A cinematic journey through seven decades
What artificial intelligence actually is, and how it grew from a two-page funding proposal in 1955 into the defining technology of 2026.
Strip away the mythology, and artificial intelligence is software that learns its own rules from data instead of following rules a programmer wrote by hand.
A conventional program is a recipe: a human anticipates every situation and writes an instruction for it. A learning system is different. It is shown examples, millions or trillions of them, and it adjusts millions or trillions of internal numbers called weights until its outputs start matching the patterns hidden in those examples. Nobody writes the rule for recognizing a cat, translating a sentence, or finishing your code. The rule emerges from the data.
That is why the field's history swings between euphoria and disappointment. Learning from data sounds simple, but it took seventy years of mathematics, hardware, and sheer accumulated data before it worked at scale. Most of what today's systems can do was not programmed into them. It was found, by an optimization process, inside oceans of text, images, and sound.
And that is also what AI is not. A large language model is not a database looking up answers, and it is not a conscious mind. It is a prediction machine of astonishing scale: given everything so far, it estimates what plausibly comes next, one small piece at a time. Everything below, from a 1958 machine the size of a room to the agents of 2026, is a variation on that single, strange, powerful idea.
"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."
The Dartmouth proposal · August 31, 1955
Beneath every modern AI system sits the same architecture of ideas: the neural network, an old invention that the world kept giving up on.
Neurons and weights. Each node receives numbers, multiplies them by its weights, and passes the result on. One node knows almost nothing. Stack thousands of layers of them and the network can represent language, images, protein structures, and strategy.
Training. The network makes a guess, the guess is compared with the right answer, and an algorithm called backpropagation traces the error backward, slightly correcting every weight. Repeat billions of times and competence slowly condenses out of randomness.
Inference. When you talk to a model, no learning is happening. The frozen weights simply transform your input into a prediction. Modern "reasoning" models spend extra computation at this stage, drafting and checking their own thoughts before answering.
Scale. The recipe barely changed after 1986. What changed is scale: more data, more parameters, more compute. The bitter, repeated lesson of this history is that scale beats cleverness more often than anyone expected.
1950 - 1955
Before the machine could think, someone had to ask whether it ever could. The question arrived politely, in an English philosophy journal, and it has never left.
1950
Alan Turing publishes "Computing Machinery and Intelligence" and sidesteps philosophy with a test: if a machine's conversation cannot be distinguished from a person's, the question of whether it "really" thinks loses its meaning. The imitation game, later the Turing test, gives the unborn field its north star.
1955
On August 31, John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon propose a summer workshop at Dartmouth College, choosing a deliberately ambitious name for its subject: artificial intelligence. They ask the Rockefeller Foundation for funding for ten researchers, and predict that significant progress can be made in one summer.
1955
Even before the workshop convenes, Allen Newell and Herbert Simon are building the Logic Theorist, a program that proves theorems from Russell and Whitehead's Principia Mathematica. Machine reasoning stops being speculation and becomes running code.
1956 - 1965
One summer in New Hampshire, a handful of scientists founded a discipline. Within a few years they had languages, learning machines, and a dangerous surplus of confidence.
1956
Through the summer, researchers drift in and out of Dartmouth College for weeks of open-ended argument about thinking machines. Little is solved, but a field is born: the attendees, McCarthy, Minsky, Newell, Simon, Samuel, Selfridge and others, will dominate AI research for the next two decades.
1958
Frank Rosenblatt unveils a machine that learns to recognize simple shapes by adjusting its own connection strengths: the first artificial neural network to capture the public imagination. The New York Times reports the Navy expects it will one day "walk, talk, see, write, reproduce itself and be conscious of its existence." The hype problem is born the same day as the technology.
1958
John McCarthy creates LISP, a programming language built for manipulating symbols rather than numbers. It becomes AI's native tongue for over thirty years, and the second-oldest language still in use today.
1959
IBM's Arthur Samuel, whose checkers program improves by playing against itself, defines machine learning as the study that "gives computers the ability to learn without being explicitly programmed." His program eventually defeats respectable human players, a first glimpse of self-improvement through experience.
1965 - 1974
Money flowed, demos dazzled, and the founders promised human-level machines within a generation. The demos were real. The promises were not.
1965
Herbert Simon declares that "machines will be capable, within twenty years, of doing any work a man can do." Marvin Minsky is similarly bullish. Government money, especially from the Pentagon's ARPA, flows into MIT, Stanford, and Carnegie Mellon with few strings attached.
1966
Joseph Weizenbaum builds ELIZA, a simple pattern-matching chatbot that parrots users' words back as questions. To his horror, people confide in it and insist it understands them. The "ELIZA effect," our readiness to see minds in machines that merely echo us, is documented sixty years before it matters most.
1966
SRI begins work on Shakey, the first mobile robot to reason about its own actions, combining perception, planning, and movement. It is glacially slow and works only in prepared rooms, but its planning algorithms, including A* search, outlive it everywhere.
1969
Minsky and Seymour Papert publish a rigorous critique proving what single-layer perceptrons cannot do, including the simple XOR function. The result says nothing about deeper networks, but funding agencies read it as a verdict. Neural network research enters a fifteen-year exile.
"In from three to eight years we will have a machine with the general intelligence of an average human being."
Marvin Minsky, Life magazine, 1970
1974 - 1980
The machines of the 1970s had the memory of a paragraph and the speed of a pocket calculator. Reality invoiced the field for its promises, and the funding collapsed.
1973
Mathematician James Lighthill, commissioned by the British government, concludes that AI has utterly failed to deliver on its "grandiose objectives" and that its methods collapse outside toy problems. UK funding is cut to almost nothing, and the skepticism crosses the Atlantic: DARPA follows suit.
1974
The term "AI winter" enters the vocabulary: a season when funding, students, and respectability all vanish at once. The underlying problem is brutally simple. Approaches that work on small examples explode combinatorially on real ones, and 1970s hardware cannot brute-force the gap.
1979
Out of the spotlight, the future is germinating. Backpropagation is described in Paul Werbos's overlooked 1974 thesis. Hans Moravec's Stanford Cart inches across a room, an ancestor of every self-driving car. Winters, it turns out, are when the roots grow.
1980 - 1987
AI came back as a business. If general intelligence was out of reach, narrow expertise was not: bottle a specialist's rules, and sell the bottle.
1980
Digital Equipment Corporation deploys XCON, a program encoding thousands of if-then rules from human experts to configure computer orders. It reportedly saves the company tens of millions of dollars a year. By mid-decade, two thirds of Fortune 500 companies are running expert systems, and an industry worth billions appears.
1982
Japan launches the Fifth Generation Computer Systems project, a ten-year, half-billion-dollar national program to build massively parallel reasoning machines. Washington and London, terrified of losing the race, reopen the funding taps they had closed in the winter.
1982
Physicist John Hopfield shows that a network of simple neurons can store and retrieve memories, giving the exiled field new mathematical respectability. A small community, including a persistent Geoffrey Hinton, keeps working on learning machines while the mainstream chases rules.
1986
Rumelhart, Hinton, and Williams publish the paper that finally makes deep learning practical: backpropagation lets multi-layer networks learn internal representations, dissolving the XOR objection of 1969. Nearly every system in the chapters that follow is trained with this algorithm.
1987 - 1993
The expert-system economy was a castle of hand-written rules. It could not learn, could not adapt, and could not survive the desktop computer.
1987
Specialized AI workstations costing tens of thousands of dollars are suddenly outrun by ordinary Sun and Apple desktops. The half-billion-dollar LISP machine industry evaporates within a year, taking investor confidence with it.
1990
Companies discover the fatal flaw of rule-based systems: they do not learn. Every exception demands another hand-written rule, maintenance costs balloon, and systems fail silently at the edges of their knowledge. "AI" becomes such a tainted word that researchers rebrand their work as informatics, analytics, or machine learning.
1992
Japan's Fifth Generation project ends without its promised reasoning machines. The field absorbs a hard lesson that will define its next thirty years: systems built from human-authored rules plateau, while systems that learn from data keep improving as data and computers grow.
1993 - 2011
AI stopped promising minds and started shipping statistics. Under other names, it spread into checkbooks, search engines, and living rooms, while three ingredients quietly accumulated: data, compute, and patience.
1997
IBM's chess machine beats world champion Garry Kasparov in a six-game match, the first computer victory over a reigning champion under tournament conditions. It wins by brute-force search, not learning, but the symbolism is seismic: a fortress of human intellect has fallen.
1998
Yann LeCun's convolutional network, LeNet, is deployed by banks to read handwritten digits on millions of checks. It is perhaps the first large-scale commercial triumph of deep learning, years before anyone calls it that.
2005
Stanford's robot Stanley crosses 132 miles of Mojave desert unaided and wins the DARPA Grand Challenge, one year after every entrant had failed. Machine learning, not hand-coded rules, is the difference, and the modern autonomous-vehicle industry traces its lineage to this race.
2009
Fei-Fei Li releases ImageNet, millions of labeled images across thousands of categories, and launches an annual recognition contest. It is the proving ground the coming revolution needs: for the first time, there is enough data to show what deep networks can really do.
2011
IBM's Watson defeats the two greatest Jeopardy! champions on live television, and Apple puts Siri into the iPhone. AI is back in public view, now as a consumer product rather than a laboratory promise.
2012 - 2019
Two gaming graphics cards, one old algorithm, and a million images. In 2012 the exiled idea returned and, within five years, rewrote every field it touched.
2012
Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton enter the ImageNet contest with a deep convolutional network trained on two consumer GPUs. Their error rate is almost half the runner-up's. Overnight, thirty years of skepticism about neural networks ends, and every major lab pivots to deep learning.
2014
Ian Goodfellow's generative adversarial networks pit two networks against each other, one forging images, one detecting forgeries, until the forgeries become convincing. Generative AI, machines that create rather than classify, becomes a serious research program.
2016
DeepMind's AlphaGo defeats Go legend Lee Sedol four games to one. In game two, its 37th move is so alien that commentators assume it is a mistake; it is genius. Two hundred million people watch a machine display something indistinguishable from intuition.
2017
Eight Google researchers publish the Transformer, an architecture that processes sequences by letting every word attend to every other word at once. It is scalable in a way nothing before it was. Virtually every system in the next two chapters, GPT, BERT, Claude, Gemini, is a Transformer.
2018
Google's BERT and OpenAI's early GPT models show that one network, pretrained on vast unlabeled text, can be adapted to almost any language task. In 2019 OpenAI deems GPT-2 risky enough to release in stages. The age of the general-purpose language model has begun.
2020 - 2023
Scale stopped being an engineering detail and became the discovery itself. Then, one November evening, a research demo handed the whole story to the public.
2020
OpenAI's GPT-3, with 175 billion parameters, writes essays, code, and poetry from a plain-text prompt, learning new tasks from a few examples with no retraining. Researchers formalize "scaling laws": capability grows predictably with model size, data, and compute. Bigger, it turns out, is genuinely different.
2020
DeepMind's AlphaFold 2 predicts the 3D structure of proteins from their amino acid sequences with near-experimental accuracy, cracking a grand challenge of biology. Its predictions for hundreds of millions of proteins are released freely, and drug discovery is permanently changed.
2022
DALL-E 2, Midjourney, and the openly released Stable Diffusion turn text descriptions into photorealistic images. Art, design, and an uneasy debate about authorship arrive together. Generative AI becomes something you can see.
2022
On November 30, OpenAI releases a chat interface to its models as a "research preview." It reaches a million users in five days and an estimated hundred million within two months, the fastest-adopted consumer application in history at the time. AI stops being an industry story and becomes a household one.
2023
GPT-4 arrives with striking reasoning ability, Anthropic releases Claude with a focus on safety and helpfulness, and Meta's Llama models put strong AI in open hands. Thousands of researchers sign letters urging caution; governments convene their first AI safety summits. Capability and anxiety rise together.
"We can only see a short distance ahead, but we can see plenty there that needs to be done."
Alan Turing, 1950
2024 - 2026
The question stopped being whether machines could talk, and became whether they could work: plan, use tools, check their own reasoning, and act on our behalf.
2024
New "reasoning" models think before answering, spending extra computation to draft and verify chains of thought, and abruptly improve at mathematics and science. The Nobel Prize in Physics goes to John Hopfield and Geoffrey Hinton for neural networks; the Chemistry prize honors AlphaFold. The same year, the EU AI Act, the first comprehensive AI law, enters into force.
2025
AI systems graduate from chat windows to workplaces: coding agents work in terminals and repositories for hours at a stretch, models operate computers, browse, and call tools, and open standards like the Model Context Protocol let them plug into the world's software. Open-weight reasoning models make frontier techniques cheap, and adoption spreads from developers to everyone.
2026
Frontier systems now sustain long, multi-step work: researching, writing, coding, and coordinating with other software with less and less supervision. The open questions have shifted from capability to trust: how to verify what these systems do, align what they optimize for, and share what they create. Seventy-one years after Dartmouth, the summer project is still open.
Artificial intelligence is not magic, and it is not a mind. It is the seventy-year-old craft of getting rules to emerge from data, an idea that failed, twice, for decades, until scale made it work. Its history is a lesson in humility running in both directions: the pioneers wildly underestimated the difficulty, and the skeptics wildly underestimated the ceiling.
What happens next is not written. The machines will keep improving; the questions that matter, about truth, work, power, and trust, belong to the people who use and govern them. The next chapter of this exhibit does not exist yet. We are living inside it.
Exhibit closes here · July 3, 2026