The History of Ai: How We Got Here

Transcript

Chapter 1: The Ancient Dream of Thinking Machines

Imagine the quiet glow of firelight falling across a figure made of bronze. Nearby, water passes through narrow pipes inside an elaborate clock, lifting one vessel, lowering another, and setting a carved bird into motion. A hidden wheel turns. A wooden musician raises an arm. For a moment, the boundary between the living and the made seems less certain.

Long before electronic computers, people wondered whether human skill could create something that moved, performed, or appeared to act on its own. The history of artificial intelligence begins, in one sense, with these questions. Yet it does not begin with a thinking machine. It begins with stories, craftsmanship, and curiosity about life, imitation, agency, and control.

Agency means the capacity to choose or act intentionally. A falling stone moves, but it does not decide to fall. A person may choose to lift a cup, write a sentence, or change direction. Between simple movement and intentional action lies a distinction that occupied storytellers, philosophers, and engineers for centuries.

Ancient myths often explored that uncertain border. In Greek mythology, Hephaestus was the divine smith, associated with fire, metal, and astonishing devices. Stories described automatic bellows in his workshop and golden assistants that could support him and respond to his wishes. These figures were anthropomorphic: they had humanlike forms or personalities, although they were imagined as manufactured rather than born.

Another Greek story told of Talos, a bronze guardian who watched over Crete. He was powerful, tireless, and bound to a purpose. The tale raised enduring questions. Who gives such a guardian its task? Does it merely obey, or possess intentions? What happens when strength and obedience are combined in a being that resembles life?

Talos and the golden assistants were not constructed examples of artificial intelligence. They belong to mythology, not engineering. But myths reveal the questions people found compelling: manufactured life, commanded action, faithful service, and the possibility that humanlike behavior might emerge from something made by skilled hands.

Alongside these stories, real makers developed devices that moved through physical mechanisms. An automaton is a machine that performs a predetermined sequence of movements. Its actions may appear spontaneous, but they are arranged in advance by the structure of the machine.

Ancient and medieval automata could be powered by falling weights, flowing water, air pressure, ropes, gears, and concealed mechanisms. A descending weight could pull a cord. Water entering a vessel could cause it to tip. Air could force sound through a pipe. One motion could release another, creating a sequence that seemed lifelike despite arising from ordinary physical causes.

Hero of Alexandria described many such devices, including mechanical theaters in which miniature figures moved through prepared scenes. Ropes wound around parts of the mechanism, weights descended, and carefully arranged components produced changes in movement. Doors opened, figures turned, and a scene unfolded without a visible human operator.

Hero also described devices powered by air and water. To observers who could not see the machinery, their motions might have seemed mysterious. Yet every movement followed the path prepared by its maker. The figures did not understand their scenes or possess goals of their own; they offered an outward imitation of purposeful action.

That difference between apparent intention and intention itself remained important throughout the history of artificial intelligence. A device may do something associated with thought without possessing thought in the human sense. It may produce a meaningful result through a process unlike our own.

Centuries later, engineers in the medieval Islamic world advanced the craft of automata, creating sophisticated water clocks and other moving devices. Among the most celebrated was al-Jazari, who worked in the early thirteenth century. His clocks contained moving figures while water flowed through chambers and vessels, regulating motion, sounding instruments, and marking the hours.

Al-Jazari also designed a mechanical musical ensemble whose performance could be altered by changing the arrangement of pegs. As part of the machine rotated, the pegs triggered movements and produced a prepared rhythmic sequence. This arrangement is sometimes described, cautiously, as programmable because repositioning the pegs changed the performance.

It was not a computer. It could not interpret broad symbolic instructions, reason about music, or invent a new performance. Its actions remained constrained by its physical design. Even so, the adjustable mechanism expressed an important idea: a machine's behavior did not always have to be fixed permanently when the machine was built.

European clockmakers also created increasingly intricate mechanical figures. Clockmaking demanded precision. Gears divided time into regular intervals, springs stored energy, levers transferred movement, and cams shaped repeated actions. A cam is a shaped rotating part that converts rotation into a particular motion.

As a cam turns, another piece resting against it moves up and down or from side to side. By shaping it precisely, a maker can determine when a mechanical hand rises, when a bird lowers its head, or when a musician appears to breathe.

In the eighteenth century, Jacques de Vaucanson became famous for automata that imitated living creatures and skilled performers. One of his best-known creations was a mechanical flute player, designed to imitate aspects of playing through controlled air and coordinated movement. It could perform selected motions and produce music, a striking imitation of refined human action.

Yet it did not hear the melody as a listener did. It did not understand a performance, choose an interpretation, or learn from an imperfect note. The distinction between skillfully reproduced behavior and understanding remained clear beneath the wonder.

Between roughly 1768 and 1774, another extraordinary group of automata was created by Pierre Jaquet-Droz, his son Henri-Louis Jaquet-Droz, and Jean-Frédéric Leschot. Known as the Writer, the Draughtsman, and the Musician, these figures displayed mechanical craftsmanship of remarkable delicacy.

The Writer could form letters with a pen and produce prepared text. The Draughtsman created pictures through coordinated movements of the hand. The Musician, seated at an instrument, pressed the keys and moved like a human performer. Intricate internal mechanisms governed every action.

Cams and other components guided the figures through arranged paths. Their movements appeared fluid and intentional because the makers had encoded human gestures into mechanical parts. A viewer might watch the Writer bend over the page and feel, briefly, that the figure knew what it was saying.

But the Writer did not understand the words or consider a better phrase. The Draughtsman did not imagine a picture, and the Musician did not feel a melody. Their apparent liveliness came from the insight and patience of their makers.

Their fascination lay partly in that appearance. People often recognize intelligence through behavior: we listen to words, observe gestures, and notice responses. An automaton complicates this familiar process. It offers the gesture without intention, handwriting without thought, and melody without awareness.

To see a mechanical figure write is to encounter two things at once. There is the visible performance, graceful and familiar. Beneath it lies the hidden mechanism, where every turn and pause has been prepared. The surface suggests a living presence; the machinery tells another story.

This contrast does not make the accomplishment less remarkable. It reveals how carefully a complex human action can be studied. To reproduce a gesture mechanically, a maker must notice its smaller parts: a hand rises, shifts, pauses, and changes direction. What feels natural must be translated into metal, wood, tension, and rotation.

Still, dividing an action into movements is not the same as understanding it. A machine may trace a word without knowing its meaning, produce music without hearing, or resemble a person without possessing a person's agency.

There was no straight road from Talos to the modern computer. Myths, mechanical theaters, water clocks, musical machines, and clockwork figures belonged to different cultures and answered different needs. Some entertained. Some measured time. Some displayed craftsmanship or inspired wonder.

They were not early computers waiting to become modern artificial intelligence, nor imperfect versions of later machines. Each belonged to its own place, purpose, and moment. Together, however, they reveal a long continuity of curiosity.

Could motion be organized so that matter seemed alive? Could a sequence be preserved inside a mechanism? Could music, handwriting, drawing, or gesture be reproduced by wood, metal, air, and water? If the imitation became convincing enough, how would an observer distinguish lifelike behavior from genuine thought?

For

Chapter 2: Logic, Numbers, and the Idea of a Programmable Machine

The possibility of changeable instructions began quietly, among brass wheels, numbered dials, punched cards, careful diagrams, and patient columns of ink.

Long before electronic computers, arithmetic demanded a great deal of human effort. Long calculations had to be completed one step at a time, and a small error could pass unnoticed into all the work that followed. Inventors began to wonder whether some of this labor might be entrusted to machinery.

In the seventeenth century, the French mathematician Blaise Pascal created a mechanical calculator to help automate arithmetic. Its interlocking wheels could add numbers, carrying a value from one column to the next, much as a person carries a one when calculating on paper. Later, the German thinker Gottfried Wilhelm Leibniz developed another calculating machine capable of performing several kinds of arithmetic.

These devices were ingenious, though they were not general-purpose computers. Their abilities were largely determined by their construction. A calculator built for arithmetic could not be turned toward an entirely different task simply by receiving a new set of instructions.

Even so, the early calculators demonstrated an important principle. A procedure that took place in the mind, or on paper, could sometimes be divided into small, definite steps and embodied in a mechanism. The movement of wheels and gears could represent the movement from one stage of a calculation to the next.

Today, such a procedure is often called an algorithm. An algorithm is a defined series of steps for completing a task. A written method for long division is an algorithm. A carefully stated recipe can be one as well, provided that each instruction is clear enough to follow.

An algorithm does not have to be performed by a machine. People have followed ordered methods for a very long time. But once a process is described precisely, another question becomes possible. Could a mechanism follow those same steps?

A different kind of machine offered another part of the answer.

In the early nineteenth century, the loom associated with Joseph-Marie Jacquard used punched cards to control textile patterns. Each card represented part of a design. Holes in particular positions allowed certain mechanical actions to take place, while unpunched spaces did not. As the loom worked, a sequence of cards directed a sequence of movements among the threads.

The cards were physical instructions. They could be held in the hand, placed in order, removed, preserved, and replaced. By changing the cards, the same loom could weave a different pattern.

The Jacquard loom was not a general computer, and it should not be treated as a direct cause from which modern computing simply followed. It was a textile machine. Yet it provided a vivid example of how encoded instructions could guide a complex sequence. The mechanism remained in place while the pattern changed.

This is close to the central idea of programmability. A programmable machine can follow changeable instructions rather than performing only one fixed operation. The instructions determine what the machine does, and a different set of instructions can guide it through a different task, or a different version of the same task.

The loom’s cards made this idea almost tangible. Information appeared not as spoken language or handwriting, but as a pattern of holes and solid spaces. An abstract design had been given a physical form that a machine could respond to.

In this world of machinery and calculation, the English mathematician and inventor Charles Babbage began designing engines of his own.

His first great design was called the Difference Engine. It was intended to produce mathematical tables. Instead of asking people to carry out every stage of the calculation, Babbage imagined an arrangement of precisely coordinated mechanical parts that could complete the necessary operations.

The Difference Engine was ambitious, but Babbage’s thoughts soon moved toward an even broader design.

He called this later machine the Analytical Engine. Unlike the Difference Engine, which was intended to produce particular kinds of mathematical tables, the Analytical Engine was conceived as a more general machine. In principle, it could perform different operations according to the instructions supplied to it.

One part of the design was called the “store.” This was Babbage’s own term for the place where numbers and intermediate results would be held. In modern language, the store can be understood as a form of memory. Information could remain there until another part of the machine required it.

Babbage called that other part the “mill.” The word suggests a workshop in which something enters, is transformed, and then emerges in a new form. In the Analytical Engine, the mill would perform operations on numbers. Values would be drawn from the store, worked upon in the mill, and then returned or used in later stages of a calculation.

Instructions for the machine would be supplied through punched cards. Different arrangements of cards could guide the engine through different procedures. The design therefore brought together two important ideas: a mechanism able to carry out operations, and instructions that could be changed without rebuilding the entire mechanism.

The plans also included repeated sequences. If the same group of steps had to be performed many times, the engine could be directed to carry out those steps again. In later programming, such repetition would commonly be described as a loop.

Conditional operations appeared in the design as well. A conditional operation is an instruction that depends on whether a particular condition has been met. Its basic shape is gentle and familiar: if this is true, do one thing. Otherwise, do something else.

This kind of choice allows the path through a procedure to depend on a result produced along the way. Rather than moving through every instruction in one unchanging order, the machine can be directed toward one sequence or another, according to a defined condition.

Babbage revised the Analytical Engine repeatedly. Plans changed as his ideas developed. Diagrams and notations accumulated, offering different views of a machine that remained unfinished. The complete Analytical Engine was never built during his lifetime.

There is a quiet sense of incompletion around it. The engine existed in detailed thought, in mechanisms considered and reconsidered, and in pages covered with symbols and lines. It stood between a physical machine and a powerful idea, carefully conceived but never completed in its intended form.

The broader meaning of that idea became clearer through the work of Augusta Ada King, Countess of Lovelace, now usually known as Ada Lovelace.

Lovelace became deeply interested in Babbage’s Analytical Engine. In 1843, she translated an account of the proposed machine written by the Italian engineer Luigi Menabrea. To this translation, she added extensive notes of her own. Her notes were not merely brief comments. They examined how the engine might work, how its operations could be arranged, and what its abilities might mean.

In Note G, Lovelace presented a detailed procedure for calculating Bernoulli numbers on the proposed engine. Bernoulli numbers are a sequence used in parts of mathematics. Her procedure described how the engine could be guided through the calculation, one ordered stage after another.

This procedure is often described as the first published computer program.

That description is conventional, but the related title of “first programmer” is also debated. The Analytical Engine was never completed, so Lovelace’s procedure was not run on the full machine. Babbage had developed algorithms for his designs as well. Historical work does not always fit comfortably into a single title or belong neatly to only one person.

Lovelace’s achievement is best understood through what her notes actually did. She published a carefully organized procedure for a proposed general machine, and she explained its wider possibilities with unusual clarity.

Her deepest insight concerned what numbers might represent.

At first, a calculating engine appears to deal only with quantity. One number may represent a count, a measurement, or a value in a mathematical table. Yet numbers can also be used to stand for other things. A symbol does not have to remain tied to ordinary counting. It can represent an element within a larger formal system.

If relationships among symbols can be expressed through precise rules, then a general machine might operate upon those symbols. It could arrange, compare, combine, or transform their representations according to its instructions.

This is symbol manipulation: operating on marks or representations by following formal rules. The machine does not need to understand a symbol in the way a person understands it. It needs a clear procedure that determines what should happen to that symbol.

Lovelace allowed this thought room to expand. She considered the possibility that numbers might represent musical relationships. If those relationships could be expressed in a suitable formal way, she imagined that an engine might produce elaborate pieces of music.

This was not a detailed prediction of modern artificial intelligence. Lovelace was identifying a more foundational possibility. A sufficiently general machine might manipulate symbols representing more than numerical quantities. Its operations could apply to any structured relationships that people were able to express formally.

At the same time, she offered an important caution. The engine did not originate ideas by itself. It carried out operations that people knew how to specify. The machine could extend a human procedure through many intricate steps, but the instructions and formal relationships still came from human thought.

This balance is part of what makes her account so enduring. She saw that a machine’s reach might extend beyond ordinary arithmetic, yet she did not confuse formal operation with independent imagination. The engine could follow a method. It could transform representations. But it did not decide, on its own, what method ought to exist.

The old dream of lifelike machinery had now taken on a quieter and more abstract form. The central question was no longer only whether gears could imitate the movement of a hand, a musician, or a mechanical figure. It was whether a machine could preserve and carry out a method.

Mechanical calculators had shown that arithmetic could be embodied in wheels and gears. The Jacquard loom had shown that physical instructions could direct a complex sequence while the main mechanism remained unchanged. Babbage brought together a store, a mill, punched-card instructions, repeated sequences, and conditional choices in the design of a general engine. Lovelace then looked beyond numerical calculation and recognized that numbers might stand for a wider world of structured relationships.

Much of this remained on paper: inked diagrams, tables, notes, and plans for a machine never completed during Babbage’s lifetime. Yet an important threshold had been crossed in thought. A machine could be imagined not merely as a fixed arrangement of motion, but as a mechanism whose activity was shaped by changeable instructions.

The movement from mechanism to instruction was gradual. First, a wheel embodied a rule of arithmetic. Then a card encoded part of a woven pattern. At last, in the plans for the Analytical Engine, cards could direct a general sequence of operations, while the store held values and the mill transformed them.

The materials were still tactile: brass, paper, ink, and punched card. But the idea held within them was becoming abstract. Instructions could be separated from the mechanism that followed them. Symbols could stand for many kinds of things. A sequence could be repeated, and a choice could depend upon a condition.

For an engine to follow such instructions, however, every step had to be expressed with exceptional precision. Each operation, repetition, and choice needed a definite form. Nothing essential could be left to intuition.

And so the question began to shift once more. If machines could follow formal instructions, what kinds of instructions were possible? What could be represented by symbols? What could a precise procedure truly accomplish?

These questions would lead onward, away from the soft turning of brass wheels and toward the work of mathematicians and logicians, who would ask how far rules, symbols, and carefully defined procedures could carry the idea of computation.

Chapter 3: The Foundations of Computation

The designs of Charles Babbage and the notes of Ada Lovelace had revealed a far-reaching possibility. A machine might follow instructions, preserve intermediate results, repeat operations, and change its path according to a rule. Yet before such a machine could become truly general, its instructions needed unusual clarity.

Ordinary language is rich with suggestion, ambiguity, and several meanings at once. A mechanical procedure allows less freedom. Each action must be precise, each symbol must keep its agreed meaning, and every rule must identify what happens next. During the nineteenth and early twentieth centuries, mathematicians began to study this precision itself, establishing the logical and mathematical foundations of computation.

One early figure was the English mathematician George Boole. Decades before electronic computers, Boole developed an algebraic treatment of logic, showing that methods resembling algebra could represent statements and their logical relationships.

Consider two statements: the lamp is lit, and the door is open. Each may be treated as true or false, then joined or altered through defined operations. AND asks whether both are true. OR asks whether at least one is true. NOT reverses a statement: if the lamp is NOT lit, it is unlit.

From these simple elements, Boolean logic can build dependable chains of reasoning. Much later, engineers applied related principles to digital circuits. Electrical switches could represent contrasting conditions, such as on and off, while networks of switches performed logical operations. Boole did not design his algebra for electronics; the machines that used it did not yet exist. An idea developed for one purpose had found an unexpected home.

Boole’s work belonged to a wider effort to make reasoning rigorous. Gottlob Frege developed precise systems for logical relationships. Giuseppe Peano created influential mathematical notation. Bertrand Russell and Alfred North Whitehead explored how mathematics might be grounded in formal logic, while David Hilbert sought clear foundations and reliable methods for resolving mathematical questions.

Formal logic is reasoning expressed through defined symbols and rules. A formal system identifies its assumptions, permitted symbols, and exact steps from one statement to another. This drive toward precision raised a natural question: if reasoning could be written as exact steps, could every logical question be settled by the right procedure?

Hilbert brought this into focus through the decision problem, which asked whether a mechanical procedure could determine the validity of every statement in a formal logical system. Here, mechanical meant a procedure exact enough to require no inspiration or guesswork; each step followed a definite rule.

Kurt Gödel’s incompleteness results had already revealed unexpected boundaries: under particular conditions, sufficiently powerful formal systems contain true mathematical statements that cannot be proved from within those systems. His work concerned proof, not computation; the subjects are related but not interchangeable.

In the 1930s, Alonzo Church and Alan Turing approached mechanical procedure through different mathematical methods. Their work helped establish computability: the study of what algorithms can and cannot calculate. An algorithm is a definite sequence of ordered steps. Unlike an everyday recipe that may say “season to taste,” a formal algorithm specifies exactly what comes next.

In 1936, the young British mathematician Alan Turing described an abstract machine for examining such procedures. It was not a blueprint for a physical computer, but a mathematical model reduced to essential parts.

Imagine a potentially unlimited tape divided into squares, each able to hold a symbol. A reading and writing head inspects one square at a time. It can write, erase, or replace a symbol, then move one square left or right. The machine also has a finite set of internal states. Its current state and the symbol being read determine which exact rule it follows.

Read, write, move, repeat.

It reads a square, may write a symbol, moves, and repeats, one precise step at a time.

Despite its spare design, this machine could represent any procedure reduced to explicit steps. Different rules made it behave like different calculating machines. Turing then described a universal machine, which could receive a description of another machine with its data and imitate its operation. In modern language, one general machine could run many programs.

This clarified the distinction between a physical machine and its instructions. The same general device could perform arithmetic, arrange symbols, or carry out another well-defined task when given the appropriate description and data. The idea later echoed through stored-program computers, though Turing’s machine was not an engineering plan. It was a way to define mechanical procedure and examine what it could do.

Just as importantly, Turing’s work revealed limits. Some problems have no general algorithm that always gives a correct answer; they are undecidable. This does not mean merely difficult or awaiting a clever solution. It means no algorithm can solve every possible case correctly.

The best-known example asks whether an arbitrary program will eventually stop or run forever. For some programs, inspection makes the answer obvious; for others, it does not. Yet no single general algorithm can correctly decide, for every possible program and input, whether it will halt.

This boundary is not caused by inadequate imagination or machinery. Faster processors and larger memories cannot remove it; the limitation belongs to computation’s logical structure. Discovering what cannot be computed was therefore clarifying rather than discouraging. A map is more useful when it marks both open roads and genuine boundaries.

The Church–Turing thesis proposes that anything calculable by an effective mechanical procedure can be computed within these formal models. It concerns the meaning and reach of effective calculation; it should not be enlarged into the claim that every process in the universe is simply a computer program.

By defining computation carefully, Church, Turing, and their contemporaries made it possible to reason about algorithms as a general class. They could ask not only whether one calculation was possible, but whether any exact procedure could perform a given task. Sometimes the answer was no. This did not diminish computation’s power; it gave that power a clearer shape.

Turing’s questions did not end with tapes and abstract rules. In 1950, he published “Computing Machinery and Intelligence,” opening with the question, “Can machines think?” He observed that machine and think were difficult to define in terms everyone would accept. Debate could become trapped in definitions before it began.

Instead, he proposed an observable conversational exercise called the imitation game. In a familiar version, a human evaluator communicates by text with unseen participants and tries to distinguish a machine from a person. Appearance and voice are excluded, so judgment rests on written replies.

The exercise became widely known as the Turing test, but that phrase can oversimplify his argument. His paper was not merely a proposed contest, nor did it claim that one brief performance could settle every question about intelligence or mind.

Turing considered many objections to machine intelligence, as well as learning machines and future possibilities. Rather than trying to construct a complete artificial adult mind, he suggested beginning with something like a child machine and educating it. Some abilities might be acquired gradually through examples, correction, and experience.

He also considered uncertainty, mistakes, originality, and the limits of conclusions drawn from behavior. Success in an imitation game would show that, under particular conditions, a machine could produce conversation difficult to distinguish from human conversation. It would not by itself prove consciousness, inner experience, understanding, or feeling.

Those are different questions, difficult even when asked about living beings. A system may imitate conversation without possessing a mind in the human sense. Observable performance gives evidence about ability, but does not reveal what, if anything, is experienced within.

Still, Turing made an important shift. Instead of awaiting agreement about thought, researchers could study observable abilities. Could a machine use language, learn from correction, solve a problem, adapt its behavior, or respond appropriately to context? These narrower questions could be examined more carefully.

Across the work of Boole, Frege, Peano, Russell, Whitehead, Hilbert, Gödel, Church, Turing, and many others, an old dream gained precise foundations. Logic could be represented through symbols. Procedures could be studied mathematically. A universal machine could imitate many specialized procedures, and computation possessed both remarkable reach and definite limits.

The path was not a simple march toward the modern computer. These mathematicians pursued different aims, worked in different traditions, and answered different questions. Together, however, their ideas created a language in which general computation could be imagined clearly.

For the moment, much of the machine remained on paper. Its tape stretched quietly through the imagination, square after square, while its reading and writing head passed patiently along the row.

Read, write, move, repeat.

With symbols, states, and rules, the abstract machine had defined mechanical calculation, shown how one general machine might imitate many others, and marked boundaries that speed could not erase.

The next step would be physical. Engineers would need to build electronic machines fast and reliable enough to put these carefully formed ideas to work. In their circuits, precise operations would move from mathematical symbols toward working machinery, carrying the foundations of logic and computability into a new technological age of practical, general-purpose electronic computation itself.

Chapter 4: War, Electronics, and the First Digital Computers

The abstract machine imagined by Alan Turing needed only a tape, a set of simple rules, and as much time as a problem required. A physical machine was less accommodating. It needed materials, electrical power, reliable components, skilled hands, and a clear purpose.

During the Second World War, governments found urgent purposes for faster calculation. Resources flowed into communications, codebreaking, ballistics, radar, and electronic engineering. Problems that might once have developed slowly through universities and laboratories were pursued under intense pressure.

This acceleration cannot be separated from the suffering of the period. New machines and mathematical methods emerged amid destruction, displacement, and loss. Wartime demands brought people and institutions together, often under strict secrecy, and asked them to solve difficult problems quickly. Technical progress did accelerate, but it did so within a period of profound human cost and ethical complexity.

One important center of wartime codebreaking was Bletchley Park in Britain. Its work depended on a large and varied community. Mathematicians worked alongside linguists, engineers, clerks, operators, and many others. Messages had to be received and organized. Machines had to be designed, built, maintained, and operated. Results had to be checked and interpreted. Much of this work remained secret for years, and many individual contributions were not widely recognized.

A cipher is a method of transforming a message so that it cannot be read without the right key or procedure. A sentence may be rearranged or altered according to a system, producing a sequence that appears meaningless. A recipient who knows the system and possesses the correct key can reverse the transformation and recover the message.

One well-known cipher machine was Enigma. Alan Turing played a major role in developing methods and machinery used against it. Yet Turing did not work alone. His ideas formed part of a broad, coordinated effort involving colleagues, equipment, established routines, and carefully gathered information. Neither the achievement nor the machines involved should be credited to a single person.

The same is true of another remarkable wartime machine, Colossus. Designed principally by the engineer Tommy Flowers, Colossus first became operational in the 1940s. It was built to assist in breaking the Lorenz cipher, which protected certain German communications.

Colossus used vacuum tubes. A vacuum tube was an electronic component used for switching and amplification before transistors became available. Electrical signals could be directed through these tubes with great speed. A tube might help represent one of two distinct conditions, such as on and off.

This use of clearly separated states is part of what makes a system digital. Digital information is represented through discrete states rather than through a continuously changing range. A simple comparison might be a row of lamps. Each lamp can be either dark or illuminated. By arranging many such choices together, a machine can represent and process more complicated information.

Vacuum tubes made rapid electronic operation possible, but they also brought practical difficulties. They consumed substantial electricity, produced heat, and could fail. A large machine containing many tubes required careful engineering and regular attention. Electronic speed did not mean effortless operation.

Colossus was electronic, and it was programmable in a limited sense. Its settings could be adjusted so that it carried out particular operations. However, it was not a general-purpose stored-program computer. It was a specialized codebreaking system, closely shaped around the work it had been built to perform.

Its operation depended continuously on people. Information had to be prepared and supplied to the machinery. Operators managed repeated runs, watched for faults, and checked the process. Engineers maintained the equipment. Other workers examined and interpreted the results.

Colossus could process patterns far more quickly than a person working by hand, but it did not understand the messages it helped to examine. It did not know their language, their context, or their human consequences. The machine carried out electronic operations. Meaning remained with the people around it.

Codebreaking provided valuable intelligence during the war, but estimates of precisely how much it shortened the conflict remain uncertain. Historical causes cannot easily be separated and measured. There is no second version of the past that can be run again without one particular contribution. Claims about an exact amount of time saved should therefore be treated cautiously.

Across the Atlantic, another large electronic machine was taking form. ENIAC, the Electronic Numerical Integrator and Computer, was completed in the United States in 1945. It was publicly announced in 1946.

ENIAC was an electronic, general-purpose digital computer. The phrase general-purpose meant that it could be arranged to perform many different kinds of numerical calculation, rather than being permanently limited to one narrow task. This flexibility was an important achievement, although changing from one task to another was still far from easy.

Early ENIAC reprogramming required extensive manual work. Cables had to be connected in carefully chosen arrangements, and switches had to be set correctly. Preparing a new calculation could involve studying the planned movement of numbers through the machine, dividing the problem into manageable steps, and checking every connection.

Among the first ENIAC programmers were Kay McNulty, Betty Jennings, Betty Snyder, Marlyn Wescoff, Fran Bilas, and Ruth Lichterman. They translated mathematical procedures into physical arrangements that ENIAC could follow. Because electronic programming was new, their work required patience, precision, and a deep understanding of both the problem and the machine.

For many years, the contribution of these programmers was underrecognized. Their names and responsibilities were often absent from popular accounts of ENIAC. Yet the machine’s operation depended upon their knowledge and labor. Their story is a quiet reminder that the history of technology is not only a sequence of devices. It is also a history of people whose work may be essential even when it is not immediately celebrated.

ENIAC filled a room. Its vacuum tubes consumed substantial electricity and produced considerable heat. Components were fragile, and maintenance was a continuing responsibility. The machine required teams of programmers, operators, engineers, and technicians.

The sight of such a system must have been impressive. Cabinets and panels stood in long rows. Cables joined one part to another. Switches waited to be set. Warm air gathered around equipment that had to be watched and maintained. The machine performed calculations with remarkable speed, yet that speed rested upon a wide foundation of human care.

Neither ENIAC nor Colossus independently understood the task it was given. ENIAC did not wonder why a calculation was needed. Colossus did not know that it was assisting with an encrypted message. Each followed arrangements created by people. Their speed extended human procedures, but it did not provide human comprehension.

This is one reason the phrase “the first computer” can lead to disagreement. The answer depends on which qualities are considered essential.

Does the word computer mean electronic? Must the machine be digital? Must it be programmable? Must it be general-purpose? Must its instructions be stored in memory?

Different machines were first according to different combinations of these features. Colossus was electronic, digital, and programmable in a limited sense, but specialized. ENIAC was electronic, digital, and general-purpose, but its early programming depended on extensive rewiring and switch setting. The history is therefore better understood as a cluster of developments than as one perfectly clear moment of invention.

An especially important development was the stored program. In a stored-program computer, instructions are held in electronic memory along with the data being processed. The machine does not need to depend entirely on a physical arrangement of cables and switches to define each new task. Instead, both instructions and data can be represented within the machine.

The idea was subtle, but powerful. Imagine a musical instrument that does not need to be rebuilt for each new melody. A new score can simply be placed before it. In a similar way, a stored-program computer can receive a different sequence of instructions and begin another kind of calculation.

This did not remove the need for human labor. Someone still had to understand the problem, prepare the instructions, enter them, check them, and examine the result. Yet it changed the nature of reprogramming. Gradually, changing a machine’s task could become less a matter of moving cables and more a matter of preparing symbolic commands.

In June 1948, a small experimental machine known as the Manchester Baby ran a stored program. Its successful operation demonstrated a principle that would become central to modern computing. Instructions could reside in electronic memory, ready to guide the machine without requiring the whole device to be reconstructed for each new task.

The transition was not immediate or tidy. Early electronic memories were limited. Designs varied from one project to another. Vacuum tubes remained fragile. Computers continued to fill rooms, consume large amounts of electricity, and produce considerable heat. They still demanded teams of operators and technicians.

Even the word automatic could be misleading. A machine might proceed through a sequence of operations without a person performing every calculation by hand. But people still selected the task, prepared the instructions, supplied the data, maintained the equipment, and judged whether the output was useful. The computer operated within a carefully constructed human setting.

The early digital computer was therefore both powerful and limited. It could calculate rapidly, repeat procedures without tiring, and manage patterns too extensive for convenient work by hand. At the same time, it possessed no independent understanding of what those patterns meant. It did not recognize the hopes, dangers, or purposes surrounding its work.

After the war, electronic computing began to move into universities, government laboratories, and other institutions. The machines remained rare, expensive, and demanding. Still, they were becoming more flexible. Ideas that had once existed mainly in mathematical papers could now be explored through working electronic equipment.

This new hardware did not emerge from one person, one machine, or one moment. It grew through the combined labor of mathematicians, linguists, engineers, clerks, operators, programmers, technicians, and many others. Some became well known. Many remained in the background. Together, they transformed abstract procedures into physical processes of switches, signals, memory, and motion.

By the late 1940s, a new possibility had come into view. Machines could store instructions and perform many kinds of calculation. They remained enormous and fragile, and they understood nothing of their tasks. Yet they offered researchers a practical tool with which to explore questions that had once seemed purely philosophical.

Could computation be used for more than arithmetic? Could a machine follow a chain of reasoning, recognize a pattern, work with language, or learn from experience?

The physical foundations were now beginning to take shape. In the years ahead, researchers would gather around these questions with renewed confidence. And before long, the old ambition to imitate aspects of intelligence would receive a new and enduring name.

Chapter 5: When Artificial Intelligence Received Its Name

By the middle of the 1950s, electronic computers had begun to change the character of an old question.

For centuries, people had imagined artificial beings that could move, calculate, speak, or reason. Philosophers had examined the nature of thought. Mathematicians had developed formal systems of logic. Engineers had built mechanical automata. More recently, wartime laboratories had shown that electronic machines could follow long sequences of instructions at remarkable speed.

Now, these separate streams of thought were beginning to meet.

During the summer of 1956, a small and changing group of researchers gathered at Dartmouth College in Hanover, New Hampshire. The setting was quieter than its later reputation might suggest. We might picture papers spread across tables, diagrams taking shape in chalk, and green summer leaves visible through open windows. Conversations unfolded in seminar rooms and offices. Ideas passed between people who arrived at different times, stayed for different lengths of time, and often followed their own lines of inquiry.

It was not a tightly organized conference in the modern sense. Records of attendance and daily activity are incomplete. Participants came and went throughout the summer, and later recollections did not always agree on every detail.

Yet the gathering became a landmark. It helped bring several existing traditions together beneath a new name: artificial intelligence.

That name had appeared the year before. In 1955, John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon prepared a proposal for the Dartmouth Summer Research Project on Artificial Intelligence. The term artificial intelligence is generally credited to McCarthy.

So, the phrase was not invented spontaneously during a conversation at Dartmouth in 1956. It had already been written into the proposal that helped make the summer project possible.

The name was broad enough to create a shared territory. Artificial intelligence could be understood as a field concerned with machines performing tasks associated with intelligence. Those tasks might include learning, reasoning, understanding language, recognizing patterns, solving problems, or selecting an action.

But the apparent simplicity of the phrase concealed a difficult question.

What, exactly, was intelligence?

To one researcher, intelligence might mean the ability to follow a valid chain of logic. To another, it might mean learning from experience. Someone else might emphasize language, creativity, perception, or the ability to act successfully in an uncertain setting.

The new name brought these interests into one field, but it did not resolve their differences. Intelligence already meant different things to different people, and the researchers gathered around artificial intelligence did not begin with a single definition.

The organizers expressed their hopes plainly. They suggested that aspects of learning and intelligence might be described so precisely that machines could simulate them.

A simulation is an artificial model that reproduces selected behavior of another system. It does not need to reproduce every detail. It attempts to capture certain patterns or processes that are important for a particular purpose.

In this sense, a computer program might reproduce one part of intelligent behavior without being identical to a human mind. It might follow a chain of reasoning, recognize a limited pattern, or choose among several possibilities. Each of these would represent only a selected aspect of intelligence, modeled within the machine.

This possibility seemed newly practical because electronic computers now existed. These computers were large, costly, and difficult to operate. Yet they could store instructions and manipulate symbols at speeds no unaided person could match.

They gave researchers something earlier philosophers had lacked: a working instrument on which ideas about reasoning, learning, and problem-solving could be tested.

Several neighboring fields contributed to this atmosphere of possibility.

One was information theory, developed especially through the work of Claude Shannon. Information theory is the mathematical study of information, communication, and uncertainty. It examines how messages can be represented and transmitted, and how information can be recovered even when a communication channel contains noise.

Information theory did not offer a complete account of meaning, understanding, or thought. Still, it provided precise mathematical tools for discussing signals, communication, and uncertainty. These tools encouraged researchers to imagine that processes once described only in broad philosophical language might be examined in more exact ways.

Another important influence was cybernetics.

Cybernetics is the study of communication, control, and feedback in machines and living systems. Feedback occurs when the result of an action is used to guide what happens next.

A thermostat provides a simple example. It measures the temperature, compares that measurement with a desired setting, and adjusts the heating in response. The new temperature then becomes part of the next adjustment. In this quiet cycle, the outcome of one action helps shape the following action.

More complex forms of feedback could guide machines, describe regulation in living systems, or suggest ways behavior might adapt over time. Cybernetics therefore offered another vocabulary for thinking about purposeful action without assuming that every step had to be directed from outside.

Mathematical logic supplied a further foundation. Logic showed how statements and rules could be expressed formally. With suitable programming, computers could manipulate these expressions step by step.

Early programs were beginning to perform activities associated with reasoning. A machine might search through possible deductions, test alternatives, and arrive at a conclusion by following explicit rules. Such a program did not possess the broad understanding of a person. Yet within a carefully described problem, it could carry out a sequence that resembled one part of reasoning.

The Dartmouth project brought together people from these overlapping worlds.

John McCarthy was central to the proposal and to the naming of the field. Marvin Minsky pursued questions about learning machines, neural models, and the mind. Claude Shannon brought the influence of information theory, along with an interest in how machines could carry out tasks such as playing games. Nathaniel Rochester brought experience with electronic computers.

Other figures associated with the gathering included Allen Newell and Herbert Simon, whose work on computer problem-solving was producing striking early results. Oliver Selfridge explored ideas connected with pattern recognition. Ray Solomonoff considered mathematical approaches to prediction and machine learning.

Still others formed part of the wider circle. Some were present for only a portion of the summer. Others were connected through shared problems and ongoing conversations. Because the records are incomplete, the Dartmouth gathering is better imagined as a loose summer project than as one room filled with the same people from beginning to end.

The participants did not arrive with one agreed method. Nor did they leave with a single theory of intelligence.

Some believed intelligent behavior could be built from symbols and formal rules. In this approach, a computer might represent facts, goals, and logical relationships, then search through them to find a solution.

Others were interested in models inspired loosely by networks of neurons. Some focused on learning and asked whether a machine could improve through examples rather than receiving every rule in advance. Language, creativity, abstraction, and perception also appeared among the proposed areas of study.

These approaches could sometimes support one another. At other times, they competed.

A system designed to prove a theorem might look very different from one intended to recognize a visual pattern. A program that succeeded in a carefully defined puzzle might provide little help with an ordinary conversation. Even the word learning could refer to several different processes.

For this reason, Dartmouth is best understood not as the instant invention of artificial intelligence, but as the naming and consolidation of a field.

Many of the essential ideas had come earlier. Mathematical logic, theories of computation, neural models, feedback systems, and electronic computers all possessed histories of their own. Cybernetics remained an influential movement with its own language and priorities. British research into machine intelligence, shaped in part by Alan Turing and others, also continued outside the Dartmouth circle.

The field called artificial intelligence grew from all these currents. Its boundaries were never perfectly fixed.

Still, names can matter.

A name allows researchers to recognize common interests, even when they disagree about methods. It makes it easier to organize meetings, form laboratories, seek support, and describe a shared ambition. Once artificial intelligence had a name, it could become an institutional field as well as a collection of questions.

University groups could gather around it. Students could enter it. Governments and companies could identify and support projects connected with it. Researchers working on logic, learning, language, and perception could see themselves as participants in a wider undertaking.

The optimism surrounding the young field was sincere.

Researchers had watched computers develop from unusual experimental devices into programmable machines capable of carrying out complex procedures. Problems once thought to require uniquely human reasoning were beginning to yield, at least in limited settings, to computation.

Logic could be encoded in programs. Machines could examine alternatives. Early systems could perform tasks that seemed to contain small but recognizable pieces of reasoning. With these examples before them, it was reasonable to wonder how far the progress might continue.

Yet some early predictions greatly underestimated the difficulty of ordinary, real-world intelligence.

Human beings recognize familiar faces under changing light. They understand unfinished sentences. They move through crowded rooms, respond to uncertain situations, and draw upon years of experience without consciously examining every step.

These abilities often feel effortless because they are deeply woven into perception, memory, language, bodies, and social life. Describing them precisely enough for a machine proved far harder than early demonstrations suggested.

During that summer at Dartmouth, however, the depth of these difficulties was only partly visible. The immediate challenge was to turn broad hopes into working programs.

Researchers began with worlds that could be clearly described. They chose logical proofs, board games, puzzles, and limited sets of possible actions. Inside such boundaries, a computer could examine alternatives in an orderly way. It could follow rules, compare choices, and search through branching possibilities.

The summer gathering had given the endeavor a name, but a name alone could not demonstrate what machines might do. That would require programs, experiments, failures, revisions, and patient work.

And so, the history moved gently onward, from proposals and seminar-room conversations to small, concrete demonstrations. The new field would first make itself visible through systems that proved theorems, played games, and searched through paths of possibility.

They were narrow programs, built for carefully defined settings. Yet they gave tangible form to a broad and enduring hope: that at least some portions of reasoning might be described clearly, expressed as instructions, and carried out in code.

Chapter 6: Programs That Reasoned and Played

After the conversations at Dartmouth, the broad idea of artificial intelligence began to take shape in working programs. Researchers did not begin with the whole, untidy world of human experience. Instead, they chose smaller domains whose boundaries could be drawn clearly. A theorem written in symbolic logic. A board game with fixed rules. A puzzle with a known goal.

These limited worlds offered something valuable. The allowed actions could be defined. Results could be checked. Experiments could be repeated. Within these carefully arranged spaces, researchers could ask a modest but important question. Could a machine carry out some of the structured operations that people call reasoning?

In 1955 and 1956, Allen Newell, Herbert Simon, and Cliff Shaw developed a program called the Logic Theorist. It was designed to prove theorems in symbolic logic. A theorem is a statement that can be shown to follow from accepted starting points by using agreed rules. Symbolic logic expresses statements and rules in a precise form. In this way, some kinds of reasoning can be treated as careful operations performed upon symbols.

The Logic Theorist began with known statements and permitted transformations. It then searched for a sequence of steps leading toward a desired conclusion. In a very simple case, this might seem straightforward. Yet even a small collection of rules can produce many possible next steps. Each choice leads to further choices, and those choices lead to still more.

It can help to imagine a tree slowly spreading outward. The trunk is the starting problem. Each branch represents a possible step. From every branch, smaller branches may grow, representing the choices that become available next. Those branches divide again. Before long, the tree can become so broad and dense that inspecting every twig would take far too much time.

This branching map of possibilities is called a search tree. It does not need to be drawn on paper. It is simply a way of describing how a program considers possible actions and the outcomes that may follow from them. Search trees became one of the central ideas in early artificial intelligence because they turned choice into something that could be explored in an orderly fashion.

The number of possible branches, however, could grow remarkably quickly. A program with limited time could not always inspect every path. It needed some way to decide which branches appeared worth exploring first.

The Logic Theorist therefore used heuristics to favor promising paths. A heuristic is a practical rule of thumb. It can save time by directing attention toward choices that seem likely to help, although it does not guarantee the shortest route or the best possible answer.

People use similar shortcuts in ordinary life. When looking for a misplaced book, someone might first check beside the chair where the book was last read. They would probably not begin by examining every shelf, cupboard, and drawer in a fixed order. The first guess may be wrong, but it is reasonable. It uses a small clue to make the search more manageable.

In much the same way, a program can use features of a problem to choose which branches to examine first. Instead of treating every possible path as equally useful, it can spend more of its limited effort on paths that appear promising.

Using search and heuristics, the Logic Theorist proved theorems in symbolic logic. Its achievements were striking because they gave visible form to an old idea. At least some portions of reasoning could be represented as a sequence of operations and carried out by a machine.

Still, the word reasoning needs gentle care here. In this context, reasoning means structured operations performed over representations. The Logic Theorist worked with symbols, goals, and rules. It did not understand mathematics in the way a human mathematician does. It did not bring a lifetime of experience to a problem, wonder why a result mattered, or feel curiosity about an unexpected pattern. Its world consisted of the formal materials its designers had provided.

This broad approach became known as symbolic AI. In symbolic AI, the relevant parts of a problem are represented explicitly. A statement may be stored as a fact. A desired result may be represented as a goal. Rules describe which transformations are permitted. The system proceeds by operating on these clearly defined elements.

One advantage of this explicit structure was that some of the program’s reasoning steps could be inspected. Researchers could follow a path and observe which rule had been applied. They could see how one symbolic state led to another. Within a carefully defined world, the process could have a satisfying clarity.

Yet the same clarity carried a burden. Someone had to decide in advance which symbols to use, which facts mattered, and how the relevant world should be represented. In formal logic, these decisions could be manageable because the starting statements and permitted rules were already defined. In everyday life, the boundaries were much less certain.

Ordinary situations contain details that may be incomplete, changing, or difficult to express. What matters in one moment may be unimportant in another. Before a symbolic system can reason about a world, researchers must first decide how that world will appear inside the machine. This problem of representation would remain one of the enduring challenges of artificial intelligence.

Newell and Simon later worked on another ambitious project called the General Problem Solver. Its name reflected the hope that methods used in one structured task might reveal something broader about problem solving itself. It was an attempt to describe problem solving in a more general way, rather than creating an entirely separate procedure for every individual problem.

The aim was broad, but generality proved difficult. A method that works neatly for one symbolic puzzle may not contain the knowledge needed for a very different task. Rearranging symbols according to explicit rules does not automatically provide an understanding of language, perception, or ordinary life. The General Problem Solver was historically important, but it did not become a universal solver.

Board games offered another orderly setting in which to explore machine reasoning. They had clear rules, defined legal actions, measurable outcomes, and repeatable experiments. A program could play again and again under the same basic conditions. Researchers could compare different methods and observe whether performance changed.

A game board was, in a sense, a small and complete world. The pieces had defined positions. The available actions were limited by rules. A game ended with an outcome that could be measured. These qualities made games useful laboratories for studying search, choice, and adaptation.

Even so, the number of possibilities could be immense. In 1950, Claude Shannon described the problem in influential work on computer chess. A machine could not inspect every possible continuation indefinitely. After one possible move came several possible replies. Each reply opened further continuations, and each of those led to still more.

Once again, the image of the branching tree was useful. Near the trunk, the number of branches might appear manageable. A little farther out, they divide repeatedly. Very soon, there are more paths than the machine can examine within a practical amount of time.

A chess program therefore needed to search selectively. It had to choose some continuations for closer attention while leaving others less deeply explored. It also needed to stop searching at some limited point. When it could not follow a path all the way to the end of the game, it required a way to estimate the position it had reached.

This estimate was provided by an evaluation function. An evaluation function is a method for judging how favorable a position or state may be. It does not provide complete certainty. Rather, it offers a useful estimate that allows the program to compare possibilities when a full search is impossible.

Search and evaluation worked together. Search opened a view of possible futures. Heuristics helped direct attention toward promising branches. The evaluation function offered a provisional judgment when the program could not continue to the final outcome. The machine did not have to inspect everything. It needed a practical way to choose among more possibilities than it could fully explore.

Checkers became another important field of study. Arthur Samuel developed a checkers program that became an influential early example of machine learning. His program played games, assessed positions, and adjusted aspects of its play through experience. It could also gain experience through self-play, by playing against another version of itself.

Samuel’s work suggested that not every useful preference had to remain exactly as a programmer first supplied it. Some numerical settings could be altered in response to experience. The program could adjust the importance it assigned to selected features of a position. Through repeated games, its internal preferences could change.

This was learning in a real technical sense, but it was learning within a narrow and carefully defined setting. Samuel’s program adjusted selected numerical preferences while playing checkers. It was not a general learner. It could not step beyond the board and naturally apply its experience to language, perception, or the varied circumstances of daily life.

That limitation did not make the achievement unimportant. Instead, it helped clarify what machines could do when a task offered stable rules, visible choices, repeatable trials, and clear feedback. Games allowed researchers to investigate planning and learning without first having to describe the whole surrounding world.

There was something quietly satisfying about these limited domains. Within them, uncertainty could be reduced. Rules could be written down. Choices could be arranged into branches. Outcomes could be compared. A small world could be represented clearly enough for a machine to move through it.

At the same time, a game board was not everyday life. Success in chess or checkers did not automatically transfer to conversation, vision, or ordinary action. Beyond the board, the available choices may not be completely known. Goals can be uncertain. Important information may be missing. The meaning of a situation may depend on context that is difficult to state as a fixed rule.

The early reasoning and game-playing programs therefore offered both encouragement and instruction. They showed that machines could search through possibilities, apply explicit rules, estimate uncertain positions, and sometimes improve selected aspects of their performance. They also revealed how strongly these achievements depended on the way a problem had been defined.

A carefully represented world could be made understandable to a machine. Yet deciding what to represent, and how to represent it, remained a human task. The symbols, goals, rules, and evaluation methods did not appear on their own. They were chosen and shaped by researchers.

One path toward artificial intelligence emphasized these explicit structures. It described facts, goals, symbols, and rules, then guided the machine through a branching space of possibilities. Alongside it, another path was slowly developing. This second path asked whether a machine might improve from examples, changing its own internal settings through experience.

Chapter 7: Teaching Machines to Learn

The earliest programs for reasoning and game playing often depended on carefully prepared instructions. Human researchers described the rules, represented the possible choices, and designed methods for searching among them. Within a well-defined world, this approach could produce striking results.

Yet it also raised a quiet and persistent question. Did every useful rule have to be written by a person, or could a machine gradually change its own behavior through experience?

This question opened another path through the history of artificial intelligence. It drew inspiration from the nervous system, though always in a greatly simplified form. The brain contains vast networks of biological neurons, living cells connected through intricate chemical and electrical processes. Early researchers did not try to reproduce all this complexity. Instead, they asked whether a few abstract features of neurons could be expressed through mathematics.

In 1943, Warren McCulloch and Walter Pitts published a mathematical model of simplified artificial neurons. Their artificial units received signals and produced one of two possible outputs. A unit was either on or off, rather like a small lamp that could be lit or dark.

This was not a complete description of a living brain cell. Biological neurons behave in far richer ways. They are influenced by timing, chemistry, physical structure, and many kinds of interaction. The McCulloch–Pitts model set nearly all of that detail aside.

Its value came from this very act of simplification. By reducing a neuron to an idealized mathematical unit, McCulloch and Pitts could study networks of such units using the tools of formal logic.

A unit might receive several signals. Depending on which signals were present, it would either activate or remain inactive. Arranged carefully, these units could perform logical operations. One arrangement might activate only when two incoming signals were both present. Another might activate when either signal was present.

In this way, networks of idealized neurons created a connection between the language of nervous activity and the language of logic. Complicated patterns of decision could, in principle, emerge from many simple units working together.

The idea was both modest and profound. Yet these early units did not provide a practical, general method for learning. Their connections and patterns of operation were still largely designed in advance. The network could process its signals, but it did not necessarily adjust itself after experience.

A further source of inspiration appeared in 1949, when the Canadian psychologist Donald Hebb published *The Organization of Behavior*. Hebb was interested in how biological nervous systems might change through experience. He proposed that connections between biological neurons could become stronger when those neurons were repeatedly active together in a coordinated way.

This principle is sometimes summarized by the familiar phrase, “Cells that fire together wire together.” Those were not Hebb’s exact words, but the phrase captures the broad idea. If the activity of one cell repeatedly helps another cell become active, the connection between them may grow stronger. Over time, repeated experiences could therefore shape a network of connections.

Later machine-learning researchers drew inspiration from Hebb’s principle. It offered a way to imagine memory and adaptation as changes distributed across a network, rather than as separate instructions stored in one special place.

Still, Hebb did not invent the training methods used by modern neural networks. His work offered a theory about biological learning. The path from that theory to practical machine learning required many additional ideas.

By the late 1950s, one of the most visible attempts to build a learning machine was led by the American psychologist Frank Rosenblatt. Rosenblatt developed a model called the perceptron. Its name suggested a system concerned with perception, receiving signals from the world and learning to sort them into categories.

To understand a perceptron, it helps to consider one idea at a time.

First, there are inputs. An input is simply a piece of information supplied to the model. In a basic visual experiment, the inputs might come from a grid of light sensors. Each sensor indicates something about a small region within the image before it.

Next, each input is given a weight. A weight is a number that controls how much influence an input has on the final result. A large positive weight allows a signal to push strongly toward activation. A negative weight pushes in the other direction. A weight close to zero gives that input relatively little influence.

The perceptron combines all these weighted inputs into a total. It then compares the total with a threshold. The threshold is the level the combined signal must cross before the unit activates. If the total is high enough, the perceptron produces one output. If the total remains below the threshold, it produces another.

The individual steps are simple. Signals enter. Their influence is adjusted by weights. The weighted values are combined. The result is compared with a threshold. Then an output appears.

The important question is how the machine finds useful weights.

Imagine a quiet laboratory panel covered by a grid of light-sensitive points. A card is placed before it. One card carries a simple pattern, perhaps a shape leaning toward one side. Another card carries a different arrangement of light and dark.

At first, the machine’s weights may not separate the cards correctly. A sensor near an upper corner may have too much influence, while a more useful sensor near the center may have too little. The machine looks at the pattern through its grid of sensors, combines the signals, and gives an answer.

Then it receives feedback.

If the answer is wrong, the weights are adjusted. Some influences become stronger, while others become weaker. Another card is presented, and the machine tries again. The cycle continues gently and repeatedly. There is an input, an answer, a correction, and then another input.

This repeated adjustment is what training means in machine learning. The machine is not given a spoken explanation of the pattern. Instead, its numerical parameters are changed according to examples or feedback.

A parameter is simply an internal value that can be adjusted. In a perceptron, the weights are parameters. The threshold also helps determine when an output will activate. By changing these internal values, the system changes the way it responds to future inputs.

Learning, in this technical sense, does not imply awareness. The machine does not feel surprise when it is wrong, satisfaction when it is right, or curiosity about the cards. It does not comprehend light and shape as a person does. The word *learning* describes a measurable change in performance produced by adjustments made from data or feedback. It does not imply emotion, understanding, or conscious experience.

Rosenblatt’s ideas were explored in a physical experimental system called the Mark I Perceptron. It included an array of light sensors and components used to create and adjust connections. From the outside, it did not resemble a biological brain. It looked like laboratory equipment, with wires, switches, panels, and a field of sensors facing the patterns it was meant to classify.

Within carefully controlled experiments, the Mark I Perceptron could learn certain visual classifications. This was genuinely important. A researcher did not have to specify every detailed rule for recognizing a pattern. Instead, the perceptron could adjust its weights using examples and gradually form a useful dividing boundary between categories.

Suppose each card can be described by a collection of measurements. The perceptron attempts to separate those measurements into two groups. For some arrangements, this is like drawing a straight line between two clusters of marks on a page. The marks on one side belong to one category, and those on the other side belong to another.

When a simple separating boundary exists, a single-layer perceptron can learn to distinguish the categories. Its repeated corrections gradually move the boundary into a more useful position.

This achievement attracted considerable attention. Some early press reports suggested that perceptrons might soon see in a humanlike way, recognize people, or even reproduce themselves. Such descriptions went far beyond what the experiments had demonstrated.

The actual systems were much narrower. They could classify certain limited patterns under controlled conditions. They were not general visual intelligences, and they were not mechanical people.

Even so, the early enthusiasm is understandable. Computers were still new and unfamiliar. A machine that changed its own connections after receiving examples seemed to offer a new relationship between instruction and behavior. It suggested that not every decision had to be anticipated and described in advance.

But the limitations were real. A single-layer perceptron can solve only certain kinds of separable patterns. If two categories cannot be divided by a simple boundary, this basic model cannot learn the distinction, no matter how many examples it receives.

Some more complex patterns require additional structure. They may need more layers of processing, allowing a system to form intermediate distinctions before reaching a final answer. Other patterns may require different methods altogether.

This limitation did not make the perceptron useless. Instead, it revealed something important about machine learning. Learning depends both on experience and on structure. Examples alone are not enough if the model has no suitable way to represent the pattern hidden within them.

A map needs appropriate symbols if it is to show rivers, roads, or mountains. In a similar way, a learning system needs an architecture capable of expressing the relationship it has been asked to discover. Training can adjust what the system already has the capacity to represent, but training cannot make every possible structure emerge from an unsuitable model.

The comparison with biology also required care. Artificial neural networks were, and remain, only loosely inspired by brains. Their units are simplified mathematical operations, not living cells. Their weights are numbers, not complete biological synapses. Their training procedures are engineered methods of numerical adjustment.

The language of neurons and connections is useful because it points toward a source of inspiration. Yet it should not suggest that a miniature brain is quietly living inside a machine. The abstraction captures only a few selected ideas while leaving most biological complexity aside.

Even with these cautions, early neural-network research introduced a lasting change in perspective. Intelligence might be approached not only by writing facts and rules, but also by constructing systems that adjusted themselves in response to examples.

The distinction was never absolute. Learning systems were still designed by people. Researchers chose their structures, goals, training examples, and methods of correction. The environments were prepared, and the desired categories were defined. Yet some of the system’s detailed behavior could now emerge through repeated adjustment rather than through direct instruction alone.

The perceptron therefore rested between promise and constraint. Its successes were narrow, but its central lesson endured. A machine could improve at a task by changing the numerical influence of its inputs. Across many examples, small corrections could accumulate into a useful pattern of response.

Recognizing a simple card in a controlled experiment, however, was only one small part of what people called intelligence. Beyond the laboratory waited sentences whose meanings changed with context, rooms filled with ordinary objects, conversations shaped by memory, and bodies moving through an uncertain world.

Researchers would next try to bring machines into these richer settings. They would explore language, objects, conversation, vision, and motion, discovering once again how different the open world could be from the careful order of an experiment.

Chapter 8: Language, Robots, and Early Ambition

Beyond the laboratory waited sentences whose meanings shifted with context, rooms filled with ordinary objects, conversations shaped by memory, and bodies moving through an uncertain world. In the 1950s and 1960s, researchers began guiding computers into these more complicated territories. They asked machines to translate languages, take part in conversations, answer questions, and move through physical space.

Each effort produced moments that seemed to offer a glimpse of broad intelligence. Yet each also revealed how much hidden knowledge people rely upon during even the most ordinary exchange. What appeared effortless to a person could demand layers of information that a machine did not possess.

Language was among the earliest ambitions. During the Cold War, governments and institutions wanted rapid translation, especially between Russian and English. Human translators brought deep knowledge to their work, but careful translation took time. A computer, it was hoped, might process documents quickly and make their information more readily available.

In 1954, Georgetown University and IBM presented a public demonstration of machine translation. A computer translated a small collection of Russian sentences into English. To many observers, the event suggested that automated translation might soon become routine. Words entered in one language and emerged in another, apparently transformed by the machine.

The demonstration generated considerable enthusiasm. Yet it had been carefully prepared. Its vocabulary was small, its grammar was restricted, and its sentences had been selected in advance. It did not show unrestricted translation of unfamiliar writing. The difference was rather like following a marked path through a garden, instead of finding a route through an unknown forest.

Translation proved difficult because words do not carry meaning alone. Their meanings depend on neighboring words, the subject under discussion, shared customs, and knowledge of the world. A single word may refer to several different things. An idiom may mean something quite different from the literal meanings of its parts. Pronouns may point back to people or objects mentioned earlier. Even a simple sentence may require an understanding of what is physically possible, or what is socially expected.

People resolve much of this ambiguity without noticing. They draw upon years of experience and a vast store of background knowledge. Early computers had neither. They could apply dictionaries and grammatical rules, but they struggled whenever meaning depended on context.

The enthusiasm inspired by the Georgetown–IBM demonstration therefore met a slower and more complicated reality. Translation was not simply the replacement of one word with another. It involved interpretation. To choose the right words, a translator often had to recognize the larger situation in which those words were being used.

A different experiment in language appeared at the Massachusetts Institute of Technology during the 1960s. There, Joseph Weizenbaum created a program called ELIZA. In a paper published in 1966, he described ELIZA as a system that identified keywords in a person’s input and applied transformation rules to produce a reply.

ELIZA became best known through a script called DOCTOR, which imitated the conversational style of a nondirective therapist. Such a therapist often encourages a person to continue speaking, rather than offering firm instructions. This style was especially well suited to a program built from simple rules.

If a user entered a statement about being unhappy, the program might detect a word related to emotion and respond with a general question. If someone wrote that their mother worried them, it might rearrange part of the statement and ask them to say more about their mother. When no useful keyword appeared, it could offer a broad prompt, inviting the person to continue.

The exchange could feel surprisingly natural. DOCTOR returned attention to the user, turning statements into questions or encouraging further explanation. Because the user supplied most of the details, the conversation could seem personal, even when the program’s replies remained general.

Yet ELIZA did not understand sadness, families, memory, or care. It was not a therapist, and it had no awareness of the feelings described to it. Its responses came from keywords, pattern matching, prepared prompts, and transformation rules.

Even so, some people spoke with ELIZA as though it understood them. They attributed empathy and judgment to a mechanism far simpler than they imagined. This tendency became known as the ELIZA effect. It described the human inclination to perceive understanding in a computer’s responses, especially when those responses were presented in familiar, conversational language.

Weizenbaum became concerned by how readily people treated computer output as human judgment. The lesson reached beyond one program. Language carries a powerful social presence. When words are arranged in the form of a conversation, people naturally look for a mind behind them. A system may therefore appear wiser, more caring, or more attentive than its underlying process warrants.

ELIZA showed that the appearance of conversation did not necessarily require deep understanding. A few well-chosen patterns could encourage a person to supply coherence and meaning on the program’s behalf. The machine produced the words, but the human reader often supplied the sense of connection.

Not long afterward, another language project demonstrated what could be achieved when a conversation was placed inside a very small and carefully organized world. Between about 1968 and 1970, Terry Winograd developed SHRDLU. The program operated in a simulated setting containing simple geometric blocks.

Within this blocks world, objects could have different shapes, colors, sizes, and positions. They could be moved, stacked, or placed beside one another. The program did not need to discuss every possible subject. It needed only to work with the objects and actions that belonged to this limited environment.

A person could type commands asking SHRDLU to move a block. The program could also respond to questions about the scene and about what had happened. It could keep track of the limited arrangement and use information from earlier parts of an exchange. This made its language appear fluid and purposeful.

The achievement was genuine. SHRDLU connected words with objects and actions inside its simulated environment. A command was not treated merely as a row of symbols. It could be related to a particular block, a particular location, and a possible change in the scene.

But SHRDLU’s apparent fluency depended upon strict boundaries. Its vocabulary was limited. The number of objects was small. The possible actions had been carefully defined. There was no weather in the blocks world, no shifting social intention, and no endless variety of things that a word might mean.

SHRDLU did not demonstrate a broad command of English. Instead, it showed that language could be handled impressively within a narrow, well-described domain. The blocks world functioned like a quiet miniature stage. Every object had a known role, and nearly everything that could happen had been anticipated.

This small world avoided many of the difficulties that appear in ordinary conversation. In daily life, people refer to things indirectly. They change topics. They make assumptions, tell jokes, use idioms, and leave thoughts unfinished. They depend upon memories and shared experiences that may never be stated aloud. SHRDLU’s environment was much simpler. Its fluency was supported by the careful order surrounding it.

Physical robots brought another kind of difficulty. A simulated block remains where the program says it is. A real object may be partly hidden, poorly lit, or difficult to identify. Floors may be uneven. Wheels can slip. Camera images contain shadows. Actions do not always produce exactly the expected result.

From 1966 to 1972, researchers at the Stanford Research Institute, or SRI, worked on a robot named Shakey. Its name reflected its somewhat unsteady movement. Shakey was a mobile machine equipped with cameras and sensors. It could gather information about its surroundings, form a plan, and carry out a sequence of actions.

This combination was important. Earlier machines often performed fixed motions, but Shakey attempted to connect perception, planning, and movement. When given a goal, it could consider intermediate steps instead of merely producing one immediate response. It might need to move through a room, approach an object, or alter its route before reaching the desired location.

The rooms used for Shakey were arranged to make these tasks manageable. Objects were simple, and the environment was controlled. Even then, progress could be slow. A camera did not deliver a tidy sentence announcing that there was a box near a wall. It delivered an image that had to be analyzed.

The robot then had to connect uncertain visual information with an internal description of its surroundings. Its plan had to account for where it could travel and what its actions might change. Perception, reasoning, and movement all had to work together.

This was a substantial challenge. A plan could be neat and orderly when represented inside a computer. The physical room was less obedient. Motors, wheels, surfaces, distance, and light all introduced uncertainty. A small difference between the internal model and the real setting could affect what happened next.

Shakey demonstrated that planning could guide physical behavior. It also made plain the difference between a clean internal model and the untidy material world. In a program, an action might be represented as a precise change from one state to another. In a room, the same action unfolded through machinery, surfaces, light, and chance.

Machine translation, ELIZA, SHRDLU, and Shakey were very different projects. Yet together they carried a shared lesson. A machine could seem remarkably capable when the task, vocabulary, or environment was carefully bounded. Within those boundaries, rules and plans could produce memorable results. Beyond them waited ambiguity, common sense, physical uncertainty, and countless details that no one had thought to specify.

These demonstrations were not empty illusions. They established useful methods and raised enduring questions. Machine translation exposed the importance of context and world knowledge. ELIZA revealed how easily people could attribute understanding to conversational output. SHRDLU showed the power of connecting language to a limited, structured environment. Shakey joined perception and planning with movement through physical space.

Still, public impressions often traveled farther than the systems themselves. A translated sentence could suggest universal translation. A reflective prompt could suggest empathy. A conversation about blocks could suggest mastery of language. A robot crossing a prepared room could suggest general independence.

The distance between a demonstration and a general ability would become one of the central themes in the history of artificial intelligence. A successful demonstration showed that something important was possible under particular conditions. It did not always show that the same method would work in the open world, where conditions were less orderly and meanings less certain.

Early ambition had opened several doors. Beyond each threshold, however, lay more complexity than expected. Language required more than dictionaries. Conversation required more than plausible replies. Physical action required more than a clear plan. Again and again, researchers found that intelligence seemed to depend upon a broad background of context, experience, and ordinary knowledge.

As the 1970s approached, these difficulties became harder to overlook. Researchers faced slower computers, limited data, and goals that had sometimes been described too confidently. The field would not disappear, and work would continue. But after the bright promise of early demonstrations, a quieter period of reassessment was drawing near.