The Story of AI
How humanity's oldest dream became its newest industry
The Dream Before the Machine
Long before circuits, there was bronze. In Greek myth, [[Talos|talos]] — a giant forged by Hephaestus, the divine blacksmith — marched around the island of Crete three times a day, hurling boulders at enemy ships. He was imagined some two and a half thousand years ago: a manufactured being doing a sentient being's job. Hephaestus's mythic workshop was full of such things — self-moving tripods, golden mechanical assistants. The Greeks did not have the technology. They had the *idea*, fully formed.
The dream traveled. Jewish tradition gave us the [[golem|golem]] of Prague, a servant shaped from clay and animated by sacred words — and by rules that, mishandled, end in catastrophe. In the 12th century, the Islamic engineer [[al-Jazari|aljazari]] built actual programmable automata: a mechanical serving girl, a robotic band whose drumbeats could be reprogrammed with pegs. Hero of Alexandria had already built a programmable cart a millennium earlier.
Notice what every version of the story carries: a maker, a made mind, a job to do — and an anxiety about control. The plot of the Golem is the plot of *Frankenstein* is the plot of a thousand modern AI op-eds. Before we ask what AI is, the record suggests an answer to why it is: humans have wanted to build minds for as long as we have known we had them.
The Words We Chose
Every field inherits its destiny partly from its vocabulary. The word *robot* entered the world's languages in 1921, through a Czech stage play — Karel Čapek's *R.U.R.* The word was actually suggested by his brother Josef, from the Czech [[robota|robota]]: forced labor, the work a serf owed a feudal lord, rooted in the Old Slavic *rab*, slave. The word for our mechanical servants has slavery in its bones — and in the play, the robots revolt. The anxiety was in the etymology from day one.
*Artificial intelligence* is younger still, and it was a marketing decision. In a [[1955 funding proposal|proposal]] to the Rockefeller Foundation, a young Dartmouth mathematician named John McCarthy — with Marvin Minsky, Claude Shannon, and Nathaniel Rochester — coined the phrase to name a summer workshop. McCarthy reportedly chose a [[new term|naming]] partly to escape the gravity of existing fields like cybernetics and automata theory, and the agendas of the men who ran them.
So the field's two flagship words encode its two permanent tensions: *robot* carries the labor question (who works, who owns the worker), and *artificial intelligence* carries the ambition question — the audacious claim, written into the 1955 proposal itself, that intelligence 'can in principle be so precisely described that a machine can be made to simulate it.' Everything that follows is people arguing about that sentence.
A Summer in New Hampshire
In the summer of 1956, a shifting cast of mathematicians and engineers gathered at Dartmouth College for the workshop McCarthy had proposed. In its classrooms and corridors, 'artificial intelligence' was used, debated, and defined for the first time as the name of a field. The attendees — McCarthy, Minsky, Shannon, Rochester, plus visitors like Herbert Simon and Allen Newell — would dominate AI research for the next two decades.
What is striking in hindsight is the [[founding conjecture|conjecture]]: not a research question but a bet — that intelligence is describable, and whatever is describable is mechanizable. The workshop produced no breakthrough that summer. What it produced was a *frame*: intelligence as symbol manipulation, mind as program, brain as hardware.
The decade that followed ran on [[extraordinary optimism|optimism]]. Programs proved theorems, played checkers, solved word problems. Predictions of human-level machines within a generation flowed freely. The geography mattered too: AI was born American, Cold-War funded, English-speaking — a fact that shaped what 'intelligence' was taken to mean (chess, logic, language) for half a century.
The First Winter
The optimism died in two stages, each delivered on paper. In 1969, Minsky and Papert's book [[Perceptrons|perceptrons]] proved mathematical limits on what single-layer neural networks could compute. The proofs were correct but narrowly scoped; the *effect* was sweeping — mainstream neural-network research in the West largely stopped for a decade. One half of AI's solution space went dark, partly on the strength of an argument that didn't actually apply to multi-layer networks.
Then in 1973, the British mathematician Sir James Lighthill delivered a [[report to Parliament|lighthill]] concluding that AI had failed its 'grandiose objectives' and that its discoveries amounted to little. The UK gutted AI funding at all but two universities; the chill spread to DARPA in the US. The years 1974–1980 became the first [[AI winter|winter]] — the field's own meteorological metaphor for collective abandonment.
The winter's lesson is uncomfortable for every era, including ours: funding follows narrative, and narrative overshoots in both directions. The claims of the sixties were too big; the verdict of the seventies was also too big. Both were errors of confidence, not of evidence.
Boom, Bust, and the Hype Curve
AI thawed in the 1980s by getting narrow. [[Expert systems|expert]] — programs encoding the rules of human specialists — diagnosed infections (MYCIN), configured computers (XCON), and saved real companies real money. Japan announced its Fifth Generation Computer project; the US and UK panicked and funded counter-programs. A billion-dollar industry of LISP machines and knowledge engineers bloomed.
It collapsed within a decade. Expert systems proved brittle — they could not learn, choked at the edges of their rules, and cost fortunes to maintain. The specialized hardware market evaporated in 1987. By the early 1990s the term 'AI' was so toxic that researchers rebranded their work as 'informatics', 'machine learning', 'analytics' — anything else. This was the [[second winter|winter2]].
Step back and the shape of the story becomes visible: ambition, boom, disillusion, winter, rebrand, repeat. Twice now the field had mistaken a working *demo* for a working *theory of intelligence*. The pattern is worth carrying forward as we enter the era everyone knows — because the central question of 2026 is whether this time is different, and the only honest answer requires knowing the curve.
The Quiet Revolution
While symbolic AI boomed and busted, the neural-network heretics kept working. The 1986 popularization of [[backpropagation|backprop]] showed multi-layer networks *could* be trained — answering Perceptrons seventeen years late. But networks stayed marginal for two more decades, waiting on two things nobody fully controlled: data and compute.
Both arrived around 2010. The internet produced labeled data at planetary scale — crystallized in [[ImageNet|imagenet]], a dataset of fourteen million images. Video games, of all things, produced the hardware: graphics cards built to render explosions turned out to be ideal for the matrix math of neural networks. In 2012, [[AlexNet|alexnet]] — a deep network trained on two consumer GPUs by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton — crushed the ImageNet competition. Before it, almost no leading computer-vision papers used neural nets. After it, almost all of them did.
The deeper lesson of 2012 is about *how* fields turn: not by argument but by demonstration. Hinton had been making the same case since the seventies; what changed was not the idea but the substrate. The 'golden decade' that followed — 2012 to 2022 — was largely the old connectionist program finally running at the scale it had always needed.
Attention, and Everything After
In 2017, eight Google researchers published a paper with a Beatles-pun title: 'Attention Is All You Need.' The [[Transformer|transformer]] architecture it introduced processed language not word-by-word but all-at-once, letting every word attend to every other word in parallel. It was built for translation. It turned out to be a universal engine for sequence prediction — text, code, images, proteins.
OpenAI's bet was [[scale|scaling]]: take a Transformer, make it bigger, feed it more of the internet, and capabilities emerge that nobody programmed — translation, arithmetic, rudimentary reasoning. GPT-2 (2019) wrote plausible paragraphs. GPT-3 (2020) wrote plausible essays. Each step was quantitative; the effects felt qualitative.
Then came the interface. In November 2022, [[ChatGPT|chatgpt]] wrapped a tuned GPT-3.5 in a chat box, free to anyone. A hundred million users arrived in two months — the fastest-adopted consumer application in history to that point. Seventy years of AI research collapsed into a single cultural moment, and the public re-asked, all at once, every question this story has already visited: Will it take our jobs (1921)? Does it really understand (1980)? Is it overhyped (1973, 1987)? The vocabulary of the debate was older than the technology.
The Voices That Said No
The critique is as old as the field. Philosopher Hubert Dreyfus argued from the 1960s that human intelligence rests on a *background* of embodied, unspoken know-how that can never be fully captured in symbols — a claim symbolic AI spent thirty years inadvertently confirming. In 1980, John Searle's [[Chinese Room|chineseroom]] argued that even a perfectly fluent program understands nothing: syntax is not semantics, simulation is not the thing simulated.
The modern critics split into camps that disagree with each other as much as with the optimists. The [[stochastic parrots|parrots]] school (Bender, Gebru and colleagues, 2021) argues language models are pattern-mimics without understanding — and that the real harms are present-tense: bias, environmental cost, concentration of power. Gary [[Marcus|marcus]] argues deep learning hits walls — data hunger, brittleness, no causal reasoning — and needs symbolic structure bolted back in. And a third camp argues the opposite worry: that these systems may work *too well*, posing [[existential risk|xrisk]]. Hinton himself left Google in 2023 to say so freely.
Hold the camps side by side and notice the strange shape of the moment: AI is simultaneously accused of being a parlor trick and a potential extinction event. Both cannot be fully right. A serious answer to 'what is the story of AI?' has to keep all three critiques on the table — because each one has, at some point in this history, been the correct call.
The Roads Taken and Not Taken
Seen from above, AI history is a tree of bets. The *symbolic* branch bet that intelligence is logic: represent knowledge, apply rules. It gave us theorem provers, expert systems, and most of computer science's tooling — then hit the wall Dreyfus predicted: the world refuses to be fully written down. The *connectionist* branch bet that intelligence is statistics learned by brain-like networks. Dismissed in 1969, vindicated in 2012, it now runs the world — while hitting walls of its own: hallucination, opacity, and a strained relationship with truth.
The roads not (yet) fully taken are where the next chapter likely lives. [[Neurosymbolic|neurosymbolic]] AI tries to weld the branches: neural perception with symbolic reasoning — notably, today's reasoning models writing chains of thought and calling external tools are a step in roughly this direction, smuggled in through the back door. *Embodied* approaches argue intelligence needs a body and a world, not just text — robotics is where that bet is being settled. And a contrarian camp holds that the [[scaling hypothesis|scalinghyp]] is simply, boringly right, and architecture debates are noise against a trillion-fold compute curve.
The honest reading of the solution space: every branch has been declared dead at least once, and no branch has stayed dead. The field does not progress by replacing wrong ideas with right ones — it progresses by re-rating old ideas when the hardware, the data, or the money changes. That, more than any single architecture, is the mechanism of this story.
The Answer, For Now
Where the timeline ends, today: frontier models from a handful of labs [[specialize and reason|now2026]] — 'thinking' models generate long reasoning traces before answering, agents carry out multi-step work, and training compute has grown about a trillion-fold since the deep-learning era began. The vocabulary of 1956 ('can a machine think?') has become operational ('which model should file my taxes?'). And the 2024 Nobel Prizes quietly canonized the field — physics for Hopfield and Hinton's networks, chemistry for protein-folding AI.
So, the answer. *Thread one — the dream is ancient.* Talos to golem to robot to GPT is one continuous act of imagination: making a mind, and fearing the made mind. AI is not a technology with a culture attached; it is a culture that finally got its technology. *Thread two — the words were never neutral.* 'Robot' encodes slavery; 'artificial intelligence' encodes a 1955 funding bet and an academic rivalry. The field's perpetual boom-bust cycle is partly a linguistic inheritance: its very name promises more than any decade could deliver. *Thread three — progress came from the losing side.* The architecture running the world today spent forty years as the dismissed branch. The field's true engine is not any single idea but the re-rating of old ideas when substrate and scale change.
Which yields the forecast discipline this story demands: we are either inside the first AI spring that never ends, or near the peak of the largest hype curve ever drawn — and the people best positioned to know [[genuinely disagree|openq]]. The story of AI is a 2,500-year-old dream, running on video-game hardware, named by a grant proposal, built by heretics, and still arguing about whether it understands a single word it says. That argument — older than the field, older than the word — is the story.
Sources and research
Linguistics — the words
## The Language of Artificial Minds
### robot (1921)
From Czech **robota** — forced labor, serf-work — rooted in Old Slavic *rab*, 'slave'. Suggested by Josef Čapek for his brother Karel's play *R.U.R.* (Rossum's Universal Robots), Prague, 1921. The word displaced 'automaton' and 'android' in most world languages within a decade. **robotics** was coined by Isaac Asimov (1941).
- The play's robots are biological workers, not metal machines — and they revolt.
- The labor anxiety is not a modern addition to the word; it is its root meaning.
Sources: [Science Friday](https://www.sciencefriday.com/segments/the-origin-of-the-word-robot/), [MIT Press Reader](https://thereader.mitpress.mit.edu/origin-word-robot-rur/), [NPR](https://www.npr.org/2025/09/11/nx-s1-5532095/where-word-robot-came-from-rur)
### artificial intelligence (1955)
First written in the September 2, 1955 proposal by **McCarthy, Minsky, Rochester & Shannon** for the 1956 Dartmouth summer workshop. McCarthy reportedly avoided 'cybernetics' partly to escape Norbert Wiener's intellectual orbit.
- The founding conjecture: *'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.'*
Sources: [Dartmouth](https://home.dartmouth.edu/about/artificial-intelligence-ai-coined-dartmouth), [Dartmouth Alumni Magazine](https://dartmouthalumnimagazine.com/articles/birth-artificial-intelligence)
### Other vocabulary with embedded history
- **AI winter** — coined ~1984 by analogy to *nuclear winter*.
- **machine learning / informatics** — partly rebrands adopted when 'AI' became toxic after the second winter.
- **stochastic parrot** (2021), **hallucination**, **alignment** — each term is itself an argument about what these systems are.
Deep time — myths & automata
## Artificial Beings Before Computers
### Myth
- **Talos** (Greece, ~700 BCE in oral tradition): bronze giant built by Hephaestus, guarded Crete, destroyed via his single vein of ichor — the first single point of failure. ([Stanford Report](https://news.stanford.edu/stories/2019/02/ancient-myths-reveal-early-fantasies-artificial-life), [Smithsonian](https://www.smithsonianmag.com/history/was-talos-the-bronze-automaton-who-guarded-the-island-of-crete-in-greek-myth-an-early-example-of-artificial-intelligence-180986467/))
- **Pandora**, golden handmaidens, self-moving tripods: Hephaestus's workshop as the first robotics lab of the imagination.
- **The Golem** (Jewish folklore; Prague, 16th c.): clay servant animated by sacred words; the control problem in theological form.
### Engineering
- **Hero of Alexandria** (1st c. CE): programmable cart, automated theatre, steam toys.
- **al-Jazari** (1136–1206): *Book of Knowledge of Ingenious Mechanical Devices* — water clocks, a mechanical serving girl, a peg-programmable robotic band. ([Discover](https://www.discovermagazine.com/robots-and-artificial-intelligence-have-ancient-mythology-origins-44067))
- **The Mechanical Turk** (1770): chess 'automaton' with a human hidden inside — the original AI demo that was secretly humans all the way down.
### Why it matters
Adrienne Mayor (*Gods and Robots*, 2018): the Greeks already explored autonomous machines' promise and peril. Every modern AI debate — labor, control, deception, hubris — has a pre-industrial rehearsal.
Timeline — milestones & winters
## The Canonical Timeline
- **1950** — Turing, *Computing Machinery and Intelligence*: the imitation game.
- **1955** — 'Artificial intelligence' coined in the Dartmouth proposal.
- **1956** — Dartmouth Summer Research Project: the field's founding event. ([Wikipedia](https://en.wikipedia.org/wiki/Dartmouth_workshop))
- **1957–69** — Golden years: theorem provers, checkers, Perceptron hype; Simon predicts machines doing 'any work a man can do' within 20 years.
- **1969** — Minsky & Papert, *Perceptrons*: single-layer limits; neural research defunded.
- **1973** — **Lighthill Report** to UK Parliament: 'grandiose objectives' unmet; UK funding gutted. ([AI Innovations Unleashed](https://www.aiinnovationsunleashed.com/a-chilly-history-how-a-1973-report-caused-the-original-ai-winter/))
- **1974–80** — **First AI winter**. ([Wikipedia](https://en.wikipedia.org/wiki/AI_winter))
- **1980–87** — Expert-systems boom: MYCIN, XCON, Japan's Fifth Generation, LISP machines.
- **1987–93** — **Second AI winter**: LISP market collapse, DARPA cuts; 'AI' becomes a dirty word.
- **1986** — Backpropagation popularized (Rumelhart, Hinton, Williams).
- **1997** — Deep Blue defeats Kasparov.
- **2012** — **AlexNet** wins ImageNet on two gaming GPUs; deep learning era begins. ([Timeline](https://vrungta.substack.com/p/timeline-of-deep-learnings-evolution))
- **2016** — AlphaGo defeats Lee Sedol.
- **2017** — **Transformer**: 'Attention Is All You Need'. ([Wikipedia](https://en.wikipedia.org/wiki/Transformer_(deep_learning_architecture)))
- **2018–20** — GPT series; BERT; scaling laws formalized.
- **2022** — **ChatGPT**: 100M users in two months.
- **2023** — GPT-4; CAIS extinction-risk statement; Hinton leaves Google.
- **2024** — Nobel Prizes: physics (Hopfield & Hinton), chemistry (protein folding).
- **2025–26** — Reasoning models, agents, specialization; training compute ~10^12× since 2012. ([AISI](https://www.aisi.gov.uk/frontier-ai-trends-report))
The critics — full dossier
## Four Generations of Refusal
### Philosophical (1965–1990)
- **Hubert Dreyfus**, *What Computers Can't Do*: intelligence rests on embodied background know-how that cannot be formalized. Symbolic AI's failures largely vindicated him.
- **John Searle**, the **Chinese Room** (1980): perfect symbol manipulation without understanding; syntax ≠ semantics. Key replies: systems reply, robot reply, the Churchlands' connectionist reply. ([SEP](https://plato.stanford.edu/entries/chinese-room/))
### Institutional (1969–1973)
- **Perceptrons** (Minsky & Papert) and the **Lighthill Report**: technically narrow critiques with field-wide chilling effects. Lesson: critique, like hype, overshoots.
### Contemporary — harms now (2018–)
- **Stochastic Parrots** (Bender, Gebru, McMillan-Major, Mitchell, 2021): LLMs as fluent mimics; real harms are bias, environmental cost, power concentration. ([paper](https://dl.acm.org/doi/pdf/10.1145/3442188.3445922))
- **Gary Marcus**: ten walls of deep learning — data hunger, brittleness, no causal/compositional reasoning; prescribes neurosymbolic hybrids. ([IEEE Spectrum](https://spectrum.ieee.org/gary-marcus))
### Contemporary — risks later (2014–)
- **Existential risk**: Bostrom, CAIS 2023 statement ('global priority alongside pandemics and nuclear war'), Hinton & Bengio signatories. Skeptics: LeCun ('turbojet safety in 1920'), Mitchell. Expert probability estimates span orders of magnitude. ([Wikipedia](https://en.wikipedia.org/wiki/Existential_risk_from_artificial_intelligence))
### The tension
The parrot critique and the x-risk critique pull in opposite directions — too dumb to trust vs. too capable to control. Any honest synthesis must explain how both communities, looking at the same systems, reached opposite fears.
Solution space — alternatives
## The Tree of Bets
### Symbolic AI (1956–)
Intelligence as logic and explicit knowledge. Wins: theorem proving, planning, expert systems, databases. Wall: the frame problem — the world won't be fully written down.
### Connectionism (1943, 1986, 2012–)
Intelligence as statistics in brain-inspired networks. Declared dead 1969; resurrected by backprop (1986), data (ImageNet) and GPUs (2012). Currently dominant. Walls: hallucination, opacity, data/compute appetite.
### Neurosymbolic (ongoing)
Weld perception (neural) to reasoning (symbolic). Today's reasoning models — chain-of-thought traces, tool calling, code execution — are a de facto neurosymbolic turn smuggled into pure neural stacks. ([MIT Tech Review](https://www.technologyreview.com/2019/09/27/65250/we-cant-trust-ai-systems-built-on-deep-learning-alone/))
### Embodiment
Brooks ('Intelligence without representation'), Dreyfus: minds need bodies and worlds. Being settled now in robotics, where progress lags language conspicuously.
### Pure scaling
The bet that compute + data + one architecture suffices. Evidence for: a decade of scaling laws. Evidence against: cost curves, data exhaustion, persistent brittleness. ([Compute trends, arXiv](https://arxiv.org/pdf/2202.05924))
### Meta-observation
No branch of this tree has stayed dead. The field advances by **re-rating old ideas when substrate changes** — the single most reliable pattern in seventy years.
Now — June 2026 snapshot
## State of Play (researched June 2026)
- A handful of frontier labs ship competing model families; the defining 2026 trend is **specialization** — fast models for routine work, 'thinking' models spending inference-time compute on long reasoning traces. ([TeamAI overview](https://teamai.com/blog/large-language-models-llms/the-2026-ai-frontier-model-war/))
- **Agents** — models that plan, browse, call tools, and execute multi-step work — are the commercial frontier; token usage and pricing are rising accordingly. ([Citrini Research](https://www.citriniresearch.com/p/state-of-the-themes-june-2026))
- Training compute has grown roughly **a trillion-fold** across the deep-learning era, with further thousand-fold increases projected within years. ([AISI Frontier AI Trends Report](https://www.aisi.gov.uk/frontier-ai-trends-report))
- Multimodality (text, image, documents) is a floor, not a differentiator.
- Benchmarks are saturating faster than they can be written; evaluation itself is now a research bottleneck. ([Kili](https://kili-technology.com/blog/ai-benchmarks-guide-the-top-evaluations-in-2026-and-why-theyre-not-enough))
### Epistemic note
This chapter describes a fast-moving present (verified against June 2026 sources, linked). Treat specific model names and rankings as perishable; the structural claims — specialization, agents, compute growth, evaluation crisis — are the durable layer.