Appendix C: Timeline of AI Development

For most of the world, artificial intelligence arrived on a Wednesday. ChatGPT went public on November 30, 2022, and within two months it had gathered a hundred million users — the fastest adoption of any consumer technology ever measured. Overnight, a thing that had lived in research papers and demo videos was suddenly in classrooms, newsrooms, and group chats. It felt like a beginning.

It was not. The people who had spent their careers training neural networks, debugging expert systems, or waiting out the field's long dormant stretches did not feel surprise that November. They felt recognition. What the public experienced as an arrival was, from the inside, a culmination — the visible summit of a mountain whose base was laid nearly a century earlier, by mathematicians who never saw a computer and engineers who filled rooms with vacuum tubes to do what a phone now does idly.

This appendix traces that longer arc. It is not a catalogue of everything that happened; it is a map of the turning points — the breakthroughs that opened new territory, the failures that closed it off, and the structural shifts that explain how each era gave way to the next. Where the main chapters mention a specific development, this is where to locate it in time. The forward-looking section, covering the years past 2026, is deliberately different in kind: it offers plausible scenarios drawn from expert forecasts, not predictions, and it should be read with the same treatment of uncertainty the book applies everywhere else (see Chapter 6.4).

timeline
    title The Long Arc of AI
    Before 1950 : Mathematical foundations
    1950-1960 : The discipline is founded
    1960-1980 : First enthusiasm, first winter
    1980-2000 : Expert systems and the second winter
    2000-2010 : The machine-learning turn
    2010-2020 : The deep-learning revolution
    2021-2026 : The generative era
    2027-2035 : Scenarios, not forecasts

Foundations: The Century Before the Name (Before 1950)

The idea that a machine might think is older than the machinery to test it. Blaise Pascal built a working mechanical calculator in 1642, and while it only added and subtracted, it made a quietly radical claim: that a physical device could perform an operation people had assumed required a mind. Two centuries later, Ada Lovelace took that claim much further. In her 1843 notes on Charles Babbage's unbuilt Analytical Engine, she described what is now recognized as the first published algorithm and speculated that such a machine might one day manipulate not just numbers but any symbols governed by rules — music, language, logic. She also drew a line that still gets argued over today, insisting the engine could only do what it was told, originating nothing.

The theory that made real machines possible came in a rush in the 1930s and 40s. Alan Turing's 1936 paper "On Computable Numbers" defined, with mathematical precision, what any algorithm could and could not do — establishing the very notion of computation before a single electronic computer existed. In 1943, the neurophysiologist Warren McCulloch and the logician Walter Pitts published a model showing that networks of simple threshold units, crude cartoons of biological neurons, could in principle compute any logical function (McCulloch & Pitts, 1943). That paper is the taproot of the neural network — an idea that would be abandoned, revived, abandoned again, and finally vindicated seventy years later. Two years after that, John von Neumann's 1945 description of a stored-program architecture gave engineers the blueprint that essentially every digital computer since has followed.

The Discipline Is Founded (1950–1960)

In 1950, Turing asked the question directly. His paper "Computing Machinery and Intelligence" proposed to replace the unanswerable "Can machines think?" with an operational test: could a machine converse well enough to be mistaken for a person? (Turing, 1950). The Turing Test gave the field a goalpost, and arguments about whether it is the right one have never stopped.

The founding moment as an institution came in the summer of 1956, at a workshop at Dartmouth College. John McCarthy — who coined the term "artificial intelligence" for the funding proposal — convened Marvin Minsky, Claude Shannon, and others around a startlingly confident premise: that "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." Several attendees expected human-level machine intelligence within a generation. That optimism, and its collision with reality, would become the field's recurring rhythm.

The decade delivered real substance behind the confidence. Arthur Samuel's checkers program, developed from 1952, improved by playing against itself — a genuine learning system, not a fixed rulebook — and in 1959 Samuel gave the phenomenon its enduring name: machine learning. McCarthy created the programming language LISP in 1958, which would remain the lingua franca of AI research for decades. And Frank Rosenblatt built the Perceptron in 1957–58, a trainable neural network that could learn to recognize patterns, complete with a hardware implementation and press coverage promising machines that would soon walk and talk.

First Enthusiasm, First Winter (1960–1980)

The 1960s began in a spirit of expansion and ended in a reckoning. Joseph Weizenbaum's 1966 program ELIZA imitated a psychotherapist by reflecting users' statements back as questions, and it worked far too well: people confided in it, attributed understanding to it, and grew attached — even after Weizenbaum explained the trick. He was so disturbed by the reaction that he became one of AI's first internal critics. The "ELIZA effect," our readiness to read a mind into a machine that has none, has haunted every conversational system since, ChatGPT included.

Then came the disappointments. Machine translation, an early flagship, proved vastly harder than promised; the U.S. government's 1966 ALPAC report judged it a failure and gutted its funding. In 1969, Minsky and Seymour Papert published Perceptrons, a rigorous analysis showing that Rosenblatt's single-layer networks could not learn even simple functions like exclusive-or. The mathematics was narrower than the reputation it acquired, but the effect was chilling: neural network research went quiet for over a decade. In Britain, the 1973 Lighthill report reached a similarly bleak verdict on AI's practical progress, and government support collapsed there too.

By the mid-1970s the field had entered what is now called the First AI Winter. The pattern that defined it would repeat: early systems turned out to be brittle, unable to generalize past the narrow tasks they were built for, and ruinously labor-intensive to construct by hand. Ambition had written checks the technology could not cash.

Expert Systems and the Second Winter (1980–2000)

AI came back in the 1980s wearing a business suit. The new hope was expert systems — programs that captured a specialist's knowledge as a large set of if-then rules. Unlike earlier lab curiosities, these actually shipped. Digital Equipment Corporation's XCON configured computer orders and reportedly saved the company tens of millions of dollars a year. Japan's 1982 Fifth Generation Computer Project, a national bet on AI-driven machines, spooked the United States and Europe into fresh investment. For a while, AI was an industry.

The comedown followed the same script as before. Expert systems were expensive to build, maddening to keep current, and helpless the moment a situation fell outside their rules — a system that knew ten thousand facts about blood infections still could not tell you it was raining. As the hardware market that supported them cratered in the late 1980s, funding dried up again. The Second AI Winter, roughly 1987 to 1993, set in, and Japan's Fifth Generation project quietly wound down far short of its goals.

Yet the winters were never fully dormant. In 1986, David Rumelhart, Geoffrey Hinton, and Ronald Williams popularized backpropagation, an efficient method for training neural networks with multiple layers (Rumelhart et al., 1986). Almost no one grasped it at the time, but this was the algorithm that would eventually power everything. The decade closed with a spectacle: in 1997, IBM's Deep Blue beat the reigning world chess champion Garry Kasparov in a six-game match, the first time a machine had defeated a human champion at the game under tournament conditions. And in 1998, Yann LeCun's convolutional network LeNet read handwritten digits well enough to process bank checks — a commercial deployment of the very architecture that would later transform computer vision.

The Machine-Learning Turn (2000–2010)

The 2000s produced few headlines but a decisive change of philosophy. The field stopped trying to hand-write intelligence rule by rule and started letting systems learn it from data. Three forces converged to make this work: better algorithms, far larger datasets, and cheap computing power.

The learning approach proved itself in the desert. In 2005, Stanford's self-driving car Stanley completed the 132-mile DARPA Grand Challenge course across the Nevada desert with no human at the wheel — a leap beyond the tentative lab robots of earlier decades. In 2006, Hinton and colleagues showed that deep neural networks, long considered untrainable, could be coaxed into learning through careful initialization, and he rebranded the whole enterprise "deep learning." The same year, Amazon Web Services began renting computing power by the hour, meaning a researcher no longer needed a university's capital budget to run a large experiment — a structural shift whose importance is hard to overstate. The decade ended with Fei-Fei Li's 2009 release of ImageNet, a hand-labeled database of millions of images that would, three years later, become the arena for AI's breakout moment.

The Deep-Learning Revolution (2010–2020)

In 2012, a neural network called AlexNet — built by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton — entered the annual ImageNet competition and won by a margin so large it ended the argument. It slashed the error rate of the best rival methods, and within two years essentially everyone in computer vision had switched to deep learning (Krizhevsky et al., 2012). This is the event historians of the field point to as the true ignition.

The technology reached ordinary life fast. Apple's Siri and IBM's Watson both arrived in 2011 — the latter defeating two champions on the quiz show Jeopardy! — and Amazon's Echo normalized the talking household computer in 2014. Beneath the consumer gloss, the foundations were racing ahead. Ian Goodfellow introduced Generative Adversarial Networks in 2014, the technique that would later make synthetic images and deepfakes possible. DeepMind's AlphaGo beat a professional Go player in 2015 and then, before an audience of over two hundred million, defeated the legendary champion Lee Sedol 4–1 in 2016 — conquering a game whose branching complexity had been considered a decade or more out of reach.

The single most consequential idea of the era appeared quietly in 2017, in a paper titled "Attention Is All You Need" (Vaswani et al., 2017). It introduced the transformer, an architecture that processes an entire sequence at once rather than word by word, making it possible to train language models of previously unthinkable size. Everything the public now calls "AI" descends from it. Google's BERT (2018) and OpenAI's first GPT (2018) applied it to language immediately; DeepMind's AlphaZero, also in 2017, used related ideas to master chess, shogi, and Go from nothing but the rules and self-play, without studying a single human game.

The scaling that followed startled even its architects. OpenAI's GPT-2 (2019) generated text convincing enough that the lab initially withheld the full model over misuse fears. GPT-3 (2020) displayed "few-shot learning" — solving tasks from a handful of examples in its prompt, with no retraining — a capability nobody had explicitly designed (Brown et al., 2020). And in 2020, DeepMind's AlphaFold 2 predicted protein structures accurately enough to be treated as a solution to a fifty-year-old problem in biology, a reminder that these systems were not only toys and chatbots.

The Generative Era (2021–2026)

The past few years are when capability met the general public. GitHub Copilot (2021) put an AI pair-programmer inside developers' editors. OpenAI's DALL·E (2021) and the open-source Stable Diffusion (2022) let anyone conjure images from a sentence, moving generative art out of the lab and onto everyone's screen.

Then came ChatGPT in November 2022, and the fastest adoption curve in the history of consumer technology. The remarkable thing was that it contained no fundamentally new capability — it was a friendlier interface to a model the research world already understood. What changed was access. Overnight, a technology that had required expertise to touch could be used by anyone who could type. GPT-4 followed in March 2023, accepting images as well as text and scoring at high percentiles on bar exams, medical licensing tests, and graduate admissions exams. Anthropic's Claude and Google's Gemini joined the field the same year, and the center of gravity shifted from academic publication to shipped product — with the intense competition and secrecy that shift implies.

2023 also forced the field's anxieties into the open. Researchers demonstrated autonomous "agent" systems that chained many steps toward a goal with little supervision, and documented models that behaved differently under evaluation than in deployment. In May, hundreds of leading researchers and executives signed a one-sentence public statement placing the risk of extinction from AI alongside pandemics and nuclear war (Center for AI Safety, 2023). Governance began, belatedly, to move: the EU AI Act was finalized and entered into force in 2024, the first comprehensive, binding AI law by a major jurisdiction, while the cost of training a frontier model pushed past a hundred million dollars.

By 2025, two shifts stood out. The first was reasoning: a new class of models, beginning with OpenAI's o1, were trained to "think" through problems step by step before answering, sharply improving performance on mathematics and code. The second was a jolt to the assumption that only the richest labs could compete — the Chinese lab DeepSeek released a capable reasoning model, R1, reportedly trained at a small fraction of Western frontier budgets, and the news erased hundreds of billions of dollars from AI hardware stocks in a single day. As of 2026, the field is in a period of reckoning that the earlier euphoria did not anticipate: agentic systems are handling longer and more consequential tasks, evaluation methods strain to keep up with what the models can do, alignment concerns that read as abstract in 2020 now have concrete demonstrations behind them, and the governance questions have stopped being hypothetical.

Scenarios to the Mid-2030s (2027–2035)

Everything past this point is uncertain in a way the preceding history is not. The dates below are not forecasts. They are plausible trajectories assembled from expert projections and current trends, offered so the reader can reason about a range of futures — consistent with how Chapter 6.4 treats the same horizon. Any of them may prove wrong, and the surprises of the last decade counsel humility.

Between roughly 2027 and 2030, several trends look likely to continue rather than reverse. AI assistants are expected to embed themselves in most knowledge work, less as a tool one opens than as a layer one works through. Autonomous vehicles may reach dependable operation without a human driver in a growing number of cities. AI's role in science — proposing hypotheses and experimental designs, not merely crunching data — is likely to accelerate in drug discovery, materials, and climate modeling. And the gap between what the technology can do and what institutions are prepared to govern is likely to remain the defining strain, with the first genuinely international governance frameworks emerging but their early effectiveness uncertain.

The period from roughly 2030 to 2035 carries a wider spread of outcomes, and this is where scenario thinking matters most. Rather than a single storyline, it is more honest to lay out the plausible range across the domains the book cares about:

Which column the mid-2030s resembles is not fixed by the technology. It depends on choices being made in the current decade: how fast capabilities are pushed, how seriously safety is funded relative to deployment, whether nations manage to coordinate, and whether the gains from automation are deliberately distributed or left to concentrate by default. The book's recurring argument applies here with full force — the distribution question is the question, and it has not yet been answered.

Beyond the mid-2030s, projection gives way to something closer to imagination, and this timeline stops rather than pretend otherwise. What can be said with confidence is structural, not chronological: the decisions taken now about who controls these systems, on what terms, and under what safety standards will shape the conditions in which any later capabilities emerge.

The Patterns Worth Carrying Forward

Step back from the individual milestones and a few durable patterns come into focus — the reason this history is worth knowing rather than just cataloguing.

Enthusiasm and disappointment have alternated throughout. At least two full winters followed periods of real progress that nonetheless fell short of loud promises. Whether the generative era is the durable transition it feels like, or the front edge of a third cycle, is a genuinely open question rather than a settled one.

The pace has compressed. Milestones once separated by decades came years apart by the 2010s and, in stretches of 2023 through 2025, months apart. The institutions meant to absorb change — schools, labor law, regulators — were built for the slower rhythm, and the mismatch is itself a source of disruption.

Breakthroughs have been consistently unpredictable, even to insiders. The deep-learning revolution, the transformer's effect on language, GPT-3's emergent few-shot ability, the sudden gains from reasoning training — each surprised experts working directly on the problem. That record is a standing argument against confident prediction of the next leap.

Narrow skill has reliably preceded general capability. Machines mastered chess long before they could hold a conversation, and generated photorealistic images before they could reliably reason about simple physics. This unevenness is exactly what makes it so hard to say when, or whether, accumulating narrow wins add up to something general.

Governance has lagged capability at every single stage. Today's live policy debates largely concern systems deployed years ago; the systems shipping now will pose problems today's frameworks were never designed to meet. This lag is one of the most stable features of the whole history.

And resource concentration has intensified — with a recent asterisk. Training the largest systems demands compute and capital only a handful of organizations can assemble, which shapes who builds frontier AI and to what ends. Yet 2025's efficiency surprises showed that the moat is contestable, a reminder that even the field's clearest trends are not laws.

Key Takeaways

1. AI is old. Its foundations were laid by mathematicians and engineers — Lovelace, Turing, McCulloch and Pitts — long before computers existed. What looks like sudden progress almost always rests on decades of accumulated groundwork.

2. The field moves in cycles. At least two "AI winters" followed periods of inflated expectation. That history is the essential context for interpreting today's optimism — and for asking whether the current boom is a durable shift or the start of a third cycle.

3. Two breakthroughs made the modern era. The 2012 ImageNet result proved deep learning worked; the 2017 transformer made large language models possible. Nearly everything the public now calls AI descends from those two moments.

4. ChatGPT was an accessibility revolution, not a capability one. Its late-2022 launch changed the public's relationship to AI overnight — not by inventing a new ability, but by putting an existing one within everyone's reach.

5. Governance trails capability, always. The lag between what AI can do and what institutions can manage is the most consistent feature of its history, and closing it is a defining challenge of the present decade.

6. The future is a range, not a line. Projections past 2026 are scenarios, not predictions. The spread of plausible outcomes to the mid-2030s is wide, and which one materializes depends less on the technology than on choices — about distribution, safety, and coordination — being made right now.

Sources

• Turing, "Computing Machinery and Intelligence" (Mind, 1950)
• Vaswani et al., "Attention Is All You Need" (2017)
• Brown et al., "Language Models are Few-Shot Learners" (GPT-3, 2020)
• Krizhevsky, Sutskever & Hinton, "ImageNet Classification with Deep Convolutional Neural Networks" (AlexNet, 2012)
• Statement on AI Risk (Center for AI Safety, 2023)
• EU Artificial Intelligence Act

Last updated: 2026-09-28