Skip to content

1 Introduction to AI

Introduction to AI

Instructors

  • Marwa Abdulhai — PhD @ Berkeley, MS + UG @ MIT
  • Ezinne Nwankwo — PhD @ Berkeley, Lecturer in EECS; BA + MS @ Harvard & Duke

NOTE! Do not cheat. Do not miss discussions.

AI's Real-World Impact

AI is touching everything — the economy, politics, law, labor, sciences, education, public imagination:

  • AlphaFold predicts protein 3D structure from sequence
  • DeepMind trained an agent to control nuclear fusion

On writing: LLM edits shift the meaning and tone of essays, potentially homogenizing our voices and eroding individual writing fingerprints.

What Should Machines Do?

Think Act
Like people Cognitive modeling Turing test framing
Rationally Laws of thought, logic This course

Rational here means a specific thing:

  • Goals expressed as utility of outcomes
  • Maximize utility toward pre-defined goals
  • World is uncertain → use expected utility

Perspectives on Intelligence

Skills-based — "Intelligence is a collection of skills" (play chess, use language, learn from experience)

Embodiment (Rodney Brooks) — You don't need to play chess. Intelligence requires a physical body existing in an environment.

Psychometrics (François Chollet) — Measure abilities, not skills, across a broad range of tasks — including tasks previously unknown to the system and its developers.

Human-compatible (Stuart Russell) — Maximizing utility isn't enough. Maximize human utility. Observe human behavior, infer preferences over time, learn reward functions from principals.

"A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying…"

Specialization is for insects.

Brain vs. Machine

  • Brains make good rational decisions but aren't perfect and can't be fully reverse-engineered
  • AI can outperform brains at specific tasks (e.g. math)
  • Brains are to intelligence as wings are to flight — great inspiration, not the only solution
  • Human brain: ~100T synapses | GPT-4: ~1.8T parameters

Brief History of AI

1940–50 — McCulloch's perceptron Boolean circuit model (1943). Turing proposes the Turing test.

1950–70 — Samuel's checkers program, Logic Theorist, geometry engine. Dartmouth meeting: "AI" name coined. They planned to solve AI in two months.

1970–90 — Perceptrons can't learn XOR. Multi-layer networks + backpropagation were invented. Expert Systems industry busts → first AI winter.

1990–2010 — Statistical approaches, resurgence of probability and uncertainty. TD-Gammon reaches human-level Backgammon. Deep Blue defeats Kasparov.

2010–17 — Siri. AlexNet wins ImageNet. DeepMind's DQN masters Atari. AlphaGo defeats Lee Sedol. Google Translate goes neural.

2017– — Google invents the Transformer. DeepStack/Libratus defeats poker champs. AlphaFold predicts protein structure. Scaling era begins.

Rational Agents

An agent perceives its environment through sensors and acts through actuators. The environment's properties largely determine agent design.

Sensors  <──── percepts ──── Environment
   ↓
Condition-action rules
   ↓
Actuators ───── actions ───> Environment

Environment types:

  • Partially observable — agent needs internal memory
  • Stochastic — must plan for uncertainty and contingencies
  • Multi-agent — may need randomized behavior
  • Static — agent has time to compute a rational decision
  • Continuous — agent operates as a continuous controller
  • Unknown physics — must explore
  • Unknown preferences — must observe/interact with a human principal

Course Topics

  1. Search and planning
  2. Probability and inference
  3. Supervised learning
  4. Reinforcement learning

plus applications and impact on science, technology, and society.