: Utility, game theory, decision theory—how rational agents choose

• Neuroscience: McCulloch-Pitts neurons (1943), brain-inspired computation

• Control theory: Feedback, stability, optimal control

• Linguistics: Grammar, semantics, natural language understanding

• 1943: McCulloch & Pitts — artificial neurons

• 1950: Turing — “Computing Machinery and Intelligence”

• 1956: Dartmouth Conference — “Artificial Intelligence” coined

• Founders: McCarthy, Minsky, Shannon, Newell, Simon

Early Enthusiasm (1952–1969)¶

• Logic Theorist (1956): First AI program, proved theorems

• General Problem Solver (1957): Means-ends analysis

• Lisp (1958): AI programming language

• Microworlds: Blocks world, limited domains

A Dose of Reality (1966–1973)¶

• Machine translation failed to deliver

• Combinatorial explosion: Search spaces too large

• Knowledge bottleneck: Programs lacked common sense

• Lighthill Report (1973): Critical of AI progress

AI Winters: The Pattern¶

• Boom: Enthusiasm, bold claims, funding flows

• Bust: Overpromise, failure to deliver, funding cuts

• First winter (1974–1980): Lighthill, machine translation, knowledge bottleneck

• Second winter (1987–1993): Expert systems limits, neural net hype

• Lesson: Manage expectations; incremental progress is real

• DENDRAL: Chemical analysis

• MYCIN: Medical diagnosis

• Knowledge engineering: Encode expert knowledge as rules

• Commercial success in narrow domains

The Return of Neural Networks (1986–present)¶

• Backpropagation (Rumelhart et al.): Training multilayer networks

• Connectionism: Distributed representations

• Limitations: Data and compute requirements

• Second AI Winter (1987–1993): Funding dried up again

Probabilistic Reasoning & ML (1987–present)¶

• Bayesian networks: Uncertainty representation

• Hidden Markov models: Speech recognition

• Machine learning: Learn from data rather than program

From Symbolic to Statistical¶

• Symbolic AI (1956–1980s): Logic, rules, explicit knowledge representation

• Statistical AI (1990s–): Learn from data, probabilistic reasoning

• Hybrid: Modern systems combine both—neural nets with symbolic reasoning

• Shift: From “program the knowledge” to “learn from examples”

• Scale: Millions of examples

• Web data: Text, images, user behavior

• Storage and processing: Cloud, distributed systems

Deep Learning (2011–present)¶

• ImageNet (2012): AlexNet breakthrough

• Deep networks: Many layers, automatic feature learning

• Applications: Vision, speech, NLP, games

• Transformers (2017): Attention mechanism

Examples of AI in Practice¶

• Recommendation: Netflix, Spotify, YouTube—what you’ll like next

• Search: Google, Bing—finding the right information

• Assistants: Siri, Alexa, ChatGPT—conversation, task completion

• Medical: Diagnosis from images, drug discovery, personalized treatment

• Autonomous: Self-driving (partial), drones, warehouse robots

The State of the Art¶

• Game playing: Chess, Go, poker — superhuman

• Natural language: Translation, question answering, generation

• Perception: Image recognition, speech

• Robotics: Manipulation, autonomous vehicles (hardest—physical world)

Current Challenges and Open Problems¶

• Robustness: Systems fail on out-of-distribution inputs

• Interpretability: Why did the model decide that?

• Efficiency: Huge compute requirements; can we do better?

• Generalization: Transfer learning, few-shot, zero-shot

• Safety and alignment: Value alignment, robust beneficial AI

• Healthcare: Diagnosis, drug discovery, personalized treatment

• Education: Adaptive tutoring, accessibility tools

• Scientific discovery: Protein folding, materials, climate modeling

• Automation: Dangerous or tedious tasks

• Assistive: Personal assistants, accessibility for people who need it

Risks of AI¶

• Jobs: Displacement, augmentation, which tasks are affected?

• Bias: Amplifying inequalities in hiring, loans, criminal justice

• Safety: Value alignment, do systems do what we intend?

• Misuse: Deepfakes, autonomous weapons, misinformation at scale

Summary: Key Takeaways¶

1. AI = building rational agents that act optimally

2. Multiple foundations: philosophy, math, neuroscience, etc.

3. History: cycles of enthusiasm and “AI winters”

4. Current: Deep learning, big data, broad applications

5. Future: Consider benefits and risks carefully

What’s in This Course¶

• Part I: Intelligent agents (Ch. 2)

• Part II: Problem-solving—search, games, constraints (Ch. 3–6)

• Part III: Knowledge and reasoning—logic, planning (Ch. 7–11)

• Part IV: Uncertainty—probability, decision making (Ch. 12–18)

• Part V: Machine learning—from examples to deep learning (Ch. 19–22)

• Part VI: Perception and action—NLP, vision, robotics (Ch. 23–26)

• Part VII: Philosophy, ethics, future (Ch. 27–28)

How to Use This Course¶

• Read the chapter before or after the lecture

• Do the exercises—that’s where learning happens

• Implement algorithms yourself; aima-python as reference

• Connect ideas across chapters—they build on each other

• Ask questions; the field rewards curiosity

• Russell & Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Ch. 1

• AIMA website

• Exercises

Questions?¶

Next lecture: Intelligent Agents (Chapter 2)