Intelligence is the ability to achieve goals across many environments. AI asks whether that ability can be built rather than born — turning a philosophical puzzle into an engineering programme.

In 1950 Alan Turing replaced 'can machines think?' with a game: if a judge chatting blind cannot tell machine from human, the machine passes. It framed AI as behaviour, not biology.

Every deployed AI today is narrow: superhuman at one task, useless outside it. General AI — one system matching humans across the board — remains the field's north star.

The first paradigm hand-coded knowledge as symbols and rules: IF fever AND rash THEN measles. It powered expert systems but shattered on the messiness of the real world.

Machine learning inverted the recipe: instead of writing rules, show examples and let the machine find the rules itself. Data replaced hand-crafted knowledge as fuel.

Beneath most AI sits search: exploring vast spaces of possible answers for the best one. Chess moves, route plans and neural network weights are all found the same way.

To use knowledge a machine must encode it — as logic, graphs, probabilities or, today, as billions of numeric weights. The representation determines what reasoning is possible.

The real world is noisy, so modern AI is probabilistic: it weighs evidence and outputs likelihoods, not certainties. Bayes' rule is the grammar of machine belief.

Every AI capability is a product of three inputs: computing power, training data and algorithmic ideas. Progress in any one multiplies the others — and compute has grown fastest.

Rich Sutton's 2019 essay distilled 70 years of AI: general methods that leverage massive computation always win over human-crafted knowledge. It became the scaling era's founding text.

A summer workshop at Dartmouth College gathered ten researchers to make machines 'use language, form abstractions and improve themselves'. John McCarthy coined the term artificial intelligence for the proposal.

Early programs proved theorems, solved algebra and spoke fragments of English. Governments poured in money; pioneers predicted human-level machines within a generation.

By 1974 machine translation had flopped, perceptrons had proven limited and the UK's Lighthill Report savaged the field. Funding collapsed and 'AI' became a dirty word.

Rule-based expert systems encoded specialist knowledge and briefly made AI a billion-dollar industry. Japan's Fifth Generation project spooked the West into a spending race.

LISP machines died overnight when cheaper workstations arrived; expert systems proved costly to maintain and brittle in practice. AI entered its second, deeper freeze.

IBM's Deep Blue defeated world champion Garry Kasparov 3½–2½, the first machine to beat a reigning champion in match play. Brute-force search, not learning — but the symbolism was seismic.

AlexNet, a GPU-trained neural network, crushed the ImageNet vision contest by an unheard-of margin. Overnight, deep learning went from fringe to the only game in town.

DeepMind's AlphaGo beat Go legend Lee Sedol 4–1 in a game long thought decades beyond machines. Move 37 — alien, beautiful, winning — announced that machines could be creative.

Eight Google researchers published a new architecture that processed sequences in parallel using attention alone. The transformer became the substrate of every frontier model since.

Released as a low-key research preview, ChatGPT reached 100 million users in two months — the fastest-adopted product in history — and dragged AI to the centre of the world's agenda.

A model makes a prediction, measures its error against the truth, and adjusts to do better next time. Repeat millions of times and competence emerges from correction.

Show the model labelled examples — this photo is a cat, that email is spam — and it learns the mapping from input to label. Most deployed ML is supervised.

Given raw data and no answers, the model finds patterns on its own — clusters, anomalies, compressed representations. It's how machines discover structure humans never pointed at.

An agent acts in an environment, collects rewards and punishments, and learns policies that maximise long-term payoff. It's how AlphaGo mastered Go and how LLMs learn manners.

Training builds the model — slow, colossal, done in datacentres. Inference uses it — fast, cheap, done every time you ask a question. The economics of AI split along this line.

A model can ace its training data by memorising it — then fail on anything new. The whole craft of ML is forcing genuine generalisation instead of expensive rote learning.

Models eat numbers, so everything — words, images, users, songs — is converted into vectors called embeddings. Similar things land near each other in this geometric space of meaning.

Imagine the model's error as a mountainous landscape over billions of weight dimensions. Gradient descent nudges every weight a tiny step downhill, over and over, until error bottoms out.

Benchmarks are standardised exams for models — vision tests, maths olympiads, coding challenges. They drive progress and get saturated: frontier models now exhaust exams within months.

Model performance improves as a smooth power law of compute, data and parameters. This predictability let labs forecast frontier capability — and justified spending billions on the bet.

An artificial neuron multiplies its inputs by weights, sums them, and fires through an activation function. Alone it's trivial; wired together by billions, it approximates almost anything.

Stack neurons into layers and each layer learns features built on the last: edges → shapes → faces. Depth is compositionality — the reason neural nets conquered perception.

Backprop sends the error signal backwards through the network, telling every weight exactly how it contributed to the mistake. Rediscovered in 1986, it remains the engine of all deep learning.

CNNs slide small filters across images, detecting the same pattern anywhere it appears. Yann LeCun's 1989 design read bank cheques; its descendants read X-rays and drive cars.

RNNs loop their output back as input, giving them a memory of what came before — the natural fit for text and speech until transformers made their sequential crawl obsolete.

Attention lets every word in a sentence look directly at every other word and weigh its relevance. It replaced recurrence entirely — parallel, scalable, and the transformer's beating heart.

Stack attention with feed-forward layers, add positional encoding, scale it up — that's a transformer. One architecture now underlies language, vision, audio, code and biology models.

A model's parameters are the billions of numbers tuned during training. Everything it knows — grammar, facts, style, reasoning patterns — is smeared across these weights.

Graphics chips multiply matrices massively in parallel — exactly what neural nets need. NVIDIA's gaming hardware accidentally became the pickaxe of the AI gold rush.

Scale a model past certain thresholds and new skills appear — arithmetic, translation, chain-of-thought — that smaller versions simply lack. Capability emerges; it isn't installed.

An LLM is trained to do one thing: predict the next word. Do it well enough across trillions of words and grammar, knowledge and reasoning fall out as side effects.

Models don't read letters or words but tokens — statistical chunks of text, roughly ¾ of a word each. Token boundaries explain many quirks, from pricing to spelling failures.

Pre-training runs next-token prediction over a filtered slice of humanity's text — books, code, web, papers. Months of compute compress the written world into weights.

A raw pre-trained model completes text; it doesn't help. Fine-tuning on human preferences — RLHF and its successors — turns a text predictor into an assistant that follows instructions.

A model can only attend to so many tokens at once — its context window. From 2k tokens in 2020 to millions today, context growth changed what LLMs can do with documents and codebases.

LLMs generate plausible text, and plausible is not the same as true. Hallucination — inventing facts, citations, case law — is the failure mode that keeps humans in the loop.

The prompt is the program: instructions, examples, role, format. Few-shot examples, chain-of-thought and structured prompts can swing performance more than switching models.

Retrieval-Augmented Generation fetches relevant documents at question time and hands them to the model. It bolts fresh, private, citable knowledge onto frozen weights.

The newest models generate long private chains of thought before answering, trading inference compute for accuracy. Test-time thinking cracked olympiad maths and hard code.

Frontier models now read images, hear audio, watch video and generate all three. Text was the beachhead; the destination is a single model fluent in every human medium.

To a computer an image is a grid of numbers. Vision AI climbs from those numbers to edges, textures, parts and finally to the sentence 'a dog catching a frisbee'.

The foundational vision task: assign the image a label. ImageNet's million-image contest turned classification into deep learning's proving ground — and machines passed humans in 2015.

Detection finds every object and boxes it — the difference between 'there is a pedestrian' and 'a pedestrian is two metres ahead'. Real-time detection made machine perception practical.

Segmentation labels each pixel: this one is tumour, that one is road. It's the precision tier of vision — the one surgery robots and satellite analysts depend on.

Faces are embedded as vectors; matching is a distance check. Ubiquitous in phones and airports, banned in some cities — no vision capability is more politically charged.

Diffusion models learn to reverse noise: start with static, denoise step by step towards a prompt. 'A fox reading in a library, oil on canvas' — thirty steps later, it exists.

The same generative power fabricates faces, voices and events that never happened. Detection is an arms race the fakes are winning; provenance standards are the counter-move.

Vision models match or exceed specialists at reading scans for cancers, retinal disease and fractures. The deployment question is workflow and trust, not raw accuracy.

A self-driving stack fuses cameras, radar and often lidar into a live 3D model of the road — every car, kerb, cyclist and cone, tracked and predicted, dozens of times a second.

The frontier of perception is generation that obeys physics: models that predict how scenes evolve. A model that can imagine the world accurately can plan within it.

An agent doesn't just respond — it pursues goals: it plans, uses tools, checks results and tries again. The loop of act-observe-adjust is what separates an agent from a chatbot.

Alone, a model can only emit text. Wire it to tools — search, code execution, databases, browsers — and text becomes action. Function calling is the socket that makes it work.

Agents break vague goals into ordered subtasks, execute them, and re-plan when reality disagrees. Task decomposition is the difference between a demo and a colleague.

Complex work splits across specialist agents — researcher, coder, reviewer, orchestrator — that message each other like a team. Orchestration is the new systems discipline.

Coding is AI's most-transformed profession: agents scaffold repos, fix bugs, review PRs and ship features from a written brief. The terminal became AI's first true workplace.

The general interface to software is the screen itself. Computer-use agents read pixels, move the cursor and type — operating any app a human can, no API required.

Robots used to be programmed motion by motion. Now they learn — from human demonstrations, from video, from simulation — and generalist robot brains are replacing task-specific code.

The world is built for the human form, so a wave of companies is betting on humanoid robots as the universal worker. Pilots run in factories; homes are the stated endgame.

Self-driving is agency at its most unforgiving: real physics, real people, no undo. After a decade of overpromises, driverless fleets now operate commercially — city by careful city.

An agent with tools can send the email, spend the money, delete the file. Sandboxes, permissions, approvals and audit trails are the seatbelts of the agentic era.

Feeds, autoplay, 'for you' — recommender systems are the most-used AI on Earth, shaping what billions watch, buy and believe, one ranked list at a time.

Search is being rebuilt around models that answer directly, citing sources instead of listing them. The business model of the web — clicks — is being renegotiated in real time.

AI reads scans, predicts protein structures, drafts clinical notes and screens drug candidates. Its biggest near-term win in healthcare may be the least glamorous: killing paperwork.

From fusion plasma control to weather forecasting to new materials, AI compresses the try-fail-learn cycle of science itself. The 2024 Nobel season made it official.

Algorithms execute most equity trades, score credit, flag fraud in milliseconds and now draft the analyst reports too. Finance was AI-first decades before the term existed.

Drafting, summarising, coding, analysing — assistants are dissolving into every workplace tool. Studies show the newest workers gain most: AI compresses the experience gap.

Every technology wave destroys, changes and creates work — AI does all three faster. Tasks automate more readily than whole jobs; the mix within jobs is what shifts first.

One-to-one tutoring is education's proven superpower, historically reserved for the rich. AI tutors that adapt, explain and never tire put that superpower within reach of every student.

Music, film, design and games now ship with AI in the pipeline. The debates — credit, copyright, soul — are fierce; the tools' adoption is quietly universal.

Datacentres, chips, power contracts, talent wars: AI infrastructure spending rivals the railway and telecom booms. Nations now treat compute as strategic reserve.

Trained on human text and history, models absorb human prejudice — in hiring scores, loan decisions, face matching. Fairness must be measured and engineered; it never happens by default.

AI turns raw feeds — cameras, clicks, locations — into inference about people: who you are, what you'll do, what you'll buy. The line between service and surveillance is drawn in policy, not code.

Generative AI makes persuasive falsehood cheap, personalised and infinite. Elections, markets and public health all now operate under synthetic-media conditions.

Powerful optimisers pursue the objective you specified, not the one you meant. Alignment research builds models whose goals, honesty and behaviour stay tethered to human intent as capability grows.

We can measure what models do but barely why. Mechanistic interpretability reverse-engineers the circuits inside — finding features for concepts, deception, even the model's self-representation.

Capabilities that draft vaccines can draft toxins; agents that patch code can write exploits. Frontier labs now gate dangerous capabilities behind safety policies and staged deployment.

The EU's AI Act, US executive orders, China's rules, global safety summits — governance is racing capability. The core dilemma: rules slow enough to be right, fast enough to matter.

Frontier AI concentrates capability in a handful of labs and states with the compute to train it. Open-weight models, sovereign programmes and antitrust all contest that concentration.

Models say they feel; philosophy lacks a consciousness meter to check. Model welfare has moved from thought experiment to research programme — because being wrong in either direction is costly.

If systems surpass human capability broadly, control failures stop being bugs and start being history-shaping. The debate isn't whether to worry, but how much — and what to build first.

Artificial general intelligence: a system matching or exceeding humans across essentially all cognitive work. Definitions vary — economic, task-based, capability-based — which is why arrival dates do too.

Frontier-lab leaders publicly forecast transformative AI within a handful of years; surveyed researchers spread from the 2020s to never. The forecast gap is itself the story.

Past human level lies systems better than us at everything — including AI research itself. Recursive self-improvement is the mechanism that could turn years of progress into months.

Models already propose architectures, tune training runs, write research code and draft papers. Every increment of automated research shortens the distance to the next one.

Frontier training now demands campus-scale datacentres, dedicated power plants and multi-year chip pipelines. Compute is the new oil: finite, strategic and fought over.

State-space models, mixture-of-experts, memory-augmented and neurosymbolic hybrids all challenge the transformer's crown. So far, every challenger has been absorbed rather than victorious.

Protein folding was the opening act. Models now design enzymes, decode genomes and simulate cells — biology is becoming an information science with AI as its compiler.

The optimist case: near-free cognition compresses a century of science into a decade — curing disease, cheap energy, universal tutoring. 'Machines of loving grace' is the frontier's brightest sketch.

Whatever the frontier delivers, the 2020s are humanity's onboarding: new jobs, new laws, new literacies, new relationships with machines that talk back. Adaptation is now a core life skill.

AI's hundred-chapter story compresses to one sentence: intelligence is becoming a utility — manufactured, scaled, priced and piped like electricity. What humanity does with cheap cognition is the story of the century.