Prologue

Before the field had a name

The field was named at Dartmouth in 1956. The idea is older. Three figures across three centuries set up the question the field would inherit.

The first was Gottfried Wilhelm Leibniz. Leibniz was a polymath in the modern sense — co-inventor of calculus, independently of Newton and with the notation we still use, plus philosopher, diplomat, and inveterate organiser of all human knowledge. The piece of him that mattered for AI was a particular conjecture he turned over for most of his adult life. Reasoning is a kind of arithmetic. If every concept could be reduced to a primitive symbol, and if every argument could be expressed as a formal manipulation of those symbols, then philosophical disagreement would be a kind of computational error. Calculemus, Leibniz proposed: let us calculate, instead of arguing. The dream had two parts: a characteristica universalis, a symbolic language in which thought could be precisely encoded, and a calculus ratiocinator, a mechanical procedure that would settle disputes by computation rather than by debate. He did not build the machine. The machine he imagined would not be built in any form for another two and a half centuries. The bet he placed — that thought is a formal operation amenable to mechanisation — is the bet that AI inherited.

The second was Ada Lovelace, working with Charles Babbage in 1840s England. Babbage had designed the Analytical Engine, a fully mechanical computer specified down to the level of gears and never actually built. Lovelace, translating the Italian engineer Luigi Federico Menabrea’s notes on the Engine for English readers, appended commentary of her own. Her Notes — three times longer than the original article — became the founding document of programming. One passage of them is the one that matters here. The Engine, Lovelace observed, could in principle act on anything the operator could encode as a symbol, not only on numbers. It could “compose elaborate and scientific pieces of music of any degree of complexity or extent,” if music could be reduced to a representation the Engine could manipulate. A hundred and twenty years before the first generative model, Lovelace had named the move that makes them possible: anything representable is computable. She was the first person on record to have seen what a computer was actually for.

The third was Alan Turing. The Turing whose 1950 paper opens the next section is the one most readers know. The Turing that mattered first was younger — a twenty-four-year-old mathematician who in 1936 settled an open problem in mathematical logic by inventing an abstract machine. The Turing machine — a strip of tape, a read-write head, a finite table of rules — was not a proposed engineering project. It was a definition. This is what to compute means. From the definition Turing showed both that any conceivable computational procedure could be expressed in his framework and that certain well-posed questions could not be settled by any such procedure. He had drawn the boundary of computation as a mathematical object before the first electronic computer existed. A few years later, at Bletchley Park, he would help design the machines that broke the German Enigma cipher and shorten the Second World War by an estimated two years. The theoretical work and the cryptographic work are usually told as separate stories. For AI they are the same story. By 1945 Turing had a precise mathematical account of what computation was and the practical experience of having watched real machines execute it.

Three centuries, three figures. Leibniz argued that thought could be formalised. Lovelace argued that anything formalised could be computed. Turing made computation itself a precise mathematical object and lived long enough to ask what the resulting machines might do.

The field was now possible. It needed only to begin.

Two seeds, one summer

Both seeds were planted in the same five-year window. The history of AI is the history of which one got watered.

In 1943, two researchers in Chicago published a paper that no one quite knew what to do with. Warren McCulloch was a neurophysiologist; Walter Pitts was a mathematician who at one point in his life had been a literal teenage runaway. Together they proposed a mathematical model of a neuron.¹ Not the gloopy, electrochemical, half-comprehensible neuron from biology class, but an idealized version: a unit that takes some inputs, sums them with weights, and fires if the sum crosses a threshold. Build a network of these, they argued, and you could in principle compute anything a digital computer could compute. The paper was twelve pages long. The empiricist branch of AI starts there.

In 1950, Alan Turing — the same Turing of Bletchley Park and the Enigma cipher — published a paper in Mind called “Computing Machinery and Intelligence.”² It opens with one of the most famous questions in the history of science. Can machines think? Turing then refuses to answer the question and replaces it with a different one. Forget what thinking means; ask whether a machine could behave so much like a thinking person that a careful observer could not tell them apart. The setup he proposes — a human interrogator chatting via teletype with two unseen interlocutors, one human and one machine — became known as the Turing test. The test is foundational not for one camp but for both, since either could in principle pass it.

In 1956, ten researchers spent two months at Dartmouth College in New Hampshire trying to build thinking machines.³ The proposal that pulled the workshop together, drafted the year before by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon, contains the first written occurrence of the phrase artificial intelligence. The plan was ambitious to the point of absurdity: a small group of smart people, working through one summer, would crack the core problems of language, abstraction, and self-improvement. Sixty-something years later, we are still working on those problems. But the Dartmouth proposal set a template — symbolic, top-down, logic-oriented — that defined the field’s rationalist core for the next two decades.

In 1957, Frank Rosenblatt, a psychologist at Cornell, built a physical machine called the Perceptron.⁴ It was the first device that learned from examples rather than from instructions. You showed it pictures, told it whether each picture was a triangle or a circle, and over many trials it adjusted its internal weights until it could classify new pictures on its own. The New York Times described it as a machine that would soon be able to walk, talk, see, write, reproduce itself, and be conscious of its existence. The Navy, which had funded the work, was less restrained still. The empiricist branch had a public face. It also had a publicity problem we will get to in a minute.

Both seeds were now in the ground. Dartmouth was the rationalist founding myth. The Perceptron was the empiricist counter-myth. For the next decade, both branches grew in parallel, drawing from the same small pool of researchers, the same small pool of funding, and the same small pool of computers. The argument had begun.

The symbolic ascendancy

The rationalists won the first round.

It is not hard to see why. In the late 1950s, compute was tiny. Memory was tiny. Data, in the modern sense of millions of labelled examples, did not exist; it could not have been stored if it had. What the rationalist branch could do, and the empiricist branch could not, was write a program that did something specific and inspect every step of its behaviour. You could read the code, debug it, prove things about it. For two decades, the symbolic camp had not just the better results but the better tools.

In 1956, the same year as Dartmouth, Allen Newell, Herbert Simon, and Cliff Shaw built the Logic Theorist — a program that proved mathematical theorems from Principia Mathematica by searching through a tree of possible proof steps.⁵ It was the first program to do something that, when humans did it, was unambiguously called reasoning. A few years later, the same team built the General Problem Solver, which extended the trick to any problem expressible as a search through states.⁶ The thesis behind both programs was simple: cognition is search. If you can write down the goal and the rules for moving from one state to another, then thinking is a matter of exploring the resulting space efficiently. Chess, theorem proving, planning a route through a city, planning a sequence of actions to stack blocks — all of these became, briefly, the same problem.

In 1958, John McCarthy designed LISP, a programming language built around lists, recursion, and the radical idea that programs could be data and data could be programs.⁷ LISP would remain the lingua franca of symbolic AI for the next thirty years. If you wanted to write a system that reasoned about its own reasoning — a meta-program — LISP was the only practical choice. To a symbolic AI researcher in 1965, LISP was not a tool. It was a worldview.

In 1966, Joseph Weizenbaum at MIT wrote a small program called ELIZA.⁸ It pretended to be a Rogerian psychotherapist. You typed a sentence; ELIZA matched it against a small set of patterns and produced a response that mostly consisted of rephrasing your sentence as a question. I am feeling depressed today. — Why do you say you are feeling depressed today? ELIZA did not understand a single word it processed. It was a four-page program with no model of anything. And yet, Weizenbaum reported in horror, people who knew it was a four-page program with no model of anything still treated it as a confidant. They asked it for advice. They wanted privacy with it. ELIZA was the first machine to teach the field a lesson it would have to relearn six more times over the next sixty years: the surface of intelligence and the substance of intelligence are different things. We will come back to this when we discuss large language models.

The high-water mark of pure symbolic AI was a 1970 program by Terry Winograd called SHRDLU.⁹ SHRDLU lived in a simulated world consisting of coloured blocks on a table. You could talk to it in English about the blocks. You could ask it to pick up the red block and put it on top of the green pyramid. You could ask it whether the blue block was supporting anything; it would think about it and tell you. You could even ask it to define a new word — call the tall block “Bert” — and it would remember. SHRDLU spoke and understood. Inside its tiny, closed, ten-block world, it really did seem to think.

The catch, of course, was the world. SHRDLU’s blocks-world was a place where every object was knowable, every relationship enumerable, every action reversible. The real world is not like this. The real world contains rain, and grandparents, and resentment, and the smell of coffee, and the precise way light falls through a window at five in the afternoon. SHRDLU could not have known what any of these things were, and could not have been taught, because to teach it would have required listing all the rules.

In the closed world of symbols, symbols were enough. The next decade was about discovering, painfully, that the real world is not closed.

Minsky, Papert, and the burial of the perceptron

In 1969, two researchers at MIT — Marvin Minsky, one of the Dartmouth founders, and Seymour Papert, his collaborator — published a small book called Perceptrons.¹⁰ The book was a mathematical analysis of Rosenblatt’s machine. The headline result was that a single-layer perceptron, no matter how many inputs you gave it, could not compute even simple logical functions like exclusive-or.

The proof was correct. It was also, in retrospect, narrow. Minsky and Papert acknowledged in the book that a multi-layer network — perceptrons stacked on top of perceptrons — could in principle compute anything you wanted. But they noted, fairly, that nobody at the time knew how to train such a network. The book did not say neural networks were impossible. It said the version of them that anyone could currently train was sharply limited.

But the field was hungry for clarity. The Perceptron’s New York Times-flavoured hype had set up a backlash, and Perceptrons delivered it. Funding for connectionist research collapsed. Rosenblatt himself died on his 43rd birthday in July 1971, in a boating accident on Chesapeake Bay.¹¹ Backpropagation — the algorithm that would resurrect his branch of AI by making multi-layer training actually work — would not arrive at scale until 1986. Seventeen years passed in which the empiricist branch was, for practical purposes, dead.

Modern AI is built on the work of people who were not yet born when Minsky and Papert published. The reason they had to be born late, the reason their work came so much later than it could have, is that the field they would eventually return to had been near-dead for almost two decades. The symbolic camp’s victory in 1969 was real. The field paid for it.

The cost of winning the argument too hard is one of the recurring lessons of this chapter. The field will pay it again.

Expert systems and the knowledge era

By the mid-1970s, the symbolic branch had its own crisis of confidence. The General Problem Solver was not, as it turned out, a general problem solver; it required the user to hand-encode the problem in a form so contorted that you had usually already solved the problem to do so. Symbolic AI needed something to point at — something that worked, in a domain people cared about, that brought in money.

It found expert systems.

The first one to really matter was MYCIN, built at Stanford between 1972 and 1980.¹² MYCIN diagnosed bacterial infections of the blood. You typed in the patient’s symptoms; MYCIN walked through a few hundred hand-coded rules — if the patient has a fever, and the patient is over 12, and the gram stain is positive, then consider these three antibiotics — and produced a recommendation. In controlled studies, MYCIN matched or outperformed human infectious-disease specialists. It was never deployed clinically, partly for liability reasons and partly because the rule encoding was a nightmare to maintain. But it had been built, it had worked, and the news got around.

By the late 1970s, expert systems were the thing. DENDRAL identified organic molecules from mass-spectrometry data.¹³ XCON configured DEC mainframe orders and saved DEC tens of millions of dollars a year doing it.¹⁴ A new profession was born — knowledge engineer — whose job was to sit down with a domain expert, extract their decision-making rules, and translate those rules into a form a rule engine could execute. The thesis behind the whole enterprise was clean and seductive: intelligence is rules plus facts. Hire the expert. Extract the rules. Encode the rules. Ship the system.

What made expert systems feel like the future was that they were legible. You could read every rule. You could ask the system to explain why it had drawn a particular conclusion, and the explanation was a literal trace of the rules that had fired. You could audit it. You could certify it. If the system was wrong, you could find the rule that was wrong, fix it, and move on. (This is a property modern large language models do not have. We will return to it in the chapter on alignment.)

There were two problems.

The first was that common-sense knowledge — the kind of knowledge a five-year-old has and an adult mostly takes for granted — is unrepresentable in rules. Water is wet. Birds fly, except for penguins, except baby penguins, except dead penguins. Hot things burn you if you touch them, but not always, and not in all the ways the word “burn” can mean. In 1969, the cognitive scientist Satoshi Watanabe had proved a theorem he called the Ugly Duckling theorem: there is no universal feature representation under which two objects are objectively more similar than two other objects.¹⁵ Every classification depends on an inductive bias. The inductive bias of human common sense was, it turned out, not something you could write down.

The second problem was named Cyc, and it is the saddest story in the history of AI. In 1984, Doug Lenat — one of the most respected symbolic researchers alive — proposed a project to manually encode all of common-sense human knowledge into a single coherent logical knowledge base.¹⁶ He estimated it would take ten years. As of this writing, in 2026, Cyc has been under continuous development for forty-two years. It contains millions of assertions. It is, in some narrow domains, useful. It is also a monument — possibly the most thoroughly humbling monument in the history of cognitive science — to how much of what we know cannot be said.

By the end of the 1980s, expert systems had failed to generalise beyond narrow domains, knowledge engineering had revealed itself as a bottomless time sink, and the funding tide had turned. In the UK, the Lighthill Report of 1973 had already warned that AI was overpromising.¹⁷ The American collapse hit in the late 1980s, when corporate buyers of expert-systems hardware discovered the maintenance costs and quietly stopped buying.¹⁸ The Second AI Winter had begun. The rationalists had won their decade and then some. The field was tired of them.

The empirical rebellion

The pendulum started to swing back, slowly, in 1986.

That year, three researchers — David Rumelhart, Geoffrey Hinton, and Ronald Williams — published a paper in Nature describing an algorithm called backpropagation.¹⁹ The math had been worked out independently several times before, by several people, going back to the 1960s. What Rumelhart, Hinton, and Williams did was demonstrate that you could use it to train multi-layer neural networks, and that when you did, the networks learned interesting internal representations of their input. The XOR problem from Perceptrons dissolved. The field of neural networks was technically resurrected, even if it would remain marginal for another two decades.

Through the 1990s and 2000s, the empiricist branch did not score one big victory. It scored a thousand small ones. Support vector machines arrived in 1995 and quietly displaced hand-engineered classifiers across one application domain after another.²⁰ Random forests arrived in 2001 and did the same.²¹ AdaBoost arrived in 1995 and proved that an ensemble of weak learners could beat any single strong learner, given enough data.²² None of these methods was a deep neural network. All of them were statistical. All of them shared a common philosophy: do not try to figure out the rules. Try to figure out the function that maps inputs to outputs, and let the data tell you what that function looks like.

The clearest demonstration that the wind had changed came from a domain that had been the symbolic camp’s home turf — natural language processing. For three decades, NLP had been the preserve of computational linguists, who built elaborate hand-crafted grammars and parsers in an attempt to capture the structure of human speech. Then, in the late 1980s, the team at IBM working on speech recognition started replacing the linguists with statistical methods. Frederick Jelinek, who led the team, was later quoted saying — possibly never in those exact words, but in some form — that every time we fire a linguist, the system improves.²³ By the late 1990s, the textbook of the field was Manning and Schütze’s Foundations of Statistical Natural Language Processing, and the linguists, mostly, had been fired.²⁴

Why was the empiricist branch winning now? Three things changed in parallel, slowly. Compute grew, as Moore’s Law had been promising it would. Data started to appear, in volumes no one had previously imagined, because the world had built a thing called the internet. And the empiricist methods themselves — kernel methods, ensembles, eventually neural networks — were simple enough to scale with both. The methods that win, it would later be argued, are the methods that scale. Sutton’s Bitter Lesson, which we will get to in two sections, names this principle. In the 1990s, no one had yet named it. It was just happening.

By the year 2000, AI as a field had effectively split. The empiricist branch had been renamed machine learning and was doing the productive work — the work people paid for. The symbolic branch had been renamed good old-fashioned AI — GOFAI, sometimes with affection, sometimes with a sneer — and was doing the philosophical work. The argument was unresolved. But the two camps no longer shared a room.

The internet era and the data awakening

Between 2000 and 2012, empiricism was winning, but quietly. There was no public moment. There were a thousand private moments.

Search engines worked because of statistical ranking. Spam filters worked because of statistical classification. Recommender systems — for movies, for songs, for purchases — worked because of statistical factorisation. Machine translation worked, when it worked at all, because of statistical alignment between corpora in two languages. Ad auctions worked because of statistical click prediction. None of this was news. None of this was AI, in the public mind. It was just software.

The clearest public marker of the era is the contrast between two IBM showcases. In 1997, Deep Blue beat Garry Kasparov at chess.²⁵ Deep Blue was the last great public victory of the symbolic camp: alpha-beta search through a game tree, with hand-tuned position evaluations encoded by grandmasters. There were almost no learned components in it. It was a triumph of search and a triumph of hardware, but it was not a triumph of learning. Fourteen years later, in 2011, Watson won Jeopardy.²⁶ Watson looked symbolic — it answered questions in natural language, after all — but inside, it was mostly information retrieval and statistical NLP, wearing symbolic clothes. The transition between the two systems is the transition the field was making in plain sight.

A few specific data milestones from this period are worth naming. In 1998, two Stanford graduate students named Sergey Brin and Larry Page published a paper describing PageRank, the algorithm that turned Google from a research project into the company that ate the internet.²⁷ In 2006, Netflix offered a million-dollar prize to anyone who could improve its movie recommendation system by ten percent.²⁸ The three years of the Netflix Prize trained a generation of data scientists in the discipline of beating benchmarks — measuring everything, iterating fast, trusting the numbers. In 2009, Fei-Fei Li and collaborators at Stanford released ImageNet, a dataset of fourteen million labelled images across twenty thousand categories.²⁹ ImageNet looked like a research curiosity. It would turn out to be the lit fuse.

By 2010, the empiricist branch had built the data, the infrastructure, and the methods. The compute had quietly arrived in the form of graphics-card chips repurposed as numerical accelerators. Everything was in place. Something was about to give.

The deep learning earthquake

In September 2012, three researchers — Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, all at the University of Toronto — entered the ImageNet Large Scale Visual Recognition Challenge.³⁰ Their entry was a deep neural network with eight learnable layers, trained on two consumer-grade graphics cards, using a then-obscure technique called ReLU activation³¹ and a regularisation trick called dropout.³² The system was called AlexNet. Its top-five error rate on the ImageNet test set was 16 percent. The second-place entry, a more conventional hand-engineered classifier, scored 26 percent.

A ten-point gap in machine-learning benchmarks is not just an improvement. It is a different category of result. Within six months, every serious computer-vision lab in the world had pivoted to deep neural networks. Within five years, no one trained a state-of-the-art image classifier any other way. AlexNet is, by a wide margin, the single most consequential paper of modern AI. It is the moment the empiricist branch stopped being one approach among several and became the approach.

The cascade was almost too fast to track. In 2014, Sutskever and two colleagues at Google published the sequence-to-sequence architecture, which turned machine translation from a statistical alignment problem into a deep-learning problem overnight.³³ The same year, Ian Goodfellow at Montreal proposed generative adversarial networks, in which two networks — a generator and a discriminator — trained against each other until the generator could produce realistic images.³⁴ In 2015, the DeepMind group in London published DQN, a deep neural network that learned to play Atari games at a superhuman level by reinforcement alone, given nothing but the raw pixels.³⁵ Year after year, one more domain fell to the same recipe: a deep neural network, gradient descent, a lot of data.

The most public moment came in 2016. AlphaGo, also from DeepMind, played Lee Sedol — one of the strongest Go players in the world — over five televised games in Seoul.³⁶ Go had been considered, for decades, too combinatorially deep for any pure-search approach. The board has more legal positions than there are atoms in the observable universe. AlphaGo won 4–1. Inside the program were two ingredients: a deep neural network that learned to evaluate positions from millions of human games, and a Monte Carlo Tree Search loop that planned moves over the network’s evaluations. The first ingredient was empiricist. The second was symbolic. The synthesis we will land on at the end of this chapter is already visible inside AlphaGo. In 2016, almost no one was looking at it that way.

In 2017, eight researchers at Google Brain published the architecture that would underwrite the next decade. The paper was called “Attention Is All You Need” and it introduced the Transformer.[³⁷{.g key=attention}-is-all-you-need] The Transformer replaced the recurrent connections of older sequence models with a mechanism called attention, which allowed every token in a sequence to interact with every other token in parallel. Trained models converged faster. Larger models trained at all. Everything that comes after — every model in your chat window, every assistant on your phone, every generative image system — descends from this one paper. If you read only one paper from the history of AI, it should probably be this one.

By 2017, the empiricists were ascendant in every benchmark that mattered. The symbolic branch had been reduced to a holdout faction of philosophical objections, which the field, busy winning, was disinclined to take seriously.

Two years later, in 2019, Richard Sutton — one of the founders of reinforcement learning, a researcher who had spent his career on the empiricist side of the argument — published a short blog post called The Bitter Lesson.³⁸ The argument fits on one page. Over seventy years of AI research, Sutton observed, the same pattern recurs. A team of researchers spends years building a system that encodes their hard-won domain expertise. The system works, for a while. Then someone with more compute and a simpler, more general method beats it. The general method, given enough compute and data, wins every time. The lesson is bitter because it tells researchers that the clever, careful, hand-engineered insights they spend careers producing are routinely steamrolled by someone with more GPUs and less taste. Sutton called this the field’s most important pattern. The empiricist branch had its manifesto.

The Bitter Lesson is the standard the rest of this book will, gently, complicate. It is mostly right. It is not entirely right. And the thing that complicates it, as we are about to see, is what symbolic AI was good at all along.

AlphaFold, the unsung crown jewel

The most consequential AI system of the modern era is not a chatbot. It does not write poems. It does not pass any version of the Turing test. It is also not in your phone, or your browser, or your editor. It is in the bloodstream of structural biology, where it has, in the past five years, quietly compressed a fifty-year-old problem into a footnote.

The problem is this. Proteins are the working machinery of every living cell. Each protein begins life as a linear chain of amino acids — a sequence, like a sentence written in an alphabet of twenty letters. As soon as the chain is built, it folds, in milliseconds, into a precise three-dimensional shape. The shape determines the protein’s function. Get the shape right and you have an enzyme, an antibody, a signalling molecule. Get the shape wrong and you have a disease — Alzheimer’s, Huntington’s, mad cow.

Predicting the shape from the sequence is one of the oldest and most important problems in biochemistry. Christian Anfinsen received the 1972 Nobel Prize in Chemistry for arguing that the sequence alone, in principle, determines the shape.³⁹ In the half-century that followed, the field tried to make good on that promise. The rationalist approach was to simulate the physics: write down the energies of every chemical bond, every electrostatic interaction, every entanglement of the surrounding water, and search the configuration space for the minimum-energy fold. This approach was beautiful and almost completely intractable. Even for a small protein, the configuration space is astronomical and the energy landscape is treacherous.

Starting in 1994, the field ran a biennial competition called CASP — Critical Assessment of protein Structure Prediction.⁴⁰ Teams were given amino-acid sequences whose structures had been experimentally determined but not yet published, and asked to predict the structures. The scores were given in GDT, a measure of how close the predicted shape was to the experimental ground truth. By 2016, after twenty years of competition, the best methods on the hardest tier of targets — free-modelling targets, where no related structure existed to copy from — had plateaued at GDT scores around forty. The field was stuck.

In 2018, DeepMind entered CASP with a system called AlphaFold.⁴¹ It was a deep neural network — empiricist to its bones — that had absorbed every protein structure ever experimentally determined. Its CASP13 score on the hard targets was around sixty. The field took notice.

In 2020, DeepMind entered AlphaFold 2. Its median GDT on the CASP14 free-modelling targets — the same tier where the field had been stuck at forty — was above ninety-two.⁴² The fifty-year-old grand challenge of structural biology was, for practical purposes, solved. The peer-reviewed paper appeared in Nature in 2021. The predictions were released as an open database covering essentially every protein known to science.⁴³ Pharmaceutical labs, vaccine designers, and drug-target researchers worldwide woke up to a problem that had defined their working lives and discovered it had been finished while they slept.

In 2024, Demis Hassabis and John Jumper — DeepMind’s lab director and the lead author of the AlphaFold 2 paper — shared the Nobel Prize in Chemistry.⁴⁴ It is, to date, the only AI work to have produced a Nobel-level scientific breakthrough. The Nobel citation is worth reading. It is not a citation about AI as a technology. It is a citation about a problem that was solved.

The reason AlphaFold matters more than most people outside biology realise, and the reason I am giving it a section of its own in a chapter about the history of AI, is that AlphaFold is what empiricist AI looks like when it is genuinely useful rather than genuinely viral. The chatbots get the headlines. The image generators get the lawsuits. The protein folder gets the world. The book will return to AI in scientific discovery in Part II, but the seed is here: the most consequential AI of the modern era is not the one in your chat window.

The generative turn

The chat window’s turn came next.

OpenAI, a research lab founded in San Francisco in 2015, spent the late 2010s building successively larger versions of a particular kind of model — a Transformer trained on a giant pile of internet text, with a single training objective: predict the next word, given the words that came before. The first model, GPT-1, had a hundred and seventeen million parameters and was, mostly, a curiosity.⁴⁵ The second, GPT-2 in 2019, had a billion and a half; OpenAI initially declined to release it on the grounds that it was too dangerous, a decision that did not age well.⁴⁶ The third, GPT-3 in 2020, had a hundred and seventy-five billion.⁴⁷ Each iteration was the same architecture and the same objective. Each iteration surprised the field with capabilities — translation, summarisation, arithmetic, coding, mediocre poetry — that nobody had explicitly trained for. The phenomenon got a name: emergent capabilities. Scale produced behaviour that smaller versions of the same model did not exhibit.

Google’s parallel line, BERT in 2018, ran on a slightly different recipe — a bidirectional encoder rather than a left-to-right decoder — but the pattern was the same.⁴⁸ Within a year, BERT was the workhorse of mainstream natural-language processing across the industry. The empiricist paradigm was now operationalised: pretrain a Transformer on a mountain of text, fine-tune it for whatever you cared about, ship.

Then images joined the party. DALL·E in 2021, DALL·E 2 and Stable Diffusion and Midjourney in 2022.⁴⁹⁵⁰⁵¹ You could now type a sentence and get a picture. The pictures were, by any historical standard, miraculous. They were also, by any honest standard, occasionally cursed. The aesthetic discourse of the next three years would be about figuring out which.

Quietly, in 2020, two papers had given the empiricist branch something it had never had before: a theory. Kaplan and collaborators at OpenAI published the first paper on scaling laws — empirical curves showing that loss decreased predictably with model size, dataset size, and compute, over many orders of magnitude.⁵² In 2022, the Chinchilla paper from DeepMind refined the recipe, arguing that most large models had been over-parameterised and under-trained, and that the optimal balance gave you more capability per dollar.⁵³ For the first time, the field could predict what a bigger model would do before training it. Empiricism, the part of AI that had always run on intuition and luck, now had something resembling physics.

The generative era reached escape velocity on November 30, 2022, when OpenAI released a free chatbot wrapped around its latest model. The product was called ChatGPT.⁵⁴ It crossed a hundred million users in two months. It made every newspaper. The post-ChatGPT world is the world the rest of this book lives in.

The reasoning pivot

ChatGPT’s quality was not, mostly, about scale. By late 2022, several labs had models the same size as GPT-3.5. None of them felt like ChatGPT. The thing OpenAI had figured out, and that the rest of the field would spend the next year frantically reverse-engineering, was the post-training stack.

A raw pretrained language model is a very strange object. It has read everything; it has opinions on nothing. Ask it a question and it might answer, or it might continue the question, or it might generate a list of related questions, or it might break into a story about how some character once asked a similar question. It is a fluent mimic of all of internet text, including the parts where people ask questions and the parts where people fail to. To turn this mimic into something useful, you have to teach it to behave like a helpful assistant in particular.

The first move is instruction tuning — fine-tune the model on a curated set of question-answer pairs that exemplify the behaviour you want. The second move, which is where ChatGPT departed from its predecessors, is reinforcement learning from human feedback (RLHF).⁵⁵ Show the model a question; have it generate several candidate answers; have humans rank the candidates by preference; use reinforcement learning to push the model toward the higher-ranked candidates. A small amount of human preference data, multiplied by RL, turns a competent autocomplete into a useful assistant. The post-training stack now has cousins — DPO, RLAIF, constitutional methods — each of which is a different recipe for the same trick.⁵⁶⁵⁷ But the family resemblance holds.

For a year after ChatGPT, the field built on this stack and got incremental gains. Then, in late 2022 and early 2023, a separate research thread quietly converged on something stranger. Jason Wei and collaborators at Google noticed that if you prompted a language model to think step by step before answering, it answered measurably better on hard problems.⁵⁸ The trick was free at inference time; you just spent more tokens. The trick was also strange, because nothing in the model’s training had told it that thinking step by step was something it could do.

In September 2024, OpenAI released o1, the first commercial reasoning model.⁵⁹ o1 was a language model that had been further trained — using reinforcement learning on verifiable reasoning traces — to think before it answered. In some regimes, it spent more compute at inference time than it had spent at training. A new scaling law arrived: performance scaled not just with training compute but with test-time compute. The reasoning paradigm, almost overnight, became the new frontier.

In January 2025, DeepSeek-R1 arrived from a Chinese lab.⁶⁰ It was an open-weight reasoning model, trained for a fraction of o1’s reported cost, released with full technical details. The reasoning paradigm escaped the closed labs in a single afternoon. Within three months, every major lab had its own reasoning model.

By 2026, the reasoning paradigm is the default. OpenAI’s GPT-5.5, Google’s Gemini 3.5, Anthropic’s Claude Opus 4.7 — these are the frontier.⁶¹⁶²⁶³ Non-reasoning models are now the budget tier.

What is happening inside an RL-trained reasoning model is not, as far as anyone can tell, reasoning in the symbolic sense. The model has been trained to produce token sequences that look like reasoning, because those sequences correlate with correct answers, because the reward signal during RL was tied to verifiable correctness. Whether the result is reasoning or a very convincing performance of reasoning is a question we will pick up in Part III. For now, what matters is that the trick works.

The three big ones, and the open insurgency

By 2026 the institutional landscape of frontier AI has condensed.

Three labs lead. OpenAI, founded in San Francisco in 2015, runs the GPT line and its reasoning successors. Anthropic, spun out of OpenAI in 2021, runs the Claude line and has built its identity around a research-first focus on alignment and safety. Google DeepMind, formed from the 2023 merger of Google Brain and DeepMind, runs the Gemini line and has the deepest stack — chips, infrastructure, search, Android — of the three.⁶⁴ Each lab has a flagship line, a distinct design philosophy, and broadly converging capabilities at the frontier. Other capable closed labs exist — xAI, Cohere — but the gravitational centre is the three.

Training a frontier model is now a capital-intensive sport. By 2025, the training cost for a single frontier model was in the hundreds of millions of dollars, with credible reports of approaches to a billion.⁶⁵ Inference is its own infrastructure problem; some labs are running data centres whose marginal kilowatt is more valuable than the marginal kilowatt of a small city.

The other side of the institutional story is the open-weight insurgency. Meta has released the Llama series — increasingly capable open-weight models, dating from 2023 — that have anchored a vast community of fine-tuners, evaluators, and downstream researchers.⁶⁶ Mistral, a French lab founded in 2023, has shipped a series of compact and high-quality open-weight models.⁶⁷ Qwen, Alibaba’s model line, has done the same from China.⁶⁸ DeepSeek, of R1 fame, has done both. By 2026, open-weight models lag the closed frontier by perhaps six to twelve months on the strongest benchmarks. On many benchmarks the gap has closed. On all benchmarks, the open models can be run locally, fine-tuned freely, audited deeply.

Nothing in this section is the point of the chapter. But you cannot understand what is happening on the agentic frontier without knowing that three labs are pulling everyone else along, and that an open-weight insurgency is following six months behind with a different set of capabilities and a different politics. The book returns to both threads in Part II’s policy chapter and Part III’s chapters on risk.

The agentic turn

The technical centrepiece of the current moment, and the place where the seventy-year argument finally gets its synthesis, is the rise of agents. The word matters. So does its history.

The word agent did not come from machine learning. It came from classical AI. In the 1970s and 1980s, the symbolic-AI literature converged on a particular abstraction for an intelligent system: an entity that perceives its environment, deliberates over possible actions, selects one, and acts — then perceives the consequences and loops. The early planning systems were the seed; the STRIPS planner at SRI in 1971 was an early case in point. The architecture got its formal statement in Michael Bratman’s 1987 book on practical reasoning, which introduced the belief-desire-intention (BDI) framework, and in the 1991 paper by Anand Rao and Michael Georgeff that turned BDI into an engineering paradigm. By 1995, the Russell-and-Norvig AI textbook had consolidated the intelligent agent abstraction as the organising concept for the whole field.⁶⁹

The symbolic-agent project never quite worked. The shell was right; the brain was missing. The architecture — perceive, deliberate, act, observe — is sound. The agent needs to model the world to deliberate over it, and pure symbolic computation could not model the world. Cyc, again, is the long sad demonstration. The shell sat there for decades, structurally correct and operationally empty.

The empiricist branch, when it reintroduced the word agent in the 1990s and 2000s, meant something different. In reinforcement learning, an agent is a learned policy — a function that maps states to actions, trained by trial and error against a reward signal. DQN was an agent. AlphaGo was an agent. Robots in OpenAI Gym were agents. The architecture (perceive, decide, act) was the same shape; the cognition slot was now filled by a neural network rather than a logic engine. A new brain, slotted into the old shell. Useful, sometimes spectacular, but narrow. An AlphaGo cannot make you a sandwich. A DQN cannot answer a customer-service email. The brain was specific to the task.

In 2024, the cognition slot inside the agent architecture got filled a third time. This time, the brain was a general-purpose reasoning language model. The architecture is still the seventy-year-old symbolic-agent frame: perceive (read files, observe outputs, examine state), deliberate (chain-of-thought reasoning), select an action from a named, semantically-typed action space (read_file, run_tests, send_email), act, observe consequences, loop. The brain is now an LLM. This is the synthesis.⁷⁰

The reason the synthesis is more than the sum of its parts is worth stating slowly.

From the empiricist side, the LLM inherits flexibility. The model has read enough of the world that you do not have to tell it what a file is, or what a test is, or what an angry user sounds like, or what a customer who has been on hold for six minutes is probably going to say next. Common sense — the thing Cyc spent forty-two years failing to enumerate — has been silently absorbed.

From the symbolic side, the harness inherits structure. The actions in the action space have names. Each action has semantics the system designer chose. The set of possible world-impacts is bounded; the agent cannot run rm -rf unless someone explicitly wired that action into the action space and granted permission for it. The agent’s trajectory is auditable, because the actions are discrete and named, and the deliberation in between is in plain text.

The model gets to think in fluid, statistical, half-formed thoughts. The harness gets to act in crisp, named, accountable steps. The model can hallucinate; the system cannot, because the harness is the gatekeeper. The 1970s symbolic agent could never reason. The 2010s RL agent could never generalise. The 2026 LM-plus-harness agent does both — badly, often clumsily, but for the first time at the same time. This is the great synthesis.

The first concrete instantiation was tool use. In mid-2023, OpenAI introduced function calling in its API.⁷¹ You could now define a set of tools — Python functions, in effect — and pass them to the model. The model could choose to call one, observe its return value, and continue. The other labs followed within months. Tool use is the moment the LLM gets handed the action space that the symbolic-agent architecture had been waiting for decades to give to something competent.

From tool use, the trajectory has been steady. Chatbot, then chatbot with tools, then chatbot with a planning loop, then chatbot with planning and memory and a file system and a budget. Each step adds more symbolic structure around the empiricist core. The modern name for the structure around the core is the harness, and the field of harness engineering has its own emerging best practices.⁷²

A small research benchmark with a large implication is worth flagging here. The METR research group has tracked the length of task — measured in human-minutes — that a frontier AI agent can autonomously complete at a fixed reliability threshold. The result, fit over data from 2019 to 2025, is a doubling time of approximately seven months.⁷³ In early 2024, the frontier task length was about thirty minutes. By 2026, it is approaching the half-workday. If the trend continues — and we will discuss in Part III whether it should — the frontier task length crosses a full workday within two years.

The synthesis took as long as it did because both halves had to be strong enough to help each other. The symbolic shells of the 1970s had to do all the heavy lifting, because the empiricist substrates underneath were too weak to contribute. The empiricist substrates of the 2010s had no symbolic shells, because they did not need them — pure pattern-matching beat the benchmarks. Both halves had to be competent in their own right before a hybrid could be more than the sum. That moment is now.

Software development as the canary

The agents are visible in many domains by 2026. They are most visible — and most consequential, for now — in software development.

There are three reasons software development is the leading edge. First, the environment is legible to the agent: code is text, errors are text, success is a passing test. The agent can read everything in its environment. Second, the feedback loops are fast: run the code, see what happens, adjust. A loop that took a human pair-programmer ten minutes now takes the agent thirty seconds. Third, the domain has clean verifiers: compilers, type checkers, test suites, linters — programs whose job is to tell you, deterministically, whether some other program is correct. Verifiers are what make reinforcement learning with verifiable rewards work, and reinforcement learning with verifiable rewards is what made reasoning models work.

The trajectory in software is a clean curve. GitHub Copilot launched in 2021, powered by an early OpenAI model also called Codex.^{74 75} Copilot was autocomplete on steroids; it suggested the next few lines of code, and you accepted or ignored. ChatGPT-for-code arrived in 2023, and it was conversational pair-programming — paste a snippet, describe a bug, get a fix. Then, in 2024 and 2025, the agents arrived. Cognition Labs’ Devin in early 2024 was the first system to advertise itself as a fully autonomous AI software engineer.⁷⁶ Cursor turned its IDE into an agentic environment that year as well.⁷⁷ Anthropic shipped Claude Code in early 2025.⁷⁸ OpenAI shipped a cloud SWE agent — also called Codex, reusing the name but as a different beast — in May 2025.⁷⁹ Google’s Gemini CLI followed.

What these systems share is the architectural pattern. An LLM at the core — empiricist; trained on most of the publicly available code on the internet, plus a great deal of natural language. A harness around it — symbolic; exposes a file system, a shell, a small set of tools (read, write, search, run). A constraint frame around the harness — the test suite, the type checker, the linter; verifiable rewards in the literal sense. A human in the outermost loop, reviewing the diff before it merges. All four layers of the synthesis live in a system you can run on your laptop today.

What you should take from this section is not that you should download an agent and try it tomorrow, though by all means do. What you should take is that software development is the canary. The patterns that work here are spreading. They will reach research first, then education, then creative work, then law, then medicine — the domains in roughly the order in which their environments are legible and their feedback loops are fast. What you see agents doing in code in 2026 is what you will see them doing in your own field, give or take a few years.

The synthesis

The seventy-year argument did not produce a winner. It produced a synthesis.

The empiricist branch produced a substrate that is, by any reasonable measure, the most capable cognitive artefact ever built — a language model that has absorbed most of the written record of humanity and can pattern-match across it in real time. The rationalist branch produced an architecture — agent shells, action spaces, planning loops, explicit procedures — that for fifty years had no brain capable enough to make it useful. The combination, finally, is. Bolt the symbolic shell onto the empiricist substrate and you get an agent in the full classical sense: a system that perceives, deliberates, selects an action with a name and a meaning, acts on the world, and observes what happened.

The synthesis has three layers, and it is worth being explicit about each.

The first layer is the learning substrate — a model trained on the data of human civilisation, generative in form, statistical in nature. This is what people mean when they say AI in 2026, and most of the public discourse stops at this layer. Most of this book is about the other layers.

The second layer is the symbolic shell — the harness, the tools, the explicit procedures, the named action space, the verifiers. This layer is what makes the substrate accountable. It is the difference between a model that can describe how to send an email and a system that can send the email, named, logged, and reversible.

The third layer is the human frame — the ethical, legal, epistemological boundaries that decide what the system is for and how to evaluate whether it has done its job. The third layer is where this book lives. It is what the techno-pragmatist insists on, and what the manifesto in the previous chapter named: the future is not predetermined. We have the power to decide, and the responsibility to decide well. The third layer is also where every honest argument about AI in 2026 takes place. The first two layers are engineering. The third layer is the question of what we want the engineering to do.

The rest of this book opens the synthesis layer by layer. Part I, which begins on the next page, takes the first two — the empiricist substrate and the symbolic shell — and unpacks them mechanism by mechanism. Part II shows what the synthesis can do, in the domains I happen to know best. Part III names what it can break, and what we should do about it.

The synthesis exists. What we do with it is still up to us.