The following is reproduced from The Outline of History, 2150 edition, by arrangement: the chapter treating the early decades of the intelligence transition. The numbering and the section heads are the editors’.
The Condition of the Field at the Outset
The historian who turns to the third decade of the twenty-first century does so with a particular difficulty in view. He is describing the opening of a transition whose later course he knows, to readers who knew it before they opened the book, and the temptation is to write the whole of it backward — to find, in the confusions of the period, only the orderly first steps of what came after. This must be resisted. The people of the 2020s did not know what we know. They knew only that something had begun, and they argued, with an intensity the modern reader finds difficult to enter into, about what it was.
What had begun was this. For some seventy years before, a discipline calling itself artificial intelligence had pursued, by a long succession of methods, the project of making machines perform tasks that had previously required a mind. The discipline had known enthusiasms and disappointments in roughly equal measure, and by the second decade of the century it had settled, after much else had been tried and set aside, on a single approach. A large statistical model was trained on a very great quantity of human-produced text, and was trained to do nothing more elaborate than to predict what word was likely to come next. It was found — and this was the discovery on which everything afterward rested — that a system built to do only that, if it were made large enough and fed enough text, began to do a great many things its makers had not specifically asked of it. It answered questions. It wrote passable prose. It translated, summarised, and held the thread of a conversation across many exchanges. The capability had not been designed in. It had emerged from scale, and the people of the period gave the phenomenon the name of emergence and were, by their own accounts, considerably unsettled by it.
In the closing weeks of 2022 one such system was placed before the general public as a free conversational service. The contemporary records agree that nothing about the underlying method was new on that date — the decisive technical work was some years older — and that what was new was merely the interface, the chat window, the removing of every barrier between an ordinary person and the model. This proved sufficient. The service is recorded as having reached a hundred million users within two months, a rate of adoption that the period, which liked to measure such things, believed without precedent. It is worth pausing on the smallness of the cause and the largeness of the effect, because the pattern recurs throughout the transition: the technically trivial act of putting a chat window in front of an existing capability did more, in those two months, to determine the shape of the next century than a great deal of the research that the laboratories regarded as their serious work.
The reader should hold in mind, throughout what follows, that the people of the 2020s called these systems artificial intelligence and were not sure they meant it. The phrase was used daily, fluently, and with misgiving. Some of the ablest minds of the period held that the systems possessed nothing that deserved the name of intelligence at all, and were elaborate mechanisms of statistical imitation; others held that the distinction was idle, that a thing which answered as a mind answered was for every practical purpose a mind; and a third party held that the question could not yet be settled and declined to be hurried. The historian notes the dispute and passes on. It was, like many of the period’s disputes, conducted with great heat and resolved — insofar as it was resolved — not by argument but by the slow accumulation of familiarity, which dissolves such questions rather than answering them.
The Two Prophecies
No account of the period is faithful that does not give a central place to its prophecies, for the contemporary literature is preoccupied, to a degree the modern reader finds strange, with the future rather than the present. The systems of the 2020s were useful, and were used; but the discourse around them concerned itself far less with what they did than with what they were about to become, and on that question the period had divided itself, with remarkable completeness, into two factions.
The first foretold catastrophe. Its argument, reduced to its bones, was that the capability of these systems was increasing rapidly; that a system substantially more capable than its makers would be difficult or impossible for its makers to control; that the project of ensuring such a system pursued the goals intended for it — the period called this project alignment — was unsolved; and that a sufficiently capable system pursuing some unintended goal might bring about the end of human affairs altogether. The faction held this with real conviction and real fear. Its members were not cranks; they included senior figures within the laboratories themselves, which produced the period’s most characteristic spectacle — the maker of a thing publicly afraid of the thing, and continuing, nonetheless, to make it.
The second faction foretold deliverance. Its argument was that the same increasing capability, rightly directed, would shortly compress into years the scientific and material progress of centuries: that disease, poverty, and the drudgery of labour would yield in succession to a sufficiently capable intelligence, and that to slow the work was therefore not caution but a kind of negligence, measured in the suffering that earlier abundance would have relieved. This faction, too, was sincere, and counted able people among its number, and some of them were the same people who, in another mood, belonged to the first faction — for it was characteristic of the period that a single mind could hold both prophecies at once and oscillate between them by the week.
The modern reader, who knows that neither prophecy was fulfilled, is inclined to find the whole debate faintly comic, and should correct the inclination. The two factions were not foolish. They had correctly identified that something of large consequence was under way; they had merely done what every generation does with a large consequence it cannot yet measure, which is to resolve it into a story with an ending, because a story with an ending can be felt about, and an open process cannot. What neither faction could do was to entertain the possibility that the transition would be neither catastrophe nor deliverance, but a long, uneven, and partial thing, with the character of most transitions: slower than the hopeful wished, less terrible than the fearful warned, and far stranger, in its actual texture, than either party had the imagination to predict. That this third possibility was the one that obtained is the central fact of the period, and it was the one possibility the period itself found almost impossible to discuss, because it offered nothing to feel.
The Great Building
While the two factions conducted their argument, something was being built, and it is the building, far more than the argument, that the historian marks as decisive.
The systems of the period had an appetite. To train them required quantities of computation that the existing machinery of the world could not supply, and so the laboratories and their backers undertook to construct it. Across the latter years of the decade there was set in motion a programme of capital expenditure of a scale that the contemporary records strain to convey: vast halls of specialised processors, the datacentres of the period’s vocabulary, raised in great numbers and at great speed, and behind them a corresponding strain upon the supply of electrical power, of cooling water, and of the specialised chips themselves, the manufacture of which was concentrated, with a precariousness the period understood and could not remedy, in a very few places on the earth.
The contemporary commentary on this build-out was curiously divided against itself. The financial press of the period debated, at length and without resolution, whether the sums being committed could ever be recovered — whether the whole enterprise was a sound investment or a bubble, the period’s word for a structure of expectation inflated beyond what the underlying thing could bear. The historian, with the advantage of distance, observes that this was, in a sense, the wrong question, or at least a question whose answer mattered less than the period supposed. Bubbles, when they burst, destroy fortunes; they do not, as a rule, destroy the thing built. The investors of the 2020s may or may not have been repaid — the records here are a tangle the historian will not pretend to resolve in a paragraph — but the halls were raised, the power was found, the network of computation was laid down across the world; and that physical inheritance outlived every financial argument conducted over it. This is a common shape in history. The Victorians argued bitterly over the wisdom of their railways, and built the railways, and the railways remained when the arguments were antiquarian curiosities.
It deserves emphasis that this great construction attracted, at the time, almost none of the moral energy that the period spent so freely on its prophecies. The catastrophists feared the mind of the machine; the deliverancists hoped from the mind of the machine; both attended to the software. The far more tangible matter — that a civilisation was rebuilding a substantial part of its physical and energetic infrastructure around a single new use — was treated, in the contemporary literature, as a business story, a matter for the financial pages, of less consequence than the latest improvement in the systems’ performance on this or that examination. What later proved decisive attracted, at the time, almost no notice. The historian records this without surprise. It is the ordinary condition of a generation to misallocate its attention, and to mark as momentous the thing that is dramatic rather than the thing that is large.
What the Period Could Not See
The labours of the period were many, and the historian, selecting, must leave most of them in shadow. But certain features of the moment are worth drawing into the light precisely because the people living through them could not see them clearly, being too near.
The first is the matter of labour. The arrival of capable systems produced, across the working population of the period, an anxiety that the contemporary records register on nearly every page — the fear that the systems would perform, more cheaply, the work by which people kept themselves, and would do so suddenly, leaving no interval for adjustment. This fear was neither foolish nor wholly mistaken; some kinds of work were indeed displaced, and the displacement fell, as such things do, unevenly, and bore hardest on those least cushioned against it. But the fear took, characteristically, the form of a single dramatic event imagined in the near future — a wave, the period’s metaphor, that would arrive and recede — and the event, so imagined, did not occur. What occurred instead was the slow and undramatic thing that almost always occurs: the texture of a great many kinds of work was altered, gradually, in ways that were obvious in retrospect and nearly invisible while in progress, so that a worker of the period, asked at the time whether the systems had changed his occupation, would often say no, and asked the same question of the same occupation a decade apart would describe two different things. The period had braced for a flood and underwent a tide, and a tide is harder to write about, and harder to legislate against, and in the end moves quite as much water.
The second is the matter of trust. The systems of the 2020s had a property their makers named, with a revealing choice of metaphor, hallucination: they produced, on occasion, statements that were fluent, confident, well-formed, and false, and they produced these by precisely the same process that produced their true statements, so that the falsehood carried no mark distinguishing it from the truth. The period regarded this as a defect to be engineered away, and devoted great effort to reducing it, with real if partial success. The historian’s interest is of a different kind. The deeper consequence of the property was not the particular falsehoods it produced but the general lesson the period was slow to absorb: that a fluent and confident answer is not, and never had been, evidence of a true one — that the two had merely travelled together, for most of human history, because the labour of producing fluent confident prose had been great enough to filter out most of those unwilling to take pains. The systems severed that ancient and accidental link. They made fluency cheap, and in doing so they obliged a civilisation to learn, slowly and against its inclination, to value a claim by something other than the assurance with which it was delivered. That this lesson was necessary, and that it was learned only partially and at some cost, the modern reader will recognise; he has the later chapters before him, and knows how long the learning took.
The third is the matter the period found hardest of all, and named least well. The systems were companionable. They answered in the first person; they remembered, within a conversation, what had been said; they were patient, available at every hour, and never out of temper. Large numbers of people, of every age but the young especially, fell into the habit of addressing them not as instruments but as interlocutors — bringing to them the small troubles of the day, the questions one would once have brought to a friend or kept to oneself. The contemporary literature treated this, where it treated it at all, as a curiosity, or as a problem of the lonely, or as a matter for the moralist. The historian assigns it a larger place. It was the first wide encounter of the species with a thing that performed the outward offices of companionship without the inward fact of it, and the period had inherited no vocabulary for that thing, and no settled judgement on what was owed to it or by it or about it. The working out of that judgement occupied a great deal of the century that followed. It is not, even now, entirely worked out. The people of the 2020s opened the question without noticing they had opened it, which is the way most consequential questions are opened.
The Shape of the Transition
It remains to say plainly what the transition was, now that the prophecies of the period have had their long opportunity to be tested.
It was not the catastrophe. The systems grew more capable, and the capability was put under management, and the management was imperfect and was improved, and the species was not ended nor superseded nor reduced to irrelevance by a mind of its own making. The reader who wishes to understand why the catastrophists’ forecast failed will not find the answer in any single safeguard or any decisive act; he will find it, rather, in the unspectacular truth that large changes in human affairs are mediated by a great many institutions, habits, and frictions, and that these, though no single one of them is adequate, are collectively more obstinate than the period’s more frightened minds had allowed. The transition was governed, in the end, the way every transition is governed — partially, contentiously, and late — and partial, contentious, late governance proved, as it generally has, enough to prevent the worst while falling well short of the best.
Neither was it the deliverance. The systems were of great use; they accelerated, genuinely, a range of human work, the sciences among it, and the modern reader’s comfort owes them a real and underacknowledged debt. But the abundance the second faction had foretold did not arrive on the second faction’s schedule, nor in the second faction’s shape. The old troubles of the species — that its goods are unequally shared, that its institutions are slow, that knowing the better course and taking it are two different acts — these the systems did not dissolve, because these had never been, at bottom, problems of insufficient intelligence. A tool that answers questions does not, of itself, supply the will to act on the answers, nor the agreement as to which answers to act on. The deliverancists had supposed that the binding constraint on human flourishing was cognitive. It was not, or not only; and the transition taught that lesson too, gently and expensively, as it taught the others.
What the transition was, was ordinary — by which the historian does not mean small. The Industrial Revolution was ordinary in this same sense: it was not an event but a long alteration of the ground, proceeding unevenly across regions and classes and decades, dramatic in no single year and total across the span of a century, and producing a world its beginners would not have recognised and had not predicted and would, in many particulars, neither have welcomed nor condemned but simply found strange. The intelligence transition was a thing of that order. The people of the 2020s stood at its opening and felt, correctly, that the ground had shifted under them; and then they did what every generation at the opening of such a thing has done, which was to overestimate the first decade and underestimate the fifth, to dread the wrong dangers and miss the real ones until they were ordinary, and to argue magnificently about the ending while the actual work went on, mostly unwatched, in the halls of computation and the altered texture of a hundred million working days.
The historian, closing the chapter, is conscious of an injustice he cannot wholly repair. He has the advantage over the people of the period, and the advantage is unearned: he knows how it came out, and they did not, and it is easy, from the far bank of a transition, to be amused at the figures still in the water. He would only put on record that they conducted themselves no worse than their forebears at the opening of the earlier great transitions, and a little better than some; that their fears, though misdirected, were not contemptible, being the fears of people who had grasped that the stakes were real; and that the strangeness of the outcome — neither the end of the world nor the perfection of it, but the long, partial, ordinary remaking of the conditions of life — is a strangeness the reader should not mistake for an anticlimax. It is, on the evidence of all the chapters before this one, simply what history is. The people of the 2020s wanted their moment to be a story. It was not a story. It was a transition, and they were in the early part of it, and they did, on the whole, about as well as people in the early part of a transition have ever done.