What AI Is

This is Unit One of Artificial Intelligence: what the field actually is. The definition this course uses is deliberately narrow, and what it leaves out is the interesting part. It says nothing about brains, nothing about consciousness, and nothing about what a system is made of. It describes systems that decide what to do based on what they can sense. That silence lets the field study its subject without settling questions nobody has settled. And it comes with a working standard: a system acts intelligently when its actions fit its circumstances and goals. The standard tests doing rather than experiencing, and the course says so rather than pretending otherwise. The unit also covers the field's history, and there is a pattern in it worth carrying. Twice the field promised far more than it delivered and funding collapsed. Those episodes are the base rate behind every confident forecast you will ever read about this subject. By the end you will be able to do five things. One: state the definition, and what it deliberately omits. Two: state the reflex that keeps moving the target. Three: say what passing the famous test would establish. Four: say why competence belongs to the agent and its environment together. Five: state bounded rationality, and why it is not a lowered standard.

Three entries fixing the subject. A definition, the thing it names, and a standard for judging one — and all three are written to avoid taking positions on questions the field cannot settle.

The definition.

AI 0001 What Artificial Intelligence Is

The field that builds and studies systems deciding what to do based on what they can sense. Brains, consciousness and silicon go unmentioned — and that silence is a choice rather than an oversight.

The object.

AI 0002 The Computational Agent

One whose decisions break into small mechanical steps a physical device could carry out. Mechanical sets no upper limit on sophistication — chess engines and language models both qualify.

The standard.

AI 0003 The Behavioral Standard

Acting intelligently means acting in a way that fits the circumstances and the goals, within the limits of what can be sensed and computed. It is offered as a working standard, and it tests doing rather than experience.

Two entries. One names a habit that has quietly redefined the subject for seventy years, and the other fixes a vocabulary problem the press creates daily.

The reflex.

AI 0004 The AI Effect

Every task a machine masters gets reclassified as not really intelligence. Run for seventy years, that reflex makes the word mean whatever machines have not managed yet — which is a moving definition.

The two names.

AI 0005 Machine Learning Is Part of AI

Learning from examples is one subfield of a larger field that includes much else. Headlines use the two names interchangeably, and this course does not.

Two entries on the field's most famous proposal. What it asks, what its author actually predicted — which was far more modest than the legend — and how narrow a conclusion passing it would support.

The test.

AI 0007 The Turing Test

Whether an interrogator who cannot see the respondents can reliably tell a machine from a person by written questions. The original prediction was modest — fooling an average questioner some of the time.

What passing shows.

AI 0008 The Test Is Not a Certificate

Skill at conversational imitation — one ability, not a certificate covering the rest. And the reverse holds too: failing the test rules out about as little as passing rules in.

Two entries on the gap between talking well and knowing anything — and the second contains the sharper observation, which is about where the impression of understanding actually comes from.

The gap.

AI 0010 Fluency Without Understanding

A system can produce fluent, appropriate language with no grasp of what the words refer to. When people attributed understanding to early systems, the attribution was the listener's contribution rather than the machine's achievement.

And what follows.

AI 0011 "It Talks Like It Understands"

Appropriate language can be produced by pattern manipulation at every scale, so fluency alone certifies nothing about understanding. Crediting a system with a mind because its sentences are smooth is a move you are making, not one it earned.

Four entries of history, and the last two are the ones with predictive value. The field has a signature failure mode — a correct long-range vision attached to a wildly wrong timeline — and it has produced two collapses.

The founding bet.

AI 0012 The Dartmouth Bet

A written wager that every feature of intelligence can be described precisely enough for a machine to simulate. The bet is about describability, not about any deadline.

The first successes.

AI 0013 Early Wins in Small Worlds

Programs that reasoned, played and learned, each inside a small tidy world. Every clean-world success invited the same extrapolation — that a scaled-up version must follow.

The pattern.

AI 0014 The Overreach Pattern

The field's signature failure is a correct long-range vision priced at a wildly optimistic schedule. A confident short timeline about this subject belongs to a genre with a long record.

And the consequences.

AI 0015 AI Winters

Twice — in the mid-nineteen-seventies and the late nineteen-eighties — promises outran delivery and both money and confidence withdrew. Those episodes are the base rate behind every forecast in this field.

Two entries on where the field went. What now dominates it, and two famous machines that beat human champions — using approaches different enough that comparing them teaches more than either does alone.

The turn.

AI 0016 The Learning Turn

Systems learning from enormous collections of examples now dominate. The costs of that approach — appetite for data, difficulty seeing inside — are part of the content rather than caveats.

Two paradigms.

AI 0017 Deep Blue and AlphaGo: Two Paradigms

One machine won at chess with massive search and human-written evaluation and no learning at all. The other won at go by learning. Neither examined everything — both were selective, and what did the selecting is the difference.

Four entries introducing the framework the rest of the course runs on. The key move is treating an agent and its surroundings as a single system — which means competence is never a property of the agent alone.

One system.

AI 0021 Agent and Environment: One System

The environment delivers stimuli and the agent delivers actions back. Competence belongs to the pair — the same vacuum cleaner thrives in one house and fails in another without changing at all.

What an agent is not.

AI 0022 "AI Agents Are Robots"

A robot is one kind of agent — the kind with motors. Most agents have no body whatever, and restricting the word to things that move is like restricting musician to drummers.

What comes in.

AI 0023 Stimuli: What Comes In

Anything the environment delivers — an observation picked up, or an action imposed. What an agent cannot sense sets the boundary of its knowable world.

And what it carries.

AI 0024 The Belief State

Its running record of the world, what it has learned, and what it wants. Without one, an agent is purely reactive — the same stimulus produces the same act every time, like a doorbell.

Four entries, and three questions that between them decide what kind of problem an agent is facing. Can it see everything? Does the same action always do the same thing? And is anything else in there making decisions?

The first question.

AI 0025 Observability

Fully observable when the stimuli reveal everything relevant, partially observable otherwise. Where sight fails, an agent has to act on what it believes rather than on what it sees — and belief becomes the working material.

And the fallacy.

AI 0026 The Chess-World Fallacy

Treating game-like fully visible worlds as normal is a mistake — almost every environment worth acting in hides most of itself. Ask of any demonstration whether the system could see everything.

The second question.

AI 0027 Deterministic and Stochastic

Deterministic when an action has exactly one outcome, stochastic when the same action can lead to different ones. Acting well under spread means comparing spreads rather than lucky cases.

The third.

AI 0028 Other Minds

An environment containing other decision-makers requires reasoning about their choices. Weather is unpredictable and indifferent; an opponent is unpredictable and studying you — a different problem entirely.

Two entries closing the unit. A space of design choices along ten dimensions, and then the standard an agent operating in the real world is actually held to — which is not perfection.

The space.

AI 0029 The Design Space

Designs differ along ten dimensions — how much is visible, how uncertain effects are, whether others are present, whether it learns, how far ahead it plans. Excellence in one region transfers poorly a few dimensions away.

And the standard.

AI 0030 Bounded Rationality

Choosing the best action available given its goals, its stimuli, and its computational limits. Perfection would need unlimited time and information — and bounded is not sloppy, since choosing good shortcuts is itself a decision made well.

Five things you can now do. You can state the field's definition and what it deliberately omits — brains, consciousness, and what a system is made of — and why that silence is useful. You can state the reflex that reclassifies every mastered task as not really intelligence, and say what it produces over seventy years: a moving target. You can say what passing the famous test would show — skill at one kind of imitation — and that fluency certifies nothing about understanding. You can say why competence belongs to the agent and environment together, since the same system thrives in one setting and fails in another unchanged. And you can state bounded rationality: the best available action given real limits, where choosing a good shortcut is itself choosing well.