PROLOGUE · BEFORE WEEK 1
CHAPTER 0

What Is AI?

Before Bota can travel, it has to understand what kind of creature it is.

Manga page: Astana. A tiny robot wakes up for the first time. / People call me artificial intelligence... but what AM I? / You're an AGENT! Let's ride to Almaty. I'll teach you AI on the way. / Let's go!

Four ways to define AI

Artificial Intelligence (AI) is the field that builds computer systems able to do tasks that normally need human intelligence: perceiving, reasoning, planning, learning and acting. Russell and Norvig, authors of the main AI textbook, sort definitions along two questions. Do we care about thinking or acting? Do we measure success against humans or against an ideal of rationality?

Like a human
Rationally
Thinking
Cognitive modelling. Build programs that think the way people do, and test them against psychology experiments.
Laws of thought. Use logic to derive correct conclusions, an idea that goes back to Aristotle.
Acting
Turing test. Behave so well that a human judge can't tell the machine from a person.
Rational agent ★ Act to achieve the best expected outcome. This book and our course use this view.
IN PLAIN WORDS

An AI is a program that looks at its situation, decides what to do, and acts to reach a goal as well as it can with what it knows. It doesn't have to think like a human. It just has to make good decisions.

The Turing test (1950)

Alan Turing asked: instead of "Can machines think?", ask "Can a machine behave in conversation so that a judge can't reliably tell it from a human?" Passing such a test would need natural-language processing, knowledge, reasoning and learning. Today's chatbots can sound very human. Yet researchers no longer treat "fooling people" as the main goal, because sounding smart is not the same as being correct or useful.

AI, Machine Learning, Deep Learning: which is which?

ARTIFICIAL INTELLIGENCE search · games · logic · planning · learning MACHINE LEARNING learns rules from data DEEP LEARNING many-layer neural nets LLMs

Many AI systems contain no learning at all. The A* route planner in Chapter 2 and the chess-style minimax player in Chapter 4 work from rules and search. Machine learning (ML) is the part of AI where the program improves from data. Deep learning is ML with many-layered neural networks (Chapter 9). Large language models such as ChatGPT sit inside deep learning (Chapter 10). Generative models in general can use deep learning or simpler statistics: Chapter 10's tiny bigram text generator just counts words.

Narrow AI vs General AI

Narrow AI is excellent at one task: a spam filter, a route planner, a Go player. Artificial General Intelligence (AGI) would match humans across almost all tasks. Every system that exists today, including the most impressive chatbots, is still evaluated task by task, and whether any of them is "general" is an open debate.

A short history of AI

Tap a year to open the story. Red dots are breakthroughs; blue dots are "AI winters", when funding and hype collapsed.

1956!

Dartmouth workshop.
John McCarthy's proposal gives the field its name.

AI is older than the internet?!

AI winters
Promises ran ahead of results, so twice the money dried up.

DID YOU KNOW?

The word "robot" comes from the Czech robota ("forced labour"). Karel Čapek used it in his 1920 play R.U.R., decades before any real AI program existed.

AI Myth vs Fact

Tap a card to flip it.

MYTH"AI understands the world like a human does."
FACTMost AI systems optimise a mathematical objective on a model of the problem. They can be brilliant inside that model and fail badly outside it.
MYTH"All AI is machine learning."
FACTSearch, game-tree reasoning, logic and planning are AI too, and they need no training data. Google Maps-style routing is mostly search.
MYTH"If the AI is 99% accurate, it is safe."
FACTIf only 1% of cases are dangerous, a model that always says "safe" is 99% accurate and useless. Chapter 8 explains precision and recall.
MINI-QUIZ

1. Which approach to AI does this course (and AIMA) mainly follow?

Rational agents choose the action with the best expected outcome given their evidence. It's measurable and doesn't require copying human thinking.

2. An A* route planner that uses no data at all is…

AI ⊃ ML ⊃ deep learning. Search algorithms are classic AI without learning.

3. What happened at Dartmouth in 1956?

McCarthy, Minsky, Shannon and Rochester proposed a summer workshop on "artificial intelligence". It's considered the birth of the field.
Manga page: So you are a rational agent. / Then how do I actually see... and choose?
0