Russell and Norvig organise the definitions of AI along TWO DIMENSIONS, giving four quadrants. This 2×2 is the standard opening question of the subject. DIMENSION 1: THOUGHT vs BEHAVIOUR is the goal to replicate the internal REASONING PROCESS, or only the external ACTIONS? DIMENSION 2: HUMAN vs RATIONAL is success measured against HUMAN performance, or against an ideal standard of RATIONALITY? HUMAN RATIONAL ┌────────────────────┬────────────────────┐ THOUGHT │ Systems that THINK │ Systems that THINK │ PROCESSES │ like humans │ RATIONALLY │ │ (cognitive │ (laws of thought) │ │ modelling) │ │ ├────────────────────┼────────────────────┤ BEHAVIOUR │ Systems that ACT │ Systems that ACT │ │ like humans │ RATIONALLY │ │ (the Turing test) │ (rational agent) │ └────────────────────┴────────────────────┘ 1. ACTING HUMANLY — THE TURING TEST APPROACH Turing (1950) proposed the IMITATION GAME: a human interrogator converses by text with a human and a machine; if the interrogator cannot reliably tell which is which, the machine passes. THE CAPABILITIES IT REQUIRES: natural language processing · knowledge representation · automated reasoning · machine learning (and for the TOTAL Turing test, adding video and physical objects: computer vision and robotics) THE OBJECTION: it defines intelligence by IMITATION, which rewards deception. It also sets an anthropocentric target — aeronautical engineering does not aim to fool pigeons. 2. THINKING HUMANLY — THE COGNITIVE MODELLING APPROACH To say a program thinks like a human, you must first know how humans think. Three routes to that knowledge: INTROSPECTION — catching our own thoughts as they pass PSYCHOLOGICAL EXPERIMENTS — observing a person in action BRAIN IMAGING — observing the brain in action COGNITIVE SCIENCE is the interdisciplinary field this produces. The test is not "does it get the right answer?" but "does its reasoning trace resemble a human's, including the mistakes?" 3. THINKING RATIONALLY — THE LAWS OF THOUGHT APPROACH Aristotle's SYLLOGISMS gave patterns of argument that always yield correct conclusions from correct premises: Socrates is a man; all men are mortal; therefore Socrates is mortal. This became formal LOGIC, and the LOGICIST tradition in AI: represent knowledge as logical statements and derive conclusions. TWO OBSTACLES: · informal knowledge is hard to state with the certainty logical notation demands — most things we know are probable, not certain · a problem solvable in principle may be intractable in practice, because the search space explodes 4. ACTING RATIONALLY — THE RATIONAL AGENT APPROACH An AGENT perceives and acts. A RATIONAL AGENT acts so as to ACHIEVE THE BEST OUTCOME, or, under uncertainty, the best EXPECTED outcome. THIS IS THE DEFINITION RUSSELL AND NORVIG ADOPT, and modern AI generally follows them. Its advantages: · MORE GENERAL than the laws-of-thought approach: correct inference is one way to be rational, but not the only one — a reflex that works needs no inference at all · MORE AMENABLE TO SCIENCE than the human-based approaches, because rationality is mathematically specifiable while "human-like" is not · it does not require the agent to think at all, only to do the right thing THE DISTINCTION THAT RESOLVES MOST CONFUSION: STRONG AI the claim that a machine can genuinely be conscious and understand — a philosophical position WEAK AI the claim that a machine can act as if intelligent, producing useful behaviour — an engineering position Essentially all working AI is weak AI, and almost all public argument is about strong AI. Searle's CHINESE ROOM argument targets strong AI specifically: a person following symbol- manipulation rules could produce fluent Chinese without understanding a word, so behaviour alone does not establish understanding.

WHAT RATIONALITY DOES AND DOES NOT MEAN — a frequently examined distinction: RATIONALITY IS NOT OMNISCIENCE. An omniscient agent knows the actual outcome of its actions; no real agent does. Rationality is about maximising EXPECTED performance given what is known. Crossing a road after looking both ways is rational even if a cargo door falls from a passing aeroplane and flattens you. The action was right; the outcome was unlucky. RATIONALITY IS NOT PERFECTION. Perfection maximises actual performance; rationality maximises expected performance. RATIONALITY DEPENDS ON FOUR THINGS: 1. the PERFORMANCE MEASURE defining the criterion of success 2. the agent's PRIOR KNOWLEDGE of the environment 3. the ACTIONS the agent can perform 4. the agent's PERCEPT SEQUENCE to date INFORMATION GATHERING is part of rationality: doing actions in order to MODIFY FUTURE PERCEPTS — looking before crossing — is rational, which is why a rational agent is not merely a calculator. WHY AI PROBLEMS ARE HARD — the two properties that distinguish them from ordinary computation: 1. COMBINATORIAL EXPLOSION. The space of possibilities grows faster than any exhaustive method can search. WORKED — the 8-queens problem, three ways of counting: place 8 queens anywhere on 64 squares C(64,8) = 4,426,165,368 arrangements one queen per column 8^8 = 16,777,216 one per column AND one per row (a permutation) 8! = 40,320 THE SAME PROBLEM, and a 110,000-fold reduction obtained purely by reformulating what counts as a candidate. THIS IS THE CENTRAL SKILL OF AI PROBLEM SOLVING: the representation determines the difficulty far more than the algorithm does. 2. KNOWLEDGE-INTENSIVENESS. The behaviour depends on large amounts of knowledge about a domain, much of it never written down because humans consider it obvious — the DOMAIN REQUIREMENTS problem from software engineering, at scale. "Put the book on the table" requires knowing that books are rigid, tables are horizontal, gravity exists, and the book must not be placed on the cat. THE MORAVEC PARADOX, worth knowing because it is counter-intuitive and keeps being rediscovered: TASKS HUMANS FIND HARD (chess, integration, medical diagnosis) TURNED OUT TO BE EASIER FOR MACHINES THAN TASKS HUMANS FIND EFFORTLESS (walking over uneven ground, recognising a face, understanding a joke). The explanation is evolutionary: the effortless skills have had hundreds of millions of years of optimisation and are implemented in enormous specialised neural hardware we have no introspective access to, while chess is a few thousand years old and consciously reasoned. WHAT FEELS SIMPLE IS NOT SIMPLE.

THE SUBFIELDS, and what each is responsible for: SEARCH AND PROBLEM SOLVING finding a sequence of actions reaching a goal (ACtE0902) KNOWLEDGE REPRESENTATION encoding what is known so it AND REASONING can be used (ACtE0903) PLANNING constructing action sequences under constraints MACHINE LEARNING improving performance from experience (ACtE0905) NATURAL LANGUAGE PROCESSING understanding and generating language (ACtE0904) COMPUTER VISION interpreting images ROBOTICS acting in the physical world EXPERT SYSTEMS domain-specific reasoning from an explicit knowledge base TWO BROAD TRADITIONS, and the modern reconciliation: SYMBOLIC AI ("classical", or GOFAI — good old-fashioned AI) Intelligence as the manipulation of explicit symbols and rules. ✔ INTERPRETABLE — you can read the rules and see why it concluded what it did ✔ needs no data, only knowledge ✔ handles reasoning, planning and explanation naturally ✘ BRITTLE at the edges — a case the rules do not cover produces nonsense ✘ the KNOWLEDGE ACQUISITION BOTTLENECK: encoding enough knowledge by hand is enormously laborious CONNECTIONIST / STATISTICAL AI (neural networks, ACtE0906) Intelligence as patterns learned from data by adjusting numerical weights. ✔ learns from examples, so no hand-coded rules ✔ tolerant of noise and of cases never explicitly considered ✔ excels precisely where symbolic AI fails — perception ✘ OPAQUE: a 175-billion-parameter model cannot explain itself, and "why did it decide that?" may have no answer a human can check ✘ needs large amounts of data, and inherits whatever bias the data contains THE HONEST CURRENT POSITION: the statistical approach has won decisively on perception and language, while the symbolic approach retains the advantages of explanation, guarantees and data-free reasoning. NEURO-SYMBOLIC work attempts to combine them, and the fact that this remains an open research direction rather than a solved problem is worth knowing — the two traditions are not merely rival fashions, they have complementary strengths. A NOTE ON WHAT AI IS NOT: a program is not AI merely because it is complicated, nor because it uses statistics. The useful informal test is whether the program's behaviour depends on KNOWLEDGE OR EXPERIENCE rather than on a fixed procedure the programmer worked out in advance. A payroll system computing tax by a published formula is not AI; a system inferring which claims are likely fraudulent from past claims is.