AIMA Chapter 1 — Introduction to AI.

If Chapter 2 teaches you how to look at an agent, Chapter 1 teaches you how to look at the whole field. It is the chapter that says, “before we get lost in algorithms, what game are we even playing?” That question matters more than it first appears. People say “AI” when they mean chatbots, robots, exam cheating, productivity tools, machine learning, science fiction, surveillance, autocomplete, or automation. Russell and Norvig start more carefully. They ask what counts as intelligence, what counts as good behaviour, and whether we should judge a machine by human resemblance or by rational performance.

01 Why the term “AI” feels slippery in the first place

Artificial intelligence is one of those words that everybody recognises and almost nobody uses in exactly the same way. In casual conversation, AI can mean “a machine that talks like a person,” “software that predicts things from data,” “a robot that moves on its own,” or simply “something complicated with neural networks.” The problem is that each of those examples captures only one slice of the field. A chess engine is AI, but it does not look human. A self-driving car is AI, but it is not just “machine learning on wheels.” A theorem prover, a medical diagnosis system, and a route-planning algorithm can all qualify, even though they operate in totally different ways.

That is why AIMA does not begin with one slogan definition. Instead, it surveys the major ways people have tried to define the field. This is actually a kindness to the student. It tells you that the confusion is not because you missed a lecture; the field itself grew from different intellectual traditions. Some researchers cared about modelling human thought. Some cared about building useful behaviour. Some wanted logical correctness. Some wanted human-level imitation. The field inherited all of those ambitions, and the result is a family of definitions rather than one sentence that everyone fully agrees on.

So when you study AI, it helps to ask two separate questions every time. First: are we judging the system by what is going on inside its head, or by the action we can observe from outside? Second: are we comparing the system to how humans behave, or to how an ideal rational agent should behave? Those two questions create the famous AIMA 2×2 grid, and that grid is still the cleanest map of Chapter 1.

02 The four classic views of AI in one 2×2 grid

Why does the course settle on “act rationally”? Because the other three views are either too narrow or too fragile as a general engineering target. Acting humanly is interesting, but humans are not always a gold standard. We are inconsistent, forgetful, emotional, biased, and slow at arithmetic. Thinking humanly is useful if your goal is cognitive modelling, but many successful AI systems solve problems in a completely non-human way. Thinking rationally is attractive, but pure logic alone cannot carry the whole weight of intelligence because the world is uncertain, incomplete, noisy, time-bounded, and computationally brutal.

The rational-agent lens is also more future-proof. A vacuum-cleaning robot, a chess engine, a protein-folding model, and a route planner do not need a shared internal mechanism to all count as AI. They only need to fit the same outer template: perceive something, evaluate options, and choose an action that tends to improve performance. That is a much more flexible umbrella for the field.

03 A very short history of AI, from Turing to transformers

Chapter 1 usually gives students their first timeline of the field, and that timeline matters because AI has never moved in a straight line. It moves in waves of optimism, technical progress, disappointment, and reinvention. In 1950, Alan Turing published “Computing Machinery and Intelligence,” asking the now-famous question: can machines think? Turing immediately did something clever. He avoided endless philosophy about the word “think” and proposed the imitation game instead. If a machine can converse well enough that a human judge cannot reliably tell it from a person, then calling it intelligent becomes at least operationally defensible.

The 1950s were full of optimism. Researchers discovered that symbolic manipulation on digital computers could express reasoning steps. LISP, introduced by John McCarthy in 1958, became the iconic early AI language because it was built for symbolic expressions and recursion rather than raw numeric throughput. Programs began solving algebra word problems, proving theorems, and playing simple games. The mood in early AI was so confident that some researchers predicted human-level machine intelligence within a couple of decades.

Then reality punched back. Early systems worked well in toy environments but scaled badly. Search spaces exploded. Representation was brittle. Commonsense knowledge turned out to be enormous and painfully hard to encode. This mismatch between laboratory demos and messy real intelligence helped drive the first AI winter in the 1970s: funding cooled, hype collapsed, and many grand promises looked naive in hindsight.

The 1980s brought expert systems. These were knowledge-heavy programs that used large rule bases like “if symptom A and lab value B, then consider diagnosis C.” In restricted domains, expert systems worked well enough to create real commercial excitement. They showed that explicit knowledge representation could create value, especially when the world was stable and the expertise could be written down. But they were also expensive to build and maintain. Rules broke when the domain changed. Uncertainty handling was awkward. Knowledge acquisition became a bottleneck. When expectations outpaced delivery again, a second AI winter followed in the late 1980s and early 1990s.

The next major turn came from data-driven methods. Instead of manually encoding every rule, researchers increasingly built systems that learned patterns from examples. This is the machine learning revolution. Better algorithms, more data, and faster hardware made statistical methods practical. Decision trees, support vector machines, probabilistic models, and ensemble methods all helped AI move from brittle handcrafted logic to adaptable pattern learning.

Then came the deep learning era. Neural networks had existed for decades, but around the 2010s, big datasets, GPUs, improved optimisation, and layered architectures changed what was feasible. Suddenly image recognition got dramatically better. Speech recognition became useful at scale. Translation improved. Games once thought out of reach became tractable. A famous moment was AlphaGo defeating Lee Sedol in 2016, because Go had long been treated as a symbolic marker of human strategic depth.

The transformer wave pushed things again. Attention-based architectures transformed natural language processing, then spilled into vision, multimodal learning, code generation, biology, and scientific modelling. By 2023 and after, large language models had made AI feel personal and public. People who had never heard of sequence modelling now used chat interfaces for writing, coding, brainstorming, summarising, and tutoring. That does not mean LLMs solved intelligence in the grand sense. It means the user interface of AI became mainstream.

Period	Milestone	Why it matters
1950	Turing's paper	Gives the field its first widely cited framing question and the imitation-game benchmark.
1958	LISP	Becomes the symbolic language of early AI research.
1970s	First AI winter	Funding retreats after early optimism runs into scale and complexity limits.
1980s	Expert systems	Knowledge-based AI becomes commercially useful in narrow domains.
Late 1980s–1990s	Second AI winter	Rule-heavy systems prove brittle and expensive to maintain.
1990s–2000s	Machine learning rise	Data-driven methods increasingly outperform handcrafted rules in many tasks.
2010s	Deep learning era	Neural networks scale with data and compute, reshaping vision, speech, and language.
2023+	Transformers / LLMs	Foundation models make AI conversational, multimodal, and socially unavoidable.

The clean exam-friendly summary is this: the field starts with foundational questions, grows through symbolic reasoning, crashes when scale bites, re-emerges through expert knowledge, crashes again when brittleness bites, then explodes through learning, data, compute, and finally foundation models. If you remember the rhythm of optimism → capability → limitation → reinvention, you understand AI history better than someone who only memorised dates.

04 What the current state of the art makes possible

Students often enter Chapter 1 with “AI = ChatGPT” in their heads. AIMA tries to widen that picture. Modern AI is not one application; it is a toolbox deployed across very different tasks. In computer vision, image recognition models can classify objects, detect tumors, segment roads for autonomous driving, and inspect industrial defects at scale. In biology, systems such as AlphaFold changed protein structure prediction from a long-standing bottleneck into a much faster computational pipeline. That does not mean biology is solved, but it shows how domain knowledge plus machine learning can shift an entire scientific workflow.

In games, AlphaGo and later systems matter not because winning games is the ultimate purpose of AI, but because games give us precise rules, measurable outcomes, and hard search spaces. They are laboratories for planning, evaluation, self-play, and learning. In language, LLMs like ChatGPT changed public expectations by making AI interactive. People now experience search, summarisation, translation, tutoring, and code assistance through natural-language interfaces instead of specialist software menus.

Robotics is another frontier because it forces AI to leave the clean world of text and pixels and collide with friction, delay, uncertainty, and physical risk. A robot does not get to hallucinate a box into existence. It must perceive, model, and act under constraints. Self-driving systems are similarly hard: they combine perception, prediction, planning, uncertainty handling, and real-time control in dynamic environments full of humans and edge cases.

At the same time, state of the art is not magic. Strong results often depend on narrow benchmarks, huge compute budgets, giant datasets, and very careful system design. A model can be spectacular on average and still brittle on the cases that matter most. A chapter about AI should therefore excite you and sober you at the same time.

05 The big questions AI is trying to answer

Underneath all the applications, AI keeps returning to a set of deep questions. How should an intelligent system represent the world? How should it reason under uncertainty? How should it choose among many possible action sequences? How should it learn from limited experience? How can perception become useful action instead of just passive classification? How can multiple agents compete or cooperate? How should a machine balance speed against correctness, exploration against exploitation, or safety against ambition?

These questions sound abstract, but each maps directly onto technical subfields. “How should a system choose among many possible futures?” becomes search and planning. “How should a system represent facts, categories, rules, and relations?” becomes knowledge representation and reasoning. “How should a system improve from data?” becomes machine learning. “How should a system act when it cannot observe everything?” becomes probabilistic reasoning, filtering, and decision-making under uncertainty.

There are also philosophical questions hiding in the engineering. What counts as understanding? Is prediction enough, or do we want explanation too? How much intelligence requires embodiment? Can a system be rational without being conscious? Chapter 1 does not solve those debates, but it gives you enough structure that you can talk about them without turning every discussion into science fiction fog.

06 What rationality really means — and what it does not mean

Students often hear “rational agent” and silently translate it to “perfect agent.” That is wrong, and Chapter 1 is very careful about it. Rationality does not mean omniscience. A rational agent does not know the future. It does not mean perfection either. The agent cannot choose the perfect move using information it does not have or computation it cannot afford. Rationality means doing the best expected thing given the percepts so far, the prior knowledge available, and the computational resources in hand.

That “expected” word matters a lot. If I carry an umbrella because the forecast says 80% rain and then the sun comes out all day, the action can still have been rational. Rationality is judged at decision time, not with unfair hindsight. This is the same reason a medical diagnosis system can be rational even if a rare disease fools it, or a self-driving car can be rational even if another driver behaves unpredictably. Rationality is about sensible action under uncertainty, not guaranteed success.

Another subtlety: rationality depends on the performance measure. If the measure rewards shortest travel time only, a taxi may drive aggressively. If the measure rewards safety, legality, comfort, and fuel economy, the rational behaviour changes. “Good action” is always relative to a task specification. There is no context-free rational action floating in the air.

This is also where bounded rationality enters the conversation. In a perfect mathematical world, the ideal rational action might be defined by an exhaustive calculation. In the real world, the agent is bounded by memory, time, sensing quality, and modelling quality. So practical AI often tries to be rational enough under constraints rather than perfectly optimal under fantasy assumptions.

Once you understand this, many later chapters become easier. Search algorithms, heuristics, game-playing methods, and learning systems are all attempts to approximate rational action under different assumptions and constraints.

07 The agent perspective: perceive, decide, act

Chapter 1 starts hinting at a unifying idea that Chapter 2 will formalise properly: an agent is something that perceives its environment and acts upon that environment. That sentence looks almost too simple, but it is incredibly powerful. It lets us talk about very different systems using the same template. A thermostat, a chess engine, a robot vacuum, a stock-trading bot, and a conversational assistant can all be described in agent language. They differ in sensors, actuators, environments, and internal programs, but they share the same basic loop.

The beauty of the agent perspective is that it connects perception to consequence. A classifier by itself is not yet an agent. It becomes part of an agent when its output influences action. This is why Chapter 1 is not just a motivational chapter. It plants the organising idea that later chapters keep expanding: design an entity that maps incoming evidence to useful behaviour.

08 Why AI is hard even when the idea sounds simple

The slogan version of AI sounds suspiciously manageable: gather information, choose a good action. So why is the field so hard? First, because the number of possible situations can explode. Even a modest board game can create more possible states than brute force can touch. Real-world planning problems are worse. Second, environments are noisy and only partly visible. A camera image does not reveal everything. A sensor can fail. A user can type an ambiguous question. A driver can behave erratically. The agent often has to act before it fully understands the situation.

Third, goals can conflict. A delivery drone wants speed, safety, battery efficiency, legal compliance, and robust navigation at the same time. Optimising one objective too hard can damage another. Fourth, the world changes while the agent is thinking. In a dynamic environment, “wait until you know everything” is not a strategy; it is how you lose.

There is also the old enemy of computational complexity. Many interesting AI problems are not impossible in principle; they are impossible at the required scale and speed with naive methods. That is why heuristics, abstractions, approximations, and domain knowledge matter so much. AI is often the art of getting enough intelligence without paying the full impossible cost of exact optimal reasoning.

09 The three big branches you should keep in your head

Chapter 1 previews several areas, but for this course it helps to keep three major branches mentally separated. The first is search and planning. This branch asks: given a model of actions and outcomes, how do we find a sequence of actions that reaches a goal? Route finding, puzzle solving, game search, planning in robotics, and decision trees in action spaces all live here. Chapters 3, 4, and 5 lean strongly into this branch.

The second is knowledge representation and reasoning. This branch asks: how should an intelligent system store facts, categories, relationships, constraints, and rules so that it can infer new conclusions? Logic, ontologies, rule systems, semantic networks, and probabilistic knowledge structures belong here. Even when your course does not go deep into all of them, this branch matters because it addresses a big pain point: how to turn raw information into structured, queryable understanding.

The third is machine learning. This branch asks: how can a system improve from data instead of depending entirely on manual programming? Classification, regression, clustering, reinforcement learning, representation learning, neural networks, and large language models fit under this umbrella. Modern public excitement around AI mostly sits here, but AIMA reminds you that learning is not the entire field.

Good modern systems often blend these branches. AlphaGo used deep learning and search. Medical systems combine learned prediction with explicit workflows and safety rules. Code assistants mix language modelling, retrieval, and tool use. So the branches are conceptual buckets, not isolated kingdoms.

10 Where AIMLCZG557 sits inside all this

This course does something pedagogically smart: it does not try to cover the entire modern AI universe in one sweep. Instead, it gives you a strong base in classical AI through AIMA Chapters 1 to 5 and then connects that base to evolutionary and swarm-based computational intelligence. That means you do not begin with “train a giant model and hope.” You begin with agents, rationality, problem formulation, uninformed and informed search, local search, game playing, and related classical ideas that teach you how intelligent behaviour can be engineered step by step.

Then the course widens into computational intelligence: evolutionary computation, ant colony optimisation, and related nature-inspired methods. This is a useful pairing. Classical AI teaches explicit search and reasoning structure. Computational intelligence teaches adaptation, stochastic optimisation, and population-based problem solving. Together they give you a more complete mental picture than a single narrow “AI = deep learning” course would.

So Chapter 1 is not just introductory fluff. It tells you why the rest of the syllabus is organised the way it is. You are about to study agents, search, heuristics, adversarial reasoning, and optimisation as different attempts to produce rational action in hard environments.

11 FAQs students usually carry after Chapter 1

Is AI the same thing as machine learning?

No. Machine learning is one major branch of AI, not the whole field. AI also includes search, planning, knowledge representation, reasoning, robotics, and decision-making under uncertainty. In public conversation, people often compress the entire field into “AI = ML,” but academically that is too narrow.

If rationality is the goal, why not just do perfect optimisation everywhere?

Because the world is too large, too uncertain, and too time-sensitive. Exact optimisation is often computationally unaffordable or requires information the agent does not have. Practical AI uses approximations, heuristics, and bounded reasoning.

Did modern LLMs make classical AI irrelevant?

No. LLMs changed the interface and widened the application space, but they did not erase search, planning, reasoning, or optimisation. In fact, many powerful systems combine foundation models with retrieval, planning, tools, constraints, or classical search.

Why does AIMA prefer “acting rationally” over “thinking like humans”?

Because the engineering goal becomes clearer and broader. Human imitation is interesting but not necessary for every AI system. A route planner or protein predictor can be excellent without reasoning like a human mind.

12 Takeaways before you move to intelligent agents

• AI has four classic definitions, but this course mainly adopts the acting rationally view.
• Rationality means choosing the best expected action from available percepts, knowledge, and compute — not being all-knowing.
• AI history moves in waves: early symbolic optimism, AI winters, expert systems, ML, deep learning, and the transformer era.
• Modern AI applications are broad: language, vision, science, games, robotics, and autonomy.
• The field's recurring hard problems are representation, search, uncertainty, complexity, and learning.
• For AIMLCZG557, Chapters 1–5 give the classical AI core before the course moves into evolutionary and swarm CI.