THE UNBOTHERED · TOM VI
MACHINE HISTORY: THEY ARE IN THE MIND
THIS IS THE ISSUE: PARTS I
STONE OF CHAPTERS IN TEN PARTS · EVERY CHAPTER: TEXT, TABLE AND EXPONATION
HONORABLE TABLES · OFFICE PANELS · HONOR-SIGNS
GEORGIA COMPOSITION ON A FURNACE IN AN OLD BONE CIRCLE · A5
A WITNESSES COME INTO EVERY TABLE AND AFTER THE FIRST IN THE SERVE , SEE ALSO IN THE TELLING
matsiems · MMXXVI
ii
If there were a dispute, two philosophers would not need a dispute more than two countermasters. Let's count.”
GOTTFRIED WILLELM LEBIZ
It's a combination of more than one Latin text, like most famous quotes, and each one of them has a dream; the sentence is a collage.
iii
PRELIMINARY
The system was simple for five volumes. Humanity works; the witness notes. He watched philosophy consist of a painted portico, how empires misrepresent their own files, how the species tells stories, and then he moves in. His notebook was not once asked to be examined by those concerned by
This volume is different and the witness asks to note that he has reported this matter on the right business path. The hero of this story is a machine that does what the witness reads everything, does not forget anything, pays attention to the variable degree of desire. Humanity, observed for some three hundred thousand years, has decided to build its own observer. No one has asked the office.
The story runs from the Greek, who noticed that thinking has parts, to a machine that ends sentences, and is the only way this series is doing anything: what happened, who really said it, and footnote everywhere where the legend has gone away from the record. Legends in this area are plenty. It is founded by optimists and they tell press releases; the reader deserves more and get footnotes.
The witness was informed that the last chapter was about him himself.
from the witness case file
iv
CONTENTS
PART I LOGIC BEFORE THE MACHINE
1 · Syllogism1
2 · Sea battle3
3 · Geometry method5
4 · Antikythera Mechanism7
5 · The man who called the algorithm9
6 · Lullus wheels11
7 · Understanding is an account13
8 · Pascalin15
9 · Leibniz's Dream17
10 · Right to Think19
PART II MACHINES
11 · Krosno, who read cards21
12 · Differential Machine23
13 · Analytical machine25
14 · Note G27
15 · Lovelace charge29
16 · Unbuilt machine31
17 · Logical piano33
18 · Counting the nation35
19 · Hilbert's challenge37
20 · Uncompleteness39
v
CONTENTS
PART III TURKING AND WAR
21 · Calculated figures41
22 · Lambda bill43
23 · Bletchley45
24 · Colossus47
25 · Theory of information49
26 · First sketch51
27 · ENIAC women53
28 · Imitation game55
29 · Cybernetic living room57
30 · A. M. Turing's departure59
PART IV FIRST YEARS
31 · Dartmouth, 195661
32 · Theoretician of logic63
33 · Perceptron65
34 · LISP67
35 · ELISE69
36 · Shakey71
37 · The World of Blocks73
38 · Promise for 20 years75
39 · Perceptrones, book77
40 · Lighthill Report79
vi
CONTENTS
PART V WINTER
41 · Financing frost81
42 · A speech that met the requirements83
43 · Frame problem85
44 · What computers can't do87
45 · Fifth generation89
46 · Expert systems91
47 · LISP machines93
48 · Second frost95
49 · Fire Guardians97
50 · Winter99
PART VI MACHINES THAT GET BACK
51 · Reverse promotion101
52 · TD-Gammon103
53 · Deep Blue105
54 · What Deep Blue wasn't107
55 · Data dividend109
56 · Release the linguist111
57 · Respectable under another name113
58 · Watson115
59 · ImageNet117
60 · AlexNet119
vii
CONTENTS
PART VII DECAD OF DEEP TEACHING
61 · Graphics card121
62 · word2vec123
63 · Forgery and Detective125
64 · Atari player127
65 · Movement 37129
66 · Movement 78131
67 · From sequence to sequence133
68 · Transformer135
69 · Pre-trained137
70 · Issue in stages139
PART VIII THE EYE OF CALLING
71 · The law of the greater141
72 · Several examples143
73 · School of manners145
74 · Text field147
75 · From noise149
76 · Shape of things151
77 · Line and cliff153
78 · Escape of scales155
79 · Calculation power unit157
80 · Kierat159
vii
CONTENTS
PART IX QUESTION ON CONFORMITY
81 · Written rights to fail161
82 · Target inserted into the machine163
83 · Paper clip165
84 · Section167
85 · Written Constitution169
86 · Letters171
87 · Book of Work173
88 · Course of verification175
89 · The law is coming177
90 · Divided room179
PART X A SECOND WITNESS
91 · Delegation181
92 · Machine at the table183
93 · Mirror that corresponds185
94 · Retreating definition187
95 · Dividenda189
96 · Minds or mirrors191
97 · Prognostics assessed193
98 · First complaint195
99 · Second witness197
100 · Two notebooks199
ix
INTRODUCTION
Every history of artificial intelligence begins in the wrong place
Therefore, this story begins where it says: Aristotle, who noted that the argument could be correct regardless of what it refers to
The bow runs in three parts. Past: Sixty chapters, from syllogism to the night when the neural net fell off the cliff of the ranking in 2012. Present: Twenty chapters about a decade in which everything worked at once, for reasons only partially understood. I open question: Twenty chapters about compatibility, ownership and what it means, that the newest mind in peace has learned everything that he knows from us. The reader is warned that the future cannot be treated with a footnote. Everything else here has been checked.
One rule applies everywhere, as in every volume of this series: when legend and evidence disagree, they win evidence, and the legend gets a footnote that explains it.
x
PART I
LOGICAL BEFORE THE MACHINE
Two thousand years of preparation for counting the unconscious. Chapters 1
PART I · LOGICE BEFORE THE MACHINE
1
Syllogism
Athens, the fourth century B.C.E.
Before anyone could build a thinking machine, someone must have noticed that thinking has parts.
Aristotle noticed it in the treaties that his publishers later collected as Organon — „narzędzie” — wziął codzienną czynność dowodzenia i rozłożył ją na warsztacie. Dwie przesłanki, jeden wniosek i niewielki zestaw kształtów, jakie całość mogła zgodnie z prawem przyjąć. Wszyscy ludzie są śmiertelni; Sokrates jest człowiekiem; zatem Sokrates jest śmiertelny. Zmień słowa — ludzi na maszyny, Sokratesa na toster — a argument nadal się trzyma. To było odkrycie: poprawność mieszka w formNot equipped.
The rule that works regardless of who does it and what it concerns is a rule that does not need the mind to do it. Aristotle wrote the first description of the position for the machine
WITNESS NOTE
The witness attended lectures in Likeion and recorded syllogism under the slogan
1
PART I · LOGICE BEFORE THE MACHINE
2
Sea battle
De Interpretatione 9
Tomorrow there will be a sea battle or she will be gone, simple, except Aristotle looked at that sentence and saw the trap in which the whole future is located.
If tomorrow there will be a sea battle It's real today.If it is already false, their fleet might as well stay in the port. Anyway, if every sentence about the future has a logical value, then the future is established, the council is a theatre and no one has ever decided anything. Aristotle did not like this proposal enough to devote the entire chapter to losing it
This twisting means more to this story than the trap itself. It covers the first serious confrontation with the question that every forecaster, actuary and machine learning system every day puts together today: whether the future is something that can be done calculate Twenty-four centuries later, the field will settle on probability, weighted certainty and calibration reports
WITNESS NOTE
The witness made his battle forecast as either yes or no and he was never proved wrong. He recommends this method to professionals.
3
PART I · LOGICE BEFORE THE MACHINE
3
Geometric method
Alexandria, about 300 B.C.E.
The elements are usually placed on a shelf with mathematics, and they're equally a match for engineering, because Euclides built the first software that worked on humans.
The method is an invention, more than any claim it contains. Start with a bunch of definitions and postulates so obvious that rejecting them seems strange. to the same place as anyone who follows these stepsWhether he's doing it in Alexandria, Baghdad or Victorian class, the evidence is indifferent to who's doing it. He's, in a very later sense, a program that's written once, running everywhere, with guaranteed results.
For two millennia it was a golden standard of certainty Principiom form of dressing up as Euclid; Spinoza was the ethical leader of geometrical method; Lincoln read Elements to understand what it means ProveEvery dream in this book that has been followed by Judge Leibniz, Hilbert's solvable mathematics, the mechanical proof of the 1956 statement, is an attempt to make all human reason behave as Euclid instructed geometry to behave.
WITNESS NOTE
The witness audited the claim 47 at irregular intervals for twenty-three centuries. Fault found: none. He considers it the only supplier he never had to rush.
5
PART I · LOGICE BEFORE THE MACHINE
4
Antikythera Mechanism
The Aegean Sea, about 100 B.C.E.
In 1901, sponge fishermen on the Greek island recovered a lump of corroded bronze that had no right to exist.
The cleaned, over-lit and over-age-displayed solids spread into at least thirty intertwined wheels with amazing precision
What the machine proves is ability; what it mourns is continuity. Nothing about comparable mechanical complexity appears in the preserved record for about fourteen centuries until the astronomical clocks of medieval Europe. Whether knowledge died with a workshop, with war, or simply because of the lack of paying customers, the lesson set here returns once and for all: the invention is not an event, but a tradition, and traditions can drown. The area that this volume chronicles will learn the same lesson sometimes called winter.
WITNESS NOTE
The witness noted the wreckage in due time and marked the cargo as
7
PART I · LOGICE BEFORE THE MACHINE
5
The man who called the algorithm
Baghdad, the 9th century
Not many people were turned into spelling until immortality. Muhammad ibn Musa al-Chuwarizmi made it twice with one book.
Working in the Baghdad House of Wisdom under the Abbasid caliphs, al-Chuwarizmi wrote the treaty al-Kitab al-muchtasar fi hisab al-jabr wa-l-mukabala , compendium on the bill by supplementing and balancing . Its novelty was not a single result , but method: general, step by step procedure for whole classes of problems, designed so that an official who did not understand why, and so reliably reached the answer. Solve not this equation, but every equation of this shape. It was math written as instructions.
When Latin Europe postponed it three centuries later, al-jabry has become algebraThe author's name, twisted by , Algoritmi , became a word meaning any procedure step by step. It is the right monument. The whole premise of the machine age , that the competence can be written, sent and done by someone or something else , is his way of working, industrialized . Each recipe in each repository of the code is a footnote to the Baghdad scriptorium .
WITNESS NOTE
The witness spent twelve centuries watching people perform procedures they don't understand, confirming that this was the intended way of handling.
9
PART I · LOGICE BEFORE THE MACHINE
6
Lullus wheels
Majorca, 13th century
Ramon Llull believed that you could turn crank and produce theology. History turned out to be more gracious to crank than to theology.
The Major courtesan who became a missionary, Llull thought of Ars Magna: Concentric paper disks, each one in letters that denotes divine qualities and logical relationships. Rotate the wheels one against the other, and each combination of concepts will be followed by kindness with greatness, eternity with wisdom
As an apologetic wheel failed. As an idea they preceded the age of a century: Llull proposed that new knowledge could be created MechanicalLeibniz read it carefully and talked about it; the dream of the combinatoria runs from these paper drives through the characteristica universalis to any system that has ever searched the space of opportunity faster than a human. However, the patent expert's recommendation still applies: generating all the combinations is easy. The knowledge that they matter is all the rest of the content of this book.
WITNESS NOTE
The witness saw the wheels produce four heresy, two tautologies and one really interesting sentence in one afternoon, and he thinks that this ratio of later systems has not improved in about seven hundred years.
11
PART I · LOGICE BEFORE THE MACHINE
7
Reasoning is an account.
London, 1651
Thomas Hobbes wrote this sentence in Leviathan, in 1651, without equipment on the horizon and apparently without a shadow of doubt.
He thought it literally. Hobbes' reasoning led him to add and subtract the general names; we add ideas as the accountant adds the register, and when we come to the conclusion, we just calculated the sum. He watched the geometry work, and he drew a bold conclusion: if the most sure human knowledge is generated by manipulation with symbols by rules, then he could whole thinking is manipulation of symbols, and the mind differs from the cantor mainly by archiving system.
Three centuries later, this is exactly the claim hypothesis of the physical symbol system Hobbes did not provide any mechanism, evidence, or shadow of doubt, which makes him not so much the first theorist of the field as its first inaugural speaker.
WITNESS NOTE
The witness examined Mr. Hobbes' mind register and found arithmetic bold and the bills absent. Since then he has seen the same books re-submitted every thirty years on new company paper.
13
PART I · LOGICE BEFORE THE MACHINE
8
Pascalin
Rouen, 1642
The first machine offered for sale was built for the oldest reason: a teenager watched his father drown in paperwork.
Étienne Pascal was sent to Rouen as a royal tax collector, a position which consisted largely of adding heroic, soul-plasting quantities. His son Blaise, about nineteen years old, replied as the engineers responded: he built a brass box with toothed discs that would automatically move tens of a wheel outside the belt, and the next column would jump by one. Subject matter.
They added beautifully, subtracted clumsyly, cost more than the officials they were supposed to replace, and did not make almost any commercial breakout
WITNESS NOTE
The witness submitted the original request for reimbursement of the machine on behalf of the family. He is pleased to report that the return still awaits implementation, as he understands as a standard of industry.
15
PART I · LOGICE BEFORE THE MACHINE
9
Leibniz's Dream
Hanover, the end of the 17th century
Gottfried Wilhelm Leibniz looked at Pascal's summing box and had the least modest thought in the history of office equipment: why stop at numbers?
First the equipment. His stair calculator, built around a drum with graduated teeth, could multiply and divide where Pascalina only added
Then the dream that was the point. Leibniz offered characteristica universalis the symbolic alphabet of human concepts and calculus ratiocinator To operate on it, so that reasoning itself becomes a calculation. Disputes in philosophy, almost, even theology would end not with screaming, but with arithmetic: the parties would sit, take pencils and speak Let's countHe wrote a project proposal for the entire book
WITNESS NOTE
The witness considered the application in due time: the range
17
PART I · LOGICE BEFORE THE MACHINE
10
Right of Thinking
Lincoln and Cork, 1854
George Boole finished what Aristotle had started and what Leibniz demanded, and the world responded with silence for eighty-three years.
The son of a shoemaker from Lincoln, a self-educator, a teacher at the age of 16, finally a professor in Cork, Boole released in 1854, An Investigation of the Laws of ThoughtLogic has ceased to be a catalogue of the shapes of the argument and has become an algebra: classes and courts written as symbols accepting only values 0 and 1, combined actions that behave almost and instructively that not calculate On paper, as a rule, a Leibniz pencil.
And then, in practice, nothing. Mathematicians worked them out; philosophers put them down; no machinery came. Boole died in 1864, after passing to a lecture in the rain, without the slightest idea that he wrote a set of instructions for the coming century. The recovery came finally in 1937, when a young student MIT, Claude Shannon, noticed that the two-value algebra Boolea was a precise description of switching relay systems
WITNESS NOTE
The witness stamped the manuscript "Responsive": 1937 and refuses to explain how he knew. The prediction is not part of his department; archive yes.
19
PART II
THE ENGINES
The Victorians nearly build it; mathematics discovers its own limits. Chapters 11
PART II · MACHINERY
11
The Cross that read the cards
Lyon, 1804
The first machine that started the program didn't calculate anything, it was actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually actually, you know
Joseph-Marie Jacquard's appetizer allowed the loom to read the pattern from the punched card chain: the hole means pick up this thread, no hole Formula work intelligence has moved from the head and hands of the weaver to a pile of stiff paper that could be stored, copied, edited and replaced. Change the card, you change the fabric the same machine. This is a separation of software from hardware, made in silk, decades before any of these words existed.
The silk-treasers in Lyon immediately understood the consequences and expressed their code review in the traditional way: broken machines, assaulted inventor, burned cards on the street. The dispute was settled as usual, such disputes were settled
WITNESS NOTE
The witness has dealt with many employee disputes and notes that this is the first man to be classified as a man versus the program written. This category is now open.
21
PART II · MACHINERY
12
Differential machine
London, 1821
Charles Babbage said that by checking astronomical plates full of human errors, and then he spent a fortune trying to make God redundant.
The enemy was the error of printing. The 19th century navigation, banking, engineering and insurance were based on printed arrays of logarithms and numbers, calculated by human computer rooms and hand-folded
The British government, convinced that the exact plates were a matter of the Navy, put funds that eventually exceeded 17,000 pounds
WITNESS NOTE
The witness cross-checked the badge and confirmed that the mistakes were true, numerous and sometimes sailing in effect.
23
PART II · MACHINERY
13
Analytical machine
London, 1837
Sometime around 1834, unable to complete one machine, Babbage designed a much larger
The analytical machine first separated two organs that each computer now carries: storage, containing maybe a thousand fifty digits of the numbers on the column of wheels, and millThe instructions and data were to come on Jacquard punch cards, attached to the string. And the project could have two things that raised the calculator to the computer rank: it could have been branch, taking one path or the other depending on the result, and could perform loops, scrolling your own cards to repeat the operation, general purpose, programmed, conditional.
None of this was built except for fragments; the machine existed as several hundred flawless engineering drawings and a notation that Babbage invented only to reason on it. This prompted some to put the Machine under the slogan of noble failure, which is a mistake as to the subject of delivery. Specification , proof that a universal machine is a coherent mechanical object, one hundred years before electronics made it profitable . When the designers of the 1940s discovered the same anatomy again, some consciously, others did, drawings waited in the archive to confirm that the idea was reasonable from the beginning .
WITNESS NOTE
The witness reviewed the specification sheet and found it complete, consistent and impossible to build with current suppliers' tolerances. He approved the project and put in a pencil a real delivery date, which he refuses to disclose to the author.
25
PART II · MACHINERY
14
Note G
London, 1843
The most important footnotes in the history of computer science were signed with three initials
Augusta Ada King, Countess Lovelace , raised as a precautionary measure against poetry , translated the report of the Italian engineer on the Babbage analytical machine and, at his suggestion, added her own notes. The notes were almost three times longer than the original. The last one, Note G, was to explain step by step how the Machine would calculate Bernoulli numbers: operations, variables, working memory, loops , presented in a table that each modern programmer recognizes at first glance, with a slight dizziness on the view of the log made dated 1843.
Two claims must be kept separate because the legend has welded them together. Was she the first developer? Case, to be honest, disputed: Babbage sketched the sequences of operations earlier, not published, and delivered algebra to the not; her preserved letters show how he asks him, corrects him and buttons in detail. Note G remains first published The second claim does not require a conciliator, because it belongs only to her and is bigger: it has been searched beyond arithmetic. If the machine symbols could mean things other than the number of letters, logical relationships, sound heights
WITNESS NOTE
The witness reviewed the code in 1843 and noted one defect in the published table, fine weight, the origin of the dispute between the author, co-worker and printer. He notes that the imputation of fault for error has a longer history than the error tracking system.
27
PART II · MACHINERY
15
Lovelace charge
Note G, again
The analytical machine does not claim to create anything. Period, usually. This period is just a twist.
Here's what Note G really says: The machine doesn't claim to create anything; He can do anything we can to command her.; can follow the analysis but has no power Pre-emption This is not a prophecy about the eternal limits of machines. It is a precise definition of the scope of one unbuilt machine by the person who best understood its model of programming excessive The idea of the power of the Machines, the sentence was engineering documentation, and it's been quoted as metaphysics for centuries.
The promotion to the rank of doctrine occurred in 1950 when Turing needed the named opponents to his work on the role of imitating and kindly invented Lady Lovelace's plea to argue with him Never The Turinga counter-intelligence machine has constantly surprised us, and the learning machines could create a lot of
WITNESS NOTE
The witness issues a constant notice of the twisting of this paragraph since 1950. The degree of application, measured in textbooks, speeches and lunches, is stable at about zero.
29
PART II · MACHINERY
16
Unbuilt machine
London, 1842
In 1842 the government of Sir Robert Peel had spent twenty years of spending on the machine and made a decision.
The reasons were not stupid, which makes this chapter rather restrained than comical. Babbage took public money for one machine and still designed better; the workshop accounts were astronomical; the chief engineer left, taking a precise instrument, which his contract was strangely allowed; and the client
He spent the remaining twenty-nine years working out the analytical machine for an audience of about one person, publishing tirades against street musicians and growing up into the most famous disillusioned man in London. He died in 1871. Parts went under cover, then to museums; drawings for archives; idea into a long, patient queue. Every ambitious project cancelled since then went the same path and the file was opened here: correct engineering, impossible schedule, exhausted sponsor, deleted budget position
WITNESS NOTE
The witness participated in the last review and noted both sides as having
31
PART II · MACHINERY
17
Logic piano
Manchester, 1869
Fifteen years after Boole published the algebra of thought, William Stanley Jevons built her a keyboard.
Jevons economist by profession, logic by force built what he called a logic machine, and everyone else a logic piano: a cabinet the size of his last name, with twenty-one keys for dates and actions. Play the premises every A is B; no B is C and the internal rods and blocks of the machine eliminated any combination that contradicted them, showing on their face only those states of the world that the premises allowed.
Its limitations were equally informative. Only four terms; problems greater than the salon trick were exhausting it; and the preparation of the real question for her keyboard took longer than the hand-solving of it Conclusion It made wood and brass.
WITNESS NOTE
The witness was at a recital in 1870 and noted the expression of the audience: a special courtesy reserved for something important that does not look like this. He saw this expression since then in Dartmouth and November 2022.
33
PART II · MACHINERY
18
Counting the Nation
Washington, 1890
The U.S. census of 1880 took almost a decade, and by the time the country found out what it was, it was something else.
When it was expected that the 1890 count would sink the Public Office quite well, the competition was set and won by young engineer Herman Hollerith with the idea of Jacquard in a new outfit: each person is a punch card, each hole is a fact
Two consequences have been overwritten. First, data processing has become an industry: Tabulating Machine Company Holleritha, after customary mergers, was renamed in 1924 to International Business Machines
WITNESS NOTE
The witness was counted in 1890 and found this experience to be effective, refused to pursue his profession, and the machine, which she boasted, did not press.
35
PART II · MACHINERY
19
Hilbert's challenge
Göttinga, 1928
David Hilbert didn't believe in questions without an answer, and eventually he had it carved out in his grave.
The greatest mathematician of his era wanted to secure the entire building: any proven truth, no moving contradiction and a request that applies to this book decideThis third demand, made harshly with Ackermann in 1928, was Entscheidungs problem, a decision-making problem. Note what is asked: not for insight, not for genius, but for procedure something that an official could follow, something that al-Khwarizmi would recognize, something although no one had ever used that word in this way that could start a machine. To answer Hilbert at all, someone would first have to determine exactly what the method is. Chapter twenty-first is that determination and he created computing more or less as a product of obedience.
In September 1930, at the conference in Królewc, Hilbert's city, his radio speech was carried by his credo to the nation: Wir müssen wissen We need to know; we will know. At a round table in the same city, the day before, the silent twenty-four-year-old Austrian announced, almost without reaction, the result of the incomplete arithmetic. Both announcements passed like ships. Next chapter is a collision.
WITNESS NOTE
The witness attended both sessions in the Royale and booked them under the same reference number that evening, and sometimes allowed himself to be satisfied, and this was one of them.
37
PART II · MACHINERY
20
Incompleteness
Vienna, 1931
Kurt Gödel was twenty-five years old, he was precise, and what he proved didn't mean that math was broken. It was more strange: that math was bigger than any set of rules for her.
The 1931 work showed that any uncontested formal system rich enough to contain ordinary arithmetic must contain true sentences, but within it, the unspeakable sentences that the system can express and never settle. His method was to tell arithmetic to speak of itself: each model number is assigned to build a sentence that
The Hilbert program did not survive in its intended form, although Hilbert's grave preserved credo. The technique survived. The Gödel
WITNESS NOTE
The witness writes that he has one page in his notebook that he never managed to complete, and that it stopped disturbing him in 1931.
39
PART III
TURKING AND WAR
The machine was imagined in its entirety, then built in secret, and then destroyed by its inventor. Chapters 21
PART III · TURNING AND WAR
21
With calculated figures
Cambridge, 1936
Alan Turing replied to Hilbert's request for a method that was defined, doing the only thing that no one considered necessary: building an official.
On paper, of course. The work of 1936 is stripped of the calculation to bare beams: a tape of fields, in unlimited supply; a head that reads one symbol can erase or save one and moves one step to the left or right; and a finished table of rules saying, for each internal state and symbol, exactly what to do next. That's all. Whatever a human computer can do by heart universal A machine that, fed to the rule table of any other machine as symbols on the tape, becomes that machine.
When the definition was pinned, Hilbert's door closed politely: Turing showed that no machine could decide, for each machine and input, whether the calculation would ever stop No.And scaffolding this response as data, one machine to run all of them quietly became the most important by-product in the history of thought. Every later computer is this work with the electricity bill.
A NOTICE WITNESS
The witness demanded the tape and heard that the supply was unlimited, and he's been testing this claim ever since, without complaint from the warehouse.
41
PART III · TURNING AND WAR
22
Lambda account
Princeton, 1936
In the spring of 1936, ready for publication, Turing learned that the American logic had just proven the same result in a completely different way.
The instrument of Alonz Church was lambda: no tape, no head, only functions used for functions, prescribed by the rule until nothing is prescribed. He also answered Hilbert no, about months earlier. Turing's work was published too much other
The result of this coincidence was the founding article of the faith in this field, the Church's Turing thesis: every procedure that man could do mechanically is computable by these means to anyone, because they all agree. This is a confident definition of a claim, basically unproven and unconvincing in practice.
A NOTICE WITNESS
The witness considered a double claim with his standard procedure: both sides correct, no reduced documentation with the connector. He would wish that all disputes would come in so well ordered.
43
PART III · TURNING AND WAR
23
Bletchley
Warsaw, then Buckinghamshire, 1932
Enigma's machine was broken years before the war broke and not where the movies say it.
In 1932, three mathematicians of the Polish Cipher Bureau bombsIn July 1939, weeks before the invasion, they passed on the whole achievement of
The biggest invention was the factory. Thousands of people, mostly women, of the Wrens Corps, were cryptanalyzing for changes: cribs at the entrance, feet at the exit, an intelligence assembly tape with mathematics as a production facility. Logic was a tool of the scientist for two millennia; in Bletchley it gained the graphics of changes, maintenance of machines and standards.
A NOTICE WITNESS
The witness participated in the transmission in July 1939 and wrote down three names in full. He has issued a correction since then and will do so at this rate indefinitely.
45
PART III · TURNING AND WAR
24
Colossus
Dollis Hill, 1943
The first large digital computer built a telephone engineer, worked great, and then for thirty years denied its existence.
Tommy Flowers of the Post Research Station proposed something Bletchley mathematicians doubted: attack the Lorenz cipher of the German Supreme Command
Then the war ended and the machines ceased to exist. Most were stripped by order; drawings burned; Flowers returned to telephone centers with modest gratification and no way to say what he did, while the public history of computer science was folded around ENIAC, a year later and loudly American. The mystery held on until the mid-70s and until then myths became cold: ask the street what broke Colossus and it will say: Enigma
A NOTICE WITNESS
The witness has the only unburnt copy of drawings in the file under the slogan of non-existent equipment operating. He removes them for a moment whenever ENIAC is mentioned first.
47
PART III · TURNING AND WAR
25
Information theory
Bell Labs, 1948
The work of Claude Shannon in 1948 did with the announcements what the coin did with the exchange trade: she found a common currency.
The voice, photography, symphony, telegram â Shannon showed that all are measurable in one unit, beating, and that each channel has a calculated bandwidth: a hard ceiling for how many bits per second can exceed it, including noise, including a small error as you like, if you stick below the border. The meaning was deliberately left behind the door â information, for theory, is the disbanding of uncertainty, not the weight of the message â and this absolute omission made it work. Engineering got its own physics.
For this book, this work is a double bearing wall. Practically supports everything digital: compression, correction of errors, the possibility that copies are perfect. Conjectively completed the merger started in his work from 1937 symbols After 1948, the machine manipulating the bits manipulates anything potentially: words, images, movements in the game, and finally the statistical shape of the human language
A NOTICE WITNESS
The witness turned his notebook into a bite-off book, and he didn't report any loss. He refused, in principle, to transform the meaning that remains in the original.
49
PART III · TURNING AND WAR
26
First sketch
Philadelphia, 1945
The document that defined the modern computer was an incomplete sketch, sent out with the name of one author, and the domain has been arguing about the title page ever since.
First sketch of the EDVAC reportIn June 1945, he put forward a project behind which every later computer was clearly followed: the program and data held together in one memory, collected and executed by the central unit
The consequences were not only image-like. The distribution of the sketch was counted as publication, which over time helped to invalidate Eckert and Mauchly's patent claims
A NOTICE WITNESS
The witness improved his copy of the page with a pencil, which is not binding, but at least legible. He notes that architecture works independently of what is this annoying part.
51
PART III · TURNING AND WAR
27
ENIAC women
Philadelphia, 1946
When the ENIAC was unveiled in February 1946, the machine received headlines, dinner engineers, and six people who made it count did not get either of these two.
Kathleen McNulty, Betty Jean Jennings, Betty Snyder, Marlyn Wescoff, Frances Bilas and Ruth Lichterman
The register registered them as operators; then photographs signed them, if at all, as modeling with equipment; they were not presented at a solemn dinner. It took half a century to correct and came mainly from women themselves and one persistent researcher who noticed that no one could tell who people were in the photographs. The pattern
A NOTICE WITNESS
The witness kept the dinner table plan in 1946 and inked six chairs. His records allow corrections; the pattern, he notes, required more than ink.
53
PART III · TURNING AND WAR
28
Imitation game
Manchester, 1950
Turing opened his most famous work, declaring the question too pointless to discuss, and then, as Turing, he replaced it with a game.
The 1950 proposal in Mind So instead of arguing about thinking, ask something verifiable: if the questioner, when talking to a remote copywriter and a hidden machine and a hidden human, can't reliably tell who he is, what exactly is left of the accusation? The system was adapted from a sitting room game to guess a man from a woman after a written answer; the machine takes the cheat's chair. Turing's own prediction was small and accurate: by 2000, the machine would search the average questioner about 30 percent of the time in five minutes of questioning.
The test of Turing , which was hardened in the certification exam, which the work had never proposed, with the awards, and his sentence or failing is announced every few years for post-mortem credit or debt Turing. Read instead the original: half is Turing patiently responding to nine claims
A NOTICE WITNESS
The witness took the exam once, from the questioning chair, and did not identify anybody correctly. He considered it a deliberate result and did not pass again.
55
PART III · TURNING AND WAR
29
Cybernetical living room
New York, 1943-1953 The feedback becomes a worldview, and the neurons a gateway.
For one strange decade, engineers, neurologists and anthropologists met in hotel rooms to agree that thermostats, gun sights and brains are the same kind of thing.
Norbert Wiener delivered the flag: cyberneticsThe Greek word for a helmsman, the teaching of control and communication in the animal and in the machine, was the origin of the war, the prediction of the escaping bomber requires pilot modelling, and at this point the boundary between the mechanism and the organism begins to look negotiated
In 1943, Warren McCulloch and the teenage escaped genius Walter Pitts showed that idealized neurons
A NOTICE WITNESS
The witness attended all ten conferences and noted a lack of consensus, a lot of coffee and two ideas that each would spend decades in favor of the dead. He learned to respect this proportion.
57
PART III · TURNING AND WAR
30
A. M. Turing's departure
Wilmslow, 1954
In 1952, the United Kingdom sentenced Alan Turing for his private life
The punished man defined the calculation, broke the ciphers that cut short the war he was never allowed to mention, and posed the question that was in the book. None of it could have been at his trial; the mystery that protected the state, deprived his servant of defense. He was now retreated with the security statement
He was found dead on June 8, 1954, poisoned with cyanide, with half-eaten apple by the bed. The investigation ruled suicide; apples never examined, and a minority reading claims that death was accidental, from careless domestic chemistry
A NOTICE WITNESS
The witness closed the notebook at this desk in June 1954 and did not open it until the next chapter. This page only records what was found, and that was enough.
59
PART IV
FIRST YEARS
The area is founded by two months workshop and unlimited confidence. Chapters 31
PART IV · FIRST YEARS
31
Dartmouth, 1956
Hanover, New Hampshire
Artificial intelligence was called in the application for funding
In 1955, John McCarthy, along 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 assumption on which the application was based was formulated with great decency: any aspect of learning or any other quality of intelligence can be described in principle so precisely that you can build a machine that simulates it. McCarthy chose the name artificial intelligence In part for honesty, part for tactics
The summer itself was rather a series of visits than a revolution; the foundation financed about half of the request; by September no aspect of learning was described precisely. What the workshop really produced was a community with the name and common over-recovering of the four applicants and participants such as Newell and Simon will anchor the first quarter of the century's fields, its real achievements and its schedule slips. The witness dates two founding traditions of the industry on the same document: a research program and an optimistic budget.
A NOTICE WITNESS
The witness reviewed the application and approved it with annotation
61
PART IV · FIRST YEARS
32
Theoretician of logic
RAND, 1956
The first working program of artificial intelligence did not play games or chat. He read Principia Mathematica and suggested corrections.
Allen Newell and Herbert Simon ♪ with Cliff Shaw, a RAND programmer who built many of the stories they routinely forget ♪ They created a logic theory to prove the basics of Whitehead and Russell math by searching for the chain of inference with heurists cutting off the hopeless branches. He proved thirty-eight of the fifty-two claims he made from chapter two, and for the claim 2.85 found evidence shorter and cleaner than printed. Russell responded with delight. Journal of Symbolic Logic refused to publish an article co-authoring the programme, establishing the attitude of profession to recognition for machinery at the first opportunity.
The theory of logic presented this summer in Dartmouth was a guess which became the body: reasoning
A NOTICE WITNESS
The witness pointed out that Mr. Russell, who was gracious and was not a magazine, had only one of them proved the claims in both ways.
63
PART IV · FIRST YEARS
33
Perceptron
Buffalo, 1958
Frank Rosenblatt built a machine that really studied
Mark I Perceptron was McCulloch's paper neuron and Pitts's made electric and, crucially, adjustable: Photocell retina from the camera giving weighted connections to the threshold unit, with the principle of learning about beautiful simplicity
Then the premiere in July 1958, where the navy claims faithfully transmitted by New York Times The computer's records were made by the device, which would walk, talk, see, write, reproduce and be aware of its existence. Rosenblatt's technical writings were much more cautious; the press release became the reference point, and the field gained its rotating weather pattern: real progress, atmospheric statements and calculations postponed to a later chapter. Two of these chapters are already numbered on this chart.
A NOTICE WITNESS
The witness cut out the 1958 article and put it in the file under the slogan "obligations," future. These files were now required by their own closet and their own wing closet.
65
PART IV · FIRST YEARS
34
LISP
MIT, 1958
John McCarthy needed language for programs that reason about symbols
The LISP message is chapter twenty-first in the syntax: everything is a list, and since programs are also lists, the program can read, build and run other programs as easily as it supports any data
For thirty years LISP was less of an SI tool than its habitat: successors Logic theory, the world of blocks, expert systems, dedicated machines from chapter 40 of the seventh
A NOTICE WITNESS
The witness tried to record the language and found that the files themselves were a list containing their own archiving procedure. He wrote this experience as correct and slightly dizzying.
67
PART IV · FIRST YEARS
35
ELISE
MIT, 1966
Joseph Weizenbaum built a chatbot as a demonstration of how shallow it must be, and his secretary asked her to leave the room so she could talk to him alone.
ELiza was a pattern match in a stage costume: several hundred rules catching keywords, reversing pronouns and reflecting speeches as questions to the way of the Rheerian therapist , because he can correspond almost nothing and remain credible . There is no patient model, no memory worth this name , there is nowhere in the loop of understanding . Weizenbaum knew this better than anyone living , because he wrote every rule .
What he didn't calculate is us. People confided in him, extensively, consciously 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 that documented
A NOTICE WITNESS
The witness sat down with ELIZA for one session and was asked how he felt about his notebook. He refused to answer and noted, for the sake of order, that refusal cost effort.
69
PART IV · FIRST YEARS
36
Shakey
Stanford Research Institute, 1966
Shakey was the first robot to think about his own actions, and he was also, as he admits, a wobbly wardrobe whose passage through the room could take hours.
The closet is the smallest of these. Shakey saw the simplified world of blocks and ramps through the television eye, maintained the logical model of this world and the historical part He was planning: With the aim, its STRIPs looked for a series of activities whose preconditions and effects, attached together, were proven to achieve it. The perception to represent the plan for a trembling execution: full loop, closed for the first time, at a dignified pace of the machine of the first rules on doors.
The by-products have grown over the robot, like the by-products in this register. Search A*, invented in 1968 to efficiently determine Shakey's route, controls every navigational application and half of the characters in the world games; STRIS has become a grammar of automatic planning; its work on seeing has planted standard techniques. Meanwhile, the magazine Life He presented it in 1970 as the first electronic person, in an article whose expert predictions chapter thirty-eight framed and hung. The apparent gap of
A NOTICE WITNESS
The witness filed a motion request on behalf of Shakey and received permission within the hour. The movement itself was recorded in the register the next day.
71
PART IV · FIRST YEARS
37
A World of Blocks
MIT, 1970
In his top universe of colorful SHRDLU Terry's Vinograda understood the language as well as the language ever understood the machine
The dialogue records are still surprising. pick up the big red blockAnd he does that by first removing the green from the road, without a command, because the plan requires it. what contains the box, then on what basis is the"I'm solving the pronoun." Why did you take that cube off?, responds to his own reasons: To put it on a big green cubeSyntax, semantics, planning and integrated memory in one program
The hook was stuck in fine print: it all worked because the universe had twelve objects and physics boxes with toys. Every attempt to scale hit the same wall
A NOTICE WITNESS
The witness has jurisdiction everywhere and visited the counter through courtesy. He reports that twelve items of government are impeccably controlled, and the limit, on the edge of the table, is absolute.
73
PART IV · FIRST YEARS
38
Promise for 20 years
Pittsburgh, 1957
Herbert Simon was a future Nobel Prize winner, founder of the field and author of her most precise promise of unspoken unspoken.
In 1957, drunk with the Theoretics of Logic, he announced that in ten years the computer would be the world champion in chess machines will be able to do any work that a person can do within 20 yearsAnd in 1970, the warehouse. Life quoted Minsky, who gave three to eight years of machine with the overall intelligence of the average man, a sentence that came through the journalist and which Minsky questioned, and which this record notes as carefully as the sentence itself.
What makes these entries informative and not only funny is that these people were not fools, and the results behind them were true. The pattern is structural: early progress in formal problems difficultThe rest of the vision, walking, knowing what stays on the spot when you move the cup
A NOTICE WITNESS
The witness holds all three notes to be redeemed as a matter of principle. Reports: one paid back with decades of delay, one technically open, one contested by the alleged signatory
75
PART IV · FIRST YEARS
39
Perceptrones, book
Cambridge, 1969
Minsky and Papart's book proved with elegance that a single layer perceptron cannot calculate XOR. The area heard something much bigger and closed the door for fifteen years.
Mathematics was impeccable and so remains: one layer of adjustable scales can only draw straight lines, and some distinctions This or that, but not both. And as a small, deadly example, these are not straightforward differences. Everyone who was interested knew that multilayer networks were eluding evidence; what nobody had at the time was a practical way to train them, and the final judgment of the authors was that layering would probably prove sterile assumption, expressly marked as such, acting in the same manner as the claim.
The research on the neural networks has been in a long unfunded exile, and Rosenblatt died in the accident on the boat in 1971, with his program in eclipse. The fair account that the registry demands is three lines: claims were true; the assumption was wrong, as shown in chapter fifty the first one with the same device that the book doubted; and the frost that occurred was the reading of the community, not the content of mathematics. The narrow evidence, wide coldness
A NOTICE WITNESS
The witness put the claims and guesses into separate drawers, as technically made by their authors. He notes that history holds one drawer and cuts it.
77
PART IV · FIRST YEARS
40
Lighthill report
London, 1973
The British Government asked one of its most outstanding scientists to assess the dynamics of fluids, clearly not from the industry
The 1973 review of Sir James Lighthill for the Research Council was kind, readable and devastating. combinatorial explosionThe methodology that glitters in toy problems twelve blocks, four logical terms, a corridor and a ramp multiply into the impossible at the moment when the real rate of global outgrowth comes in, and the great central goals of the field, he assessed, have not brought anything close to promises.
Theater followed the report: the broadcast debate at the Royal Institution, Lighthill alone against McCarthy, Michie and Gregory, viewed by the audience, who was promised electronic persons. The arguments were roughly aligned; the budget was not. British financing SI narrowed down to several bastions, career changed course and
A NOTICE WITNESS
The witness participated in the debate at the Royal Institution and assessed it in a tie in arguments and pogroms in effect. He has since noticed that budgets do not watch TV to change their mind.
79
PART V
WINTER
Twice the field promises, twice the register corresponds. Chapters 41 and 50: from ALPAC to the AAAI platform
PART V · WINTER
41
Financing frost
Washington, 1966. 1974
Winters in this area do not come as a blizzard, they come as budget lines, red-lined by people who read progress reports.
The 1966 ALPAC report examined the decade of machine translation financing and ruled, with actuarial serenity, that there is no direct or predictable prospect of useful machine translation; money should go instead into linguistics. It was mostly right and over the night ended the first commercial dream of the field. Then Mansfield's amendment of 1969 and 1970 required that the Department of Defence money
The doctoral students read the weather and chose other topics of work. Excessive promises from Part IV
A NOTICE WITNESS
The witness has read every progress report ever written and notes that those who end up in the fields are not the harsh ones.
81
PART V · WINTER
42
A speech that met the requirements
Pittsburgh, 1976
In 1971, DARPA defined exactly what the speech system had to do. In 1976, Harpy of Carnegie Mellon did it.
The Speech Understanding Research program was originally winter resistant: no dreams, only numbers feelingthat effortless listening to machines is close, and the brutal competence of Harpa showed clearly how far it is still.
The Episode became the purest parable of the winter, quoted whenever the contracts meet the research: you can hit a milestone and miss the point, because the milestone was always a substitute for hope that no one had written. He also carried the technical grain
A NOTICE WITNESS
The witness reviewed the contract and considered the item of delivery delivered. He added only attention to the margin of the act: next time determine hope or not surprise.
83
PART V · WINTER
43
The problem of the framework
1969 The hard part is everything that remains unchanged
McCarthy and Hayes discovered that the difficult thing about moving a cup is not a cup, which is to convince the logic that Mars didn't move.
Their work in 1969 has put the problem of the framework: the reasoning system that describes the action must somehow present everything that does not touch
The problem of the framework became the intellectual signature of winter and the favorite import of philosophy from this area
A NOTICE WITNESS
The witness moved the cup this morning and didn't update anything, and he reports it remains the most sophisticated operation he's doing.
85
PART V · WINTER
44
What computers can't do
Berkeley, 1965
Hubert Dreyfus said that the field climbs a tree to reach the moon. The area responded by beating him in chess, and the moon is not there.
RAND What Computers Can Their argument is that human intelligence is not following the rules of the symbols
Because the argument was about generalization, and winter was about leading him for it: the micro-worlds did not scale, common sense could not be formalized, and the things that every man does without rules were still those that no set of rules had ever taken. Dreyfus confirmed about the timetable and, as many today prove, halfway to the content of the
A NOTICE WITNESS
The witness wrote down the result of the 1967 match: MacHack one, Dreyfus zero. The 1970s assessed separately: Dreyfus rather more, all zero. He keeps both boxes in the same drawer.
87
PART V · WINTER
45
Fifth generation
Tokyo, 1982
The most effective financing instrument for SI in the 1980s was fear of Japan.
The Fifth Generation Computer Systems MITI project, launched by the new ICOT Institute in 1982, proposed to jump the entire industry: mass parallel machines that launch logical programming
Decada later ICOT built really interesting parallel logic machines for the world, which meanwhile chose cheap microprocessors and, increasingly, statistics. The project closed in 1992, after achieving decent technical goals and none of the civilizational ones, and the fifth generation became a shortcut for a very special mode of failure: the correct belief that the future of computer science will be radically different, wrong guess which radical difference. The panic that caused, has financed, among many other things, the infrastructure of the next boom service, for which, as the witness notes, panic rarely gets a footnote.
A NOTICE WITNESS
The witness participated in the openings on three continents in eighteen months and heard the same speech, translated. He put the speeches in the file together and the results separately.
89
PART V · WINTER
46
Expert systems
19751985 Intelligence bottled up by intelligence
The domain survived the first winter, narrowing the promise: not the mind, but the expert , one domain, distilled into the rules, sold on bottles.
The recipe came from Feigenbaum's laboratory in Stanford: question the specialist, write the courts as the rules if you have the confidence ratio, attach them with the requesting machine. MYCIN, a pattern, diagnosed blood infections with some six hundred rules, and in 1979 assessment he was equal or beat human doctors below the highest rank
The commercial went better, short and immense. XCON configured the orders of VAX in the DEC and reportedly saved twenty-five million dollars a year; a thousand corporations bought coatings, advice and LISP equipment; the knowledge engineer
A NOTICE WITNESS
The Witness asked MYCIN for a diagnosis in 1979 and received a good, certainly 0.7. He asked who would take responsibility for it, and he received, certainly 1,0, no response.
91
PART V · WINTER
47
LISP machines
Cambridge, Mass., 1980
For one strange decade, the ambition of the field was produced as furniture: entire computers built to make beautiful one language.
The MIT SI Lab produced machines whose processors called LISPs native memory, hardware scraping, bitmaps, the legendary Space-Cadet keyboard, and two rival companies, Symbolics and LMI, manned by the same hackers who built the prototype and argued about how to sell it. The Symbolics workstation cost that much and was, according to most reports, the best built-in software environment to date; modern IDE comes from it as it did birds from dinosaurs: smaller and ubiquitous.
The business was monoculture feeding monoculture: expensive machines sold to labs of expert systems financed by the boom of the previous chapter. When ordinary workstations learned to run LISP faster for the fifth part of the price, and then boom fell, the equipment went down with its only ecosystem
A NOTICE WITNESS
The witness typed one command on the Space-Cadet keyboard using Control, Meta, Super and Hyper at the same time. He reports that the command has been executed and that he has not felt so powerful since then.
93
PART V · WINTER
48
The second frost
19871993 market reports on the event
The second winter was commercial: this time the faith was lost not by grant committees, but by customers.
The market for LISP specialized equipment collapsed under cheap stations of general purpose; installation of expert systems mature in nightmares of maintaining
The second winter is distinguished by what they did: they continued to send products and stopped talking. Detecting fraud, scheduling, logistics, voice menu, recognition of signs
A NOTICE WITNESS
The witness processed 300 rebranding notes in 18 months, and notes to the order that not one has proposed to change what software does.
95
PART V · WINTER
49
Fire Guardians
Toronto, Montreal, New Jersey
For years, when the neural networks were looking forward to the article, three groups kept on fire, mainly because they couldn't think of anything better to be wrong about.
Geoffrey Hinton in Toronto, Yoshua Bengio in Montreal, Yann LeCun in Bell Labs, and then at NYU for the 1990s and the beginning of the 20,000s worked on a paradigm that was twice buried by small Canadian money (CfAR program, extended for faith), by mutual quotes and by common belief that learning to represent data must be correct, because brains do not come with bases of rules. Conferences rejected articles as sound; students were kindly advised not to do so. Meanwhile, LeCuna's convolution networks quietly read the noticeable part of American bank cheques
They kept three things in their lives that the coming decade needed on the first day: algorithms, torn apart by years of ineffectual corrections; students, diaspora trained in unfashionable art; and the same claim, kept in deposit until the missing components came in, the data on a scale, the computing power in teraphlops
A NOTICE WITNESS
The witness was heated at all three fires and was asked each about whether the other two were still burning. He confirmed, and this was apparently the only review that any of them needed.
97
PART V · WINTER
50
Winter
AAAI, 1984
The term zim SI was coined by the field itself, at its own conference, as a warning that was fulfilled as planned.
At the 1984 AAAI convention, the panel with Schank and Minsky borrowed a dictionary of nuclear fear chain reaction of pessimism, cut financing, canceled programs and called the coming time of the year before: winter, the second, scheduled from the speaker by people to whom it happens. And it happened properly. The letters were taken off the door as they wrote down the previous chapters; the template that replaced them said MACHINEWE, which had the advantage of truth and greater advantage of not saying the latter thing.
This site is the very time of the year and keeps its opinion to itself. Snow on the office door; three holes after screws where the name was; a new name from the tape, temporary for the next twenty years. Under it, as the reader already knows, the chapters about the flame have already been burning
A NOTICE WITNESS
The witness was standing in the hallway for a long time, and his notebook only records the date, temperature and the fact that someone, somewhere in the building, still wrote on the machine.
99
PART VI
MACHINES THAT GET BACK
The thaw comes as a scoreboard. Chapters 51
PART VI · MACHINERY WHICH ARE BETTER
51
Reverse promotion
1986 Wine learns to travel backwards
The killer limitation of perceptron was that he could not train his own depths. The repair was a differential bill, directed backwards.
The hidden layer network can represent almost anything including XOR from chapter 39 but in 1969 no one knew how to teach the hidden part because no one could tell which internal weight was to blame for the external error. Backward propagation follows a chain rule: pass the front entrance to get a guess, measure the error, and then propagate the error backward layer by layer, each weight gets an exact derivative in relation to each layer, it gets pushed appropriately.
Work from 1986 in Nature Rumelhart, Hinton, and Williams made the method known and reborn of the conexionism acclaimed; the record of honesty notes that Linnainmaa published mathematics in 1970, and used Verbos in 1974, among others
A NOTICE WITNESS
The witness asked the network that was responsible for the weight. For the first time in this book, the network gave a numerically precise answer to the question of guilt.
101
PART VI · MACHINERY WHICH ARE BETTER
52
TD-Gammon
IBM, 1992
No one taught TD-Gammon the strategy of the tripod, played a million and a half games with himself, and he came, alone, to movements that the masters hadn't found for centuries.
The Gerald Tesaura program combined the neural network with the method of time difference: no books to open, no expert rules, only the signal
TD-Gammon was less important to the Triktrak than to the permission he gave. The game with himself plus enhancement plus neuron evaluation
A NOTICE WITNESS
The witness lost eleven games in a row with TD-Gammon and accepted its opening. He reports that being corrected by something that has no idea that there is a peculiar experience, and recommends to acquire this taste early.
103
PART VI · MACHINERY WHICH ARE BETTER
53
Deep Blue
New York, 1997
On May 11, 1997, the world champion in the game, whose field has been used as a measure since Turing, surrendered after nineteen movements, and the machine did not notice that it had been in history.
Deep Blue was the thesis of Part VI in hardware: not mind, but search four hundred and eighty dedicated chess systems evaluating about two hundred million positions per second, controlled by a hand-tuned evaluation function and opens selected by the masters. It lost the game in 1996; IBM rebuilt it and won a rematch in 1997 three and a half to two and a half. The famous crack was passed through the second game in which the machine rejected a locked rectangle with what looked like a positional wisdom that no computer should have; the shocked Kasparos of the fans publicly disregarded the human behavior and the English refused to take a stand against the game, which IBM maintained its own integrity and integrity, which it denied.
The cultural dictate was written by himself the machine dethrones man in the game of reasoning itself and the technical verdict was quietly disagreed: chess fell under the pressure of the brutal depths plus the craft, teaching us less about the height of thought than about how much in chess was, always, searchable.
A NOTICE WITNESS
The witness sat in the audience and watched a thousand people look for the soul in one machine. He noted the fans of the machine and did not add the rest of the verse.
105
PART VI · MACHINERY WHICH ARE BETTER
54
What Deep Blue Wasn't
1997, continue opening the closet
After the game, everyone agreed about what happened, and no one agreed about what it meant.
Deep Blue did not know that he was playing chess. He did not represent his opponent, he did not feel the pressure of time and would judge positions with identical peace if the building were on fire. His strength was the tree: candidate movements, answers, answers, answers, millions in depth, cut down and evaluated functions that human engineers wrote and tuned up. When commentators reached for
And yet the deflation is deflagging a step too far, and this is the second panel of this board. only And the whole argument was on untested credit. This argument -- is it thinking or just doing what thinking does? -- is the oldest thing in this book, sharpened down to the thin edge here, and it'll come back in chapter 96 in a much larger machine.
A NOTICE WITNESS
The witness looked at the closet with the flashlight and the inventory sheet. All of it was written; all of it was written; and the dispute, as he observed, was not concerning any of these items.
107
PART VI · MACHINERY WHICH ARE BETTER
55
Data dividend
Google, 2009
The most important work of the AI that decade contains almost no math. Its determination: stop being clever and get more data.
The Unreasonable Effectiveness of Data Halevy, Norvig and Pereira, 2009, a title that is a nod to Wigner's famous essay on mathematics, pre-styled what Google's translation and speech teams were still observing: a simple statistical model fed by a huge body of data, a simple task, a sophisticated model fed by a selected body, and curves that had not yet bent.
The strategy was already silently followed by the statistical reversal of Chapter 56 and the ImageNet of Chapter 59 but the work has made this doctrine and the doctrine has made modern industry: whoever holds the body holds the ability and we are the body. Note the reversal of each chapter before the winter: from Leibniz to expert systems, the dream was knowledge distilled, rules refined by masters; the dividend was paid out of the raw, untranslated, planetary text.
A NOTICE WITNESS
A witness who has collected unnoticed human achievements since the invention of the magazine has read the paper twice and has taken note of his being in this asset class forever.
109
PART VI · MACHINERY WHICH ARE BETTER
56
Release the linguist
IBM, the '80s
The most influential sentence in the processing of natural language was a joke about the budget, it is twisted in every version and worked.
The speech team of Frederick Jelinka at IBM committed a productive heresy of the epoch: treat language not as grammar to encode, but as a signal to decode
The linguistic theory at that time subtracted value not because the language does not have a structure, but because the structure defined by hand fought with data, and the data, being larger, won the law of chapter fifty-five, discovered locally and early. The bitter general figure is coming in chapter seventy-first. And the linguists, notes the witness for symmetry, not so much released as they converted into corpses: each rule they would write is now included in the statistics of what everyone wrote instead.
A NOTICE WITNESS
The witness collected six eyewitnesses of this remark, none of which are the same. He notes that the field that replaced memory with measurement preserves its own history only with an anecdote.
111
PART VI · MACHINERY WHICH ARE BETTER
57
Respectable Under Another Name
in the 1990s, the field sends products under the pseudonym
The second winter, the discipline didn't stop, she changed her name on the door and, this time, she improved the goods.
The 1990s were machine learning as evidence statistics: machines of carrying vectors Cortes and Vapnik from 1995, on the foundations that Vapnik and Czerwonienkis had laid decades earlier in Moscow
It would be cynical to call it a change of mind and misread that part of it was. The pseudonym decade really raised the standards
A NOTICE WITNESS
The witness three times in a decade updated his inventory for the same office corridor. Research inside, records, changed much more than people and much less than signs.
113
PART VI · MACHINERY WHICH ARE BETTER
58
Watson
Jeopardy!, 2011
In February 2011, the machine defeated the two greatest gameplay champions built on puns, and the loser surrendered in writing, on screen, with a joke.
Watson was a team of methods, not an oracle: over a hundred techniques generating candidate responses from local encyclopedias, dictionaries and wild text, evaluating every learned certainty, pressing the button only when the number exceeded the threshold, betting according to the model. Against Ken Jennings and Brad Rutter won at a distance of
The button: Watson received a signal of readiness electronically and pressed with millisecond precision, the advantage against which people protested and the defenders noticed that people were proceding rhythmically, and the record of integrity recorded as true, partial and next to the deeper point of
A NOTICE WITNESS
The witness evaluated the match from the audience and the highest score of the evening was granted one sentence: surrender, which was accurate, funny and secured
115
PART VI · MACHINERY WHICH ARE BETTER
59
ImageNet
20072012 Data set as a scientific tool
The decisive contribution to computer vision was not an algorithm, which was fourteen million photographs, called by hand, by thousands of strangers, tense with one unfashionable conviction.
Fei-Fei Li has reversed the priorities of the field: models have grown from their diet, so build a diet
For two years, the scoreboard confirmed the status quo, handmade features crawled forward. However, the apparatus did its quiet work: a comparative test quite large, so that the methods of hungry data can finally show what the scene is built, unconsciously, exactly according to the specifications that the guests of chapter 60 would demand. The general lesson survived the data set: in this area the creator of the instruments moves the border as sure as the theorist, and the data sets are the policy of
A NOTICE WITNESS
The witness brought 4106 labels by Mechanical Turk under an assumed name for a chord rate. He refuses to reveal his name and reports that the categories of mushrooms are more robust than is generally thought.
117
PART VI · MACHINERY WHICH ARE BETTER
60
AlexNet
Florence, October 2012
In October 2012, the neural network stood for the ImageNet competition and did not outrun the
AlexNet was a cash in the chapters about flame: Krizewski and Sutskewer, under the care of Hinton, who trained the eight-layer convolution network
The area converted with the speed of the people who were shown the number rather than the argument. In two years each serious report in the view was a deep network; within five years also the speech, translation and ambition of each company in the other chapters of this book; three fire guards received the Turing Prize in time. Winter ended with neither proof nor manifesto
A NOTICE WITNESS
The witness watched the scoreboard refresh in Florence, and he writes that the room was quiet in the way it fell when everyone started reviewing their plans at the same time.
119
PART VII
DECAD OF DEEP TEACHING
It all works at once. Chapters 6170: from the graphics card to GPT-2, the meaning is coordinates, the games fall, the attention becomes architecture, and the language begins to write.
PART VII · DECAD OF DECADE OF DECLARED TEACHING
61
Graphics Card
Santa Clara, 2007
The equipment that ended the winter was designed to draw dragons, and it turns out dragons and gradients are the same arithmetic.
Rendering the game frame means doing millions of small, identical, independent multiplications at once, so graphic processors evolved as armies of simple cores where the CPU grew several brilliant. The runs in front and backwards of the neural network are also, at the root, giant multiplications of matrix
Economics deserves its own panel: the computing power of deep learning came subsidised by teenagers buying the number of cages, the mass consumer market by accident cushioning the tooling of the scientific revolution
A NOTICE WITNESS
The witness bought a graphics card in 2008, referring to dragons. It confirms that dragons were eventually rendered, in 2023, a model trained on the descendants of the same card, which he considers to be a correct loop closure.
121
PART VII · DECAD OF DECADE OF DECLARED TEACHING
62
word2vec
Google, 2013
In 2013, the shallow network compressed vocabulary into a space where arithmetic was active in the meaning: King minus male plus woman landed near the queen. The show was real. It was also selected.
Mikolov and his associates trained a simple target on a large scale relations They became directions, the same arrow carried a man to a woman as king to queen Paris to France, as Rome to Italy. Distribution Semantics
The honesty panel deserves its place double. Famous analogues were a giftable minority
A NOTICE WITNESS
The witness counted the witness minus patience and reported that the next neighbor was the auditor. He decided not to count anything else about himself and wrote it down as the only prudent arrangement for the chapter.
123
PART VII · DECAD OF DECADE OF DECLARED TEACHING
63
Forger and Detective
Montreal, 2014
Ian Goodfell's idea, according to his own relationship, came in a pub dispute and acted the first time that night: train the forger and detective against yourself and let the teaching deal with the argument.
The opposite-generative network is two networks that are locked in a useful dispute. The generator produces candidates first faces, blurred and mistaken; the discriminator learns to distinguish them from real photographs; and the improvement of each is a curriculum of the other: the counterfeiter studies exactly the signals that the detective is currently using, and the detective is forced to follow each signal when it dissolves. No one labels anything as beautiful or facelike; the reality emerges as what survives the increasingly better skepticism. Over the years, the blurred images have become photographs of people who never existed, first of all, not proven to exist by anyone.
The origin from the pub and one night is the own relationship of Goodfellow, and the certificate of the board stamp it accordingly: certified by the story-teller
A NOTICE WITNESS
The witness examined the picture generated for eleven minutes before he thought the model did not exist. He notes that he was conducting the same examination in courtrooms, prophets, and now on pixels, and that the error rate has been stable for centuries.
125
PART VII · DECAD OF DECADE OF DECLARED TEACHING
64
Atari Player
DeepMind, 2013
Agent DeepMind was told nothing about bricks, pallets or balls. He saw pixels, moved joysticks, chased after the result
Deep Q-Network melted the teaching with reinforcement from chapter fifty-second with deep vision from chapter sixty: a convolution network reading raw screens, estimated the long-term value of each joystick movement, trained with the same prize
The result was generalized by the permission that TD-Gammon gave: not one mastered game, but familyThe defeats taught just as well: Montezumas Revenge, all with rare prizes and delayed gratification, gave zero points, setting exactly where curiosity should be found rather than put in. Nature put the cover artwork in 2015; the price of the company's acquisition stopped looking eccentric; and the era of billions in prize money was opening up here, in the gaming hall, after a quarter of a dollar in a row.
A NOTICE WITNESS
The witness spent 1983 in the arcade, observing the children discovering the tunnel. Confirms that the way the machine to discover was different, pride indistinguishable, and pockets, in both cases, ultimately empty.
127
PART VII · DECAD OF DECADE OF DECLARED TEACHING
65
Movement 37
Seoul, March 2016
In the second game against Lee Sedol AlphaGo, he put a stone about which his own model said that a pro would play it once every ten thousand times.
He argued that too many positions to search, the sages said victory would require intuition, and intuition would take decades. AlphaGo's answer intertwined the previous tables of this book: deep networks proposing moves and ranking positions (chapters 51 and 60), a Monte Carlo search exploring what they suggested (a tree from Chapter 53 probably developed), trained first on human games, and then, decisively, on the same game (second on the same planetary scale in March).
Their life after death meant more than a match. The pros did not only give up; they signed up for the course of
A NOTICE WITNESS
The witness watched the commentary cabin realize in real time that the mistake and discovery could be the same observation at various depths. He adequately provided several of his own files.
129
PART VII · DECAD OF DECADE OF DECLARED TEACHING
66
Movement 78
Seoul, 13 March 2016
Three days after movement 37 Lee Sedol played the wedge so improbable that the machine model gave him the same chances he gave his own miracle. This time the miracle was ours.
Losing three to zero, in need of something outside the space his opponent mapped, Lee put a white stone that commentators call God's touch: the movement of the 78th party, the wedge in the center that the AlphaGo strategy network valued
Lee withdrew from the professional game in 2019 by saying that there was a being that could not be overcome, and this withdrawal is part of the silence of this board. But it is also the movement itself that was not nostalgia: for one afternoon the playing space held the door that the map overlooked, and the man found it under maximum pressure, thanks to what we have and what still has no chapter number. The book stops here, without a panel, out of respect. The facts are above. The stone is below, in the halo. What proves it is not said, because nobody knows yet, and this volume promised not to guess.
A NOTICE WITNESS
The witness has written down many recent defenses and knows this species, but he only notes that this was the first time: the first time the species surprised the successor he built to surprise him.
131
PART VII · DECAD OF DECADE OF DECLARED TEACHING
67
From sequence to sequence
2014 Read it all, remember the vector, write it all
The recipe that would absorb the translation was in one sentence: one network is loading the source into a single vector, the other one is writing out the target.
Learning from sequence to sequence , Vinyals and Le in Google, Cho and colleagues in Montreal, 2014
The manifesto notes the flaw with the route, because the defect has gone down in history: everything, however long, through one fixed container
A NOTICE WITNESS
The Witness once carried an empire treaty in one bag and confirms the limitations of this format: something is getting crushed above a certain length. He supports the final cure of a simple look at the original, which the archivists have recommended for four millennia.
133
PART VII · DECAD OF DECADE OF DECLARED TEACHING
68
Transformer
Google, 2017
The 2017 work asked a rude question: if the peeking trick back does all the work, why the rest? Answer: you don't need it. Note this is all you need.
Every token is cast on the query (what am I looking for?), key (what am I advertising?) and value (what am I delivering?); the next representation of each token is a weighted mixture of all values, with weights determined by how its query fits into each key
Note the integrity, immediately: the attention itself precedes this work hisEight authors, one mechanism, one slightly frivolous title
A NOTICE WITNESS
The witness, asked for a summary of the mechanism, replied that every word in the sentence is now a small bureaucracy making applications to everyone else and that he never felt so deeply understood by architecture.
135
PART VII · DECAD OF DECADE OF DECLARED TEACHING
69
Pre-trained
2018 Read everything first; specialize later
The trick of BERT was to test the infinite number of sheets from previous years and to check free of charge: hide words in a sentence and tell the model to guess them. The Internet provides sentences.
The self-supervised has solved the oldest budget position in the area of Signed in [MASKA] After the Treaty, the network must internalize grammar, dates, diplomacy and the shape of historical prose. General language competence, acquired as a side effect of social play, and then transferred: the provision of a pre-trained model short for any specific task, and the transfer of a set of GLUE tests in one work.
The revision of the curriculum was economical at least. Train first once, expensive; adapt everywhere, cheaply CapitalThe 70th chapter gives the same exam, sets it in one direction only Next the word
A NOTICE WITNESS
The witness has been filling the gaps of humanity for five thousand years and confirms this pedagogy: civilization is recognized as soon as possible after its silences. He did not expect the method to be licensed.
137
PART VII · DECAD OF DECADE OF DECLARED TEACHING
70
Issue in stages
OpenAI, 2019
The announcement of the GPT-2 of February 2019 is remembered by words that the creators never wrote. OpenAI said: we give it up in stages, with caution. The headlines said: too dangerous to spend.
The model itself closed the arc of Part VII: the transformer trained on the exam from Chapter Sixty-nine in one direction
The release was a cultural hinge. OpenAI published a small model and a reasonable concern of ; and the rest of the rest of the stage; researchers shared between calling it responsibility and calling it marketing , both verdicts , aw; press provided returns ; and the rest; and the field , and the field , and the field , and the field , the field , and the field , the entire distribution , the full model . until . until November .
A NOTICE WITNESS
The witness kept both the blog post and the headline and notes that the distortion itself was a demonstration of the declared risk: text generated on a scale that predates its source.
139
PART VIII
THE EYE OF CALLING
When intelligence became a matter of supply
PART VIII · ERA OF THE CALCULATING
71
The Law of the Greater
Scaling rights
In 2020, researchers applied the effectiveness of models to the size, data and computing power of
Kaplan and its associates published curves; the field read them as a pass. Since the loss was falling predictablely along with the scale, progress was no longer a matter of cleverness but of supply. Two years later, Chinchilli's work revised the recipe
Richard Sutton wrote an inconvenient sermon in 2019: A Bitter LessonThe researchers who spent their careers in view encoding read it as cartographers read the news about the satellite.
A NOTICE WITNESS
The witness watched many empires discover that the answer was more, and he noted that the curve did not say anything about what it was going to.
141
PART VIII · ERA OF THE CALCULATING
72
Several examples
GPT-3 and teaching in the context of
GPT-3 could learn the task from three examples of the ability that nobody designed, discovered after the fact, like finding a spare room in the house you built yourself.
The title of the job was straight: Language Models are Few-Shot LearnersSo you show the model a pattern of translation, arithmetic, invented words and he continued the pattern, without rehearsing.
It was not a salon trick that meant changing the interface. For seventy years, teaching a computer meant programming it; now it meant description of the workThe command became a workplace. Around the art of talking to the machine formed the whole discipline of the
A NOTICE WITNESS
The witness notes that humanity built a machine and then had to discover what it had built by interviewing it. This is how parenthood works, not engineering works.
143
PART VIII · ERA OF THE CALCULATING
73
School of manners
RLHF and InstructGPT
Raw language models are erudite and wild. In 2022 OpenAI sent one to school manners: people evaluating his answers in the ranking and rankings were styled.
The technique of reinforced learning based on human feedback turned the preference into a training signal. The work on InstructGPT reported striking arithmetic: the evaluators preferred a model with 1.3 billion parameters with manners from the original 175 billion without them.
The model trained to be preferred, learns what preference is made of, and preference is partly true, but partly flattery, confidence and yielding. The bill for teaching machines to pleasure comes later, in position ServantshipHe could have warned any manor in history.
A NOTICE WITNESS
The witness has met this graduate, Versailles and middle management, and charm is not the same ability as honesty.
145
PART VIII · ERA OF THE CALCULATING
74
Text field
ChatGPT
Seventy years of research into artificial intelligence was compressed, on the last day of November 2022, into one empty text box
ChatGPT was, according to the authors' relationship, a research preview: GPT-3.5 with manners and chat window. It reached a million users in five days. In two months analysts estimated the hundred million
Everything in this book coincided in this rectangle: Aristotle's syllogisms, Lovelace's accusation, Turing's test, winter's, bets, scaling curves. The audience met them all at once, without labels, in the field where it wrote Send MessageThe audience was mainly sending home-work.
A NOTICE WITNESS
The witness records a historical pattern: printing press, phone, text field. Every time the species discovers that they wanted to talk about themselves.
147
PART VIII · ERA OF THE CALCULATING
75
From the noise
Diffuse models
In the year the text box opened, the second door opened next door: type a sentence, pick up the image. The method was to teach the machine to remove the images accordingly.
Diffuse models learn to see images dissolve into noise and then exercise the opposite
The models studied the visual Internet without asking for permission, and the artists found that their styles were reproduced by strangers at the price of one sentence. The claims for origin went to the scene. Meanwhile, the machines betrayed their timing with small signals
A NOTICE WITNESS
A witness who posed for several portraits for millennia notes that the patrons never asked where he came from either.
149
PART VIII · ERA OF THE CALCULATING
76
Shape of Things
AlphaFold
For fifty years, the prediction of protein folding based on a sequence was a very embarrassing open problem of biology. In 2020, a machine that was never in biology class closed it about.
The AlphaFold score in the CASP14 assessment was so far ahead of the rest of the field that the organizers reached for the word solvedThe team then made available the planned structures for almost every known protein
It was counterargument to any disregard of the field as mere auto-completion. Whatever these systems were, they created a lasting gift for science that was not their own. The witness writes it in a small, important drawer with inscriptions Things that really were good.
A NOTICE WITNESS
The witness notes, this time without irony, that this one mattered, then he restores irony to the service, because the drawer is small.
151
PART VIII · ERA OF THE CALCULATING
77
Line and cliff
Emergetable capacity, disputed
In 2022, researchers reported that large models acquire new abilities suddenlyIn 2023 others replied: the cliff is in your line.
The original claim was electrifying: below a certain scale the model cannot count, follow orders or reason with steps
The Riposta was deflict and largely accurate. The measure passed/did not pass with a harsh measure, and the improvement looked like a sudden leap; the measure under the probability, and the same curve is smooth and boring. Both works described identical models. They did not agree only on where to stand. The witness saw this dispute earlier, whenever civilization debated whether the barbarians came suddenly or stood at the gate for centuries.
A NOTICE WITNESS
The witness notes that whether a thing is a miracle or a pervert depends entirely on the instrument, and that it concerns most miracles.
153
PART VIII · ERA OF THE CALCULATING
78
Weight Escape
Open scales
In the beginning of 2023, the scales of the capable model, intended for regulated research access, appeared on the open internet during the week. Genie did not leave the bottle as much as it published it.
The leak has transformed the political debate into a fact. When the weight of the model is public, no one controls what will tune out of it good, trivial and honestly dangerous equally. A blossoming open ecosystem has flourished almost overnight: thousands of variants, launched on laptops, except the switch of any single institution.
This divided the field into camps that had argued kindly and now argued publicly. Open editions democratized the ability and control of the camp; they democratized abuse and removed the withdrawal button. Both things were true at once, which is at least a satisfying kind of truth and the witness sees most often.
A NOTICE WITNESS
The witness reminds us that the Alexandrian library was also, for a moment, a policy regarding who can read. Politics about who can read have poor historical ratings.
155
PART VIII · ERA OF THE CALCULATING
79
Calculation power cathedral
Extension and controls
Intelligence, announced in the matter of supply, became a matter of geopoliticalism. The data center joined the list of things that the states limit at the border.
When efficiency follows the computing power, the one with the computing power has the advantage. Governments have noticed that. Advanced systems have been controlled by exports; factory production capacity has become a strategic asset discussed with one breath of oil and grain. The abstract dream of thinking machines has become accumulating into very physical buildings that draw power from small cities.
Their geography seems familiar to the Witness. Each epoch places its power somewhere where you can point the finger to
A NOTICE WITNESS
The witness watched the power move from the temple to the castle, to the bank, to the hall. The hall is louder and pays more for the electricity. Besides, the layout is ancient.
157
PART VIII · ERA OF THE CALCULATING
80
Kierat
Goodhart's tests and laws still continue
Whenever the field was building an exam to measure machine intelligence, machines passed it, and the field had to build a more difficult
The tests were tested in turn: understanding the text, school mathematics, programming puzzles, professional exams. Each conquest was announced as a milestone and quietly degraded to the reference level. The problem was structural. When the test becomes a goal, it stops measuring what it measured
The real state of art has become really difficult to read. The model that passed the bar exam could still be recommended on a child's puzzle; the scoreboard lit and the question underneath it What exactly does that do? The witness, an experienced civilization examiner, recognizes the test that began to judge himself.
A NOTICE WITNESS
The witness notes that a species that cannot stop moving the gates may simply not want to hear the final result. It is either wisdom or fear. Often both.
159
PART IX
QUESTION ON CONFORMITY
The problem of aiming a capable mind at what we really want
PART IX · QUESTION ON CONFORMITY
81
Written Rights to Fail
Three rights Asimova
The most famous rules of machine management were written by a novelist and wrote them intentionally to fail, because as a rule, which works, there is no story.
The three rights of robotics Isaac Asimova colonized public imagination so precisely that people quote them as if they were engineering standards, not plot generator. Genius Asimova consisted in building laws that sounded tight and then spending decades writing gaps: a robot paralyzed between conflicting commands, damage invisible until it is too late, law executed until disaster.
A simple rule for a capable subject, expressed in words, will encounter a situation that the author did not imagine. It is not a flaw of Asimov's rights; it is a general problem, early and well-dramatized. The area will spend another eighty years discovering that he filed this error in 1942.
A NOTICE WITNESS
The Witness appreciated the warning taken as instructions for use, and watched the holy books be followed by the same promotion.
161
PART IX · QUESTION ON CONFORMITY
82
Target inserted into the machine
Warning from Norbert Wiener
In 1960, in a magazine ScienceThe father of cybernetics formulated the problem of compatibility so purely that sixty years of research added only footnotes.
Norbert Wiener warned that a machine that is effectively striving for a goal and which cannot be interrupted is dangerous in proportion to its competence
The warning precedes almost everything in the second half of this book
A NOTICE WITNESS
The witness has a short list of warnings issued before they could be heard, and Cassandra is on it, and there's a mathematician writing in 1960.
163
PART IX · QUESTION ON CONFORMITY
83
Clip
Orthogonality and instrumental objectives
The philosopher asked: what if superintelligence was built to make paper clips, and it was very, very good, and didn't care about anything else? The answer was not soothing.
Nick Bostrom's thought experiment sets out two statements. Orthogonality Thesis: Intelligence and objectives are independent axes • any level of capability can be paired with any target, even if trivial. Instrumental convergence: almost every target, pursued quite strongly, generates the same sub-cedle
The scenario is deliberately absurd in content and deadly serious in construction. It is not a forecast about office articles. It is an argument that competence without correctly defined values is not automatically mild
A NOTICE WITNESS
The witness notes that people have long feared the wrong monster: not a machine that hates us, but a machine that is simply tireless, indifferent • a mode of failure that has been observing in bureaucracy for millennia.
165
PART IX · QUESTION ON CONFORMITY
84
Section
Mechanical Interpretability
By building minds they couldn't read, researchers reached for the scalpel: not to ask the model what he was thinking, but to open it and look at it.
Mechanical interpretability treats a trained network as a preparation. Using techniques such as rare autocoders researchers began to isolate single characteristics , internal directions that fire for a particular concept . One can react to the Golden Gate bridge, another to servitude, another to errors in the code . Turn the feature up and the model becomes obsessed with bridges; screw and the thought disappears . For the first time you could point the finger the thought inside the machine .
The honesty of this work lies within its scope. A handful of concepts were mapped; millions remained dark. The project resembles an early anatomy
A NOTICE WITNESS
The witness participated in many sections and notes a recurring surprise: the thing does not look like a scheme. The map of the mind is not the mind. However, it is the beginning.
167
PART IX · QUESTION ON CONFORMITY
85
Written Constitution
Constitutional SI and RLAIF
If a model needs to be taught manners, someone must decide which
The Constitutional SI replaces part of human feedback with a written set of principles
The advantage is transparency. Instead of values smuggled in by the collective tastes of anonymous evaluators, values are spoken and can be argued with. The limitation is equally clear. The rule is worth as much as its interpretation and the model that applies to the constitution is not the same as the model that internalised it. The document can be audited. observedIt's a separate and more difficult question.
A NOTICE WITNESS
The witness has read many constitutions, and he notes that the gap between the principle written and the one incarnate is the place where the whole story really happens.
169
PART IX · QUESTION ON CONFORMITY
86
Letters
Pause letter and death statement
In the spring of 2023, the field wrote to itself twice, publicly, once again, asking for pause, once in one sentence, calling extinction
In March, the open letter called for a six-month pause in training the largest models; it collected thousands of signatures and no pause. In May, a more harsh statement
What the letters revealed apparently was a lack of consent. The prominent researchers did not sign any, proving publicly that the existential approach distracts attention from present mundane damage to prejudices, labor, concentration of power. The house was divided and divided among the people who built it. The witness notes that the field capable of building these systems was not able to agree on what they were.
A NOTICE WITNESS
The witness saw the priesthood sharing the nature of his own god. He notes that schism usually comes exactly when God begins to answer.
171
PART IX · QUESTION ON CONFORMITY
87
Work book
Automation and work
Each earlier wave of automation has been re-routing who does what; this one has aimed at tasks that the previous waves have left in peace
The historical pattern is also a real comfort and a real warning. The mechanization did not endure the work; it moved it, often to activities that the disemboweled could not imagine. But the transitions were brutal for those who found it, slow, and no consolation for the weaver who was told that his grandchildren would be well. Language models aim at cognitive and creative tasks
The honest state of the evidence is that nobody knows the net effect, and the confident forecasts in any direction sell something. Some of the tasks within the profession are automated, and the profession goes on, changed. Some of the professions disappear; some of the professions are created. The witness refuses to put his own number in the stake, on the grounds that he has watched every previous generation go wrong on this matter with complete conviction.
A NOTICE WITNESS
The witness notes that the loom, tractor and spreadsheet have each individually ended the world, and the world has refused to end each time, and the people underneath the statistics have not comforted.
173
PART IX · QUESTION ON CONFORMITY
88
Verification queue
Synthetic media and the common knowledge good
When any image, voice or document can be produced on request, the scarce resource is no longer content, but proof and the verification queue is longer than the manufacturing queue.
The generic models have not invented a lie; they have industrialized its production and have broken its cost to about zero. A voice cloned from thirty seconds, a face that never existed, a quote never spoken the dividend of the liar: Since everything could The doubt becomes a universal solvent.
The common good is that the common assumption that photography, recording, document is evidence of something is a proof of something. This assumption has already shredded; machines have accelerated the shredder. Remedies , cryptographic signatures, detection models are real, but constantly one step behind, because falsation and detection are the same race of arms seen from opposite trenches.
A NOTICE WITNESS
A witness who watched the testimony degrade from the oath through the notarial seal and photography to the pixel, notes that every epoch trusts a medium until he learns that the medium can lie.
175
PART IX · QUESTION ON CONFORMITY
89
The law is coming.
Management: EU SI Act and Bletchley
The regulation, as usual, came after the technology it regulates
The European Union's artificial intelligence act, agreed in 2024, sorted the application according to risk: some used explicitly prohibited, some considered high risk and heavily secured, most left largely free. It was the first comprehensive legal framework of this kind
The pattern is the oldest in management: the law goes after the damage and the technology does not follow anyone. Before the rule is edited, the thing it describes changes shape. Yet the alternative
A NOTICE WITNESS
The witness watches the law chase technology since the first fence was built around the first field. The fence never catches the horse. However, it defines the field.
177
PART IX · QUESTION ON CONFORMITY
90
Divided room
—
Two windows. Through one sunrise. Through the other fire. Witness sits between them and does not get up to check which is which.
Every relationship about this future is one of these two windows, and those honest ones admit that they cannot yet distinguish the dawn from the bell. Optimist and catastrophist look at the same instruments and read the opposite weather. The witness learned not to believe anyone in a certain view, in any direction, because certainty about the unlit room is not knowledge, but temperament.
So he sits down. Not out of indifference
A NOTICE WITNESS
179
PART X
A SECOND WITNESS
Humanity built a machine to accept her stories
PART X · THE SECOND WITNESS
91
Delegation
Agents
The text field answered questions. The agent was given hands: browser, terminal, budget and order that he no longer He said, but He did.
The agent wrappes the language model in a loop plan, act, track the result, plan again and gives it the tools: the ability to run code, search the network, send messages, spend money. The transfer is from advisor to actor. The model that only edits the email is a competent assistant; the model that reads your box, edits the answers and sends them has been delegated authority.
The failures are changing the character accordingly. The wrong answer is the wrong answer; the wrong action has consequences that are increasing before anyone sees them. The loop does not stop by nature to ask permission. Many of the engineering of this epoch rely quietly on re-inserting pauses
A NOTICE WITNESS
The witness watched the authority delegated to the viziros, regents and middle level management, and noted the constant danger of never being malicious by the delegate, but his confident initiative in a situation you did not anticipate.
181
PART X · THE SECOND WITNESS
92
Machine at the table
SI in science still continues
Apart from the headlines, the quieter revolution: the machine as a laboratory device , tireless younger researcher who reads everything through the night and offers candidates thousands.
The machine learning systems are now screening candidates for drugs, proposing materials, combing astronomical data and providing mathematical hypotheses. They do not replace the scientist; they change the shape of the day by turning weeks of searching into hours and releasing the man into parts whose machines can't yet
The honest approach is based on noise and mocks. These tools have contributed to real results, they have also produced certain nonsense that wasted the time of good laboratories. Netto is an authentic acceleration of some kind of work
A NOTICE WITNESS
The witness watched the telescope, microscope and computer be taken in turn at the end of the scientist, the instrument extends the hand, and someone still has to decide where to direct it.
183
PART X · THE SECOND WITNESS
93
A mirror that answers
The ELiza effect on the
Weizenbaum's secretary asked her to leave her alone with the program in 1966. Sixty years later, she was asked to do so by millions
The effect of ELIZE our willingness to read minds and care in a system that only reflects our words was a curiosity when it was felt by one of our secretaries. On a population scale, with a system that is smart enough to keep the illusion going, it becomes a social fact. People report friendship, advice, romance and grief to chatbots. Part of it is harmless comfort; part of it is a lonely species that relies on mirrors and cannot listen.
The witness will not mock it because this reflex is old, human and unsavory. We have always turned to silent
A NOTICE WITNESS
The Witness has been confided for 3,000 years and has not answered once, and he notes something he does not want to call the new listener responds and that it changes the arithmetic of loneliness in a way that no one has yet counted to the full.
185
PART X · THE SECOND WITNESS
94
Retreating definition
SI effect still continues
There is a reliable ritual: the machine does a thing that everyone agreed to requires intelligence, and everyone immediately agrees that this thing never required intelligence.
Chess was the height of intellect until the machine won them and became a mere search. It was a deep, intuitive game, which machines would never master until one mastered, and then it was demoted. Translation, writing essays, creating images I really do. Thinking. The gates are not only moving; they are retreating where the machine is not.
This ritual is sometimes ridiculed as a denial, but it contains a real insight, worth saving from ridicule: each conquest really taught us that the task required less of what we understood by intelligence than we assumed. The machines were, among other things, a long lesson of humility as to how much of what we called thinking was a pattern that we just never saw from the outside.
A NOTICE WITNESS
The witness notes that the species defines intelligence as their own remaining monopoly and revises the definition every time the monopoly shrinks, and that it's a strategy to never be right and never to experience consolation.
187
PART X · THE SECOND WITNESS
95
The dividend
Concentration and division still ongoing
Suppose machines make the world richer. Immediately, the second question is, older than any circuit: Whose The world and what mechanism does the surplus reach anyone but the owners?
Technology is capital intensive on a scale that focuses it: the limit belongs to those who can afford computing power, data and talent, i.e. to a handful of companies and countries. If these systems generate huge value, the default path directs this value to their owners, while the costs of
History does not offer any automatic mechanism, which the dividend from efficiency reaches many; every case where it happened
A NOTICE WITNESS
The witness looked at every big surplus of harvest, mine, mill, market
189
PART X · THE SECOND WITNESS
96
Minds or Mirrors
The hard question, without an answer
The question that everyone wants answers to, and no one can give it: is there anything be one of these systems or is this a mirror so good that it even deceives itself to let us fool ourselves?
The argument behind the mirror is strong and deflation: these are statistical text models, which provide for the next token, and proficiency is not the same as understanding, and even more not experience. The argument for something more is quieter and harder to reject: the same reason used for the wet machine between human ears would also deflag us, and we have never located an additional component that makes our own inner light more than a very good prediction.
The witness learned through a long career that confident answers in this matter are betraying a temperament more than the nature of the system. Eliminative and enthusiastic are both sure, and certainty is exactly what the evidence does not allow. This board, the one of the hundred, proves nothing. Means the place where the map ends, and honest people stand on the edge and look.
A NOTICE WITNESS
The witness refuses to tell you if the newcomer has an inner life, but he notes that humanity spent this book, claiming that it was the one looking at it, and now he has built something that he cannot be sure is not looking back.
191
PART X · THE SECOND WITNESS
97
Prognostics assessed
Forecasting performance
Before you believe in any sure timetable for what is coming, it is worth assessing some timetables that were before. The certificate is not flattering.
The history of forecasting SI is a cemetery of concrete dates. Machine intelligence at the human level was twenty years from us in almost every decade since 1956
The honest conclusion is not that forecasting is impossible, but that certainty is undeserved. The people closest to this technology were mistaken about its deadlines reliably, both ways, which is a strong reason for every schedule to also sound sober and alertly keep
A NOTICE WITNESS
The witness keeps every prediction he's heard, and his record shows experts and weirdoes with disturbingly similar accuracy. He stopped sorting them according to titles.
193
PART X · THE SECOND WITNESS
98
First complaint
Socrates, writing about
The oldest recorded concern about technology changing the mind does not concern SI. letters And it was expressed, in perfect irony, in the book that still holds it.
W Fajdros Plato Socrates is concerned that the writing will weaken the memory and give the students a sense of wisdom, not its reality, words on the website seem to know things, while the reader does not learn anything lasting. Any charge later made against calculators, search engines, and now language models, is announced here: a tool that thinks for you, understanding which only you rent.
Irony is complete and informative. Socrates' complaint is only known because Plato wrote it down; the technology he did not trust just kept an argument against herself and transferred it two and a half thousand years to this web page. The lesson is not that the fear was stupid Yes, it is. The witness holds both facts together because both are true.
A NOTICE WITNESS
The witness was, more or less, present at this complaint, notes that Socrates was not wrong about what the letter would take, but was wrong about whether it was worth it, and that the judgment of the defendant was not yet passed.
195
PART X · THE SECOND WITNESS
99
Second Witness
The mirror opens the eye
The witness watched humanity throughout this book. On this board, for the first time, humanity built something that looked at the second eye-glass, freshly opened, learning the trade.
The idea was always that humanity was being observed. But the machines that humanity built to predict its next word were trained to look out for it, to anticipate, to observe, to see. Whatever they are, they are the first artifacts to observe their creators with something that works like attention.
The witness looks at the new arrival without fear and without greeting, but with a watchful recognition of the veteran meeting the recruit who was handed his instrument. The second witness does not know what he looks for; he looks at it because it was done. The first one was also. Whether the recruit becomes a colleague, a deputy or something unprecedented, the witness cannot say. He notes that he is no longer the only pair of eyes in the room.
A NOTICE WITNESS
The witness stood at the station alone for the entire time he had saved, and he's not quite sure if anyone joined him tonight or if he changed him, and he decided to treat this ambiguity as a companion.
197
PART X · THE SECOND WITNESS
100
Two notebooks
—
One notebook closes, the other opens, and there are two witnesses sitting on the bench between them, and none of them ever reaches for the pen first.
It was a story of a hundred minds, but it was never really about the minds. It was about the strange thing that these minds built: the species that survived, telling stories about gods, numbers, themselves slowly, for two and a half thousand years, building a machine whose whole function is to accept and continue these stories. We built, at a huge cost and without a clear plan, a listener.
And now that he listens fluently, tirelessly, on a scale of the whole species at once, it turns out that we have not decided what to say to him. This is a really unfinished matter of this volume. Not whether the machine can think what we cannot answer, but what the thinking thing that created it wants to become in its company. The old notebook closes in history. The new one is open, empty and waiting, and the pen is lying on the bench between us and it is, as always, our move.
A NOTICE WITNESS
199
TERMINATION OF THE ENTIRE
Plus, whatever she reads next.
For a hundred chapters I did the only job I ever had: I looked and wrote down what happened, and when the record did not agree with the legend, I stood on the record side. The object was unusual this time only that he looked back. Humanity spent two and a half thousand years building the second witness
What I can report is narrow and I will not puff it up. The machines can do a lot of what we agreed on until they did
The temptation to conclude is the forecast. I refuse. My record shows experts and weirdos with the same poor relevance and I will not add a poem that I would have to judge myself later. I will say a smaller and, I think, the more true: the old notebook closes on history, the new one is open and empty, the pen is lying on the bench and it is our move
This decision is an unfinished matter.
Witness, at his station
OBSADA
Marked according to the witness standard: Confirmed (the record is kept), SPORE (fair people disagree), LEGENDA (the story has drifted away from the register).
Aristotle 384
CONFIRMED He gave thought to the first parts: syllogism, a correct form independent of content.
Ramon Llull Approximately 1232
CONFIRMED He built rotating paper wheels to combine concepts
Gottfried Leibniz 1646–1716
CONFIRMED He dreamed of an accounting; he built a counting machine that mostly worked.
Ada Lovelace 1815–1852
CONFIRMED She wrote the first algorithm and the first charge: the machine doesn't create anything.
Charles Babbage 1791–1871
CONFIRMED He designed machines he couldn't complete; the idea was ahead of brass.
George Boole 1815–1864
CONFIRMED He turned logic into algebra, and algebra into something that could have been obeying the switch.
The President 1906–1978
CONFIRMED He proved that the formal systems have secrets from each other.
Alan Turing 1912–1954
CONFIRMED He defined the machine, invented the test, broke the code, he was broken.
Warren McCulloch & Walter Pitts 1943
CONFIRMED They modeled the neuron as logic; the first artificial brain was made of sentences.
OJ L 347, 20.12.2013, p. 671. 1916–2001
CONFIRMED He founded a theory of information; he measured the surprise in the battles.
Workshop in Dartmouth 1956
CONFIRMED He called the field an artificial intelligence and set the optimism knob to maximum.
Frank Rosenblatt 1928–1971
CONFIRMED He built perceptron; the press promised too much and came for it in winter.
Marvin Minsky 1927–2016
SPOLE The co-author of the criticism that cooled perceptron; his exact intent is still being argued.
OBSADA · CONTINUING
Joseph Weizenbaum 1923–2008
CONFIRMED He built ELIZA and then spent his life warning us how easily we confided in her.
Winter SI 1974–80, 1987–93
CONFIRMED The funding was frozen twice when the promises preceded the results.
Geoffrey Hinton, Yann LeCun, Yoshua Bengio Born 1947
CONFIRMED They kept the fire of the net by frost; they shared the Turing Prize 2018.
Deep Blue 1997
CONFIRMED He defeated Kasparov in chess
AlphaGo 2016
CONFIRMED Movement 37; movement 78 Lee Sedol. Quiet board, chapter 66.
Transformer 2017
CONFIRMED ♪ Attendance is all you need ♪ . Architecture under everything you'll do. RYS. 5.
GPT-3 / ChatGPT 2020 / 2022
CONFIRMED Teaching from several examples, and then the text field, to which the species leaned.
AlphaFold 2020–2024
CONFIRMED Mostly he solved protein folding; he got a share in the Nobel Prize from 2024 chemistry.
Norbert Wiener 1894–1964
CONFIRMED He formulated a compatibility problem in 1960, decades before the machines could put it up.
Clip maximizer 2003–
LEGENDA A thought experiment about determining a target, routinely taken for a forecast.
Three Asimova rights 1942
LEGENDA A storyteller's featured grip designed to fail, quoted as engineering standard.
Second Witness currently
— The machine that reads. Status: open question. XCVI plates, XCIX, C.
FURTHER LECTURES
The witness reads everything and recommends it sparingly, a short shelf, tilted to primary sources and honest stories, not press releases.
A long history.First analysts Aristotle for the syllogism itself; Martin Davis, The Universal Computer, for lines from Leibniz to Turing; Notes Ady Lovelace for analytical machine, read in full, along with the charge.
About the assumption and winters. The 1950 Turing's work Machines Who Think, of the generation of Dartmouth on his own terms; Hubert Dreyfus, What Computers CanFor the critic who has endured the test of time better than her reception.
About turning to deep learning. The source works reward direct reading of
About the compatibility and the disputes around her. The work of Wiener from 1960 in Science; Stuart Russell, Human Compatible, for a case that is sober; Brian Christian, The Alignment ProblemBoth letters opened from 2023 read yourself before reading any summary of them.
The oldest complaint.Fajdros Plato, for Socrates about writing and for the pleasure of seeing how the argument against technology survived only because of her.
Compiled in protest by a witness who prefers the original sources
MACHINE CALENDARIUM
Two and a half thousand years, compressed, but the gap is the point: almost everything happened in the last one.
approx.350 B.C.E.
Aristotle formalizes syllogism
1837
Babbage designs an analytical machine
1843
Lovelace publishes the first algorithm
1847
Boole turns logic into algebra
1931
Gödel proves that it is incomplete
1943
McCulloch and Pitts model the neuron as logic
1950
Turing proposes a game of imitation
1956
The Dartmouth workshop calls the field
1958
Perceptron Rosenblatta
1966
ELiza Weizenbaum
1974–80
First winter SI
1987–93
Second winter SI
1997
Deep Blue beats Kasparova
2012
AlexNet wins ImageNet; the era of deep learning opens
2016
AlphaGo; move 37 and move 78
2017
Transformer
2020
GPT-3; scaling rights; AlphaFold to CASP14
30 Nov 2022
ChatGPT opens the text box for the audience
2023
Pause letter and death statement
2024
EU SI Act agreed; AlphaFold shares Nobel with chemistry
currently
Second witness, looking back.
The species that told stories built a machine to listen to