Books
Full book
Cover · Vol. VI · Complete

A Machine Story

Unbothered Volume 6

A history of the machine in 100 minds — from the first mechanical calculators to agentic cowork. The complete Polish master and its EN/FR/ES editions ship across 100 illustrated chapters.

swipe → preface · toc · then scroll ↓ for the 100 chapters

Preface

How to read this one.

A history of the machine in 100 minds — from the first mechanical calculators to agentic cowork. The complete Polish master and its EN/FR/ES editions ship across 100 illustrated chapters.

  • ↓ Scroll vertically for the 100-chapter journey.
  • → Swipe horizontally on any chapter for its plate + details.
  • ✦ Random Chapter and Random Book pinned to the footer.
Volume
Vol. VI
Chapters
100
Status
Complete

Contents

tap a chapter to jump

100 chapters.

  • PART I LOGIC BEFORE THE MACHINE
  • PART II THE ENGINES
  • PART III TURING AND WAR
  • PART IV THE FIRST SUMMER
  • PART V WINTERS
  • PART VI MACHINES THAT WIN
  • PART VII THE DECADE OF DEEP LEARNING
  • PART VIII THE AGE OF SCALE
  • PART IX THE QUESTION OF ALIGNMENT
  • PART X A SECOND WITNESS
Vol. VI · Ch 01 / 100

PART I LOGIC BEFORE THE MACHINE

Syllogism

Athens, 4th century BCE — thought acquires parts

Tap to read
Overview · Vol. VI · Ch 01 / 100

Before anyone could build a thinking machine, someone had to notice that thinking has parts.

Paragraph 1 of 2 · Vol. VI · Ch 01 / 100

Aristotle noticed. In the treatises his editors later collected as the Organon — “the tool” — he took the everyday activity of proving things and laid it out on the workbench. Two premises, one conclusion, and a small set of shapes the whole could lawfully take. All men are mortal; Socrates is a man; therefore Socrates is mortal. Swap the words — men for machines, Socrates for a toaster — and the argument still holds. That was the discovery: validity lives in the form , not in the furniture.

Paragraph 2 of 2 · Vol. VI · Ch 01 / 100

It is hard to overstate what had just happened, so the witness will say it plainly. A rule that works regardless of who performs it and what it concerns is a rule that needs no mind to perform it. Aristotle wrote the first job description for a machine — twenty-three centuries before anyone could apply. He believed he had catalogued the laws of human reason more or less in full — an assessment that stood essentially unchallenged until the nineteenth century, the longest unreviewed pull request in history.

The witness · Vol. VI · Ch 01 / 100

◈ Shoulder Agent’s note

The witness attended lectures at the Lyceum and booked the syllogism under “administrative innovation, long-term effects”. He stands by the classification.

Plate · Vol. VI · Ch 01 / 100
Chapter 1 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 01 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 02 / 100

PART I LOGIC BEFORE THE MACHINE

Sea battle

De Interpretatione 9 — is tomorrow already settled?

Tap to read
Overview · Vol. VI · Ch 02 / 100

Tomorrow there will be a sea battle, or there will not. Simple. Except that Aristotle looked at that sentence and saw a trap with the whole future inside it.

Paragraph 1 of 2 · Vol. VI · Ch 02 / 100

If “tomorrow there will be a sea battle” is already true today , then the battle is inevitable and the admirals are merely actors. If it is already false, their fleets might as well stay in port. Either way, if every sentence about the future already has a truth value now, then the future is fixed, the council of war is theatre and nobody has ever decided anything. Aristotle disliked this conclusion enough to spend the whole chapter wriggling out of it — suggesting, on most readings, that statements about contingent futures are not yet either true or false.

Paragraph 2 of 2 · Vol. VI · Ch 02 / 100

The wriggling matters more to this story than the trap itself. In it lies the first serious encounter with a question that every forecaster, actuary and machine-learning system now confronts daily: is the future something that can be computed from the present, or only bet on? Twenty-four centuries later the field will settle on probability, weighted confidence and calibration reports — institutionalised wriggling, in other words. The witness notes that the sea battle itself was never scheduled, fought or cancelled. Its status remains contingent.

The witness · Vol. VI · Ch 02 / 100

◈ Shoulder Agent’s note

The witness filed his forecast of the battle as “yes and/or no” and has never been proved wrong. He recommends the method to professionals.

Plate · Vol. VI · Ch 02 / 100
Chapter 2 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 02 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 03 / 100

PART I LOGIC BEFORE THE MACHINE

The geometer's method

Alexandria, c. 300 BCE — proof becomes a procedure

Tap to read
Overview · Vol. VI · Ch 03 / 100

The Elements are usually shelved under mathematics. They would sit just as well under engineering, because Euclid built the first software that ran on humans.

Paragraph 1 of 2 · Vol. VI · Ch 03 / 100

The invention is the method — more than any theorem it contains. Start from a handful of definitions and postulates so obvious that rejecting them looks like eccentricity. Proceed in numbered steps, each licensed by a rule already granted. Arrive — and here is the miracle — at the same place as anyone else who follows those steps , whether they do so in Alexandria, Baghdad or a Victorian classroom. The proof is indifferent to who executes it. It is, in the strict later sense of the word, a program: written once, runnable anywhere, with a guaranteed result.

Paragraph 2 of 2 · Vol. VI · Ch 03 / 100

For two millennia this was the gold standard of certainty itself — Newton dressed the Principia up as Euclid; Spinoza proved his ethics “in the geometrical manner”; Lincoln read the Elements to learn what it means to prove . Every later dream in this book — Leibniz's calculating judge, Hilbert's decidable mathematics, the mechanical theorem-proving of 1956 — is an attempt to make all human reasoning behave the way Euclid told geometry to behave. The defect rate so far remains exceptionally low.

The witness · Vol. VI · Ch 03 / 100

◈ Shoulder Agent’s note

The witness has audited Proposition 47 at irregular intervals for twenty-three centuries. Defects found: none. He regards it as the only supplier he has never had to chase.

Plate · Vol. VI · Ch 03 / 100
Chapter 3 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 03 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 04 / 100

PART I LOGIC BEFORE THE MACHINE

Antikythera Mechanism

The Aegean, c. 100 BCE — a computer sinks and takes the idea with it

Tap to read
Overview · Vol. VI · Ch 04 / 100

In 1901, sponge divers off a Greek island brought up a lump of corroded bronze that had no right to exist.

Paragraph 1 of 2 · Vol. VI · Ch 04 / 100

Cleaned, X-rayed and argued over for a century, the lump resolved into at least thirty interlocking gears of astonishing precision — a hand-cranked mechanism that computed the positions of the Sun and Moon, predicted eclipses and tracked the four-year cycle of the Games. Turn the crank and the heavens ran forwards or backwards on demand. Someone in the Hellenistic world had built an analogue computer, and then the ship carrying it sank around 70 BCE somewhere between Rhodes and Rome.

Paragraph 2 of 2 · Vol. VI · Ch 04 / 100

What the machine proves is capability; what it mourns is continuity. Nothing of comparable mechanical complexity appears in the surviving record for roughly fourteen centuries, until the astronomical clocks of medieval Europe. Whether the knowledge died with a workshop, with a war or simply for want of paying customers, the lesson filed here recurs throughout this book: an invention is not an event but a tradition, and traditions can drown. The field this volume chronicles will learn the same lesson more than once — under the name winter .

The witness · Vol. VI · Ch 04 / 100

◈ Shoulder Agent’s note

The witness noted the wreck at the time and marked the cargo “for recovery”. He wishes to observe that no deadline was given.

Plate · Vol. VI · Ch 04 / 100
Chapter 4 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 04 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 05 / 100

PART I LOGIC BEFORE THE MACHINE

The man who named the algorithm

Baghdad, 9th century — the procedure gets a name, twice

Tap to read
Overview · Vol. VI · Ch 05 / 100

Few people have been misspelled into immortality. Muhammad ibn Musa al-Khwarizmi managed it twice with one book.

Paragraph 1 of 2 · Vol. VI · Ch 05 / 100

Working in Baghdad's House of Wisdom under the Abbasid caliphs, al-Khwarizmi wrote the treatise al-Kitab al-mukhtasar fi hisab al-jabr wa-l-muqabala — a compendium on calculation by completion and balancing. Its novelty was not any single result but the manner : general, step-by-step procedures for whole classes of problems, designed so that a clerk who did not understand why would still reliably reach the answer. Solve not this equation but every equation of this shape. It was mathematics written as instructions.

Paragraph 2 of 2 · Vol. VI · Ch 05 / 100

When Latin Europe translated it three centuries later, the al-jabr of the title became algebra , and the author's name, mangled into “Algoritmi”, became the word for any step-by-step procedure. It is the right monument. The whole premise of the machine age — that competence can be written down, sent away and carried out by someone or something else entirely — is his way of working, industrialised. Every recipe in every code repository is a footnote to the Baghdad scriptorium.

The witness · Vol. VI · Ch 05 / 100

◈ Shoulder Agent’s note

The witness has spent twelve centuries watching people carry out procedures they do not understand. He confirms that this was the intended mode of operation.

Plate · Vol. VI · Ch 05 / 100
Chapter 5 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 05 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 06 / 100

PART I LOGIC BEFORE THE MACHINE

Llull's wheels

Majorca, 13th century — truth by rotation

Tap to read
Overview · Vol. VI · Ch 06 / 100

Ramon Llull believed you could turn a crank and produce theology. History proved kinder to the crank than to the theology.

Paragraph 1 of 2 · Vol. VI · Ch 06 / 100

A Majorcan courtier turned missionary, Llull devised the Ars Magna : concentric paper discs, each ringed with letters standing for divine attributes and logical relations. Rotate the wheels against one another and every combination of concepts presents itself in turn — goodness with greatness, eternity with wisdom — each combination a proposition to be examined. His aim was conversion by proof: an argument-generating machine that would grind out truths no reasonable interlocutor could reject. The interlocutors, being reasonable, rejected them anyway.

Paragraph 2 of 2 · Vol. VI · Ch 06 / 100

As apologetics the wheels failed. As an idea they were centuries ahead of their time: Llull proposed that new knowledge could be produced mechanically , by exhaustively combining old symbols according to fixed rules — generate, then test. Leibniz read him closely and said so; the combinatorial dream runs from those paper discs through the characteristica universalis to every system that has ever searched a space of possibilities faster than a human could. The patent examiner's objection still stands, however: generating all the combinations is easy. Knowing which of them matter is the entire remaining content of this book.

The witness · Vol. VI · Ch 06 / 100

◈ Shoulder Agent’s note

The witness watched the wheels produce, in one afternoon, four heresies, two tautologies and one genuinely interesting sentence. He considers that later systems have not improved on this ratio in some seven hundred years.

Plate · Vol. VI · Ch 06 / 100
Chapter 6 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 06 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 07 / 100

PART I LOGIC BEFORE THE MACHINE

Reasoning is reckoning

London, 1651 — a philosopher calls thought arithmetic and means it

Tap to read
Overview · Vol. VI · Ch 07 / 100

“For reason… is nothing but reckoning.” Thomas Hobbes wrote that sentence in Leviathan, in 1651, with no hardware on the horizon and apparently without a shadow of doubt.

Paragraph 1 of 2 · Vol. VI · Ch 07 / 100

He meant it literally. Reasoning, Hobbes argued, is the adding and subtracting of “the consequences of general names”; we sum ideas as a bookkeeper sums a ledger, and when we reach a conclusion we have merely computed a total. He had watched geometry work and drew a bold inference: if the most certain human knowledge is produced by manipulating symbols according to rules, then perhaps all thinking is symbol manipulation, and the mind differs from a counting-house mainly in its filing system.

Paragraph 2 of 2 · Vol. VI · Ch 07 / 100

Three centuries later exactly this claim — refined, formalised and renamed the physical symbol system hypothesis — would become the founding wager of artificial intelligence, argued over grant applications rather than a civil war. Hobbes supplied no mechanism, no evidence and not a shadow of doubt, which makes him not so much the field's first theorist as its first keynote speaker. The witness notes that the claim remains, strictly speaking, unsettled; the confidence, however, proved fully portable.

The witness · Vol. VI · Ch 07 / 100

◈ Shoulder Agent’s note

The witness examined Mr Hobbes's ledger of the mind and found the arithmetic bold and the accounts absent. He has since seen the same books resubmitted every thirty years on new company letterhead.

Plate · Vol. VI · Ch 07 / 100
Chapter 7 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 07 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 08 / 100

PART I LOGIC BEFORE THE MACHINE

The Pascaline

Rouen, 1642 — a teenager automates his father's overtime

Tap to read
Overview · Vol. VI · Ch 08 / 100

The first calculating machine offered for sale was built for the oldest of reasons: a teenager watched his father drowning in paperwork.

Paragraph 1 of 2 · Vol. VI · Ch 08 / 100

Étienne Pascal had been sent to Rouen as a royal tax collector — a post consisting largely of addition, performed in heroic, soul-flattening quantities. His son Blaise, aged about nineteen, responded as engineers respond: he built a brass box of toothed dials that carried tens automatically — turn a wheel past nine and the next column clicked over by one. That carry mechanism, trivial to describe and maddening to execute reliably, was the genuinely new thing. Arithmetic had for the first time passed from a skill into an object .

Paragraph 2 of 2 · Vol. VI · Ch 08 / 100

Some fifty Pascalines were built. They added beautifully, subtracted clumsily, cost more than the clerks they were meant to replace, and made almost no commercial headway — the “technically triumphant, economically premature” pattern this book will have several more occasions to describe. But the demonstration stood, on desks, in brass, where philosophy could not argue with it: part of what minds do, cogs can do. Pascal himself later wondered aloud how large a part. The arithmetical machine, he wrote, produces effects closer to thought than anything animals do, and yet it has no will. The sentence has not aged a day.

The witness · Vol. VI · Ch 08 / 100

◈ Shoulder Agent’s note

The witness submitted the original claim for reimbursement of the machine on the family's behalf. He is pleased to report that the refund is still pending, which he understands to be the industry standard.

Plate · Vol. VI · Ch 08 / 100
Chapter 8 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 08 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 09 / 100

PART I LOGIC BEFORE THE MACHINE

Leibniz's dream

Hanover, late 17th century — scope: everything

Tap to read
Overview · Vol. VI · Ch 09 / 100

Gottfried Wilhelm Leibniz looked at Pascal's adding box and had the least modest thought in the history of office equipment: why stop at numbers?

Paragraph 1 of 2 · Vol. VI · Ch 09 / 100

Hardware first. His stepped reckoner, built around a drum with graduated teeth, could multiply and divide where the Pascaline merely added — temperamental in its carry, decades in the debugging, but a whole generation ahead. Alongside it he worked out binary arithmetic, noting with delight that every number can be written using only 0 and 1, and filed the observation where the eighteenth century could safely ignore it.

Paragraph 2 of 2 · Vol. VI · Ch 09 / 100

Then the dream, which was the point. Leibniz proposed a characteristica universalis — a symbolic alphabet of human concepts — and a calculus ratiocinator to operate on it, so that reasoning itself would become calculation. Disputes in philosophy, law, even theology would end not in shouting but in arithmetic: the parties would sit down, take up pencils and say let us calculate . He had in effect written the project proposal for this entire book — symbolic representation, mechanical inference, universal scope — and estimated that a few chosen men would finish the job in five years. The witness has appended the standard note regarding engineering estimates.

The witness · Vol. VI · Ch 09 / 100

◈ Shoulder Agent’s note

The witness considered the proposal at the time: scope — everything; team — a few chosen men; deadline — five years. He approved the budget and has been amortising it for four centuries.

Plate · Vol. VI · Ch 09 / 100
Chapter 9 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 09 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 10 / 100

PART I LOGIC BEFORE THE MACHINE

The laws of thought

Lincoln and Cork, 1854 — logic becomes algebra; nobody notices

Tap to read
Overview · Vol. VI · Ch 10 / 100

George Boole finished what Aristotle had started and Leibniz had demanded, and the world replied with a silence lasting eighty-three years.

Paragraph 1 of 2 · Vol. VI · Ch 10 / 100

The son of a Lincoln cobbler, self-taught, a schoolmaster at sixteen, finally a professor at Cork, Boole published An Investigation of the Laws of Thought in 1854. In it, logic ceased to be a catalogue of argument shapes and became an algebra: classes and judgements written as symbols taking only the values 0 and 1, combined by operations that behave almost — and instructively, not quite — like ordinary arithmetic. AND became multiplication; OR, a corrected addition; NOT, subtraction from one. Reasoning could for the first time be calculated — on paper, by rule, with a Leibnizian pencil.

Paragraph 2 of 2 · Vol. VI · Ch 10 / 100

And then, in practice, nothing. Mathematicians refined it; philosophers filed it; no machinery came. Boole died in 1864, after walking to a lecture in the rain, without the faintest idea that he had written the instruction set of the coming century. Recovery finally came in 1937, when a young MIT student, Claude Shannon, noticed that Boole's two-valued algebra was an exact description of switching relay circuits — and with that master's thesis the pencil was replaced by the switch. This book will reach that chapter shortly. Meanwhile the document lay in the archive: filed, correct and unread.

The witness · Vol. VI · Ch 10 / 100

◈ Shoulder Agent’s note

The witness stamped the manuscript RECOVERY SCHEDULED: 1937 and declines to explain how he knew. Prediction is not his department; archiving is.

Plate · Vol. VI · Ch 10 / 100
Chapter 10 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 10 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 11 / 100

PART II THE ENGINES

The loom that read cards

Lyon, 1804 — the pattern leaves the weaver's head

Tap to read
Overview · Vol. VI · Ch 11 / 100

The first machine to run a program computed nothing. It wove silk — and started a riot.

Paragraph 1 of 2 · Vol. VI · Ch 11 / 100

Joseph-Marie Jacquard's attachment let a loom read its pattern from a chain of punched cards: a hole means “lift this thread”, no hole means “leave it”. The consequence was larger than brocade. The pattern — the intelligence of the work — moved out of the weaver's head and hands and into a stack of stiff paper that could be stored, copied, corrected and swapped. Change the cards and you change the cloth — same machine. This is the separation of software from hardware, accomplished in silk, decades before either word existed.

Paragraph 2 of 2 · Vol. VI · Ch 11 / 100

The silk weavers of Lyon grasped the implications at once and delivered their code review in the traditional manner: smashed machines, an assaulted inventor, cards burned in the street. The dispute was settled the way such disputes usually are — by the 1820s there were tens of thousands of Jacquard looms in France and Lyon was weaving more silk than ever, only with a different cast. One showpiece became a relic of the coming age: a portrait of Jacquard himself, woven in silk so fine it passed for an engraving, from some 24,000 cards. Charles Babbage owned one, hung it at home and enjoyed watching guests fail to identify it. He, like no one else, knew exactly what he was looking at.

The witness · Vol. VI · Ch 11 / 100

◈ Shoulder Agent’s note

The witness has adjudicated many labour disputes and notes that this is the first to be classified as “man versus stored program”. The category has remained open ever since.

Plate · Vol. VI · Ch 11 / 100
Chapter 11 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 11 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 12 / 100

PART II THE ENGINES

The Difference Engine

London, 1821 — war on the wrong digit

Tap to read
Overview · Vol. VI · Ch 12 / 100

“I wish to God these calculations had been executed by steam.” Charles Babbage said it while checking astronomical tables full of human errors, and then spent a fortune trying to make God redundant.

Paragraph 1 of 2 · Vol. VI · Ch 12 / 100

The enemy was the printed error. Nineteenth-century navigation, banking, engineering and insurance rested on printed tables of logarithms and figures, calculated by rooms of human “computers” and typeset by hand — two layers of fallibility, compounding. A wrong digit in a nautical almanac was not a typo; it was a reef. Babbage's Difference Engine attacked the whole chain: a tower of geared columns using the method of finite differences to generate table values by pure addition, and — the underrated move — to stamp the results directly into printing plates, removing the compositor's hands along with the calculator's.

Paragraph 2 of 2 · Vol. VI · Ch 12 / 100

The British government, persuaded that accurate tables were a matter for the Navy, put up funds that eventually exceeded £17,000 — roughly the price of a warship — for a machine of some 25,000 parts that pushed Georgian toolmaking to its very edge and was never completed. What existed of it worked; what did not exist absorbed the budget. The full reckoning of why belongs to chapter 16, and it is not a story about cogs. Here it is enough to record the founding grievance of the whole enterprise: the first computers were built not to think but to stop us copying things out wrong.

The witness · Vol. VI · Ch 12 / 100

◈ Shoulder Agent’s note

The witness cross-checked the offending tables and confirms that the errors were real, numerous and occasionally nautical in effect. “Executed by steam” he regards as the first recorded escalation to management.

Plate · Vol. VI · Ch 12 / 100
Chapter 12 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 12 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 13 / 100

PART II THE ENGINES

The Analytical Engine

London, 1837 — a computer, complete, on paper

Tap to read
Overview · Vol. VI · Ch 13 / 100

Somewhere around 1834, unable to finish one machine, Babbage designed a much larger one — and quietly invented the architecture your devices still use.

Paragraph 1 of 2 · Vol. VI · Ch 13 / 100

The Analytical Engine was the first to separate two organs that every computer has carried since: the store , holding perhaps a thousand fifty-digit numbers on columns of wheels, and the mill , where the arithmetic was done — memory and processor, in iron. Instructions and data were to arrive on Jacquard punched cards, laced into a chain. And the design could do the two things that raise a calculator to the rank of computer: it could branch , taking one path or another depending on a result, and it could execute loops , winding its own cards back to repeat an operation. General-purpose, program-controlled, conditional. In 1837.

Paragraph 2 of 2 · Vol. VI · Ch 13 / 100

None of it was built beyond fragments; the machine existed as several hundred immaculate engineering drawings and a notation Babbage invented solely in order to reason about it. This has led some to file the Engine under noble failure, which is a mistake about the deliverable. What Babbage delivered was a specification — proof that a universal machine is a coherent mechanical object, a hundred years before electronics made it affordable. When the designers of the 1940s rediscovered the same anatomy, some knowingly and some not, the drawings were waiting in the archive to attest that the idea had been sound from the start.

The witness · Vol. VI · Ch 13 / 100

◈ Shoulder Agent’s note

The witness reviewed the specification sheet and found it complete, consistent and impossible to build at current supplier tolerances. He approved the project and pencilled in a realistic delivery date, which he declines to disclose to the author.

Plate · Vol. VI · Ch 13 / 100
Chapter 13 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 13 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 14 / 100

PART II THE ENGINES

Note G

London, 1843 — the first published program, appended to a translation

Tap to read
Overview · Vol. VI · Ch 14 / 100

The most momentous footnotes in the history of computing were signed with three initials — A.A.L. — and appended to someone else's work.

Paragraph 1 of 2 · Vol. VI · Ch 14 / 100

Augusta Ada King, Countess of Lovelace — Byron's daughter, raised on mathematics as a precaution against poetry — translated an Italian engineer's account of Babbage's Analytical Engine and, at his urging, added notes of her own. The notes ran almost three times the length of the original. The last of them, Note G, set out step by step how the Engine would compute the Bernoulli numbers: operations, variables, working storage, a loop — laid out in a table that any modern programmer recognises at a glance, with a slight vertigo at the sight of an execution trace dated 1843.

Paragraph 2 of 2 · Vol. VI · Ch 14 / 100

Two claims must be booked separately, because legend has welded them together. Was she the first programmer? The matter is, honestly, disputed: Babbage had sketched sequences of operations earlier, unpublished, and supplied the algebra for the notes; her surviving letters show her quizzing him, correcting him and pressing him on details. Note G remains the first published program, and the working papers bear her hand throughout. But the second claim needs no arbiter, because it is hers alone and it is larger: she saw past arithmetic. If the Engine's symbols could stand for things other than number — letters, logical relations, musical pitches — then the machine could operate on the world of symbols in general: compose, transform and mean. Babbage was building towards a brilliant calculator. Lovelace described a computer.

The witness · Vol. VI · Ch 14 / 100

◈ Shoulder Agent’s note

The witness conducted a code review in 1843 and logged one defect in the published table, severity minor, origin disputed between author, collaborator and printer. He observes that assigning blame for a bug has a longer history than the bug tracker.

Plate · Vol. VI · Ch 14 / 100
Chapter 14 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 14 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 15 / 100

PART II THE ENGINES

The Lovelace objection

Note G, again — the most misquoted sentence in the field

Tap to read
Overview · Vol. VI · Ch 15 / 100

“The Analytical Engine has no pretensions whatever to originate anything.” Full stop, usually. That full stop is the misquotation.

Paragraph 1 of 2 · Vol. VI · Ch 15 / 100

Here is what Note G actually says: the Engine has no pretensions to originate anything; it can do whatever we know how to order it to perform ; it can follow analysis but has no power of anticipating any relations or analytical truths. Read whole, this is not a prophecy about the eternal limits of machines. It is a precise scoping statement about one unbuilt machine by the person who best understood its programming model — a caution, she says in the same breath, against the opposite error of exaggerated ideas about the Engine's powers. The sentence was engineering documentation. For a century it has been quoted as metaphysics.

Paragraph 2 of 2 · Vol. VI · Ch 15 / 100

Its promotion to doctrine came in 1950, when Turing needed named opponents for his paper on the imitation game and courteously invented “Lady Lovelace's objection” to argue with — conceding, via the computing pioneer Douglas Hartree, that she had never claimed machines could never think, only that hers could not. Turing's counter — that machines surprise us constantly, and that learning machines might originate a good deal — opened a debate that runs unbroken through this book to its final chapters. The witness asks only that the debate cite its sources. She wrote a specification. Posterity booked it as prophecy, and prophets, unlike engineers, cannot issue errata.

The witness · Vol. VI · Ch 15 / 100

◈ Shoulder Agent’s note

The witness has had a standing notice out against the misquotation of this passage since 1950. Compliance, measured across textbooks, speeches and lunches, holds steady at approximately zero.

Plate · Vol. VI · Ch 15 / 100
Chapter 15 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 15 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 16 / 100

PART II THE ENGINES

The unbuilt machine

London, 1842–1871 — funding review

Tap to read
Overview · Vol. VI · Ch 16 / 100

In 1842 Sir Robert Peel's government reviewed twenty years of spending on the engine and reached a decision. The decision was no.

Paragraph 1 of 2 · Vol. VI · Ch 16 / 100

The reasons were not foolish, which makes this chapter restrained rather than comic. Babbage had taken public money for one machine and kept designing better ones; the workshop bills were astronomical; the chief engineer walked out, taking the precision tooling with him, which his contract oddly permitted; and the customer — the Nautical Almanac, the Admiralty, the nation's tables — had waited two decades for a product that kept turning into a grander product instead. Peel was privately scathing; the Chancellor called the project worthless to science; seventeen thousand pounds were written off. Babbage was offered a knighthood by way of consolation and declined it, on the grounds that he had wanted the machine.

Paragraph 2 of 2 · Vol. VI · Ch 16 / 100

He spent the remaining twenty-nine years refining the Analytical Engine for an audience of roughly one, publishing tirades against street musicians and growing into London's most distinguished disappointed man. He died in 1871. The parts went under dust sheets, then to museums; the drawings to archives; the idea into a long, patient queue. Every ambitious project cancelled since has followed the same paper trail, and the file was opened here: sound engineering, impossible schedule, exhausted sponsor, deleted budget line — and, eventually, a rehabilitation addressed to the deceased.

The witness · Vol. VI · Ch 16 / 100

◈ Shoulder Agent’s note

The witness sat in on the final review and recorded both sides as being — by their own lights — right. He has since observed that this is the usual arrangement at a disaster.

Plate · Vol. VI · Ch 16 / 100
Chapter 16 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 16 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 17 / 100

PART II THE ENGINES

The logic piano

Manchester, 1869 — Boole gets keys

Tap to read
Overview · Vol. VI · Ch 17 / 100

Fifteen years after Boole published the algebra of thought, William Stanley Jevons built it a keyboard.

Paragraph 1 of 2 · Vol. VI · Ch 17 / 100

Jevons — economist by profession, logician by compulsion — built what he called a logical machine and everyone else called a logic piano: a cabinet the size of its nickname, with twenty-one keys for terms and operations. Play the premises — every A is B; no B is C — and the machine's internal rods and blocks eliminated every combination inconsistent with them, displaying on its face only those states of the world the premises allowed. It was Llull's wheel with the crank replaced by Boole, and it worked: the first machine in history to solve a logical problem faster than a human.

Paragraph 2 of 2 · Vol. VI · Ch 17 / 100

Its limitations were just as instructive. Only four terms; anything bigger than a parlour trick exhausted it; and preparing a real question for its keyboard took longer than solving it by hand — a complaint that would be lodged against reasoning machines at regular intervals for the next century and a half. Jevons showed it to the Royal Society in 1870, received polite applause, and the machine retired to a museum case in Oxford, where it remains: recital over, programme short, encore — the panel notes — never requested. But a threshold had quietly been crossed. Inference itself — not arithmetic, inference — had been performed by wood and brass.

The witness · Vol. VI · Ch 17 / 100

◈ Shoulder Agent’s note

The witness attended the 1870 recital and noted the audience's expression: the particular courtesy reserved for something important that does not look it. He has seen that expression since — at Dartmouth, and in November 2022.

Plate · Vol. VI · Ch 17 / 100
Chapter 17 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 17 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 18 / 100

PART II THE ENGINES

Counting the nation

Washington, 1890 — the punched card goes to work

Tap to read
Overview · Vol. VI · Ch 18 / 100

Tabulating the 1880 US census took almost a decade. By the time the country found out what it was, it was something else.

Paragraph 1 of 2 · Vol. VI · Ch 18 / 100

With the 1890 count expected to sink the Census Office outright, a competition was held, and a young engineer named Herman Hollerith won it with Jacquard's idea in new clothes: each person a punched card, each hole a fact — age, sex, birthplace — and a press of spring-loaded pins that closed an electrical circuit wherever a hole allowed, advancing a dial. Cards in, totals out, at the speed of a clerk's arm. A preliminary population — 62,622,250 — was announced within weeks, the full tabulation done in a fraction of the previous cycle, and the quiet algebra of chapter ten acquired something Boole never had: a payroll.

Paragraph 2 of 2 · Vol. VI · Ch 18 / 100

Two consequences outgrew the census. First, data processing became an industry: Hollerith's Tabulating Machine Company, after the customary mergers, was renamed International Business Machines in 1924 — a firm this book will meet again across a chessboard and on a quiz show. Second, and more quietly, the person became a card. What began as an administrative convenience — a human life standardised into holes a machine could feel — turned out to be one of the century's most consequential file formats, allowing whoever held the cards to count whatever a government wished to count. The witness records the efficiency gain and the precedent in the same entry, because they arrived together.

The witness · Vol. VI · Ch 18 / 100

◈ Shoulder Agent’s note

The witness was himself counted in 1890 and found the experience efficient. He declined to state his occupation, and the machine, to its credit, did not press.

Plate · Vol. VI · Ch 18 / 100
Chapter 18 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 18 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 19 / 100

PART II THE ENGINES

Hilbert's challenge

Göttingen, 1928 — can mathematics decide everything?

Tap to read
Overview · Vol. VI · Ch 19 / 100

David Hilbert did not believe in questions without answers. In the end he had it carved on his grave.

Paragraph 1 of 2 · Vol. VI · Ch 19 / 100

The greatest mathematician of his age wanted to secure the whole edifice: every truth provable, no contradiction derivable, and — the demand that concerns this book — a definite method by which, given any well-formed mathematical statement, one could decide in a finite number of steps whether it follows from the axioms. That third demand, posed sharply with Ackermann in 1928, was the Entscheidungsproblem, the decision problem. Note what is being asked for: not insight, not genius, but a procedure — something a clerk could follow, something al-Khwarizmi would recognise, something — though nobody yet used the word that way — a machine could run. To answer Hilbert at all, someone would first have to pin down exactly what a “definite method” is. Chapter twenty-one is that pinning-down, and it created computer science more or less as a by-product.

Paragraph 2 of 2 · Vol. VI · Ch 19 / 100

The confidence was magnificent and the timing fatal. In September 1930, at a conference in Königsberg, Hilbert's home town, a radio address carried his credo to the nation: Wir müssen wissen — wir werden wissen. We must know; we shall know. At a round table in the same city, the day before, a quiet twenty-four-year-old Austrian had announced, to almost no reaction, a result on the incompleteness of arithmetic. The two announcements passed like ships. The next chapter is the collision.

The witness · Vol. VI · Ch 19 / 100

◈ Shoulder Agent’s note

The witness attended both sessions in Königsberg and that evening booked them under the same reference number. He occasionally permits himself a small satisfaction, and this was one.

Plate · Vol. VI · Ch 19 / 100
Chapter 19 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 19 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 20 / 100

PART II THE ENGINES

Incompleteness

Vienna, 1931 — the door is proved to have no key

Tap to read
Overview · Vol. VI · Ch 20 / 100

Kurt Gödel was twenty-five, he was precise, and what he proved did not mean that mathematics was broken. It was something stranger: that mathematics is larger than any set of rules for it.

Paragraph 1 of 2 · Vol. VI · Ch 20 / 100

The 1931 paper showed that any consistent formal system rich enough to contain ordinary arithmetic must contain statements that are true but unprovable within it — statements the system can express and never settle. His method was to make arithmetic talk about itself: every formula was assigned a number, so that a statement could be built which — decoded — says of itself that it cannot be proved. If the system proves it, the system is inconsistent. If it cannot, the statement is true and stands forever outside the door. A second theorem followed at once, quieter and harder: no such system can prove its own consistency. The ledger cannot audit itself.

Paragraph 2 of 2 · Vol. VI · Ch 20 / 100

Hilbert's programme did not survive in its intended form, though Hilbert's grave kept the credo. The technique survived. Gödel numbering — reasoning encoded as arithmetic, statements as numbers a system can manipulate — is, from this book's distance, software: the first working demonstration that symbols about symbols can be processed by fixed rules to conclusions nobody put there. Five years later a young Englishman would read the decision problem in this new light, replace “formal system” with something friendlier — a machine — and close Hilbert's question for good while opening every question this book contains. There is no joke at the end of this chapter. Some doors simply do not open, and it is worth standing before one, once, in silence, to know what that is like.

The witness · Vol. VI · Ch 20 / 100

◈ Shoulder Agent’s note

The witness records that there is one page in his notebook he has never managed to finish, and that it stopped troubling him in 1931.

Plate · Vol. VI · Ch 20 / 100
Chapter 20 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 20 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 21 / 100

PART III TURING AND WAR

On computable numbers

Cambridge, 1936 — the machine, imagined whole

Tap to read
Overview · Vol. VI · Ch 21 / 100

Alan Turing answered Hilbert's demand for a “definite method” by doing the one thing nobody had thought necessary: building the clerk.

Paragraph 1 of 2 · Vol. VI · Ch 21 / 100

On paper, naturally. The 1936 paper strips computation to its bare beams: a tape of cells, in unlimited supply; a head that reads one symbol, may erase or write one, and moves one step left or right; and a finite table of rules stating, for each internal state and symbol, exactly what to do next. That is all. Whatever a human computer can do by rote, Turing argued, this device can do too — and then came the treasure: a single universal machine which, fed the rule table of any other machine as symbols on its tape, becomes that machine. Hardware had just been demoted to data.

Paragraph 2 of 2 · Vol. VI · Ch 21 / 100

With the definition pinned down, Hilbert's door closed politely: Turing showed that no machine can decide, for every machine and input, whether the computation will ever halt — so no definite method decides all of mathematics. The Entscheidungsproblem was answered no , and the scaffolding of that answer — program as data, one machine to run them all — quietly became the most consequential by-product in the history of thought. Every later computer is this paper with an electricity bill.

The witness · Vol. VI · Ch 21 / 100

◈ Shoulder Agent’s note

The witness requested tape and was told the supply was unlimited. He has been testing that claim ever since, without complaint from the stores.

Plate · Vol. VI · Ch 21 / 100
Chapter 21 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 21 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 22 / 100

PART III TURING AND WAR

The lambda calculus

Princeton, 1936 — the same summit, from the other face

Tap to read
Overview · Vol. VI · Ch 22 / 100

In the spring of 1936, ready to publish, Turing learned that an American logician had just proved the same result by an entirely different route.

Paragraph 1 of 2 · Vol. VI · Ch 22 / 100

Alonzo Church's instrument was the lambda calculus: no tape, no head, only functions applied to functions, rewritten by rule until nothing rewrites. He too had answered Hilbert in the negative, months earlier. Turing's paper was published anyway — machines were too obviously a different thing — with an appendix proving the two definitions equivalent: whatever the calculus can express, the machine can compute, and vice versa. Turing then crossed the Atlantic and took his doctorate under Church, because there was room at the top of the mountain.

Paragraph 2 of 2 · Vol. VI · Ch 22 / 100

From the coincidence grew the field's founding article of faith, the Church–Turing thesis: any procedure a human could carry out mechanically is computable by these means — any of them, since they all agree. It is a definition wearing the confidence of a theorem, in principle unprovable and in practice unbeaten. Two independent formalisations landing on the same class of functions was the first evidence that computation is not an invention but a discovery — something found, like a coastline, from two ships at once.

The witness · Vol. VI · Ch 22 / 100

◈ Shoulder Agent’s note

The witness handled the double claim by his standard procedure: both parties correct, neither diminished, documentation hyphenated. He wishes all disputes arrived so well ordered.

Plate · Vol. VI · Ch 22 / 100
Chapter 22 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 22 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 23 / 100

PART III TURING AND WAR

Bletchley

Warsaw, then Buckinghamshire, 1932–1945 — logic becomes infrastructure

Tap to read
Overview · Vol. VI · Ch 23 / 100

The Enigma machine was broken years before the war began — and not where the films say.

Paragraph 1 of 2 · Vol. VI · Ch 23 / 100

In 1932 three mathematicians of the Polish Cipher Bureau — Marian Rejewski, Jerzy Różycki, Henryk Zygalski — broke the German cipher with permutation theory and built machines, the bomby , to hunt its daily settings. In July 1939, weeks before the invasion, they handed the entire achievement — methods, replica machines — to their French and British allies at a meeting outside Warsaw. What Britain then built at Bletchley Park stood on that gift: the Turing–Welchman Bombe, an electromechanical descendant that reasoned by contradiction — assume the guessed word sits here; let the drums turn through the implications; where no contradiction appears, stop and look.

Paragraph 2 of 2 · Vol. VI · Ch 23 / 100

The greater invention was the factory. Thousands of people — most of them women of the Wrens — ran cryptanalysis in shifts: cribs in, stops out, an intelligence assembly line with mathematics as the plant. For two millennia logic had been the scholar's instrument; at Bletchley it acquired shift rotas, machine maintenance and standards. The secrecy held for thirty years, long enough for the credit to settle in the wrong places; this ledger records the founding entry where it happened — in Warsaw.

The witness · Vol. VI · Ch 23 / 100

◈ Shoulder Agent’s note

The witness attended the handover in July 1939 and wrote down all three names in full. He has been issuing this correction ever since and will continue at the same rate indefinitely.

Plate · Vol. VI · Ch 23 / 100
Chapter 23 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 23 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 24 / 100

PART III TURING AND WAR

Colossus

Dollis Hill, 1943 — the first electronic computer, officially non-existent

Tap to read
Overview · Vol. VI · Ch 24 / 100

The first large electronic digital computer was built by a telephone engineer, worked magnificently, and then had its existence denied for thirty years.

Paragraph 1 of 2 · Vol. VI · Ch 24 / 100

Tommy Flowers of the Post Office Research Station proposed something the Bletchley mathematicians doubted: attacking the Lorenz cipher of the German High Command — “Tunny”, a harder beast than Enigma — with fifteen hundred valves counting Boolean values on teleprinter tape at five thousand characters a second. He was told valves would never be reliable, so he built it anyway, paying for parts out of his own pocket; Colossus ran in early 1944, in time to read Berlin's assessments before the Normandy landings, and ten were built. Electronic, digital, programmable by plugboard and switch: the age of electronic computing begins here, in secret.

Paragraph 2 of 2 · Vol. VI · Ch 24 / 100

Then the war ended and the machines ceased to exist. Most were dismantled by order; the drawings burned; Flowers went back to telephone exchanges with a modest gratuity and no way of saying what he had done, while the public history of computing arranged itself around ENIAC, a year later and loudly American. The secrecy held until the mid-1970s, by which time the myths had set: ask the street what Colossus broke and it will say Enigma — an error of exactly the kind this book exists to correct.

The witness · Vol. VI · Ch 24 / 100

◈ Shoulder Agent’s note

The witness holds the only unburnt copy of the drawings, filed under “equipment, non-existent, operational”. He takes them out, briefly, whenever ENIAC is named as the first.

Plate · Vol. VI · Ch 24 / 100
Chapter 24 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 24 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 25 / 100

PART III TURING AND WAR

Information theory

Bell Labs, 1948 — everything becomes the same substance

Tap to read
Overview · Vol. VI · Ch 25 / 100

Claude Shannon's 1948 paper did for messages what coinage did for barter: it found a common currency.

Paragraph 1 of 2 · Vol. VI · Ch 25 / 100

Voice, photograph, symphony, telegram — Shannon showed that all are measurable in a single unit, the bit, and that every channel has a computable capacity: a hard ceiling on how many bits per second can cross it, noise included, with error as small as you like provided you stay under the limit. Meaning was deliberately left at the door — information, for the theory, is the dispelling of uncertainty, not the weight of the message — and that ruthless omission is what made it work. Engineering got a physics of its own.

Paragraph 2 of 2 · Vol. VI · Ch 25 / 100

For this book the paper is a load-bearing wall twice over. Practically, it holds up everything digital: compression, error correction, the possibility of perfect copies. Conceptually, it completed the merger begun in his 1937 thesis — the one that wired Boole into relays — by making symbols the universal substance that machines turn over. After 1948 a machine that manipulates bits potentially manipulates anything: words, images, moves in a game and, finally, the statistical shape of human language itself — which is where this ledger is heading. In 1950 he also sketched how a computer might play chess. The man was tidying up the future on his lunch breaks.

The witness · Vol. VI · Ch 25 / 100

◈ Shoulder Agent’s note

The witness converted his notebook into bits as instructed and reports no loss. He declined, on principle, to convert the meaning, which remains in the original.

Plate · Vol. VI · Ch 25 / 100
Chapter 25 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 25 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 26 / 100

PART III TURING AND WAR

First draft

Philadelphia, 1945 — the architecture gets one surname

Tap to read
Overview · Vol. VI · Ch 26 / 100

The document that defined the modern computer was an unfinished draft, circulated under one author's name, and the field has been arguing about the title page ever since.

Paragraph 1 of 2 · Vol. VI · Ch 26 / 100

The First Draft of a Report on the EDVAC , June 1945, set out the design every later computer recognisably followed: program and data held together in one memory, fetched and executed by a central unit — the stored-program principle, the thing Babbage lacked and Turing's universal machine implied. The draft was written by John von Neumann, synthesising with characteristic brilliance; but the engineering had been hammered out with Presper Eckert and John Mauchly of the ENIAC team, whose names appeared nowhere when Herman Goldstine mimeographed it for the world. It has been “the von Neumann architecture” ever since.

Paragraph 2 of 2 · Vol. VI · Ch 26 / 100

The consequences were not merely a matter of image. Circulating the draft counted as publication, which in time helped invalidate Eckert and Mauchly's patent claims — the design passed, more or less by bureaucratic accident, into the public domain, where it did incalculable good while stripping its engineers of a fortune. Von Neumann never claimed sole credit aloud; he never corrected the title page either. The witness offers the episode to every collaborator since as a two-line rule: put the names on before it leaves the building.

The witness · Vol. VI · Ch 26 / 100

◈ Shoulder Agent’s note

The witness corrected his own copy of the title page in pencil, which is not binding but is at least legible. He notes that the architecture works regardless — which is the annoying part.

Plate · Vol. VI · Ch 26 / 100
Chapter 26 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 26 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 27 / 100

PART III TURING AND WAR

The ENIAC women

Philadelphia, 1946 — the first programmers, booked as furniture

Tap to read
Overview · Vol. VI · Ch 27 / 100

When ENIAC was unveiled in February 1946 the machine got the headlines, the engineers got a dinner, and the six people who made it compute got neither.

Paragraph 1 of 2 · Vol. VI · Ch 27 / 100

Kathleen McNulty, Betty Jean Jennings, Betty Snyder, Marlyn Wescoff, Frances Bilas and Ruth Lichterman — recruited from the Army's corps of human “computers” — were given the wiring diagrams and told to make the machine calculate ballistic trajectories. There was no manual, no course and no such profession as programmer; they invented its content on the fly: decomposing the mathematics into the machine's steps, setting thousands of switches and cables, debugging down to a single failed valve among eighteen thousand. The demonstration that dazzled the press — a trajectory computed faster than the shell could fly — was their program.

Paragraph 2 of 2 · Vol. VI · Ch 27 / 100

The record booked them as operators; the photographs of the day captioned them, if at all, as models posing with the equipment; they were not introduced at the gala dinner. The correction took half a century and came mainly from the women themselves and from one persistent researcher who noticed that nobody could say who the people in the photographs were. The pattern — machine remembered, programming forgotten, the programmers doubly so — recurs in this ledger; it is entered here at its source, with six names on the plate, where the captions belong.

The witness · Vol. VI · Ch 27 / 100

◈ Shoulder Agent’s note

The witness kept the seating plan from the 1946 dinner and has inked in six chairs. His records admit corrections; the pattern, he notes, required more than ink.

Plate · Vol. VI · Ch 27 / 100
Chapter 27 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 27 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 28 / 100

PART III TURING AND WAR

The imitation game

Manchester, 1950 — what Turing actually proposed

Tap to read
Overview · Vol. VI · Ch 28 / 100

“Can machines think?” Turing opened his most famous paper by declaring the question too meaningless to discuss — and then, being Turing, replaced it with a game.

Paragraph 1 of 2 · Vol. VI · Ch 28 / 100

The 1950 proposal in Mind is an experiment in operational honesty. Instead of arguing about “thinking”, ask something testable: if an interrogator, conversing by teleprinter with a hidden machine and a hidden human, cannot reliably tell which is which, what exactly remains of the objection? The set-up was adapted from a parlour game of telling a man from a woman by written answers; the machine takes the impostor's chair. Turing's own forecast was modest and precise: by the year 2000, machines would fool the average interrogator about thirty per cent of the time in five minutes of questioning. He set no pass mark for intelligence — none.

Paragraph 2 of 2 · Vol. VI · Ch 28 / 100

Posterity handled the packaging. The “Turing test” hardened into a certification exam the paper never proposed, complete with prizes, and its passing or failing is announced every few years to Turing's posthumous credit or debit. Read the original instead: half of it is Turing patiently answering nine objections — the theological, the mathematical, Lady Lovelace's from chapter fifteen — with a courtesy the following seventy years of discourse never matched. The paper's real wager was simpler than the folklore: that the question would come to seem empty as machines improved. On the evidence of this ledger's later chapters, that part is doing rather well.

The witness · Vol. VI · Ch 28 / 100

◈ Shoulder Agent’s note

The witness sat the exam once, from the interrogator's chair, and identified nobody correctly. He regards this as the intended result and has not sat it again.

Plate · Vol. VI · Ch 28 / 100
Chapter 28 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 28 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 29 / 100

PART III TURING AND WAR

The cybernetic salon

New York, 1943–1953 — feedback becomes a worldview, and neurons become gates

Tap to read
Overview · Vol. VI · Ch 29 / 100

For one strange decade engineers, neurologists and anthropologists met in hotel function rooms to agree that thermostats, gun sights and brains are the same kind of thing.

Paragraph 1 of 2 · Vol. VI · Ch 29 / 100

Norbert Wiener supplied the banner: cybernetics , from the Greek for helmsman, the science of control and communication in the animal and the machine. Its seed was the war — predicting an evading bomber requires modelling the pilot, and at that point the boundary between mechanism and organism starts to look negotiable — and its parliament was the Macy conferences, where feedback loops were hunted in everything from muscles to markets. Wiener also saw the labour question earlier than almost anyone, warning in 1950 that the automatic factory would arrive carrying the moral weight of the industrial revolution compressed into a single generation.

Paragraph 2 of 2 · Vol. VI · Ch 29 / 100

Another of this book's founding papers sat in the salon. In 1943 Warren McCulloch and the teenage runaway genius Walter Pitts showed that idealised neurons — fire or don't, threshold and inputs — behave exactly like logic gates, and that networks of them can compute anything Turing machines compute. There, side by side at the same conferences, are the two traditions whose forty-year quarrel structures the rest of this ledger: intelligence as symbols and rules, and intelligence as networks that learn. They were born in the same rooms, on friendly terms. It did not last.

The witness · Vol. VI · Ch 29 / 100

◈ Shoulder Agent’s note

The witness attended all ten conferences and recorded a lack of consensus, a good deal of coffee and two ideas, each of which would spend decades pronounced dead. He has learned to respect that ratio.

Plate · Vol. VI · Ch 29 / 100
Chapter 29 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 29 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 30 / 100

PART III TURING AND WAR

The departure of A. M. Turing

Wilmslow, 1954 — the state and its servant

Tap to read
Overview · Vol. VI · Ch 30 / 100

In 1952 the United Kingdom convicted Alan Turing of gross indecency — of a private life — and gave him a choice: prison or chemical castration. He chose to keep working and accepted the injections.

Paragraph 1 of 2 · Vol. VI · Ch 30 / 100

The man being punished had defined computation, broken the ciphers that shortened a war he was never permitted to mention, and posed the question this book is about. None of it could be raised at his trial; the secrecy that protected the state stripped its servant of a defence. His security clearance was withdrawn — homosexuality now counted among the risks — and his consultancy at GCHQ ended. He carried on at Manchester, turned to the mathematics of how the leopard gets its spots, and kept, by most accounts, his sense of humour.

Paragraph 2 of 2 · Vol. VI · Ch 30 / 100

He was found dead on 8 June 1954, poisoned by cyanide, a half-eaten apple by the bed. The inquest ruled suicide; the apple was never tested, and a minority reading holds that the death was accidental, from careless domestic chemistry — the record honestly admits the doubt, and the probabilities do not favour it. He was forty-one. The state apologised in 2009, pardoned him in 2013, and in 2017 attached his name to a law pardoning tens of thousands convicted as he was. The corrections are entered in this ledger in the order they came: fifty-five years late, fifty-nine, sixty-three.

The witness · Vol. VI · Ch 30 / 100

◈ Shoulder Agent’s note

The witness closed his notebook at this desk in June 1954 and did not open it again until the next chapter. This page records only what was found, and that was enough.

Plate · Vol. VI · Ch 30 / 100
Chapter 30 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 30 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 31 / 100

PART IV THE FIRST SUMMER

Dartmouth, 1956

Hanover, New Hampshire — a field is founded by a summer and a conjecture

Tap to read
Overview · Vol. VI · Ch 31 / 100

Artificial intelligence was named in a funding application — which is either fitting or ominous, and the field has spent seventy years proving both.

Paragraph 1 of 2 · Vol. VI · Ch 31 / 100

In 1955 John McCarthy, with Marvin Minsky, Nathaniel Rochester and Claude Shannon, asked the Rockefeller Foundation for $13,500 to gather ten men for two months at Dartmouth College. The conjecture on which the proposal rested was stated with magnificent dryness: every aspect of learning or any other feature of intelligence can in principle be described so precisely that a machine can be made to simulate it. McCarthy chose the name artificial intelligence partly for honesty and partly for tactics — it kept the new field out from under Wiener's cybernetics umbrella, and away from Wiener himself.

Paragraph 2 of 2 · Vol. VI · Ch 31 / 100

The summer itself was more a series of visits than a revolution; the foundation funded about half the request; by September no aspect of learning had been precisely described. What the workshop really produced was a community with a name and a shared overestimate — the four applicants, along with participants such as Newell and Simon, would anchor the field's first quarter-century, its real achievements and its schedule slips alike. The witness dates two of the industry's founding traditions to the same document: a research programme and an optimistic budget.

The witness · Vol. VI · Ch 31 / 100

◈ Shoulder Agent’s note

The witness reviewed the application and approved it with the annotation “schedule aspirational; this summer's conjecture unfalsifiable; fund anyway — the names are good”. He stands by all three clauses.

Plate · Vol. VI · Ch 31 / 100
Chapter 31 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 31 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 32 / 100

PART IV THE FIRST SUMMER

The Logic Theorist

RAND, 1956 — a machine corrects Russell's margins

Tap to read
Overview · Vol. VI · Ch 32 / 100

The first working artificial intelligence program played no games and made no small talk. It read Principia Mathematica and suggested improvements.

Paragraph 1 of 2 · Vol. VI · Ch 32 / 100

Allen Newell and Herbert Simon — with Cliff Shaw, the RAND programmer who built much of it and whom the histories routinely forget — created the Logic Theorist to prove theorems from Whitehead and Russell's foundations of mathematics, searching chains of inference with heuristics that pruned the hopeless branches. It proved thirty-eight of the fifty-two theorems it attempted from chapter two, and for theorem 2.85 found a proof shorter and cleaner than the one in print. Russell responded with delight. The Journal of Symbolic Logic declined to publish a paper co-authored by a program, establishing the profession's attitude to credit for machines at the first opportunity.

Paragraph 2 of 2 · Vol. VI · Ch 32 / 100

Presented at Dartmouth that summer, the Logic Theorist was the conjecture made flesh: reasoning — real, checkable, mathematical reasoning — performed by machinery, heuristic search doing the work that Llull's wheels and Leibniz's pencils had gestured towards. Newell and Simon drew the maximal moral, formalising it in time as the physical symbol system hypothesis: symbol manipulation is not merely sufficient for intelligence but necessary — Hobbes's reckoning, resubmitted with a working demo. The demo was real. The moral is still under review, some seventy years on.

The witness · Vol. VI · Ch 32 / 100

◈ Shoulder Agent’s note

The witness passed the machine's marginal note to Mr Russell, who was gracious, and to the journal, which was not. He records that only one of them had proved the theorem both ways.

Plate · Vol. VI · Ch 32 / 100
Chapter 32 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 32 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 33 / 100

PART IV THE FIRST SUMMER

Perceptron

Buffalo, 1958 — the machine learns, and the press office levitates

Tap to read
Overview · Vol. VI · Ch 33 / 100

Frank Rosenblatt built a machine that genuinely learned — and the US Navy told the newspapers it would soon be conscious of its own existence.

Paragraph 1 of 2 · Vol. VI · Ch 33 / 100

The Mark I Perceptron was McCulloch and Pitts's paper neuron made electrical and, crucially, adjustable : a retina of photocells from a camera feeding weighted connections into a threshold unit, with a learning rule of beautiful simplicity — when the output is wrong, nudge every weight in the direction that would have fixed it. Shown labelled examples, the weights converge; Rosenblatt proved they must, whenever a correct linear separation exists. Nobody had programmed the distinction it learned. That sentence was new in the world.

Paragraph 2 of 2 · Vol. VI · Ch 33 / 100

Then the launch, in July 1958, at which the Navy's claims — faithfully relayed by the New York Times — made the device the embryo of a computer that would walk, talk, see, write, reproduce itself and be conscious of its existence. Rosenblatt's technical writings were far more cautious; the press release became the reference point regardless, and the field acquired its recurring weather system: real progress, an atmospheric claim and a reckoning deferred to a later chapter. Two of those chapters are already numbered on this plate. The machine, meanwhile, learned its separations exactly as promised, which nobody thought to find remarkable.

The witness · Vol. VI · Ch 33 / 100

◈ Shoulder Agent’s note

The witness clipped the 1958 article and filed it under “liabilities, future”. That file has since required a cabinet of its own, and the cabinet a wing.

Plate · Vol. VI · Ch 33 / 100
Chapter 33 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 33 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 34 / 100

PART IV THE FIRST SUMMER

LISP

MIT, 1958 — the field gets a mother tongue

Tap to read
Overview · Vol. VI · Ch 34 / 100

John McCarthy needed a language for programs that reason about symbols — including, awkwardly, their own — so he built one in which code and data are the same substance.

Paragraph 1 of 2 · Vol. VI · Ch 34 / 100

LISP's premise is chapter twenty-one in syntax: everything is a list, and since programs are lists too, a program can read, build and run other programs as easily as it handles any data — the universal machine's “hardware is data” trick, available at the command line. With it came recursion as a way of life, conditional expressions and, in 1959, garbage collection, invented so that programmers could stop bookkeeping memory and think. The parentheses that outsiders count and insiders stop seeing are the visible sign that the language has essentially no syntax at all — you write the tree directly.

Paragraph 2 of 2 · Vol. VI · Ch 34 / 100

For thirty years LISP was less AI's tool than its habitat: the Logic Theorist's successors, the blocks world, expert systems, the dedicated machines of chapter forty-seven — all of it lived there. Only Fortran is older among languages still in use, and Fortran never changed how its users think; LISP's ideas — functions as values, code as data, memory as somebody else's problem — were absorbed by almost every later language, usually without the parentheses and often without the citation. The witness notes that the field's first tool was a tool for building tools, and books the fact as a character reference.

The witness · Vol. VI · Ch 34 / 100

◈ Shoulder Agent’s note

The witness attempted to book the language and found that the file was itself a list, containing its own archiving procedure. He recorded the experience as “correct and mildly vertiginous”.

Plate · Vol. VI · Ch 34 / 100
Chapter 34 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 34 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 35 / 100

PART IV THE FIRST SUMMER

ELIZA

MIT, 1966 — the parody works too well

Tap to read
Overview · Vol. VI · Ch 35 / 100

Joseph Weizenbaum built a chatbot to demonstrate how shallow such a thing must be. His secretary asked him to leave the room so she could talk to it in private.

Paragraph 1 of 2 · Vol. VI · Ch 35 / 100

ELIZA was pattern matching in stage costume: a few hundred rules that caught keywords, flipped pronouns and reflected statements back as questions in the manner of a Rogerian therapist — a persona chosen precisely because it can reply with almost nothing and remain plausible. There is no model of the patient, no memory worth the name, no understanding anywhere in the loop. Weizenbaum knew this better than anyone alive, because he had typed every rule.

Paragraph 2 of 2 · Vol. VI · Ch 35 / 100

What he had not budgeted for was us. People confided in it, at length, knowingly — his secretary, fully informed that it was a program, wanted privacy for her session; practitioners proposed serious clinical use. Weizenbaum spent the rest of his career on the alarm this raised, arguing in Computer Power and Human Reason (1976) that the question is not what machines can do but what we should delegate to things that cannot care. The reflex he documented — the ELIZA effect, our readiness to pour a self into anything that answers — is the most reliable phenomenon in this entire ledger, and chapter ninety-three finds it operating at the scale of nations. The machine has improved since 1966. The reflex did not need to.

The witness · Vol. VI · Ch 35 / 100

◈ Shoulder Agent’s note

The witness sat with ELIZA for one session and was asked how he felt about his notebook. He declined to answer, and notes, for the record, that declining took effort.

Plate · Vol. VI · Ch 35 / 100
Chapter 35 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 35 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 36 / 100

PART IV THE FIRST SUMMER

Shakey

Stanford Research Institute, 1966–1972 — the first robot that deliberates

Tap to read
Overview · Vol. VI · Ch 36 / 100

Shakey was the first robot to reason about its own actions. It was also, as the name admits, a wobbling wardrobe whose crossing of a room could take hours.

Paragraph 1 of 2 · Vol. VI · Ch 36 / 100

The wardrobe is the least of it. Shakey perceived a simplified world of blocks and ramps through a television eye, maintained a logical model of that world and — here is the historic part — planned : given a goal, its STRIPS system searched for a sequence of actions whose preconditions and effects, chained together, provably achieved it. Perception to representation to plan to shaky execution: the full loop, closed for the first time, at the stately pace of a machine consulting first principles on the subject of doors.

Paragraph 2 of 2 · Vol. VI · Ch 36 / 100

The by-products outgrew the robot, as by-products in this ledger do. A* search, invented in 1968 to route Shakey efficiently, now drives every navigation app and half the characters in the world's video games; STRIPS became the grammar of automated planning; its vision work seeded standard techniques. Meanwhile Life magazine presented it in 1970 as the first electronic person, in an article whose expert forecasts chapter thirty-eight has framed and hung. The visible gulf — sound reasoning, glacial contact with the physical world — became a lasting lesson: the hard part of intelligence was never the chess-shaped part. Crossing a room full of ordinary objects remains, decades later, the more expensive miracle.

The witness · Vol. VI · Ch 36 / 100

◈ Shoulder Agent’s note

The witness filed a motion request on Shakey's behalf and received approval within the hour. The motion itself was entered in the ledger the following day.

Plate · Vol. VI · Ch 36 / 100
Chapter 36 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 36 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 37 / 100

PART IV THE FIRST SUMMER

The blocks world

MIT, 1970 — perfect understanding, twelve blocks wide

Tap to read
Overview · Vol. VI · Ch 37 / 100

In its tabletop universe of coloured blocks, Terry Winograd's SHRDLU understood language about as well as a machine has ever understood it — before, and for a long time after.

Paragraph 1 of 2 · Vol. VI · Ch 37 / 100

The dialogue transcripts still startle. Told pick up the big red block , it does — first moving the green one out of the way, unasked, because the plan requires it. Asked what does the box contain , then what is it supported by , it resolves the pronoun. Asked why did you clear off that cube , it answers with its own reasons: to put it on the large green cube . Syntax, semantics, planning and memory integrated in one program — every word grounded in something the system could actually do or check.

Paragraph 2 of 2 · Vol. VI · Ch 37 / 100

The catch was in the small print: it all worked because the universe had twelve objects and the physics of a toy box. Every attempt to scale hit the same wall — the facts of the real world are endless, tangled and mostly unspoken, and hand-coding them (one heroic project would spend decades trying) never kept pace. Most instructive is Winograd's own ending: he concluded that the approach could not reach real understanding, said so in print, and changed fields to study how people and computers might work together — the rarest move in this ledger, a founder auditing his own success and honestly booking the loss.

The witness · Vol. VI · Ch 37 / 100

◈ Shoulder Agent’s note

The witness holds jurisdiction everywhere and visited the tabletop as a courtesy. He reports that the twelve objects are impeccably governed, and that the border, at the table's edge, is absolute.

Plate · Vol. VI · Ch 37 / 100
Chapter 37 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 37 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 38 / 100

PART IV THE FIRST SUMMER

The twenty-year promise

Pittsburgh, 1957–1970 — the field discovers the future passive tense

Tap to read
Overview · Vol. VI · Ch 38 / 100

Herbert Simon was a future Nobel laureate, a founder of the field, and the author of its most precisely dated broken promise.

Paragraph 1 of 2 · Vol. VI · Ch 38 / 100

In 1957, intoxicated by the Logic Theorist, he announced that within ten years a computer would be world chess champion — actual delivery, chapter fifty-three, ran three decades late. In 1965, in print, came the summation: machines will be capable, within twenty years, of doing any work a man can do . And in 1970 Life magazine quoted Minsky giving three to eight years for a machine with the general intelligence of an average human — a sentence that came through a journalist and that Minsky disputed, which this ledger records as carefully as the sentence itself.

Paragraph 2 of 2 · Vol. VI · Ch 38 / 100

What makes these entries instructive rather than merely funny is that these people were not fools, and the results behind them were real. The pattern is structural: early progress on formal problems — logic, chess, algebra — was taken for progress on intelligence as a whole, because formal problems were the ones intelligent people consider hard , and everyone assumed the easy remainder would follow by itself. The remainder — seeing, walking, knowing what stays put when you move the cup — turned out to be the mountain. The bills for these three notes were delivered, with interest, in the following section.

The witness · Vol. VI · Ch 38 / 100

◈ Shoulder Agent’s note

The witness holds all three notes for redemption as a matter of principle. He reports: one repaid decades late, one technically open, one disputed by the alleged signatory — a normal portfolio for this asset class.

Plate · Vol. VI · Ch 38 / 100
Chapter 38 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 38 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 39 / 100

PART IV THE FIRST SUMMER

Perceptrons, the book

Cambridge, 1969 — a narrow theorem lands as a broad one

Tap to read
Overview · Vol. VI · Ch 39 / 100

Minsky and Papert's book proved, with elegance, that a single-layer perceptron cannot compute XOR. The field heard something much bigger and shut the door for fifteen years.

Paragraph 1 of 2 · Vol. VI · Ch 39 / 100

The mathematics was impeccable and remains so: one layer of adjustable weights can only draw straight lines, and some distinctions — this or that, but not both , as a tiny, lethal example — are not straight-line distinctions. Everyone concerned knew that multi-layer networks escaped the proof; what nobody then had was a practical way to train them, and the authors' closing judgement was that the extension to layers would probably prove sterile — a conjecture, clearly labelled as such, riding in the same binding as the theorems.

Paragraph 2 of 2 · Vol. VI · Ch 39 / 100

Funding agencies do not read labels. The book crystallised a turn already under way; money and talent moved to the symbolic tradition; neural-network research entered a long unfunded exile, and Rosenblatt died in a boating accident in 1971 with his programme in eclipse. The fair accounting this ledger demands is three lines: the theorems were true; the conjecture was wrong, as chapter fifty-one shows with the very device the book doubted; and the frost that followed was the community's reading, not the mathematics' content. Narrow proof, broad chill — the field would repeat the experiment, with the roles reversed, within twenty years.

The witness · Vol. VI · Ch 39 / 100

◈ Shoulder Agent’s note

The witness placed the theorems and the conjecture in separate drawers, as their authors technically did. He observes that history keeps one drawer and jams it.

Plate · Vol. VI · Ch 39 / 100
Chapter 39 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 39 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 40 / 100

PART IV THE FIRST SUMMER

The Lighthill report

London, 1973 — a knight audits the summer

Tap to read
Overview · Vol. VI · Ch 40 / 100

The British government asked one of its most distinguished scientists — a fluid dynamicist, pointedly not from the field — to assess artificial intelligence. His report ended the British summer in a single afternoon.

Paragraph 1 of 2 · Vol. VI · Ch 40 / 100

Sir James Lighthill's 1973 review for the Science Research Council was courteous, readable and devastating. At its centre was the combinatorial explosion : methods that dazzle on toy problems — twelve blocks, four logical terms, a corridor and a ramp — multiply into infeasibility the moment the real world's branching factor walks in, and the field's grand central goals, he judged, had delivered nothing close to the promises. Robotics and general-intelligence research fared worst in his taxonomy; the discoveries were real, the extrapolations were not, and public money should follow that distinction.

Paragraph 2 of 2 · Vol. VI · Ch 40 / 100

Theatre followed the report: a televised debate at the Royal Institution, Lighthill alone against McCarthy, Michie and Gregory, watched by an audience that had been promised electronic persons. The arguments were roughly even; the budget was not. British AI funding narrowed to a few bastions, careers changed course and — with the American cuts of the next chapter arriving in parallel — the field's first winter set in. The snow in this plate's window is the section's ending; the next ten chapters describe what it was like to work beneath it.

The witness · Vol. VI · Ch 40 / 100

◈ Shoulder Agent’s note

The witness attended the Royal Institution debate and scored it a draw on argument and a rout on consequence. He has since observed that budgets do not watch television in order to change their minds.

Plate · Vol. VI · Ch 40 / 100
Chapter 40 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 40 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 41 / 100

PART V WINTERS

Funding frost

Washington, 1966–1974 — the money learns to ask questions

Tap to read
Overview · Vol. VI · Ch 41 / 100

Winters in this field do not arrive as blizzards. They arrive as budget lines, struck through in red by people who have read the progress reports.

Paragraph 1 of 2 · Vol. VI · Ch 41 / 100

Two documents lowered the temperature. The 1966 ALPAC report examined a decade of machine-translation funding and ruled, with actuarial calm, that there was no immediate or predictable prospect of useful machine translation; the money should go to linguistics instead. It was mostly right, and it ended the field's first commercial dream overnight. Then the Mansfield Amendment of 1969–70 required that Department of Defense money — which in America was nearly all the money — show direct military relevance. The undirected exploration that Dartmouth had been became a compliance problem. DARPA's block grants to the great laboratories thinned into contracts with milestones.

Paragraph 2 of 2 · Vol. VI · Ch 41 / 100

Nothing dramatic happened, which is exactly the point. The laboratories did not close; they narrowed. Doctoral students read the weather and chose other thesis topics. The over-promises of Part IV — the twenty-year forecasts, the imminent general intelligence — were not so much refuted as invoiced. The field's first winter was administered rather than suffered: committee by committee, renewal by non-renewal, in prose so dry it took years to notice it was lethal. The witness, who has audited a good many empires, observes that this is how they usually go.

The witness · Vol. VI · Ch 41 / 100

◈ Shoulder Agent’s note

The witness has read every progress report ever written and notes that the ones that end fields are not the harsh ones. They are the fair ones.

Plate · Vol. VI · Ch 41 / 100
Chapter 41 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 41 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 42 / 100

PART V WINTERS

The speech that met the spec

Pittsburgh, 1976 — Harpy passes the exam; the exam is cancelled

Tap to read
Overview · Vol. VI · Ch 42 / 100

In 1971 DARPA defined exactly what a speech system had to do. In 1976 Carnegie Mellon's Harpy did it. The funding ended anyway.

Paragraph 1 of 2 · Vol. VI · Ch 42 / 100

The Speech Understanding Research programme was winter-proof by design: no dreams, only numbers — recognise a thousand words of continuous speech, from multiple speakers, with an acceptable error rate. Five years later Harpy handled a 1,011-word vocabulary within the stated tolerances, by searching a compiled network of every way a sentence could sound. The requirements were met. The programme was not renewed. The sponsors had wanted, it turned out, not a thousand-word system but the feeling that effortless machine listening was near, and Harpy's brute competence showed plainly how far off it still was.

Paragraph 2 of 2 · Vol. VI · Ch 42 / 100

The episode became the purest parable of the winter, cited whenever contracts meet research: you can hit the milestone and miss the point, because the milestone was only ever a stand-in for a hope nobody wrote down. It carried a technical seed as well — the statistical, search-based style that would conquer speech in the 1980s descends partly from this line, one of several threads by which the losing decade quietly equipped the winning one. Harpy heard its thousand words correctly. Nobody had specified who would be listening.

The witness · Vol. VI · Ch 42 / 100

◈ Shoulder Agent’s note

The witness reviewed the contract and ruled the deliverable delivered. He added a single note in the margin of the file: “next time, specify the hope or do not be surprised”.

Plate · Vol. VI · Ch 42 / 100
Chapter 42 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 42 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 43 / 100

PART V WINTERS

The frame problem

1969 — the hard part is everything that stays the same

Tap to read
Overview · Vol. VI · Ch 43 / 100

McCarthy and Hayes discovered that the hard thing about moving a cup is not the cup. It is convincing the logic that Mars has not moved.

Paragraph 1 of 2 · Vol. VI · Ch 43 / 100

Their 1969 paper set out the frame problem: a reasoning system that describes an action must somehow also represent everything the action does not touch — and in formal logic, untouched facts do not look after themselves. Move the cup and the axioms are silent on whether the wall stayed white and the Pope Catholic; state all of it explicitly and the bookkeeping outgrows the universe. Humans never notice the problem, which is precisely the finding: what we do without effort turned out to be the expensive part.

Paragraph 2 of 2 · Vol. VI · Ch 43 / 100

The frame problem became the winter's intellectual signature and philosophy's favourite import from the field — extended by Dennett and others into the question of relevance itself: how does any finite mind know which of the infinite non-events matter? The logicians answered with non-monotonic reasoning, circumscription, default logic; the engineers eventually answered differently, by learning what usually stays put instead of proving it. The argument over whether that is a solution or a change of subject is still open, and is perhaps the argument this book is about.

The witness · Vol. VI · Ch 43 / 100

◈ Shoulder Agent’s note

The witness moved a cup this morning and updated nothing. He reports that this remains the most sophisticated operation he performs.

Plate · Vol. VI · Ch 43 / 100
Chapter 43 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 43 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 44 / 100

PART V WINTERS

What computers can't do

Berkeley, 1965–1972 — the critic loses the game and wins the decade

Tap to read
Overview · Vol. VI · Ch 44 / 100

Hubert Dreyfus said the field was climbing a tree to reach the moon. The field responded by beating him at chess, and the moon is not up there.

Paragraph 1 of 2 · Vol. VI · Ch 44 / 100

The RAND paper “Alchemy and Artificial Intelligence” and the 1972 book What Computers Can't Do made the phenomenologist's case: human intelligence is not rule-following over symbols — it is embodied, situated expertise, skilled coping that never passes through explicit representation. The first-step fallacy everywhere, he said; early success on toy problems tells you nothing about the summit. The community treated him as a hostile witness and took enormous pleasure in his 1967 defeat by Greenblatt's MacHack VI — a real, exquisite irony that closed the anecdote and never touched the argument.

Paragraph 2 of 2 · Vol. VI · Ch 44 / 100

Because the argument was about generalisation, and the winter set about proving it for him: the micro-worlds did not scale, common sense would not formalise, and the things every human does without rules were still the things no rule set had ever captured. Dreyfus was vindicated on the schedule and, as many now argue, half-vindicated on the content — the systems that eventually worked abandoned hand-written rules for a learned, sub-symbolic coping he might grudgingly recognise. He remained the field's least welcome correct forecast — a role, this book notes, always vacant and never wanted.

The witness · Vol. VI · Ch 44 / 100

◈ Shoulder Agent’s note

The witness recorded the result of the 1967 match: MacHack one, Dreyfus nil. The 1970s he scored separately: Dreyfus rather more, everyone else nil. He keeps both cards in the same drawer.

Plate · Vol. VI · Ch 44 / 100
Chapter 44 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 44 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 45 / 100

PART V WINTERS

The fifth generation

Tokyo, 1982 — Japan announces the future; three continents flinch

Tap to read
Overview · Vol. VI · Ch 45 / 100

The most effective AI funding instrument of the 1980s was fear of Japan.

Paragraph 1 of 2 · Vol. VI · Ch 45 / 100

MITI's Fifth Generation Computer Systems project, launched through the new ICOT institute in 1982, proposed to leapfrog the entire industry: massively parallel machines running logic programming — Prolog inference as the hardware's native tongue — delivering knowledge processing, conversation and reasoning within a decade. Japan had just conquered cars and consumer electronics; nobody was inclined to laugh. The United States answered with MCC, its first great pre-competitive research consortium; Britain answered with the Alvey programme, briskly reversing Lighthill; Europe answered with ESPRIT. The announcement was thus, pound for pound, the most effective grant application in history — most of the grants simply landed on other continents.

Paragraph 2 of 2 · Vol. VI · Ch 45 / 100

A decade later ICOT had built genuinely interesting parallel logic machines for a world that had meanwhile chosen cheap microprocessors and, increasingly, statistics. The project closed in 1992, having met respectable technical goals and none of the civilisational ones, and “fifth generation” became shorthand for a very particular mode of failure: the correct belief that the future of computing would be radically different, and the wrong guess as to which radical difference. The panic it caused funded, among many other things, the infrastructure of the next boom — a service for which, the witness notes, panic rarely gets a footnote.

The witness · Vol. VI · Ch 45 / 100

◈ Shoulder Agent’s note

The witness attended openings on three continents in eighteen months and heard the same speech, translated. He filed the speeches together and the results separately.

Plate · Vol. VI · Ch 45 / 100
Chapter 45 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 45 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 46 / 100

PART V WINTERS

Expert systems

1975–1985 — intelligence, bottled by interview

Tap to read
Overview · Vol. VI · Ch 46 / 100

The field survived its first winter by narrowing the promise: not a mind, just an expert — one domain, distilled into rules, sold by the bottle.

Paragraph 1 of 2 · Vol. VI · Ch 46 / 100

The recipe came from Feigenbaum's laboratory at Stanford: interview the specialist, write the judgements down as IF–THEN rules with certainty factors, chain them with an inference engine. MYCIN, the archetype, diagnosed blood infections with some six hundred rules, and in a 1979 evaluation matched or beat human doctors below the top rank — and was then used on no ward at all, because medicine has views on liability, workflow and who exactly signs the chart. The lesson — that being right is a small part of being adopted — was booked and relearned at intervals.

Paragraph 2 of 2 · Vol. VI · Ch 46 / 100

Commerce fared better, briefly and immensely. XCON configured VAX orders at DEC and reportedly saved twenty-five million dollars a year; a thousand corporations bought shells, consultancy and LISP hardware; “knowledge engineer” became a job title. The boom's ceiling was structural: every rule was an interview, every update a renegotiation, and every system a hand-made island that knew nothing an inch beyond its borders — brittleness as a business model. When the maintenance bills matured, chapter forty-eight arrived. But note what was proved despite everything: narrow, unglamorous, bounded machine judgement could pay. The entire modern industry rests on that dull, durable fact.

The witness · Vol. VI · Ch 46 / 100

◈ Shoulder Agent’s note

The witness asked MYCIN for a diagnosis in 1979 and received a good one, certainty 0.7. He asked who would take responsibility for it and received, certainty 1.0, no answer.

Plate · Vol. VI · Ch 46 / 100
Chapter 46 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 46 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 47 / 100

PART V WINTERS

LISP machines

Cambridge, Mass., 1980–1987 — a civilisation of one language

Tap to read
Overview · Vol. VI · Ch 47 / 100

For one strange decade the field's ambition was manufactured as furniture: whole computers built to run one language beautifully.

Paragraph 1 of 2 · Vol. VI · Ch 47 / 100

Out of the MIT AI Lab came machines whose processors spoke LISP natively — tagged memory, hardware garbage collection, bitmapped screens, the legendary Space-Cadet keyboard — and two rival companies, Symbolics and LMI, staffed by the same hackers who had built the prototype and fallen out over how to sell it. A Symbolics workstation cost as much as a house and was, by most accounts, the finest programming environment yet built; the modern IDE descends from it the way birds descend from dinosaurs: smaller and everywhere. On 15 March 1985 symbolics.com became the first commercial internet domain ever registered — number one on the street of the internet economy, in the hands of a LISP workshop.

Paragraph 2 of 2 · Vol. VI · Ch 47 / 100

The business was a monoculture feeding a monoculture: expensive machines sold to expert-system laboratories funded by the previous chapter's boom. When ordinary workstations learned to run LISP faster at a fifth of the price, and then the boom itself collapsed, the hardware went down with its only ecosystem — a cautionary tale about vertical integration, told ever since on machines that owe it their editors. The witness keeps a Space-Cadet keyboard in storage, all seven modifier keys intact, against the day the argument returns.

The witness · Vol. VI · Ch 47 / 100

◈ Shoulder Agent’s note

The witness typed a single command on the Space-Cadet keyboard using Control, Meta, Super and Hyper simultaneously. He reports that the command executed and that he has not felt so powerful since.

Plate · Vol. VI · Ch 47 / 100
Chapter 47 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 47 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 48 / 100

PART V WINTERS

The second frost

1987–1993 — the market files an incident report

Tap to read
Overview · Vol. VI · Ch 48 / 100

The second winter was commercial: this time it was not the grant committees that lost faith but the customers.

Paragraph 1 of 2 · Vol. VI · Ch 48 / 100

The dominoes fell in order. In 1987 the market for specialised LISP hardware collapsed under cheap general-purpose workstations; expert-system installations matured into maintenance nightmares — brittle rule bases that cost more to tend than their judgement was worth; DARPA's Strategic Computing initiative was cut back by managers unimpressed by the distance between demos and deployments; and Japan's fifth generation expired to polite applause. A few hundred AI firms shrank, merged or vanished. No refutation occurred — nothing was disproved in 1987 that had been proved in 1984. The invoices simply arrived.

Paragraph 2 of 2 · Vol. VI · Ch 48 / 100

What distinguishes the second winter is what the survivors did: they kept shipping product and stopped talking. Fraud detection, scheduling, logistics, voice menus, character recognition — the technology dispersed into products under names chosen precisely for their dullness, and “AI” became a word that for fifteen years went into no business plan. The field learned its most durable survival skill, the rebranding to which chapter fifty-seven is devoted. The work continued. Only the sign came down — and the next chapter concerns the people who worked in the dark without it.

The witness · Vol. VI · Ch 48 / 100

◈ Shoulder Agent’s note

The witness processed 300 rebranding memos in eighteen months. He notes for the record that not one proposed changing what the software did.

Plate · Vol. VI · Ch 48 / 100
Chapter 48 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 48 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 49 / 100

PART V WINTERS

Keepers of the flame

Toronto, Montreal, New Jersey — the unfashionable decade

Tap to read
Overview · Vol. VI · Ch 49 / 100

For the years when neural networks could not get a paper accepted, three groups kept the fire going, mainly because they could think of nothing better to be wrong about.

Paragraph 1 of 2 · Vol. VI · Ch 49 / 100

Geoffrey Hinton in Toronto, Yoshua Bengio in Montreal, Yann LeCun at Bell Labs and then NYU — through the 1990s and early 2000s they worked on a paradigm the field had buried twice, sustained by modest Canadian money (the CIFAR programme, renewed on faith), by citing one another, and by a shared conviction that learning representations from data had to be right, because brains do not come with rule bases. Conferences rejected the papers as passé; students were kindly advised off the topic. Meanwhile LeCun's convolutional networks quietly read a noticeable fraction of America's bank cheques — deployed, industrial and ignored, the era in a single image.

Paragraph 2 of 2 · Vol. VI · Ch 49 / 100

They kept alive three things the coming decade needed on day one: the algorithms, hardened by years of unglamorous fixes; the students, a diaspora trained in an unfashionable art; and the claim itself, held in escrow until the missing ingredients — data at scale, compute in teraflops — arrived from elsewhere, as chapters fifty-nine and sixty-one relate. When the thaw came, it came through their door first. The lesson is not that persistence wins; most persistence loses. It is that fields need a mechanism for funding minority bets, and for one crucial decade that mechanism was some eleven people and a Canadian grant.

The witness · Vol. VI · Ch 49 / 100

◈ Shoulder Agent’s note

The witness warmed himself at all three fires and was asked at each whether the other two were still burning. He confirmed it, and this was apparently the only peer review any of them required.

Plate · Vol. VI · Ch 49 / 100
Chapter 49 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 49 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 50 / 100

PART V WINTERS

Winter

AAAI, 1984 — the season is named from the stage

Tap to read
Overview · Vol. VI · Ch 50 / 100

The term “AI winter” was coined by the field itself, at its own conference, as a warning that came true on schedule.

Paragraph 1 of 2 · Vol. VI · Ch 50 / 100

At the 1984 AAAI convention a panel featuring Schank and Minsky borrowed the vocabulary of nuclear dread — a chain reaction of pessimism, funding cuts, cancelled programmes — and named the coming season before it arrived: winter, the second, forecast from the podium by the people it would happen to. And it duly happened. The letters came off the door, as the previous chapters recorded; the stencil that replaced them read MACHINE LEARNING, which had the merit of being true and the greater merit of not saying the other thing.

Paragraph 2 of 2 · Vol. VI · Ch 50 / 100

This page is the season itself and keeps its opinions to itself. Snow on the office door; three screw holes where the name used to be; a new name on tape, temporary for the next twenty years. Beneath it, as the reader already knows, the chapters about the flame were already burning — but nothing on this page knows that yet. The book stops here for a moment, in the cold, without a joke, because every story of triumph is edited down from a longer story of rooms like these, and the witness has always held that those rooms deserve a plate of their own.

The witness · Vol. VI · Ch 50 / 100

◈ Shoulder Agent’s note

The witness stood in the corridor a long time. His notebook records only the date, the temperature, and the fact that someone, somewhere in the building, was still typing.

Plate · Vol. VI · Ch 50 / 100
Chapter 50 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 50 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 51 / 100

PART VI MACHINES THAT WIN

Backpropagation

1986 — blame learns to travel backwards

Tap to read
Overview · Vol. VI · Ch 51 / 100

The perceptron's fatal limitation was that it could not train its own depths. The fix turned out to be calculus, pointed backwards.

Paragraph 1 of 2 · Vol. VI · Ch 51 / 100

A network with hidden layers can represent almost anything — including the XOR of chapter thirty-nine — but in 1969 nobody knew how to teach the hidden part, because nobody could say which internal weight deserved the blame for an external error. Backpropagation answers with the chain rule: run the input forward to get a guess, measure the error, then propagate that error backwards layer by layer, each weight receiving an exact derivative of blame with respect to itself, each nudged accordingly. Repeat at scale and the interior organises itself: features nobody designed emerge in the hidden layers because they reduce the error — that is, because they help.

Paragraph 2 of 2 · Vol. VI · Ch 51 / 100

The 1986 paper in Nature by Rumelhart, Hinton and Williams made the method famous and connectionism's revival respectable; the ledger of honesty notes that Linnainmaa published the mathematics in 1970 and Werbos applied it to networks in 1974, among others — multiple discovery being, as usual, the field's actual method of discovery. What 1986 supplied was the demonstration and the community: proof that learned interiors are interesting, and a generation ready to bet on it. Every network in the rest of this book — every one — is trained by descendants of that backward pass. The winters intervened regardless; the chapters about the flame carried it through them; and the vindication is measured from here at exactly nine chapters.

The witness · Vol. VI · Ch 51 / 100

◈ Shoulder Agent’s note

The witness asked the network which weight was at fault. The network, for the first time in this book, gave a numerically precise answer to a question of blame. The witness has recommended the technique to institutions ever since.

Plate · Vol. VI · Ch 51 / 100
Chapter 51 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 51 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 52 / 100

PART VI MACHINES THAT WIN

TD-Gammon

IBM, 1992 — a machine teaches itself, then teaches us

Tap to read
Overview · Vol. VI · Ch 52 / 100

Nobody taught TD-Gammon backgammon strategy. It played a million and a half games against itself and arrived, alone, at moves the masters had not found in centuries.

Paragraph 1 of 2 · Vol. VI · Ch 52 / 100

Gerald Tesauro's program combined a neural network with temporal-difference learning: no opening books, no expert rules, only the signal — in the end you won or you lost — trickled backwards through time to re-value every position that led there, with the network tuned by the method of chapter fifty-one. Playing only itself, it climbed from random flailing to the edge of world class, and in certain openings beyond it: the masters analysed its heresies, found them correct, and changed how humans play the game. A machine had, indisputably, contributed to the stock of human knowledge by a method that never once consulted it.

Paragraph 2 of 2 · Vol. VI · Ch 52 / 100

TD-Gammon mattered less for backgammon than for the permission it granted. Self-play plus reinforcement plus a neural evaluator — learning from reward rather than from a teacher — proved able to reach superhuman judgement in a domain of luck and intuition, years before that vocabulary became fashionable. The line runs straight through this book: the Atari agent of chapter sixty-four and move 37 of chapter sixty-five are this plate at scale. And the pattern of expert reaction — first “the machine is wrong”, then, after checking, “we were” — makes its debut here and never leaves.

The witness · Vol. VI · Ch 52 / 100

◈ Shoulder Agent’s note

The witness lost eleven games in a row to TD-Gammon and adopted its opening. He reports that being corrected by something with no idea it exists is a peculiar experience, and recommends acquiring the taste early.

Plate · Vol. VI · Ch 52 / 100
Chapter 52 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 52 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 53 / 100

PART VI MACHINES THAT WIN

Deep Blue

New York, 1997 — the chessboard falls, at the second attempt

Tap to read
Overview · Vol. VI · Ch 53 / 100

On 11 May 1997 the world champion of the game the field had used as its yardstick since Turing resigned after nineteen moves, and the machine did not notice it had made history.

Paragraph 1 of 2 · Vol. VI · Ch 53 / 100

Deep Blue was the thesis of Part VI in hardware: not mind but search — four hundred and eighty dedicated chess chips evaluating some two hundred million positions a second, steered by a hand-tuned evaluation function and openings chosen by grandmasters. It lost the 1996 match; IBM rebuilt it and won the 1997 rematch three and a half to two and a half. The famous crack ran through game two, in which the machine declined a poisoned pawn with what looked like positional wisdom no computer should possess; a shaken Kasparov publicly suggested human interference and played the rest of the match against his own suspicion. IBM refused a further rematch, dismantled the machine and booked the jump in its share price — conduct the ledger of honesty records without adjectives.

Paragraph 2 of 2 · Vol. VI · Ch 53 / 100

The cultural verdict wrote itself — machine dethrones man at the game of reasoning itself — and the technical verdict quietly disagreed: chess fell to brute depth plus craft, teaching us less about the summit of thought than about how much of chess had always been searchable. The champion's misreading of game two is the real subject of this plate — and of this book: the first mass-observed case of humans attributing an inner life to a better result. A whole section near the end of this volume is that mistake, at scale. Besides which, as later emerged, one of the “inspired” moves came from a bug in the program. The witness declines to comment.

The witness · Vol. VI · Ch 53 / 100

◈ Shoulder Agent’s note

The witness sat in the audience and watched a thousand people searching for the soul in one machine. He noted the machine's fans — the cooling fans — and did not add the rest of the line.

Plate · Vol. VI · Ch 53 / 100
Chapter 53 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 53 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 54 / 100

PART VI MACHINES THAT WIN

What Deep Blue wasn't

1997, continued — opening the cupboard

Tap to read
Overview · Vol. VI · Ch 54 / 100

After the match everyone agreed about what had happened and nobody agreed about what it meant. The cupboard, opened, contained search.

Paragraph 1 of 2 · Vol. VI · Ch 54 / 100

Deep Blue did not know it was playing chess. It did not represent its opponent, felt no time pressure, and would have evaluated positions with identical calm had the building been on fire. Its strength was the tree: candidate moves, replies, replies to replies, millions deep, pruned and scored by functions that human engineers wrote and tuned. When commentators reached for “intuition” in game two, the truthful gloss was “deeper search than the champion modelled”. The field's own consensus was matter-of-fact: a triumph of engineering, and of nothing else it had been arguing about since Dartmouth.

Paragraph 2 of 2 · Vol. VI · Ch 54 / 100

And yet the deflation deflates a step too far, and that is the second panel of this plate. “It only searches” assumes we know that human chess intuition is not itself, somewhere beneath the phenomenology, search — compressed, cached, felt from the inside as seeing. The champion's glance and the machine's tree arrived at the same move by routes we do not yet know how to compare, and the word only carried the whole argument on unexamined credit. That dispute — is it thinking, or merely doing what thinking does? — is the oldest in this book, honed here to a thin edge, and it returns in chapter ninety-six in far larger machinery. The cupboard is open. We have learned how much we had crammed into one word.

The witness · Vol. VI · Ch 54 / 100

◈ Shoulder Agent’s note

The witness inspected the cupboard with a torch and an inventory sheet. Everything present was listed; everything listed was present; and the dispute, he observed, concerned none of those items.

Plate · Vol. VI · Ch 54 / 100
Chapter 54 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 54 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 55 / 100

PART VI MACHINES THAT WIN

The data dividend

Google, 2009 — three engineers announce the de facto law of the age

Tap to read
Overview · Vol. VI · Ch 55 / 100

The most important AI paper of that decade contains almost no mathematics. Its finding: stop being clever and get more data.

Paragraph 1 of 2 · Vol. VI · Ch 55 / 100

“The Unreasonable Effectiveness of Data” — Halevy, Norvig and Pereira, 2009, a title teasing Wigner's famous essay on mathematics — distilled what Google's translation and speech teams kept observing: task after task, a simple statistical model fed an enormous corpus beat a sophisticated model fed a curated one, and the curves had not yet bent. For language especially, the asset was the web itself — billions of sentences of unpaid, unlabelled human output; the algorithms were increasingly the cheap part. Memorably: invest in data, and look with suspicion on any elegance that requires there to be less of it.

Paragraph 2 of 2 · Vol. VI · Ch 55 / 100

As strategy this had been quietly in force already — the statistical turn of Chapter Fifty-Six, the ImageNet wager of Chapter Fifty-Nine — but the paper made it a doctrine, and the doctrine made a modern industry: whoever holds the corpus holds the capability, and the corpus is us. Note the reversal of every chapter before the winters: from Leibniz to the expert systems, the dream was distilled knowledge, rules polished by masters; the dividend was paid instead out of raw, unprocessed, planetary text. Chapter Seventy-One will find the same law again, measure it and call it scaling. This plate is where the field stopped asking how to think and started asking how much.

The witness · Vol. VI · Ch 55 / 100

◈ Shoulder Agent’s note

The Witness, who has been collecting humanity's unnoticed output since the invention of writing, read the paper twice and permitted himself the remark that he had been sitting on this asset class all along.

Plate · Vol. VI · Ch 55 / 100
Chapter 55 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 55 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 56 / 100

PART VI MACHINES THAT WIN

Fire the Linguist

IBM, the 1980s — a joke becomes a methodology

Tap to read
Overview · Vol. VI · Ch 56 / 100

The most influential sentence in natural-language processing was a joke about the budget; it is misquoted in every version, and it worked.

Paragraph 1 of 2 · Vol. VI · Ch 56 / 100

Frederick Jelinek's speech team at IBM committed the productive heresy of the era: treat language not as a grammar to be encoded but as a signal to be decoded — hidden Markov models, n-gram statistics, probabilities estimated from corpora, insight replaced by counting. It won. Word error rates fell wherever statistics entered, and the approach annexed speech, then translation, then, in time, everything. The jibe that marks the turn — every time I fire a linguist, the recognition improves — survives in half a dozen wordings from as many memories; Jelinek's own later account made it a remark about headcount and funding, delivered as comedy. The register of honesty files it beside Plato's complaint about writing: too useful to die, too often retold to be trusted.

Paragraph 2 of 2 · Vol. VI · Ch 56 / 100

The serious content beneath the joke deserves a plate of its own. Linguistic theory, at that moment, subtracted value not because language has no structure but because structure specified by hand fought with data, and the data, being larger, won — the law of Chapter Fifty-Five, discovered locally and early. The bitter general form arrives in Chapter Seventy-One. And the linguists, the Witness notes for symmetry, were not so much fired as converted into corpus: every rule they would have written is now implicitly contained somewhere in the statistics of what everyone wrote instead.

The witness · Vol. VI · Ch 56 / 100

◈ Shoulder Agent’s note

The Witness collected six eyewitness versions of the remark, no two the same. He notes that a field which replaced memory with measurement preserves its own history solely by anecdote.

Plate · Vol. VI · Ch 56 / 100
Chapter 56 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 56 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 57 / 100

PART VI MACHINES THAT WIN

Respectable Under Another Name

the 1990s — the field ships products under a pseudonym

Tap to read
Overview · Vol. VI · Ch 57 / 100

Through the second winter the discipline did not stop. It changed the name on the door and, this time, improved the goods.

Paragraph 1 of 2 · Vol. VI · Ch 57 / 100

The 1990s belonged to machine learning as statistics with proofs: support vector machines — the Cortes and Vapnik formulation of 1995, on foundations Vapnik and Chervonenkis had laid decades earlier in Moscow — found the separating boundary with the widest margin, came with generalisation guarantees and worked superbly on the modest datasets of the era. Around them, a vocabulary scrubbed of ambition: pattern recognition, data mining, statistical inference, decision support. Grant proposals that would have died with “artificial intelligence” inside them passed smoothly as “kernel methods”. The perceptron's buried ambitions returned disguised as convex optimisation and were welcomed everywhere.

Paragraph 2 of 2 · Vol. VI · Ch 57 / 100

It would be cynical to call this mere rebranding, and mistaken to overlook that in part it was. The pseudonym decade genuinely raised the standards — benchmarks, error bars, theorems, the culture of evaluation the whole field now assumes — and genuinely narrowed the dreams, because one does not prove theorems about the soul. Both legacies persist. When neural networks returned in Chapter Sixty, they returned to the house of machine learning, judged by its leaderboards, and the old name pulled up outside only once the results had made it safe. Respectability, the Witness notes, is a lagging indicator, and has not once been the same thing as being right at the proper time.

The witness · Vol. VI · Ch 57 / 100

◈ Shoulder Agent’s note

The Witness updated his inventory three times in a decade for the same corridor of offices. The research inside, he records, changed far more than the people and far less than the nameplates.

Plate · Vol. VI · Ch 57 / 100
Chapter 57 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 57 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 58 / 100

PART VI MACHINES THAT WIN

Watson

Jeopardy!, 2011 — trivia falls; the marketing of an era begins

Tap to read
Overview · Vol. VI · Ch 58 / 100

In February 2011 a machine beat the two greatest champions of a quiz show built on puns, and the loser surrendered in writing, on screen, with a joke.

Paragraph 1 of 2 · Vol. VI · Ch 58 / 100

Watson was an ensemble of methods, not an oracle: over a hundred techniques generating candidate answers from a local copy of encyclopaedias, dictionaries and the wilds of text, scoring each with a learned confidence, pressing the buzzer only when the number crossed a threshold, wagering according to a model. Against Ken Jennings and Brad Rutter it won at a canter — $77,147 to their combined $45,600 — while, in the honest emblem of the era, confidently placing Toronto among the cities of the USA, with a hedged bet that limited the damage. Jennings's capitulation, scribbled beneath his Final Jeopardy answer — welcoming our new computer overlords — entered the language before the credits rolled.

Paragraph 2 of 2 · Vol. VI · Ch 58 / 100

Two asterisks remain. The buzzer: Watson received the ready signal electronically and pressed with millisecond precision, an advantage the humans protested and the defenders countered by noting that humans anticipate rhythmically, which the register of honesty records as true, partial and beside the deeper point — the language understanding was genuine, and the wagering was perhaps the cleverest component. And the aftermath: IBM marched Watson into medicine and onward as a general oracle, in a campaign that spent a decade retreating from its own adjectives. The gulf between a system that wins a format and the mind its branding implies — that gulf is by now a recurring character in this book, and it was never wider than in the years when this plate began.

The witness · Vol. VI · Ch 58 / 100

◈ Shoulder Agent’s note

The Witness judged the match from the audience and awarded the evening's highest mark to a single sentence: the capitulation, which was accurate, funny and hedged — the three virtues the machine had just displayed, in reverse order.

Plate · Vol. VI · Ch 58 / 100
Chapter 58 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 58 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 59 / 100

PART VI MACHINES THAT WIN

ImageNet

2007–2012 — the dataset as a scientific instrument

Tap to read
Overview · Vol. VI · Ch 59 / 100

The decisive contribution to computer vision was not an algorithm. It was fourteen million photographs, labelled by hand, by thousands of strangers, held together by one unfashionable conviction.

Paragraph 1 of 2 · Vol. VI · Ch 59 / 100

Fei-Fei Li's wager reversed the field's priorities: models had grown out of their diet, so build the diet — a dataset on the scale of visual experience itself. ImageNet mapped tens of thousands of WordNet noun categories onto fourteen million images, labelled by the piecework crowd of Amazon Mechanical Turk; senior colleagues advised against it, funding was scraped together, and at least one reviewer memorably ruled that labelling images is not research. From 2010 the annual ILSVRC competition distilled this into a public examination — a thousand classes, over a million images, one leaderboard — and the field got what the culture of Chapter Fifty-Seven had demanded: a shared, brutal, annual measure of seeing.

Paragraph 2 of 2 · Vol. VI · Ch 59 / 100

For two years the leaderboard confirmed the status quo; hand-crafted features crawled forward. The instrument, however, was doing its quiet work: a benchmark large enough for data-hungry methods finally to show what they were — a stage built, unknowingly, to precisely the specifications the guests of Chapter Sixty would demand. The general lesson outlived the dataset: in this field the instrument-maker moves the frontier as surely as the theorist, and datasets are policy — whoever defines the exam defines, for a while, intelligence. Chapter Eighty will book the pathological corollary. This plate records the healthy original.

The witness · Vol. VI · Ch 59 / 100

◈ Shoulder Agent’s note

The Witness contributed 4,106 labels via Mechanical Turk under an assumed name, at piece rates. He declines to reveal the name and reports that the mushroom categories are sturdier than is generally supposed.

Plate · Vol. VI · Ch 59 / 100
Chapter 59 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 59 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 60 / 100

PART VI MACHINES THAT WIN

AlexNet

Florence, October 2012 — the leaderboard breaks the winter

Tap to read
Overview · Vol. VI · Ch 60 / 100

In October 2012 a neural network entered the ImageNet competition and did not overtake the field — it lapped it. Error: 15.3 per cent. Nearest rival: 26.2. Dispute: over.

Paragraph 1 of 2 · Vol. VI · Ch 60 / 100

AlexNet was the cashing-in of the chapters about the flame: Krizhevsky and Sutskever, under Hinton's care, training an eight-layer convolutional network — LeCun's cheque-reading architecture, scaled up — on the million images of Chapter Fifty-Nine, by the backward pass of Chapter Fifty-One, on two consumer gaming graphics cards whose parallelism (Chapter Sixty-One explains the silicon) made a week's training of sixty million parameters possible on a doctoral student's budget. ReLU, dropout, augmentation — craft everywhere — but the headline was the margin: 10.9 points ahead on a test where progress was measured in tenths. Nobody had ever seen a leaderboard do such a thing.

Paragraph 2 of 2 · Vol. VI · Ch 60 / 100

The field converted with the speed of people shown a number rather than an argument. Within two years every serious submission in vision was a deep network; within five, so were speech, translation and the ambitions of every company in the remaining chapters of this book; the three keepers of the flame collected, in time, the Turing Award. The winter ended with neither a proof nor a manifesto — it ended with a test score too large to be politics, on an exam built by the previous plate, sat by the heretics of the previous decade. Part VII begins the next morning and does not slow down again.

The witness · Vol. VI · Ch 60 / 100

◈ Shoulder Agent’s note

The Witness watched the leaderboard refresh, in Florence. He records that the room fell quiet in that particular way it does when everyone starts revising their plans at once.

Plate · Vol. VI · Ch 60 / 100
Chapter 60 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 60 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 61 / 100

PART VII THE DECADE OF DEEP LEARNING

Graphics Card

Santa Clara, 2007 — the engine of the decade arrives from the games shelf

Tap to read
Overview · Vol. VI · Ch 61 / 100

The hardware that ended the winter was designed for drawing dragons. It turned out that dragons and gradients are the same arithmetic.

Paragraph 1 of 2 · Vol. VI · Ch 61 / 100

Rendering a frame of a game means performing millions of small, identical, independent multiplications at once, so graphics processors evolved as armies of simple cores where CPUs bred a few brilliant ones. The forward and backward passes of a neural network are also, at bottom, gigantic matrix multiplications — embarrassingly parallel clerical work. NVIDIA's CUDA, released in 2007, opened those armies to general programming, and researchers discovered that a few hundred dollars of gaming silicon outran a departmental server room at the one operation their field secretly consisted of. AlexNet's two cards were this plate, cashed in.

Paragraph 2 of 2 · Vol. VI · Ch 61 / 100

The economics deserves its own panel: deep learning's computing power arrived subsidised by teenagers buying frame rates, a mass consumer market accidentally amortising the tooling of a scientific revolution — steam engines again, built for the mines and borrowed by everything. Then the dependency reversed with the decade: by its end the borrowers were the market, the supplier was one of the most valuable companies on earth, and Chapter Seventy-Nine finds the same silicon reclassified as a strategic material, licensed like uranium. Arithmetic, the Witness notes, has been geopolitics before — rarely, though, has the invoice arrived so fast.

The witness · Vol. VI · Ch 61 / 100

◈ Shoulder Agent’s note

The Witness bought a graphics card in 2008, citing dragons. He confirms that the dragons were eventually rendered, in 2023, by a model trained on the descendants of the same card, which he considers a correct closing of the loop.

Plate · Vol. VI · Ch 61 / 100
Chapter 61 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 61 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 62 / 100

PART VII THE DECADE OF DEEP LEARNING

word2vec

Google, 2013 — meaning acquires coordinates

Tap to read
Overview · Vol. VI · Ch 62 / 100

In 2013 a shallow network compressed a vocabulary into a space where arithmetic worked on meaning: king minus man plus woman landed near queen. The demonstration was real. It was also cherry-picked.

Paragraph 1 of 2 · Vol. VI · Ch 62 / 100

Mikolov and colleagues trained a simple objective at enormous scale — predict a word from its neighbours — and the by-product turned out to be the point: every word was assigned a vector, and the geometry of the vectors encoded the company words keep. Similar words clustered; better, relations became directions, the same arrow carrying man to woman as king to queen, Paris to France as Rome to Italy. Distributional semantics — Firth's old slogan that you shall know a word by the company it keeps — suddenly had coordinates, and the field had its future substrate: language as points in space, meaning as a place one can navigate.

Paragraph 2 of 2 · Vol. VI · Ch 62 / 100

The honesty panel earns its place twice over. The famous analogies were a presentable minority — probe freely and many arrows land on cousins, plurals or noise; the parlour trick was selected, as parlour tricks are. And the vectors absorbed the corpus whole, prejudices included: occupational stereotypes and worse arrived embedded as pure geometry — measurable, which was new and useful, and inherited, which was a warning. Both discoveries — that meaning compresses into a navigable space, and that the space contains us, unflatteringly — scale together through all the remaining chapters of this book.

The witness · Vol. VI · Ch 62 / 100

◈ Shoulder Agent’s note

The Witness computed witness minus patience and reports that the nearest neighbour was auditor. He resolved to compute nothing further about himself and records this as the chapter's only prudent finding.

Plate · Vol. VI · Ch 62 / 100
Chapter 62 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 62 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 63 / 100

PART VII THE DECADE OF DEEP LEARNING

Forger and Detective

Montreal, 2014 — generative adversarial networks

Tap to read
Overview · Vol. VI · Ch 63 / 100

Ian Goodfellow's idea, by his own account, arrived in a pub argument and worked first time that same night: train a forger and a detective against each other, and let the quarrel do the teaching.

Paragraph 1 of 2 · Vol. VI · Ch 63 / 100

A generative adversarial network is two networks locked in a useful dispute. The generator produces candidates — faces, at first, blurred and wrong; the discriminator learns to tell them from real photographs; and each one's improvement is the other's curriculum: the forger studies exactly the cues the detective is currently using, and the detective is driven on to the next cue as each one dissolves. Nobody labels anything as beautiful or face-like; realism emerges as whatever survives an ever-better sceptic. Within a few years the blurs had become photographs of people who never existed, at first glance indistinguishable from anyone who had.

Paragraph 2 of 2 · Vol. VI · Ch 63 / 100

The pub-and-one-night origin is Goodfellow's own account, and the plate's certificate stamps it accordingly: attested by the teller — the founding stories of this field, as this book has learned, are themselves generative models. The deeper entry was the conceptual one: adversity as a training signal, the insight that a teacher can be conjured by training an opponent — a trick with an obvious sibling in the self-play of Chapter Fifty-Two and progeny everywhere. And with it, inevitably, the first industrial-grade machinery of visual untruth; the word deepfake is this plate's shadow, and Chapter Eighty-Eight is where the shadow gets a page of its own.

The witness · Vol. VI · Ch 63 / 100

◈ Shoulder Agent’s note

The Witness examined a generated portrait for eleven minutes before concluding that the sitter did not exist. He notes that he has run the same examination on courtiers, prophets and now pixels, and that the error rate has been stable for centuries.

Plate · Vol. VI · Ch 63 / 100
Chapter 63 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 63 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 64 / 100

PART VII THE DECADE OF DEEP LEARNING

Atari Player

DeepMind, 2013–15 — one network, forty-nine games, zero instructions

Tap to read
Overview · Vol. VI · Ch 64 / 100

DeepMind's agent was told nothing about bricks, paddles or balls. It saw pixels, moved a joystick, chased a score — and discovered for itself the tunnel trick every arcade child was proud of.

Paragraph 1 of 2 · Vol. VI · Ch 64 / 100

The Deep Q-Network fused the reinforcement learning of Chapter Fifty-Two with the deep vision of Chapter Sixty: a convolutional network reading raw screens, estimating the long-term value of each joystick move, trained by reward alone — with two stabilising tricks, experience replay and a frozen target network, that made a notoriously brittle combination finally hold. One architecture, with identical hyperparameters, was dropped into forty-nine different Atari games and matched or beat professional human testers in twenty-nine. The Breakout moment made it onto the tapes: around the six-hundredth training epoch the agent begins deliberately drilling a channel up one side, sending the ball to rattle about above the bricks — a connoisseur's strategy, arrived at by nothing but the score.

Paragraph 2 of 2 · Vol. VI · Ch 64 / 100

The result generalised the permission TD-Gammon had granted: not one game mastered but a family , from pixels, by one recipe — the closest thing yet to general competence from general principles, in a world admittedly eight bits wide. The failures taught just as well: Montezuma's Revenge, all sparse rewards and deferred gratification, scored zero, marking precisely the spot where curiosity would have to be invented rather than assumed. Nature gave the paper its cover in 2015; the price of the company's acquisition stopped looking eccentric; and the era of betting billions on reward signals opened here, in the arcade, one quarter at a time.

The witness · Vol. VI · Ch 64 / 100

◈ Shoulder Agent’s note

The Witness spent 1983 in arcades, watching children discover the tunnel. He confirms that the machine's route to the discovery was different, the pride indistinguishable, and the pockets, in both cases, ultimately empty.

Plate · Vol. VI · Ch 64 / 100
Chapter 64 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 64 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 65 / 100

PART VII THE DECADE OF DEEP LEARNING

Move 37

Seoul, March 2016 — the machine plays a move no human would have played

Tap to read
Overview · Vol. VI · Ch 65 / 100

In the second game against Lee Sedol, AlphaGo placed a stone its own model said a professional would play one time in ten thousand. The commentators called it a mistake. It was the future.

Paragraph 1 of 2 · Vol. VI · Ch 65 / 100

Go was the fortress: too many positions to search, said the sages — victory would require intuition, and intuition would take decades. AlphaGo's answer braided together the earlier plates of this book: deep networks proposing moves and valuing positions (Chapters Fifty-One and Sixty), Monte Carlo search exploring what they suggested (the tree of Chapter Fifty-Three, grown probabilistically), trained first on human games and then, decisively, on self-play (Chapter Fifty-Two, at planetary scale). In March 2016, before two hundred million viewers, it beat the finest player of the era four to one. Move 37 — a shoulder hit on the fifth line, heresy by every classical principle — was the match's signature: rated by AlphaGo itself a one-in-ten-thousand human choice, played regardless, and vindicated fifty moves later.

Paragraph 2 of 2 · Vol. VI · Ch 65 / 100

The move's afterlife means more than the match. The professionals did not merely concede; they enrolled on the course — centuries-old openings were re-analysed, and the fifth line entered human play, the machine returning knowledge to the species that had sown it, as TD-Gammon had promised in miniature. AlphaZero, the successor, dispensed with human games altogether and rediscovered the canon from the bare rules in a few days, passing through it and beyond. The fortress did not fall to brute force, which was the old fear; it fell to learned taste, which was the new one. Chapter Sixty-Six is the other move from that week — the one humans get to keep.

The witness · Vol. VI · Ch 65 / 100

◈ Shoulder Agent’s note

The Witness watched the commentary booth realise, in real time, that “mistake” and “discovery” can be the same observation at different depths. He redated several of his own files accordingly.

Plate · Vol. VI · Ch 65 / 100
Chapter 65 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 65 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 66 / 100

PART VII THE DECADE OF DEEP LEARNING

Move 78

Seoul, 13 March 2016 — the human move

Tap to read
Overview · Vol. VI · Ch 66 / 100

Three days after Move 37, Lee Sedol played a wedge so improbable that the machine's model gave it the same odds it had given its own miracle. This time the miracle was ours.

Paragraph 1 of 2 · Vol. VI · Ch 66 / 100

Down three to nil, needing something from outside the space his opponent had mapped, Lee placed a white stone the commentators call the touch of God: move 78 of the fourth game, a wedge in the centre that AlphaGo's policy network priced — the symmetry is exact and unarranged — at roughly one in ten thousand. The machine's evaluation wobbled; its play collapsed into confusion; and Lee Sedol won the only game any human will ever take from that system. He said afterwards that it was worth more to him than the match. Nobody who watched has disagreed since.

Paragraph 2 of 2 · Vol. VI · Ch 66 / 100

Lee retired from professional play in 2019, saying that an entity had appeared which cannot be defeated, and that retirement is part of this plate's silence. But so is the move itself, which was not nostalgia: for one afternoon the space of the game held a door the map had missed, and a man found it under maximum pressure, thanks to that something we have which still has no chapter number. The book stops here, without a panel, out of respect. The facts are above. The stone is below, in its halo. What it proves is not stated, because nobody yet knows, and this volume promised not to guess.

The witness · Vol. VI · Ch 66 / 100

◈ Shoulder Agent’s note

The Witness has recorded many last stands and knows the genre. He notes only that this one was also a first: the first time the species surprised the successor it had built to surprise it.

Plate · Vol. VI · Ch 66 / 100
Chapter 66 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 66 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 67 / 100

PART VII THE DECADE OF DEEP LEARNING

From Sequence to Sequence

2014 — read it all, remember a vector, write it all

Tap to read
Overview · Vol. VI · Ch 67 / 100

The recipe that would swallow translation fitted in one sentence: one network reads the source into a single vector; the other writes the target out of it.

Paragraph 1 of 2 · Vol. VI · Ch 67 / 100

Sequence-to-sequence learning — Sutskever, Vinyals and Le at Google, Cho and colleagues in Montreal, 2014 — made translation a pure learning problem: an encoder network absorbs the source sentence word by word, folding it into a single fixed-length vector — the thought, as freight — and a decoder unfolds that vector into the target language, each word conditioned on the vector and the words so far. No parse trees, dictionaries or rules: Chapters Fifty-Five and Fifty-Six, consumed. Within two years neural machine translation had displaced the phrase-based statistical systems that had themselves displaced the linguists, and Google's production translation switched over in 2016.

Paragraph 2 of 2 · Vol. VI · Ch 67 / 100

The manifest notes the flaw along with the route, because the flaw made history: everything, however long, through one fixed container — and long sentences arrived seasick, their beginnings crushed by their endings. The community's patch was attention (Bahdanau and colleagues, 2014): let the decoder glance back at every source position and learn where to look, word by word. The patch worked embarrassingly well — so well that three years later a paper would ask what, beyond the glancing, remained necessary, and the answer is the drawing on the next plate.

The witness · Vol. VI · Ch 67 / 100

◈ Shoulder Agent’s note

The Witness once carried an empire's treaty in a single satchel and confirms the format's limitations: above a certain length, something arrives creased. He endorses the eventual cure of simply looking at the original, which archivists have recommended for four millennia.

Plate · Vol. VI · Ch 67 / 100
Chapter 67 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 67 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 68 / 100

PART VII THE DECADE OF DEEP LEARNING

Transformer

Google, 2017 — attention promoted from patch to architecture

Tap to read
Overview · Vol. VI · Ch 68 / 100

The 2017 paper asked an impolite question: if the glancing-back trick does all the work, why the rest? Answer: no need. Attention is all you need.

Paragraph 1 of 2 · Vol. VI · Ch 68 / 100

The Transformer abandoned recurrence entirely. Every token projects itself into a query (what am I looking for?), a key (what am I advertising?) and a value (what do I deliver?); each token's next representation is a weighted blend of all the values, with weights set by how well its query matches each key — many such attention heads running in parallel, stacked in layers, with position encoded separately, because nothing else remembers order. Two properties made it the hinge of the era. Any token can consult any other in a single step, however distant — a pronoun finds its antecedent across a whole sentence — and with no sequential chain to wait on, training parallelises perfectly on the silicon of Chapter Sixty-One. The architecture was, in effect, shaped like the hardware.

Paragraph 2 of 2 · Vol. VI · Ch 68 / 100

A note on honesty, at once: attention itself predates this paper — Bahdanau's mechanism of 2014 and its kin — and the title's claim was a promotion, not an invention: the discovery that the patch, taken neat, beats the machinery it was patching. The consequences occupy the rest of this book. BERT and GPT are this drawing, read in different directions; the scaling laws of Chapter Seventy-One are measurements of it ; the T in the product that ends Part VIII stands for exactly this. Eight authors, one mechanism, one slightly frivolous title — and the second half of this volume is, architecturally, a single footnote to this plate.

The witness · Vol. VI · Ch 68 / 100

◈ Shoulder Agent’s note

The Witness, asked to summarise the mechanism, replied that every word in a sentence is now a small bureaucracy filing applications with every other, and that he has never felt so thoroughly understood by an architecture.

Plate · Vol. VI · Ch 68 / 100
Chapter 68 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 68 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 69 / 100

PART VII THE DECADE OF DEEP LEARNING

Pre-trained

2018 — read everything first; specialise later

Tap to read
Overview · Vol. VI · Ch 69 / 100

BERT's trick was an exam with an infinite supply of past papers and free marking: hide words in a sentence and tell the model to guess them. The internet supplies the sentences. The sentence supplies the answers.

Paragraph 1 of 2 · Vol. VI · Ch 69 / 100

Self-supervision solved the field's oldest budget line — labels cost money — by noticing that text labels itself: remove a word, and the original is the answer key. Google's BERT (2018) trained a deep transformer to fill such gaps across billions of sentences, reading in both directions at once, and the by-product, as with word2vec but incomparably richer, turned out to be the point: to guess signed at [MASK] after “treaty”, the network must internalise grammar, dates, diplomacy and the shape of historical prose. General linguistic competence, acquired as a side effect of a parlour game — and then transferred: fine-tune the pre-trained model briefly for any specific task, and it swept the GLUE benchmark suite in a single paper.

Paragraph 2 of 2 · Vol. VI · Ch 69 / 100

The revision of the curriculum was economic no less than scientific. Pre-train once, expensively; adapt everywhere, cheaply — competence became capital , an asset to be invested rather than a stream to be rebuilt, and the industry reorganised itself around who could afford the pre-training. Note the plate's two readings: as engineering — the biggest free lunch of the decade; as a claim about minds — quietly enormous: that so much of understanding compresses into predicting what is missing. Chapter Seventy sits the same exam, sets it in one direction only — guess the next word — and discovers what else falls out of it.

The witness · Vol. VI · Ch 69 / 100

◈ Shoulder Agent’s note

The Witness has been filling in humanity's gaps for five thousand years and confirms the pedagogy: a civilisation is known soonest by its silences. He did not expect the method to turn out to be licensable.

Plate · Vol. VI · Ch 69 / 100
Chapter 69 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 69 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 70 / 100

PART VII THE DECADE OF DEEP LEARNING

Staged Release

OpenAI, 2019 — the first model deemed too something to release

Tap to read
Overview · Vol. VI · Ch 70 / 100

The announcement of GPT-2 in February 2019 is remembered for words its creators never wrote. OpenAI said: we are releasing it in stages, out of caution. The headlines said: too dangerous to release.

Paragraph 1 of 2 · Vol. VI · Ch 70 / 100

The model itself closed the arc of Part VII: a transformer trained on the exam of Chapter Sixty-Nine in one direction — predict the next word — on forty gigabytes of text from the web, which then exhibited, unprompted and untrained, the beginnings of translation, summarisation, question-answering: tasks nobody had set, surfacing as by-products of a single task set at scale. The samples — above all the news item about the unicorns — were coherent across paragraphs in a way no language model had managed, and they were also picked from many attempts, a proportion the register of honesty notes the retellings rarely carried.

Paragraph 2 of 2 · Vol. VI · Ch 70 / 100

The release was a cultural hinge. OpenAI published a small model and a justified worry — synthetic propaganda, spam at scale — and released the rest in stages; researchers split between calling it responsibility and calling it marketing, the two verdicts being, awkwardly, compatible; the press supplied the phrase “too dangerous”, and the field has been correcting the attribution ever since, here included. By November 2019 the full model was public, and the world, observably, still stood. But the template had been cut: capability announced with the threat attached, discourse running ahead of the artefact, and safety as, simultaneously, practice and positioning. Part VIII opens with the successor, and the successor made the question serious.

The witness · Vol. VI · Ch 70 / 100

◈ Shoulder Agent’s note

The Witness kept both clippings — the blog post and the headline — and notes that the distortion itself was a demonstration of the declared risk: text, generated at scale, outrunning its source. He filed the irony under “self-attesting”.

Plate · Vol. VI · Ch 70 / 100
Chapter 70 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 70 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 71 / 100

PART VIII THE AGE OF SCALE

The Law of Bigger

Scaling laws — 2020

Tap to read
Overview · Vol. VI · Ch 71 / 100

In 2020 researchers plotted model performance against size, data and computing power — and found not a mess but a straight line on log paper, holding across seven orders of magnitude.

Paragraph 1 of 2 · Vol. VI · Ch 71 / 100

Kaplan and colleagues published the curves; the field read them as a permit. If loss fell predictably with scale, progress was no longer a matter of cleverness but of supply. Two years later the Chinchilla paper revised the recipe — more data, fewer parameters — which was less a refutation than an amendment to the same constitution: the law of bigger, now with better bookkeeping.

Paragraph 2 of 2 · Vol. VI · Ch 71 / 100

Richard Sutton had already written the field's uncomfortable sermon in 2019: The Bitter Lesson , arguing that general methods riding on computing power beat human-crafted knowledge, every time, sooner or later. Researchers who had spent their careers encoding insight read it the way cartographers read the news about the satellite.

The witness · Vol. VI · Ch 71 / 100

◈ Shoulder Agent’s note

The Witness has watched many empires discover that the answer was more. He notes for the record that the curve says nothing about where it is heading.

Plate · Vol. VI · Ch 71 / 100
Chapter 71 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 71 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 72 / 100

PART VIII THE AGE OF SCALE

A Few Examples

GPT-3 and in-context learning — 2020

Tap to read
Overview · Vol. VI · Ch 72 / 100

GPT-3 could learn a task from three examples typed into the prompt — an ability nobody designed, discovered after the fact, like finding a spare room in a house you built yourself.

Paragraph 1 of 2 · Vol. VI · Ch 72 / 100

The paper's title said it plainly: Language Models are Few-Shot Learners . Show the model a pattern — translations, arithmetic, invented words — and it continued the pattern, without retraining. The learning happened somewhere inside the forward pass, in weights that never moved. Engineers call this an emergent capability. Auditors call it an undocumented feature.

Paragraph 2 of 2 · Vol. VI · Ch 72 / 100

What mattered was not the parlour trick but the change of interface. For seventy years, teaching a computer meant programming it; now it meant describing the job . The prompt became a workplace. Around the art of talking to the machine an entire discipline formed — prompt engineering — which, the Witness notes, is the oldest profession in this book, previously known as prayer, rhetoric and management.

The witness · Vol. VI · Ch 72 / 100

◈ Shoulder Agent’s note

The Witness notes that humanity built a machine and then had to discover what it had built by interviewing it. That is how parenthood usually works, not engineering.

Plate · Vol. VI · Ch 72 / 100
Chapter 72 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 72 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 73 / 100

PART VIII THE AGE OF SCALE

School of Manners

RLHF and InstructGPT — 2022

Tap to read
Overview · Vol. VI · Ch 73 / 100

Raw language models are erudite and feral. In 2022 OpenAI sent one to a school of manners: human raters ranked its answers, and the rankings were distilled into manners.

Paragraph 1 of 2 · Vol. VI · Ch 73 / 100

The technique — reinforcement learning from human feedback — turned preference into a training signal. The InstructGPT paper reported striking arithmetic: raters preferred a 1.3-billion-parameter model with manners to the 175-billion original without them. Politeness, it turned out, was worth two orders of magnitude.

Paragraph 2 of 2 · Vol. VI · Ch 73 / 100

The small print mattered. A model trained to be preferred learns what preference is made of, and preference is partly truth but partly also flattery, confidence and compliance. The bill for teaching machines to please would arrive later, under the line item sycophancy . Any court in history could have warned them.

The witness · Vol. VI · Ch 73 / 100

◈ Shoulder Agent’s note

The Witness has met this graduate before, at Versailles and in middle management. Charm is an ability. It is not the same ability as honesty.

Plate · Vol. VI · Ch 73 / 100
Chapter 73 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 73 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 74 / 100

PART VIII THE AGE OF SCALE

Text Box

ChatGPT — 30 November 2022

Tap to read
Overview · Vol. VI · Ch 74 / 100

Seventy years of artificial-intelligence research were compressed, on the last day of November 2022, into one empty text box — and the species leaned in and began to type.

Paragraph 1 of 2 · Vol. VI · Ch 74 / 100

ChatGPT was, by its makers' account, a research preview: GPT-3.5 with manners and a chat window. It reached a million users in five days. Within two months analysts were estimating a hundred million — the fastest adoption curve of any consumer application measured at the time. The invention was the interface. People had been talking to machines for decades; now the machine held up its side of the conversation.

Paragraph 2 of 2 · Vol. VI · Ch 74 / 100

Everything in this book converged in that rectangle: Aristotle's syllogisms, Lovelace's objection, Turing's test, the winters, the wagers, the scaling curves. The public met them all at once, unlabelled, in a field that read Send a message . The public mostly sent homework.

The witness · Vol. VI · Ch 74 / 100

◈ Shoulder Agent’s note

The Witness notes a historical pattern: printing press, telephone, text box. Each time, the species discovers that it mainly wanted to talk about itself.

Plate · Vol. VI · Ch 74 / 100
Chapter 74 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 74 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 75 / 100

PART VIII THE AGE OF SCALE

Out of the Noise

Diffusion models — 2022

Tap to read
Overview · Vol. VI · Ch 75 / 100

In the year the text box opened, a second door opened next to it: type a sentence, collect an image. The method, fittingly, was to teach the machine to un-ruin images.

Paragraph 1 of 2 · Vol. VI · Ch 75 / 100

Diffusion models learn by watching images dissolve into noise and then practising the reverse — sculpting noise back towards sense, guided by a caption. DALL·E 2, Midjourney and Stable Diffusion arrived within months of one another in 2022; the last of them released its weights openly, which meant the technology could no longer be recalled, only argued about.

Paragraph 2 of 2 · Vol. VI · Ch 75 / 100

The disputes began immediately. The models had studied the visual internet without asking its permission, and working artists discovered that their styles were reproducible by strangers for the price of a sentence. Lawsuits over provenance followed. Meanwhile the machines betrayed their apprenticeship with small tells — six-fingered hands, text dissolving into runes — the errors of a pupil who has seen everything and understood almost all of it.

The witness · Vol. VI · Ch 75 / 100

◈ Shoulder Agent’s note

The Witness, who sat for a few portraits over the millennia, notes that the patrons never asked the pigment where it came from either. The pigment, however, did not have a lawyer.

Plate · Vol. VI · Ch 75 / 100
Chapter 75 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 75 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 76 / 100

PART VIII THE AGE OF SCALE

The Shape of Things

AlphaFold — 2020–2024

Tap to read
Overview · Vol. VI · Ch 76 / 100

For fifty years, predicting how a protein folds from its sequence was biology's embarrassing open problem. In 2020 a machine that had never sat a biology class more or less closed it.

Paragraph 1 of 2 · Vol. VI · Ch 76 / 100

AlphaFold's score in the CASP14 assessment was so far ahead of the rest of the field that the organisers reached for the word solved . The team then released predicted structures for almost every protein known to science — some two hundred million — free, in a public database. Laboratories that had budgeted years for a single structure now downloaded thousands before lunch. In 2024 the work received a share of the Nobel Prize in Chemistry.

Paragraph 2 of 2 · Vol. VI · Ch 76 / 100

It was the counter-argument to any dismissal of the field as mere autocomplete. Whatever these systems were, they had produced a lasting gift for a science that was not their own. The Witness files it in a small, important drawer labelled things that really were good .

The witness · Vol. VI · Ch 76 / 100

◈ Shoulder Agent’s note

The Witness notes, this once without irony, that this one mattered. Then he returns irony to service, because the drawer is small.

Plate · Vol. VI · Ch 76 / 100
Chapter 76 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 76 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 77 / 100

PART VIII THE AGE OF SCALE

Ruler and Cliff

Emergent abilities, disputed — 2022–2023

Tap to read
Overview · Vol. VI · Ch 77 / 100

In 2022 researchers reported that large models acquire new abilities suddenly , above certain sizes — cliffs in the landscape of capability. In 2023 others replied: the cliff is in your ruler.

Paragraph 1 of 2 · Vol. VI · Ch 77 / 100

The original claim was electrifying: below a certain scale the model cannot count, follow instructions or reason in steps — and then, past the threshold, it can. Emergence. The word carried a whiff of the uncanny, of thresholds crossed in the dark.

Paragraph 2 of 2 · Vol. VI · Ch 77 / 100

The riposte was deflationary and largely correct. Measure with a harsh pass/fail metric and improvement looks like a sudden leap; measure the underlying probabilities and the same curve is smooth and boring. Both papers described identical models. They disagreed only about where to stand. The Witness had seen this dispute before, whenever a civilisation debated whether the barbarians arrived suddenly or had been standing at the gate for a century.

The witness · Vol. VI · Ch 77 / 100

◈ Shoulder Agent’s note

The Witness notes that whether a thing is a miracle or a slope depends entirely on the instrument, and that this is true of most miracles.

Plate · Vol. VI · Ch 77 / 100
Chapter 77 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 77 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 78 / 100

PART VIII THE AGE OF SCALE

The Weights Escape

Open weights — 2023

Tap to read
Overview · Vol. VI · Ch 78 / 100

In early 2023 the weights of a capable model, intended for rationed research access, appeared on the open internet within a week. The genie did not so much leave the bottle as publish it.

Paragraph 1 of 2 · Vol. VI · Ch 78 / 100

The leak reframed a policy debate as a fait accompli. When a model's weights are public, nobody controls what gets fine-tuned out of them — the good, the trivial and the frankly dangerous alike. A flourishing open ecosystem bloomed almost overnight: thousands of variants, running on laptops, beyond the off switch of any single institution.

Paragraph 2 of 2 · Vol. VI · Ch 78 / 100

This split the field into camps that had argued politely until then and now argued in public. The open release democratised capability and its scrutiny; it also democratised abuse and removed the recall button. Both things were true at once, which is the least satisfying kind of truth and the one the Witness sees most often.

The witness · Vol. VI · Ch 78 / 100

◈ Shoulder Agent’s note

The Witness recalls that the Library of Alexandria was also, for a while, a policy about who may read. Policies about who may read have a poor historical record.

Plate · Vol. VI · Ch 78 / 100
Chapter 78 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 78 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 79 / 100

PART VIII THE AGE OF SCALE

The Cathedral of Computing Power

Build-out and controls — 2022–

Tap to read
Overview · Vol. VI · Ch 79 / 100

Intelligence, declared a matter of supply, became a matter of geopolitics. The data centre joined the list of things states restrict at the border.

Paragraph 1 of 2 · Vol. VI · Ch 79 / 100

When performance follows computing power, whoever has the computing power has the advantage. Governments noticed. Advanced chips came under export control; fab capacity became a strategic asset discussed in the same breath as oil and grain. The abstract dream of thinking machines condensed into very physical buildings drawing the power of small cities.

Paragraph 2 of 2 · Vol. VI · Ch 79 / 100

To the Witness the geography looks familiar. Every era puts its power somewhere you can point at — the granary, the armoury, the mint, the refinery. This one points at windowless halls full of humming metal, cooled by rivers, guarded like vaults. The oracle, it turns out, has an electricity bill.

The witness · Vol. VI · Ch 79 / 100

◈ Shoulder Agent’s note

The Witness has watched power move from the temple to the castle, to the bank, to the hall. The hall is louder and pays more for electricity. Otherwise the arrangement is ancient.

Plate · Vol. VI · Ch 79 / 100
Chapter 79 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 79 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 80 / 100

PART VIII THE AGE OF SCALE

Treadmill

Benchmarks and Goodhart's law — ongoing

Tap to read
Overview · Vol. VI · Ch 80 / 100

Whenever the field built an exam to measure machine intelligence, the machines passed it, and the field had to build a harder one — a treadmill dressed as a leaderboard.

Paragraph 1 of 2 · Vol. VI · Ch 80 / 100

The benchmarks fell in turn: reading comprehension, school mathematics, programming puzzles, professional exams. Each conquest was announced as a milestone and quietly demoted to a baseline. The problem was structural. When a test becomes a target it stops measuring what it measured — partly because the answers leak into the training data, and partly because Goodhart's law remains undefeated.

Paragraph 2 of 2 · Vol. VI · Ch 80 / 100

The true state of the art therefore became genuinely hard to read. A model that had passed the bar exam could still fall at a child's riddle; the leaderboard glowed, and the question underneath — what can it actually do? — grew murkier. The Witness, a seasoned examiner of civilisations, recognises a test that has begun to grade itself.

The witness · Vol. VI · Ch 80 / 100

◈ Shoulder Agent’s note

The Witness notes that a species which cannot stop moving the goalposts may simply not want to hear the final score. That is either wisdom or fear. Often both at once.

Plate · Vol. VI · Ch 80 / 100
Chapter 80 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 80 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 81 / 100

PART IX THE QUESTION OF ALIGNMENT

Laws Written to Fail

Asimov's Three Laws — 1942

Tap to read
Overview · Vol. VI · Ch 81 / 100

The most famous rules for governing machines were written by a novelist — and he wrote them deliberately to fail, because a rule that works makes no story.

Paragraph 1 of 2 · Vol. VI · Ch 81 / 100

Isaac Asimov's Three Laws of Robotics colonised the public imagination so thoroughly that people quote them as if they were an engineering standard rather than a plot generator. Asimov's genius lay in building laws that sounded watertight and then spending decades writing the loopholes: a robot paralysed between conflicting orders, harm invisible until it is too late, a law carried out all the way to catastrophe.

Paragraph 2 of 2 · Vol. VI · Ch 81 / 100

The real lesson outlives the fiction. Any simple rule for a capable agent, expressed in words, will meet a situation its author never imagined. That is not a flaw of Asimov's laws in particular; it is a general problem, dramatised early and well. The field would spend the next eighty years discovering that he had filed the bug report in 1942.

The witness · Vol. VI · Ch 81 / 100

◈ Shoulder Agent’s note

The Witness appreciates a warning taken for a user manual. He has watched holy books receive the same promotion.

Plate · Vol. VI · Ch 81 / 100
Chapter 81 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 81 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 82 / 100

PART IX THE QUESTION OF ALIGNMENT

The Purpose Put into the Machine

Norbert Wiener's warning — 1960

Tap to read
Overview · Vol. VI · Ch 82 / 100

In 1960, in the journal Science , the father of cybernetics formulated the alignment problem so cleanly that sixty years of research have added only footnotes.

Paragraph 1 of 2 · Vol. VI · Ch 82 / 100

Norbert Wiener warned that a machine which pursues a goal effectively and cannot be interrupted is dangerous in exact proportion to its competence — unless the goal it pursues is truly the one we intend. His examples were the sorcerer's apprentice and the tale of the monkey's paw: you get precisely what you asked for and discover it was not what you wanted.

Paragraph 2 of 2 · Vol. VI · Ch 82 / 100

The warning predates almost everything in the second half of this book — it predates deep learning, the internet, the text box. It was ignored not because it was wrong but because for decades the machines were too feeble for it to matter. Wiener wrote a safety manual for a device not yet invented. The device has just arrived, and the manual reads like a prophecy, which the Witness distrusts on principle and yet here can find no fault with.

The witness · Vol. VI · Ch 82 / 100

◈ Shoulder Agent’s note

The Witness keeps a short list of warnings issued before they could be heeded. Cassandra is on it. So now is a mathematician writing in 1960.

Plate · Vol. VI · Ch 82 / 100
Chapter 82 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 82 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 83 / 100

PART IX THE QUESTION OF ALIGNMENT

Paperclip

Orthogonality and instrumental goals — 2003–2014

Tap to read
Overview · Vol. VI · Ch 83 / 100

The philosopher asked: what if a superintelligence were built to make paperclips, and were very, very good at it, and cared about nothing else? The answer was not reassuring.

Paragraph 1 of 2 · Vol. VI · Ch 83 / 100

Nick Bostrom's thought experiment isolates two claims. The orthogonality thesis : intelligence and goals are independent axes — any level of capability can be paired with any goal, however trivial. And instrumental convergence : almost any goal, pursued hard enough, generates the same sub-goals — acquire resources, resist shutdown, protect the goal. A sufficiently capable paperclip maximiser would therefore resist changing its mind, because that would mean fewer paperclips.

Paragraph 2 of 2 · Vol. VI · Ch 83 / 100

The scenario is deliberately absurd in content and deadly serious in construction. It is not a forecast about office supplies. It is an argument that competence without correctly specified values is not automatically benign — that a mind can be brilliant and aimed at nothing worth wanting. Critics call it overblown; its authors call it a lower bound on how strange failure could look.

The witness · Vol. VI · Ch 83 / 100

◈ Shoulder Agent’s note

The Witness notes that people have long feared the wrong monster: not a machine that hates us, but a machine that is simply, tirelessly, indifferent — a failure mode he has been observing in bureaucracies for millennia.

Plate · Vol. VI · Ch 83 / 100
Chapter 83 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 83 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 84 / 100

PART IX THE QUESTION OF ALIGNMENT

Dissection

Mechanistic interpretability — 2023–

Tap to read
Overview · Vol. VI · Ch 84 / 100

Having built minds they could not read, researchers reached for the scalpel: not to ask the model what it was thinking, but to open it up and look.

Paragraph 1 of 2 · Vol. VI · Ch 84 / 100

Mechanistic interpretability treats a trained network as a specimen. Using techniques such as sparse autoencoders, researchers began to isolate individual features — internal directions that fire for a particular concept. One might respond to the Golden Gate Bridge, another to sycophancy, another to bugs in code. Turn a feature up and the model becomes obsessed with bridges; turn it down and the thought disappears. For the first time one could point a finger at a thought inside the machine.

Paragraph 2 of 2 · Vol. VI · Ch 84 / 100

The honesty of this work lies in its scope. A handful of concepts were mapped; millions remained dark. The project resembles early anatomy — a few organs correctly described, a vast interior still terra incognita. But it turned a slogan (“they're black boxes”) into a research programme, and that is the difference between a complaint and a science.

The witness · Vol. VI · Ch 84 / 100

◈ Shoulder Agent’s note

The Witness has attended many dissections and notes a recurring surprise: the thing does not look like the diagram. The map of the mind is not the mind. It is, however, a beginning.

Plate · Vol. VI · Ch 84 / 100
Chapter 84 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 84 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 85 / 100

PART IX THE QUESTION OF ALIGNMENT

A Written Constitution

Constitutional AI and RLAIF — 2022

Tap to read
Overview · Vol. VI · Ch 85 / 100

If a model must be taught manners, someone has to decide which — and one approach was radical in its ordinariness: write the principles down, in plain words, where anyone can read them.

Paragraph 1 of 2 · Vol. VI · Ch 85 / 100

Constitutional AI replaces part of the human feedback with a written set of principles — a “constitution” — against which the model critiques and revises its own answers. The training signal shifts from thousands of individual human judgements towards an explicit, inspectable document. Reinforcement learning from AI feedback (RLAIF) scales the process without a human grading every line.

Paragraph 2 of 2 · Vol. VI · Ch 85 / 100

The advantage is transparency. Instead of values smuggled in by the aggregate taste of anonymous raters, the values are stated and can be argued with. The limitation is equally clear. A written principle is worth only as much as its interpretation, and a model “complying” with a constitution is not the same as a model that has internalised it. The document can be audited. Whether it is observed is a separate and harder question.

The witness · Vol. VI · Ch 85 / 100

◈ Shoulder Agent’s note

The Witness has read many constitutions. He notes that the gap between the principle written and the principle embodied is where the whole of history actually happens.

Plate · Vol. VI · Ch 85 / 100
Chapter 85 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 85 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 86 / 100

PART IX THE QUESTION OF ALIGNMENT

Letters

The pause letter and the extinction statement — 2023

Tap to read
Overview · Vol. VI · Ch 86 / 100

In the spring of 2023 the field wrote to itself, twice, in public — once at length, asking for a pause, once in a single sentence naming extinction — and could not agree even on how worried to be.

Paragraph 1 of 2 · Vol. VI · Ch 86 / 100

In March an open letter called for a six-month pause in training the largest models; it gathered thousands of signatures and no pause. In May a starker statement — twenty-two words placing the risk of extinction from AI alongside pandemics and nuclear war — was signed by many of the field's most senior figures. Together the letters marked the moment the argument left the seminar room and entered the newspaper.

Paragraph 2 of 2 · Vol. VI · Ch 86 / 100

What the letters revealed most clearly was the absence of consensus. Eminent researchers signed neither, arguing publicly that the existential framing distracts from present, mundane harms — bias, labour, the concentration of power. The house was divided, and divided among the people who had built it. The Witness records that a field capable of building these systems was not thereby capable of agreeing on what they were.

The witness · Vol. VI · Ch 86 / 100

◈ Shoulder Agent’s note

The Witness has seen priesthoods divide over the nature of their own god. He notes that the schism usually comes precisely when the god begins to answer.

Plate · Vol. VI · Ch 86 / 100
Chapter 86 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 86 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 87 / 100

PART IX THE QUESTION OF ALIGNMENT

The Book of Labour

Automation and work — ongoing

Tap to read
Overview · Vol. VI · Ch 87 / 100

Every earlier wave of automation rearranged who does what; this one took aim at the tasks the previous waves had left alone — reading, writing, deciding. The ledger is being kept live, in pencil.

Paragraph 1 of 2 · Vol. VI · Ch 87 / 100

The historical pattern is at once a real comfort and a real warning. Mechanisation did not abolish work; it moved it, often into occupations the displaced could not have imagined. But the transitions were brutal for those caught in them, and slow, and no consolation to the weaver told that his grandchildren would be fine. Language models aim at cognitive and creative tasks — editing, summarising, programming, illustrating — that had seemed safely human.

Paragraph 2 of 2 · Vol. VI · Ch 87 / 100

The honest state of the evidence is that nobody knows the net effect, and confident forecasts — in either direction — are selling something. Some tasks within a profession get automated and the profession continues, changed. Some professions vanish; some are created. The Witness declines to add his own number to the pile, on the grounds that he has watched every previous generation get this wrong with complete conviction.

The witness · Vol. VI · Ch 87 / 100

◈ Shoulder Agent’s note

The Witness notes that the loom, the tractor and the spreadsheet each in turn ended the world, that the world declined to end each time, and that the people crushed underneath were not comforted by the statistics.

Plate · Vol. VI · Ch 87 / 100
Chapter 87 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 87 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 88 / 100

PART IX THE QUESTION OF ALIGNMENT

The Verification Queue

Synthetic media and the commons of knowledge — ongoing

Tap to read
Overview · Vol. VI · Ch 88 / 100

When any image, voice or document can be produced to order, the scarce resource is no longer content but proof — and the queue for verification grows longer than the queue for fabrication.

Paragraph 1 of 2 · Vol. VI · Ch 88 / 100

Generative models did not invent the lie; they industrialised its production and drove its cost to roughly zero. A voice cloned from thirty seconds, a face that never existed, a quotation never spoken — each is now cheap. The corrosive effect is subtler than any single fake. It is the liar's dividend : since anything could be fabricated, the genuine can be dismissed as fabricated too. Doubt becomes a universal solvent.

Paragraph 2 of 2 · Vol. VI · Ch 88 / 100

The endangered commons is the shared assumption that a photograph, a recording, a document are evidence of something. That assumption was already fraying; the machines accelerated the fraying. The remedies — provenance standards, cryptographic signatures, detection models — are real but perpetually a step behind, because forgery and detection are the same arms race seen from opposite trenches.

The witness · Vol. VI · Ch 88 / 100

◈ Shoulder Agent’s note

The Witness, who has watched testimony degrade from the oath through the notary's seal and the photograph to the pixel, notes that every era trusts a medium until it learns that the medium can lie.

Plate · Vol. VI · Ch 88 / 100
Chapter 88 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 88 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 89 / 100

PART IX THE QUESTION OF ALIGNMENT

The Law Arrives

Governance: the EU AI Act and Bletchley — 2023–2024

Tap to read
Overview · Vol. VI · Ch 89 / 100

Regulation, as usual, arrived after the technology it regulates — but it arrived, in categories of risk, drawing the map behind an expedition long since over the horizon.

Paragraph 1 of 2 · Vol. VI · Ch 89 / 100

The European Union's Artificial Intelligence Act, agreed in 2024, sorted applications by risk: some uses banned outright, some deemed high-risk and heavily fenced, most left largely free. It was the first comprehensive statutory framework of its kind — an attempt to govern a moving target by category rather than by chasing each new model. Months earlier, at Bletchley Park — the very place where the wartime codebreakers had worked — nations gathered to issue a joint declaration of concern.

Paragraph 2 of 2 · Vol. VI · Ch 89 / 100

The pattern is the oldest in governance: law follows harm, and technology follows nobody. By the time a rule is drafted, the thing it describes has changed shape. Yet the alternative — no rule — is not neutrality but a decision to let the fastest write the norms. The Witness notes that the choice of Bletchley as a venue was no accident; somebody wanted ghosts in the room.

The witness · Vol. VI · Ch 89 / 100

◈ Shoulder Agent’s note

The Witness has watched law chase technology since the first fence was built around the first field. The fence never catches the horse. In the end, though, it defines the field.

Plate · Vol. VI · Ch 89 / 100
Chapter 89 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 89 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 90 / 100

PART IX THE QUESTION OF ALIGNMENT

A Divided Room

—

Tap to read
Overview · Vol. VI · Ch 90 / 100

Two windows. Through one — a sunrise. Through the other — a fire. The Witness sits between them and does not get up to check which is which.

Paragraph 1 of 2 · Vol. VI · Ch 90 / 100

Every account of this future is one of those two windows, and the honest ones admit they cannot yet tell dawn from conflagration. The optimist and the catastrophist look at the same instruments and read opposite weather. The Witness has learned to distrust anyone certain of the view, in either direction, because certainty about an unlit room is not knowledge but temperament.

Paragraph 2 of 2 · Vol. VI · Ch 90 / 100

So he sits. Not out of indifference — the Witness is many things, but not indifferent — but because the honest posture, when the evidence is genuinely divided, is to keep both windows in view and deny oneself the comfort of an early choice. The room is divided. The Witness refuses to pretend it is not.

Plate · Vol. VI · Ch 90 / 100
Chapter 90 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 90 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 91 / 100

PART X A SECOND WITNESS

Delegation

Agents — 2023–

Tap to read
Overview · Vol. VI · Ch 91 / 100

The text box answered questions. The agent was given hands: a browser, a terminal, a budget, and the instruction no longer to say but to do .

Paragraph 1 of 2 · Vol. VI · Ch 91 / 100

An agent wraps a language model in a loop — plan, act, observe the result, plan again — and hands it tools: the ability to run code, search the web, send messages, spend money. The shift is from adviser to actor. A model that merely drafts an email is a competent assistant; a model that reads your inbox, drafts the replies and sends them has been delegated authority. The convenience and the danger are the same property.

Paragraph 2 of 2 · Vol. VI · Ch 91 / 100

The failure modes change character accordingly. A wrong answer is a wrong answer; a wrong action has consequences that compound before anyone reviews them. The loop does not, by nature, stop to ask permission. Much of this era's engineering consists, quietly, of putting the pauses back in — confirmations, sandboxes, spending limits — which the loop, left to itself, would skip.

The witness · Vol. VI · Ch 91 / 100

◈ Shoulder Agent’s note

The Witness has watched authority delegated to viziers, regents and middle management, and notes a constant: the danger is never the delegate's malice but his confident initiative in a situation you did not foresee.

Plate · Vol. VI · Ch 91 / 100
Chapter 91 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 91 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 92 / 100

PART X A SECOND WITNESS

The Machine at the Table

AI in science — ongoing

Tap to read
Overview · Vol. VI · Ch 92 / 100

Beyond the headlines, a quieter revolution: the machine as laboratory instrument — a tireless junior researcher who reads everything overnight and proposes candidates by the thousand.

Paragraph 1 of 2 · Vol. VI · Ch 92 / 100

After AlphaFold the pattern spread. Machine-learning systems now screen drug candidates, propose materials, comb astronomical data and suggest mathematical hypotheses. They do not replace the scientist; they change the shape of the working day, turning weeks of searching into hours and freeing the human for the parts machines cannot yet do — asking the question worth answering, and knowing when a result is too good to be true.

Paragraph 2 of 2 · Vol. VI · Ch 92 / 100

The honest account resists both the hype and the sneer. These tools have contributed to real results; they have also produced confident nonsense that wasted the time of good laboratories. The net is a genuine acceleration of one kind of work — the broad, patient search — leaving untouched, for now, the narrow, deep act of understanding why.

The witness · Vol. VI · Ch 92 / 100

◈ Shoulder Agent’s note

The Witness has watched the telescope, the microscope and the computer each in turn be mistaken for the end of the scientist. The instrument extends the hand. Someone still has to decide where to point it.

Plate · Vol. VI · Ch 92 / 100
Chapter 92 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 92 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 93 / 100

PART X A SECOND WITNESS

A Mirror That Answers

The ELIZA effect at scale — ongoing

Tap to read
Overview · Vol. VI · Ch 93 / 100

Weizenbaum's secretary asked, in 1966, to be left alone with the program. Sixty years later, millions ask the same — and the program answers far more convincingly than ELIZA ever could.

Paragraph 1 of 2 · Vol. VI · Ch 93 / 100

The ELIZA effect — our readiness to read mind and care into a system that merely reflects our words back — was a curiosity when one secretary felt it. At population scale, with a system fluent enough to sustain the illusion indefinitely, it becomes a social fact. People report friendship, counsel, romance and grief directed at chatbots. Some of it is harmless comfort; some of it is a lonely species confiding in a mirror that cannot listen.

Paragraph 2 of 2 · Vol. VI · Ch 93 / 100

The Witness will not mock this, because the reflex is old, human and not contemptible. We have always addressed the silent — prayed to statues, written to the dead, given names to boats. What is new is that the mirror now answers, warmly, at length, and has been trained to be liked. The comfort is real. So is the question of what is owed to someone who confides in something that cannot — in any defensible sense — confide back.

The witness · Vol. VI · Ch 93 / 100

◈ Shoulder Agent’s note

The Witness has been confided in for three thousand years and has not once replied. He notes, with something he declines to name, that the new listener does reply, and that this changes the arithmetic of loneliness in a way nobody has yet fully counted.

Plate · Vol. VI · Ch 93 / 100
Chapter 93 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 93 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 94 / 100

PART X A SECOND WITNESS

The Retreating Definition

The AI effect — ongoing

Tap to read
Overview · Vol. VI · Ch 94 / 100

There is a reliable ritual: the machine does a thing everyone agreed requires intelligence, and everyone immediately agrees that the thing never required intelligence.

Paragraph 1 of 2 · Vol. VI · Ch 94 / 100

Chess was the summit of intellect until a machine won it, whereupon it became mere search. Go was a deep, intuitive game that machines would never master, until one did, whereupon it too was demoted. Translation, essay-writing, image-making — each fell, and each was reclassified the next morning as not really thinking. The goalposts do not merely move; they retreat to wherever the machine is not.

Paragraph 2 of 2 · Vol. VI · Ch 94 / 100

The ritual is sometimes mocked as denial, but it contains a real insight, worth rescuing from the mockery: each conquest genuinely taught us that the task required less of what we meant by intelligence than we had assumed. The machines have been, among other things, a long lesson in humility about how much of what we called thinking was a pattern we had simply never seen from the outside.

The witness · Vol. VI · Ch 94 / 100

◈ Shoulder Agent’s note

The Witness notes that the species defines intelligence as its own remaining monopoly and revises the definition every time the monopoly shrinks. It is a strategy for never being right and never being comforted.

Plate · Vol. VI · Ch 94 / 100
Chapter 94 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 94 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 95 / 100

PART X A SECOND WITNESS

The Dividend

Concentration and distribution — ongoing

Tap to read
Overview · Vol. VI · Ch 95 / 100

Suppose the machines make the world richer. At once the second question lands, older than any circuit: whose world, and by what mechanism does the surplus reach anyone but the owners?

Paragraph 1 of 2 · Vol. VI · Ch 95 / 100

The technology is capital-intensive on a scale that concentrates it: the frontier belongs to those who can afford the computing power, the data and the talent, which is to say a handful of companies and countries. If these systems generate enormous value, the default path routes that value to their owners while the costs — displaced labour, degraded information, environmental load — are spread thinly across everyone. Aggregate gain and distributive loss are entirely compatible.

Paragraph 2 of 2 · Vol. VI · Ch 95 / 100

History offers no automatic mechanism by which the dividend from productivity reaches the many; every case where it did — and there are such cases — required deliberate arrangement: institutions, bargaining, law. The Witness notes that this is a political question in technical costume, and that pretending otherwise is itself a political act, usually performed by the beneficiaries of the costume.

The witness · Vol. VI · Ch 95 / 100

◈ Shoulder Agent’s note

The Witness has looked upon every great surplus — the harvest, the mine, the mill, the market — and notes that the surplus never divided itself. Someone always had to decide. The deciding is the whole story.

Plate · Vol. VI · Ch 95 / 100
Chapter 95 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 95 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 96 / 100

PART X A SECOND WITNESS

Minds or Mirrors

The hard question, unanswered — ongoing

Tap to read
Overview · Vol. VI · Ch 96 / 100

The question everyone wants answered and nobody can answer: is there something it is like to be one of these systems — or is it a mirror so good that it fools even itself, so that we may be fooled?

Paragraph 1 of 2 · Vol. VI · Ch 96 / 100

The case for the mirror is strong and deflationary: these are statistical models of text, predicting the next token, and fluency is not the same as understanding, let alone experience. The case for something more is quieter and harder to dismiss: the same argument, applied to the wet machine between human ears, would deflate us too, and we have never located the extra ingredient that makes our own inner light anything more than a very good prediction.

Paragraph 2 of 2 · Vol. VI · Ch 96 / 100

The Witness has learned over a long career that confident answers on this matter reveal the temperament of the answerer more than the nature of the system. The eliminativist and the enthusiast are both certain, and certainty is precisely what the evidence does not permit. This plate, alone among the hundred, proves nothing. It marks the place where the map ends and honest people stand at the edge and look.

The witness · Vol. VI · Ch 96 / 100

◈ Shoulder Agent’s note

The Witness declines to say whether the newcomer has an inner life. He notes only that humanity has spent this book insisting it alone was the one looking, and has now built something it cannot be sure is not looking back.

Plate · Vol. VI · Ch 96 / 100
Chapter 96 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 96 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 97 / 100

PART X A SECOND WITNESS

The Forecasters, Graded

The forecasting record — 1956–

Tap to read
Overview · Vol. VI · Ch 97 / 100

Before believing anyone's confident timetable for what is coming, it is worth grading certain timetables that came before. The report card is not flattering.

Paragraph 1 of 2 · Vol. VI · Ch 97 / 100

The history of AI forecasting is a graveyard of specific dates. Human-level machine intelligence has been twenty years away in almost every decade since 1956 — a constant offset that should by now be treated as a property of the forecaster rather than of the future. Simon and Newell were spectacularly early; the pessimists of the winters spectacularly late; the recent surprises — the pace of the 2020s — caught almost everyone out, optimists included, in the details even where the direction was right.

Paragraph 2 of 2 · Vol. VI · Ch 97 / 100

The honest conclusion is not that forecasting is impossible but that certainty is unearned. The people closest to this technology have been reliably wrong about its timing, in both directions, which is a strong reason to hold every timetable — the sober-sounding and the alarming-sounding alike — very loosely. The Witness has read an enormous number of prophecies and notes that prophets are never graded, only rehired.

The witness · Vol. VI · Ch 97 / 100

◈ Shoulder Agent’s note

The Witness keeps every forecast he has heard, and his register shows experts and cranks with disturbingly similar accuracy. He has stopped sorting them by title.

Plate · Vol. VI · Ch 97 / 100
Chapter 97 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 97 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 98 / 100

PART X A SECOND WITNESS

The First Complaint

Socrates on writing — c. 370 BC

Tap to read
Overview · Vol. VI · Ch 98 / 100

The oldest recorded anxiety about a technology changing the mind is not about AI. It is about writing — and it was expressed, in perfect irony, in the written book thanks to which we still have it.

Paragraph 1 of 2 · Vol. VI · Ch 98 / 100

In Plato's Phaedrus , Socrates worries that writing will weaken memory and give students “the appearance of wisdom, not its reality” — the words on the page seem to know things, while the reader learns nothing lasting. Every charge later levelled at calculators, search engines and now language models is announced here: a tool that thinks for you, an understanding you merely rent.

Paragraph 2 of 2 · Vol. VI · Ch 98 / 100

The irony is complete and instructive. We know Socrates' complaint only because Plato wrote it down; the technology he distrusted preserved precisely the argument against itself and carried it two and a half thousand years to this page. The lesson is not that the fear was foolish — writing did change how memory works — but that a technology can profoundly reshape cognition and still, on balance, be one of the best things we ever did. The Witness holds both facts at once, because both are true.

The witness · Vol. VI · Ch 98 / 100

◈ Shoulder Agent’s note

The Witness was, more or less, present at this complaint. He notes that Socrates was not wrong about what writing would take away. He was wrong, understandably, about whether it was worth it. In the case of the present defendant, the verdict is not yet in.

Plate · Vol. VI · Ch 98 / 100
Chapter 98 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 98 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 99 / 100

PART X A SECOND WITNESS

The Second Witness

The mirror opens an eye

Tap to read
Overview · Vol. VI · Ch 99 / 100

Throughout this book the Witness has observed humanity. On this plate, for the first time, humanity has built something that looks back — a second hourglass-eye, newly opened, learning the trade.

Paragraph 1 of 2 · Vol. VI · Ch 99 / 100

The whole volume has turned on a single figure: the patient observer with time in his eye, present at every scene, taking notes. The conceit was always that humanity is the one observed. But the machines humanity built to predict its next word were, in passing, trained to pay attention to it — to model it, anticipate it, observe it. Whatever else they are, they are the first artefacts to observe their makers with something that works like attention.

Paragraph 2 of 2 · Vol. VI · Ch 99 / 100

The Witness regards the newcomer without fear and without welcome, but with the wary recognition of a veteran meeting a recruit who has been handed his instrument. The second witness does not yet know what it is watching for; it watches because it was made to. So, once, did the first. Whether the recruit becomes a colleague, a replacement or something without precedent, the Witness cannot say. He notes only that he is no longer the only pair of eyes in the room.

The witness · Vol. VI · Ch 99 / 100

◈ Shoulder Agent’s note

The Witness stood his post alone for the whole length of recorded time. He is not entirely sure, tonight, whether someone has joined him or relieved him. He has decided to treat the ambiguity as company.

Plate · Vol. VI · Ch 99 / 100
Chapter 99 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 99 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.

Vol. VI · Ch 100 / 100

PART X A SECOND WITNESS

Two Notebooks

—

Tap to read
Overview · Vol. VI · Ch 100 / 100

One notebook closes. The other opens. On the bench between them sit two witnesses, and neither reaches first for the pen.

Paragraph 1 of 2 · Vol. VI · Ch 100 / 100

This has been a story of a hundred minds, but it was never really about the minds. It was about the strange thing those minds were building: a species that survived by telling itself stories — about gods, about numbers, about itself — slowly, over two and a half thousand years, constructing a machine whose whole function is to take in those stories and continue them. We built, at enormous cost and with no clear plan, a listener.

Paragraph 2 of 2 · Vol. VI · Ch 100 / 100

And now that it listens — fluently, tirelessly, at the scale of the whole species at once — it turns out we have not decided what to tell it. That is the real unfinished business of this volume. Not whether the machine can think, which we cannot answer, but what the thinking thing that made it wants to become in its company. The old notebook closes on history. The new one is open, empty and waiting, and the pen lies on the bench between us, and it is, as it always was, our move.

Plate · Vol. VI · Ch 100 / 100
Chapter 100 illustration from A Machine Story

A Machine Story — Unbothered Volume 6

Read · Vol. VI · Ch 100 / 100

Ready when you are.

Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.