There is no single universally agreed definition of AI. Here are the four most cited, each emphasising a different aspect:

The common thread across all four: AI systems are designed to optimise for a goal. Not magic — sophisticated optimisation machines trained on data. Whether the goal is to win a chess game, classify an image, or predict the next word in a sentence, the mechanism is the same: minimise a loss, maximise a reward.

All AI systems — from the simplest spam filter to GPT-4 — can be understood through four fundamental capabilities. These aren't arbitrary categories: they map directly to the architectural components you'll encounter in every system throughout this curriculum.

Perceive is the entry point. In large language models, the perception layer is the tokeniser and embedding lookup — raw text transformed into high-dimensional vectors the network can process. In computer vision systems, it's the image pipeline: resize, normalise, encode. In robotics, it's sensor fusion from cameras, accelerometers, and LiDAR.

Reason sits at the heart of what makes AI feel intelligent. In a transformer model, attention mechanisms allow every token to reason about every other token in the context window. In a decision tree, reasoning is the sequence of feature comparisons that route an input to a leaf node prediction.

Learn is the mechanism that makes modern AI powerful. Traditional software is programmed with explicit rules. Learning systems instead infer rules from examples. A loss function quantifies error, and an optimiser updates parameters to reduce it — repeated millions of times across millions of examples.

Act closes the loop. For a language model, acting means sampling the next token from a probability distribution. For a recommendation engine, surfacing the top-k items. For a robotic system, issuing motor commands. The action taken modifies the environment, producing new perception data.

These terms are often used interchangeably in the press, but they have precise relationships — each is a subset of the one above. Consider ChatGPT: it is simultaneously correct to call it AI, ML, and Deep Learning — each statement is true and increasingly specific.

However, the reverse implication does not hold. Not all AI is ML — expert systems encode human knowledge directly as rules, without learning from data. Not all AI is deep learning — a random forest or naive Bayes classifier are ML but not deep learning. Data Science is a separate discipline focused on extracting business insight via statistics, visualisation, and ML — it overlaps AI but is not a true subset.

AI systems can be classified along two axes: scope (what range of tasks?) and capability level (how does it compare to human intelligence?).

Public discourse around AI is saturated with misconceptions — drawn from science fiction, corporate marketing, and media sensationalism. Before going deeper into mechanisms, it is worth explicitly naming what AI is not.

Long before the first computer was built, mathematicians and philosophers dreamed of mechanising thought. Gottfried Wilhelm Leibniz (1646–1716) envisioned a Characteristica Universalis — a universal symbolic language that could encode all human knowledge — and a calculus ratiocinator that could reason over it mechanically. This dream of reducing reasoning to symbol manipulation would resurface, unchanged in spirit, at the Dartmouth Conference three centuries later.

George Boole (1854) formalised logic as algebra — reducing true/false statements to 0s and 1s and defining the operations AND, OR, NOT. This was the mathematical bedrock everything else was built on. Ada Lovelace (1840s), writing annotations on Babbage's Analytical Engine, described the first published algorithm intended for a computing machine — and crucially, articulated both its power and its limits: the machine can only do what it is told. Claude Shannon (1938) showed that Boolean algebra could be implemented in electronic circuits, connecting Boole's logic to physical hardware. And finally McCulloch and Pitts (1943) proposed the first mathematical model of a neuron — a binary threshold unit that fires when its weighted inputs exceed a threshold. Every modern neural network traces its ancestry directly to that 1943 paper.

These five contributions — symbolic reasoning, Boolean logic, algorithms, logic circuits, and artificial neurons — were the intellectual raw materials that the Dartmouth group assembled into a new field in 1956.

In 1950 Alan Turing published "Computing Machinery and Intelligence" — opening with the question "Can machines think?" He proposed the Imitation Game as a pragmatic operational test: if a machine can convince a human judge, communicating only via text, that it is human, we have sufficient practical grounds to attribute intelligence to it. Turing was careful not to define intelligence — he proposed the test precisely to sidestep that philosophical quagmire. He also described the "child machine" concept: rather than programming adult intelligence directly, build a machine that learns. This was the first clear articulation of what we now call machine learning.

Six years later, in the summer of 1956, John McCarthy, Marvin Minsky, Claude Shannon, and Herbert Simon convened at Dartmouth College and officially founded the field of Artificial Intelligence. Their proposal was breathtakingly optimistic: "Every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." Minsky later claimed that "within a generation the problem of creating AI will substantially be solved." That generation passed without resolution. The Dartmouth group were brilliant researchers who massively underestimated three things: the importance of data, the role of perception and embodiment, and the sheer computational depth required.

The decade following Dartmouth was genuinely remarkable. Early AI programs demonstrated things no one had seen before: a machine proving mathematical theorems, a program that could hold a conversation, a system that learned to play checkers by playing against itself. In 1959 Arthur Samuel's checkers program — which improved by self-play — gave the field a new term: machine learning. The optimism was not irrational — the demos were real. The problem was that toy domains don't scale.

The AI winters are the most important episodes in AI history for developing intuition about the current moment. They were not caused by bad science. They were caused by a structural gap: researchers promised capabilities that required data and compute that didn't yet exist, and funders cut support when those promises weren't kept on schedule.

While AI winters suppressed funding, a quieter revolution was accumulating. Statistical machine learning — SVMs, random forests, gradient boosting — replaced brittle symbolic systems with principled, mathematically-grounded methods. Meanwhile, the web was producing data at a scale no previous generation could have imagined, and GPU hardware was becoming cheap enough to train meaningfully large networks.

AlexNet's ImageNet victory in 2012 was the moment the field changed permanently. Krizhevsky, Sutskever, and Hinton's network achieved 15.3% top-5 error vs. the runner-up's 26.2% — an 11-point gap that shocked the computer vision community. The model was trained on two consumer GPUs in five days. Within 18 months, every major research lab had pivoted entirely to deep neural networks. The pattern: one shocking result, then near-universal adoption.

The Transformer deserves a dedicated note. "Attention Is All You Need" (Vaswani et al., 2017) replaced recurrent networks with a single elegant mechanism: self-attention, which allows every token to attend to every other token in parallel. This unlocked two things simultaneously — full context access (no sequential bottleneck) and massive parallelism (scale with compute). The result: BERT, GPT, T5, LLaMA, Whisper, DALL-E, AlphaFold, and virtually every transformative AI system since 2018 is built on Transformers. It is arguably the most consequential single paper in AI history.

ChatGPT launched on November 30, 2022. It reached 1 million users in 5 days and 100 million users in 60 days — the fastest consumer technology adoption in recorded history. Instagram took 2.5 years to reach the same milestone. Twitter took 5 years. This wasn't a product launch — it was a cultural event. For the first time, a general-purpose AI system was accessible to anyone with a browser, and it worked well enough to be genuinely useful for everyday tasks.

Unlike the boom-bust cycles of the past, the current AI wave is reinforced by a self-amplifying economic flywheel. Better models attract more users. More users generate more revenue and data. More revenue funds more compute. More compute trains better models. The cycle has no obvious brake — which is precisely why the pace of improvement has been relentlessly accelerating since 2012, and why projection-based estimates consistently underestimate where we will be in three years.

The flywheel has been spinning faster every year since 2012. Training compute for frontier models has grown roughly 4× per year. GPT-4's training run cost an estimated $100M. Next-generation frontier models are projected at $1B+. This isn't reckless spending — the returns on better models are large enough that the economics reinforce continued investment. Unlike the AI winters, there is no plausible external shock that would stop this cycle — only the discovery of fundamental capability limits, which has not yet materialised.

The philosopher John Haugeland coined the term GOFAI (Good Old-Fashioned AI) in 1985 to describe the dominant paradigm that had governed AI research since Dartmouth. Its core assumption: "Intelligence is symbol manipulation according to rules." Feed a machine symbols representing the world, give it logical rules for manipulating those symbols, and intelligence will emerge. The physical symbol system hypothesis (Newell & Simon, 1976) formalised this: "A physical symbol system has the necessary and sufficient means for general intelligent action." This claim was bold, wrong in the absolute sense, and enormously productive for 30 years.

Two schools developed within GOFAI. The logic-based school (Newell, Simon) grounded AI in formal logic and theorem proving — intelligence as deduction. The knowledge-based school (Minsky, McCarthy) focused on representing domain knowledge richly enough that a system could reason over it — intelligence as structured lookup and inference. Both shared the same fatal assumption: that the hard part of intelligence is the reasoning engine, not the knowledge itself.

Before an AI system can reason, it must represent knowledge. GOFAI researchers developed four major representational frameworks — each with a distinct philosophy about what knowledge is and how it should be stored.

Expert systems were the first commercially successful AI technology. The idea: interview a domain expert, encode their knowledge as production rules, build an inference engine to fire those rules. In the late 1970s and throughout the 1980s, this actually worked — inside carefully bounded domains with well-defined rules.

Logic is the mathematics of valid inference. GOFAI researchers hoped to give AI the ability to reason with the same rigor as formal proof — deriving new knowledge from existing knowledge with absolute certainty. Three levels of expressive power matter for AI.

If knowledge representation is how GOFAI stores the world, search is how it reasons through it. AI as search through a state space was the dominant problem-solving paradigm from 1956–1986. Define a state (current configuration), a set of actions (legal transitions), and a goal (desired configuration). The AI finds a sequence of actions from start to goal. Simple in principle — catastrophically hard in practice as state spaces grow.

GOFAI wasn't bad science — it produced real results in bounded domains. It failed because three fundamental problems proved impossible to solve within its paradigm. These aren't engineering problems. They are conceptual limits on what rule-based symbol manipulation can express.

Alan Turing’s 1950 paper “Computing Machinery and Intelligence” is the founding document of AI philosophy. He proposed the Imitation Game as a pragmatic replacement for the unanswerable “Can machines think?” A human interrogator communicates via text with two participants — one human, one machine. If the interrogator cannot reliably distinguish which is which, the machine has passed. Turing predicted that by 2000, a machine would fool 30% of judges after 5 minutes. He was roughly right about the timeline, but the test itself proved easier to pass than it was to mean anything.

In 1980 philosopher John Searle published “Minds, Brains, and Programs” — arguably the most influential and most debated paper in AI philosophy. His thought experiment: you are locked in a room. Through a slot, people pass slips of paper with Chinese characters. You have a rulebook: “When you see symbol sequence X followed by Y, write Z and pass it back.” You follow the rules perfectly. From outside, your responses are indistinguishable from those of a fluent Chinese speaker. But you understand nothing. You are just manipulating symbols according to formal rules.

Searle’s conclusion: syntax is not sufficient for semantics. A program that processes symbols according to rules — no matter how sophisticated — does not thereby understand those symbols. It has the form of language processing without the content. By analogy: AI programs process language according to learned statistical rules. They produce outputs indistinguishable from understanding. But this doesn’t mean they understand.

Philosopher David Chalmers (1995) distinguished easy problems of consciousness — explaining cognitive functions like attention, memory, and behaviour — from the Hard Problem: why does all this processing feel like something from the inside? Why isn’t cognition just computation happening “in the dark”, without any inner experience? This question may be permanently beyond empirical investigation, because any physical description of a brain state leaves open why there is subjective experience associated with it.

Stevan Harnad (1990) identified a fundamental problem with purely symbolic AI: symbols in a dictionary are defined in terms of other symbols. “Cat: a small domesticated carnivorous mammal with soft fur…” — circular definitions all the way down. For humans, symbols are grounded in sensorimotor experience. You know “red” because you have seen red. You know “hot” because you have felt heat. For large language models, symbols are grounded only in other symbols — the statistical contexts in which words appear across trillions of tokens of text.

This is why multimodal models (CLIP, GPT-4V, Gemini) and embodied robotics are active research frontiers — they attempt to ground language in perceptual or physical experience. Whether statistical grounding in text is “sufficient” for understanding is precisely the empirical question the field is now testing at scale.

From Russell & Norvig's Artificial Intelligence: A Modern Approach — the dominant textbook in AI education — an agent is "anything that perceives its environment through sensors and acts upon that environment through actuators." The word "agent" is deliberately broad: it covers thermostats, chess programs, autonomous vehicles, and GPT-4 equally.

What makes an agent rational? For each possible percept sequence, a rational agent should select an action expected to maximise its performance measure, given the evidence provided by the percept sequence and whatever built-in knowledge the agent has. Rationality ≠ omniscience. Rationality ≠ perfection. Rationality = expected utility maximisation given available information.

Why does this abstraction matter? Because it unifies everything — a thermostat perceives temperature and acts by switching heating, GPT-4 perceives your text and acts by generating tokens, a self-driving car perceives the road and acts by steering. Every AI system in this documentation can be analysed through this lens.

PEAS stands for Performance measure, Environment, Actuators, Sensors. It is the standard tool for formally specifying any AI agent task. Before building any AI system, PEAS forces you to answer four questions: What counts as success? What will the agent interact with? How can it act? What can it perceive?

Not all AI environments are created equal. The type of environment an agent operates in determines which algorithms are suitable, how much memory the agent needs, and how complex its decision-making must be. AIMA identifies six key dimensions along which environments vary.

Russell & Norvig define five progressively sophisticated agent architectures. Each adds a new capability layer. Understanding this hierarchy maps cleanly to the spectrum from a smoke detector to a fully autonomous AI assistant.

Before an agent can search for a solution, it must formally define the problem. AIMA's five-component problem formulation is the standard method. It forces precision: what exactly is a "state"? What exactly is an "action"? What exactly is the "goal"?

Connectionism is the paradigm that won. Inspired by the structure of the biological brain, it proposes that intelligence emerges from networks of simple connected units — and that knowledge is stored not in explicit rules, but in the strengths of connections between units. A single neuron knows nothing; billions of them, connected with learned weights, produce language, vision, and reasoning.

The lineage runs from McCulloch & Pitts (1943) — the first mathematical neuron — through Rosenblatt's Perceptron (1957), through Rumelhart & Hinton's backpropagation (1986) which made multi-layer training practical, through LeCun's LeNet (1998) and AlexNet (2012), all the way to the Transformer (2017) and the LLM era. Every step was the same insight applied at larger scale with more data and compute.

For three decades the battle between symbolicism and connectionism was not a polite academic disagreement — it was personal, bitter, and high-stakes. Minsky's famous dismissal of perceptrons (1969) killed neural network funding for a decade. Connectionists returned fire in 1986 with backpropagation. The symbolicism camp called neural networks "black box statistics". The connectionist camp called expert systems "fragile toy programs". Funding, careers, and the direction of the entire field were at stake.

The third major paradigm takes a different starting point: the world is fundamentally uncertain, and any intelligent system must represent and reason about that uncertainty explicitly. Where symbolicism asks "what is true?", probabilistic AI asks "how confident am I, and how should new evidence update that confidence?" This is Bayes' theorem as the engine of intelligence: P(hypothesis | evidence) ∝ P(evidence | hypothesis) × P(hypothesis).

Probabilistic AI never dominated the way symbolicism or connectionism did, but it remains indispensable in specific contexts. Robotics relies heavily on Bayesian state estimation (Kalman filters, particle filters). A/B testing and causal inference are Bayesian at their core. Modern LLM calibration — how well a model's confidence scores match actual accuracy — is a probabilistic question. RLHF uses Bayesian reasoning about reward models. The probabilistic paradigm's contribution is the rigorous treatment of uncertainty that pure neural approaches often lack.

The frontier of AI research in 2026 is increasingly defined by the attempt to combine the strengths of both paradigms: neural perception and fluency with symbolic precision and compositionality. Neither pure approach is sufficient — neural networks hallucinate and fail at systematic reasoning; symbolic systems can't perceive raw inputs or handle uncertainty. The hybrid is not a compromise — it is a new paradigm that can do things neither can do alone.

The five most important neuro-symbolic integrations in production AI today:

• ①Tool use — LLMs call Python interpreters, calculators, and search APIs for symbolic precision. Neural fluency + symbolic correctness.
• ②Chain-of-thought prompting — forcing explicit intermediate reasoning steps mimics symbolic deduction chains. Each step is verifiable.
• ③AlphaGo / AlphaZero — the neural value network supplies learned intuition; MCTS provides principled symbolic tree search. Neither alone reaches superhuman level.
• ④RAG (Retrieval-Augmented Generation) — a knowledge graph or vector store (symbolic) is queried by a neural LLM. Grounds generation in verifiable facts.
• ⑤Reasoning models (o1, DeepSeek-R1) — extended internal "thinking" with verifiable intermediate steps before producing output. The most direct implementation of neuro-symbolic reasoning in frontier LLMs.

The most important structural shift in AI since 2020 is the emergence of Foundation Models — large models pre-trained on broad, internet-scale data that can be adapted to almost any downstream task via prompting, fine-tuning, or RLHF. The old paradigm: train a separate model for each task (one for translation, one for summarisation, one for classification). The new paradigm: one model does all of them, often better than the specialised predecessors.

This shift was driven by neural scaling laws (Kaplan et al., 2020): performance on downstream tasks improves predictably as a power-law function of compute, parameters, and data — with no sign of diminishing returns until very recently. The implication: more scale reliably buys more capability, making the economics of frontier model training self-reinforcing.

AI capability is not uniform across tasks. Understanding where AI already surpasses humans, where it approaches parity, and where it still falls short is essential for calibrating realistic expectations — and for knowing which research problems remain open.

Benchmarks are the field's primary mechanism for measuring progress — and its primary mechanism for self-deception. The benchmark lifecycle is predictable: a new benchmark is created to test a genuine capability gap → models improve rapidly → benchmark saturates → researchers create a harder benchmark. The cycle repeats every 12–24 months.

Despite the extraordinary progress of the last decade, fundamental problems remain unsolved. These are not engineering challenges awaiting more compute — they are conceptual gaps that may require new paradigms to address. Understanding them prevents over-hyping current capabilities and points toward where research is most needed.