AI Foundations & History

Key ideas

Weak vs strong AI. Weak (narrow) AI is built for one task or a narrow family of tasks (a spam filter, a chess engine, a chatbot). Strong AI or AGI would handle any intellectual task a human can; no such system exists yet¹.

Two approaches¹:

• Top-down (symbolic reasoning): extract knowledge and rules from experts and encode them (logic, expert systems). Struggles with tasks that resist explicit rules, such as guessing a person's age from a photo.
• Bottom-up (neural / connectionist): model the brain with networks of simple neurons that learn from examples. Deep learning is the branch of machine learning that stacks many such layers to learn representations.

Defining intelligence.

• The Turing test asks whether a judge chatting by text can tell the machine from a person. Criticisms: it tests human-likeness, not intelligence — a machine must also fake human weaknesses (typing errors), and a superhuman machine would have to hide its abilities; Russell and Norvig compare it to defining aeronautics as building machines that fool pigeons².
• Searle's Chinese room does not claim a program cannot answer like a native speaker; it claims that producing the right answers by symbol manipulation would not show understanding².
• The textbook definition used in most courses is acting rationally: an agent perceives its environment and chooses the action expected to maximise its performance measure, given its percepts and knowledge³. Rationality is judged by the expected outcome, not by luck: a well-informed decision that ends badly is still rational.

Agents and task environments. An agent senses through sensors and acts through actuators³; a PEAS description lists Performance measure, Environment, Actuators, Sensors (for an automated taxi: safe, fast, legal trips; roads and traffic; steering, brakes, horn; cameras, GPS, speedometer). Actuators are the agent's outputs, sensors its inputs. Environments are classified along these dimensions⁴:

A deterministic environment can still be uncertain if it is only partially observable.

Agent designs (Russell & Norvig)³:

Simple reflex agents fail in partially observable worlds (they can loop forever); adding memory (a model) fixes this³.

Formulas

The rational choice is the action with the highest expected performance:

Worked example

Classify an automated taxi's environment. Its cameras cannot see round corners → partially observable; other drivers and the weather make outcomes uncertain → stochastic; each manoeuvre changes the later situation → sequential; traffic moves while it decides → dynamic; speed and steering vary smoothly → continuous; other road users act → multi-agent. A crossword solver is the opposite on almost every axis: fully observable, deterministic, static, discrete, single-agent — but still sequential, since each filled word constrains the next.

Common pitfalls

• Weak AI ≠ "bad AI". Weak means narrow (one task); "solves many tasks" describes strong AI.
• Passing the Turing test ≠ intelligence. Fooling some users shows human-likeness of the replies; the test also penalises machines that are smarter than people².
• Chinese room is about understanding, not ability: Searle grants the program answers perfectly².
• Rational ≠ successful. Judge a decision by what was expected given the available information, not by the actual result³.
• History dates. John McCarthy introduced the term "artificial intelligence" for the 1956 Dartmouth workshop (1950s, not 1940s or 1960s); its 1955 proposal conjectured that every feature of intelligence can be described precisely enough for a machine to simulate it⁵. The first AI winter was in the 1970s (1974–1980): limited compute, combinatorial explosion, overpromising, and the discovery that perception and movement — easy for people — are the hardest to program (Moravec's paradox)⁵.
• Other history traps⁵: SHRDLU worked in the blocks world, a micro-world of blocks to stack; "scruffy" labs hacked programs until they worked while "neat" labs relied on formal logic; Minsky and Papert's Perceptrons (1969) showed single-layer limits, but multi-layer networks trained with backpropagation (popularised 1986) revived the field; expert systems such as R1/XCON were abandoned because they were costly to maintain, could not learn and were brittle; deep learning took off around 2012 (AlexNet on ImageNet) thanks to large labelled datasets and fast hardware^1,5.
• Physical symbol system hypothesis: "necessary" means human thinking itself must be symbol manipulation; "sufficient" means a machine with the right symbol system can be intelligent⁶.