Lost Memories

How the Americas invented and forgot the foundations of artificial intelligence

Yesterday — The Cords and the Ashes

There exist memories that are burned and memories that are knotted. The Americas have known both.

On July 12, 1562, in the city of Mani in the Yucatan, a Franciscan monk named Diego de Landa ordered a great fire to be lit. Into this blaze were thrown twenty-seven Maya codices — books made of fig tree bark, folded like accordions, covered in glyphs painted red and black. These books contained centuries of astronomical observations, eclipse prediction tables, calculations of the Venus cycle of stunning precision. De Landa judged that they contained "superstitions and lies of the devil." He also burned five thousand statues and an unknown number of human bodies.

The historian George Stuart estimates that "hundreds, perhaps thousands" of books were destroyed that day and in the years that followed. We will never know exactly how many. Of the entire Maya civilization, only four codices survived: Dresden, Madrid, Paris, and the Grolier Codex. Four books to bear witness to an entire library.

The irony of this destruction is dizzying. De Landa, the arsonist, later became one of the principal documentarians of Maya culture. His work, Relacion de las cosas de Yucatan, remains today an invaluable source. One specialist calculated that "99% of what we know about the Maya, we know because of what de Landa told us." But the same man "burned 99 times more knowledge than he gave."

Let us look at what was lost in the flames.

The Maya had developed a mathematical system of remarkable sophistication. Their numeration used base twenty — where we count in tens, they counted in twenties. Three symbols sufficed: a dot for one, a bar for five, a shell for zero. Zero. The Maya had invented it independently, centuries before it was imported to Europe from India by Arab mathematicians.

More still: they used a positional notation system. As in our way of writing numbers, where the digit "5" can mean five, fifty, or five hundred depending on its position, Maya symbols changed value according to their place in the column. This invention — fundamental to any complex calculation — is so rare in human history that we can count on the fingers of one hand the civilizations that developed it.

This mathematical precision served an objective: to measure time. The Maya had developed not one, but three interlocking calendars. The Tzolk'in, a sacred calendar of 260 days, governed religious ceremonies and divination. The Haab, an agricultural calendar of 365 days, guided planting and harvesting. And the Long Count, a chronological calendar, allowed any event to be situated on a time scale spanning millennia.

The combination of these three systems produced the "Calendar Round," a cycle of approximately 52 years unique in human history. Maya astronomers could predict eclipses with a margin of error of a few minutes. Their tables of Venus cycles, preserved in the Dresden Codex, are valid for several centuries. Entire cities — Teotihuacan, Chichen Itza — were arranged to reflect the order of celestial movements.

Here we have a system of calculation, prediction, and modeling of the natural world. The Maya would not have recognized the word "algorithm," but they practiced its logic: a series of repeatable operations, producing predictable results, allowing anticipation of the future. This future, de Landa threw into the fire.

Meanwhile, further south, in the high valleys of the Andes, another memory system was thriving — a system that the Spanish conquistadors did not burn, because they never truly understood what it was.

The quipu — from the Quechua word for "knot" — is a device made of knotted cords. At first glance, it resembles an ornament, a multicolored fringe hanging from a main cord. But this fringe is a computer.

Each hanging cord can carry dozens of knots. The position of the knot on the cord indicates its positional value: units at the bottom, tens above, hundreds higher still. The type of knot — simple, long, or figure-eight — specifies the digit. The color of the cord indicates the category of information. The most complex quipus had more than a thousand cords, capable of storing staggering quantities of data.

The Incas used this system for everything: censuses, inventories, tax collection, military organization, ritual calendars. The quipu was not a calculator — it did not serve to perform operations — but a device for storing and retrieving information. An external memory. A portable database.

To make this memory function, specialists were required: the quipucamayocs. These "keepers of the knots" spent years in the yachaywasi, the "houses of learning," mastering the color codes, knot types, and placement conventions. They also memorized the oral narratives that gave meaning to the encoded data. The quipu was the memory; the quipucamayoc was the processor.

The conquistadors destroyed thousands of quipus, which they considered instruments of idolatry. But unlike the Maya codices, the system survived. Seven hundred and fifty-one quipus have been catalogued around the world by Harvard's Khipu Database Project. And in remote regions of the Andes, herders still use quipus to count their flocks.

The destruction did not stop in the sixteenth century. It continued, in other forms, into the twentieth century.

In Canada, between 1883 and 1996, at least 150,000 First Nations, Inuit, and Metis children were torn from their families and placed in residential schools run by Christian churches and funded by the government. The objective, explicitly stated, was to "kill the Indian in the child." Children were punished — sometimes beaten, sometimes worse — if they spoke their mother tongue. They were forced to abandon their names, their clothes, their hairstyles, their spiritual practices.

The Truth and Reconciliation Commission of Canada called this system "cultural genocide." The Pope recognized it as genocide in 2022. More than 4,000 child deaths have been documented; actual estimates exceed 6,000. And the effects continue: 70% of Indigenous languages in Canada are now threatened with extinction.

What these residential schools destroyed was not only languages and customs. It was entire knowledge systems — ways of measuring time, counting, transmitting information from one generation to the next. The Powhatan tribes of Virginia, for example, used knotted cords and notched sticks for their accounts, following a base-ten system remarkably similar to the Andean quipu. How many of these systems were lost, not in flames, but in the forced silence of children separated from their grandparents?

The Americas are a continent of lost memories. From burned codices to extinct languages, from confiscated quipus to interrupted knowledge, systematic destruction erased centuries of computational thinking. The Maya had zero and positional notation. The Incas had portable databases. The peoples of the North had astronomical calendars and sophisticated counting systems.

All of this was judged "primitive," "superstitious," "diabolical." All of it was condemned to disappear to make way for "civilization."

The official history of computing begins elsewhere. It begins in American laboratories of the twentieth century, with machines and men. But this history, too, has its lost memories.

Today — The Cards and the Shadows

In 1880, the United States Census Bureau faced an arithmetic problem. The Constitution required a population count every ten years. In 1790, with fewer than four million inhabitants, the exercise was manageable. A century later, with 63 million Americans, it was becoming nightmarish. The 1880 census had taken more than eight years to process. At this rate, the 1890 census would not be finished before the 1900 census began.

A young German-American statistician named Herman Hollerith proposed a solution. He had observed Jacquard looms, those weaving machines programmed by punch cards. What if census data were transcribed as holes in cards? What if an electromechanical machine could read these cards, count the holes, tabulate the results?

In 1888, the Census Bureau organized a competition. Three systems were tested on a data sample. The first two took 144 and 100 hours respectively. Hollerith's machine did the job in 72 hours.

The system worked like this: each person counted became a card. Their characteristics — age, sex, birthplace, occupation — were transcribed as holes at precise positions. The machine pushed metal pins through the holes. When a pin passed through a hole, it plunged into a cup of mercury, closing an electrical circuit. This circuit advanced a counter by one notch. In seconds, a card was read and its information tabulated.

The 1890 census was completed in six months. Full data processing took two years. Hollerith's system saved the American government five million dollars. Countries around the world adopted his technology: Austria, Canada, Cuba, France, Norway, the Philippines, Russia.

In 1896, Hollerith founded the Tabulating Machine Company. In 1911, he sold it. The company merged with others and was renamed in 1924. Its new name: International Business Machines Corporation. IBM.

Hollerith's punch cards are the direct ancestors of computer memory. The holes in the cardboard — presence or absence, one or zero — foreshadow the binary system. The card's structure — rows and columns, fixed positions — foreshadows memory addressing. The electrical circuit that "reads" the card foreshadows the processor.

We are reinventing the quipu. The knot on the cord becomes the hole in the card. The fiber's color becomes the column's position. The quipucamayoc who memorizes the narratives becomes the operator who programs the machine. But this continuity, no one sees it. The codices have burned. The quipus are in museums. The residential schools did their work.

The next generation of pioneers knew nothing of Maya mathematics or Andean systems. At MIT, in the 1930s, an engineering professor named Vannevar Bush built a room-sized machine. His differential analyzer, completed in 1931, was a tangle of gears, shafts, and disks driven by electric motors. It could solve differential equations with eighteen variables — a feat impossible to accomplish by hand.

Bush's analyzer was an analog computer: it used physical quantities (rotations, positions) to represent numbers. A young mathematics student was hired to operate it. His name was Claude Shannon. In manipulating the machine's complex circuits, Shannon had an insight that would change the world.

In 1937, at twenty-one, Shannon submitted his master's thesis: "A Symbolic Analysis of Relay and Switching Circuits." The idea was simple, and revolutionary. Boolean algebra — that branch of mathematics that manipulates truth values (true/false, 1/0) — could serve as the basis for designing electrical circuits. An open or closed switch was a bit. Switches connected in series or parallel were logical operations. Any logical reasoning could be translated into circuitry.

A historian of science called this thesis "probably the most important master's thesis of the century." It transformed digital circuit design "from an art to a science." It laid the theoretical foundations for everything that would follow: computers, networks, artificial intelligence systems.

Three years later, on September 11, 1940, a Bell Labs researcher named George Stibitz gave a demonstration that prefigured our connected world. At a meeting of the American Mathematical Society at Dartmouth College, he installed a teletype connected by telephone line to a computer located in New York, several hundred kilometers away. The mathematicians present — among them John von Neumann and Norbert Wiener — proposed equations. Stibitz typed them on the teletype. In less than a minute, answers arrived.

It was the first demonstration of remote computing in history. The audience was described as "stunned." Such remote access would not be repeated for ten years.

Then came the war, and with it, acceleration.

The American army needed to calculate ballistic trajectories — the curves that shells and bombs follow. Each trajectory depended on dozens of variables: initial velocity, firing angle, air resistance, Earth's rotation. A manual calculation took between twenty and forty hours. A complete firing table could require months of work.

At the Moore School of Engineering at the University of Pennsylvania, two professors — John Mauchly and J. Presper Eckert — proposed building a fully electronic machine capable of performing these calculations in seconds. The army agreed. In 1945, ENIAC (Electronic Numerical Integrator and Computer) was completed. Thirty tons of metal, 18,000 vacuum tubes, power consumption capable of dimming lights throughout the neighborhood.

ENIAC could calculate a trajectory in thirty seconds. It was, by all standards of the era, a technological marvel.

But who had programmed it?

The army had hired more than two hundred women as "computers" — the word then designated a job, not a machine. These women, mathematics graduates, performed ballistic calculations by hand. Six of them were selected to program ENIAC: Betty Holberton, Kay McNulty, Marlyn Wescoff, Ruth Lichterman, Betty Jean Jennings, and Fran Bilas.

Programming ENIAC was colossal work. The machine had no software in the modern sense. To make it work, one had to physically plug cables, configure switches, establish connections between different units. The six programmers spent weeks studying the machine's plans, understanding how it worked, designing the necessary operation sequences.

When the army presented ENIAC to the press in February 1946, the six women were not on the stage. They were not introduced. They were not named. Photographs sometimes show them in the background, but captions do not identify them. For decades, the history of computing forgot them completely.

This was not an isolated accident. At the Harvard Observatory, since the 1880s, entire teams of women — the "Harvard Computers" — had analyzed photographic plates of stars, classified spectra, catalogued galaxies. They were paid less than half men's salaries. Among them, Henrietta Swan Leavitt discovered the relationship between the brightness of variable stars and their pulsation period — a discovery that made it possible to measure cosmic distances. Annie Jump Cannon developed the stellar classification system still used today. Florence Cushman spent fifty years cataloguing data.

Their names rarely appear in textbooks. When people speak of the "Harvard Computers," they often mention the observatory director, Edward Pickering. The women who did the work remain in the shadows.

Grace Hopper, a mathematician and Navy officer, programmed the Harvard Mark I during the war. She invented the first compiler — a program capable of translating human instructions into machine code. She developed the COBOL language, which remained for decades the standard for business programming. She became an admiral and remained active past eighty.

Her story is known. But how many others remained invisible?

Beyond — Memories to Come

There is a troubling parallel between the two forms of erasure we have traversed.

On one side, Indigenous knowledge: burned codices, confiscated quipus, extinct languages, knowledge systems judged "primitive" and condemned to disappear. On the other, women's work: calculations performed but not credited, programs written but not signed, essential contributions made invisible.

In both cases, the same logic operates. What is destroyed or forgotten is what does not correspond to the image one wishes to project. The conquistadors could not admit that the Maya had invented zero and predicted eclipses with precision superior to Europe's. Military officials could not admit that the first computer programmers were women. History is written according to the prejudices of those who tell it.

But lost memories have a way of resurfacing.

In 1988, a researcher named Kathy Kleiman came across a photo of ENIAC. In the background, women were manipulating cables and switches. Who were they? She undertook research, found the surviving programmers, recorded their testimonies. Her work has helped to recognize Betty Holberton, Kay McNulty, and their colleagues as the pioneers they were.

In 2005, Harvard's Khipu Database Project began systematically cataloguing all known quipus. Researchers discovered unsuspected complexities: some quipus seem to contain not only numbers but also narrative accounts. The system may have been even more sophisticated than previously thought.

In the remote regions of Guatemala and Chiapas, Maya communities continue to use the traditional 260-day calendar. The glyphs have been deciphered. Millions of Maya language speakers still inhabit their ancestral lands. What de Landa burned has not entirely disappeared.

What do these recovered memories teach us about the artificial intelligence we are building today?

First, that the history of computing is older and more diverse than the official narrative suggests. The binary system does not begin with Leibniz. Positional notation does not begin with the Arabs. Information storage does not begin with punch cards. Everywhere in the world, civilizations invented ways to calculate, predict, memorize. Most of these inventions were destroyed or forgotten. But they existed.

Second, that invisibilization is not an accident. It is the product of power relations. Indigenous knowledge was destroyed because it threatened the authority of conquistadors and missionaries. Women's work was made invisible because it threatened the image of masculine and military computing. What is recognized as "foundational" depends on who writes the history.

Finally, that lost memories continue to influence the present. The artificial intelligence we build today bears the traces of its origins. It inherits the biases of those who designed it, the gaps of those who documented it, the silences of those who were erased. Current machine learning systems often reproduce the prejudices of their creators — including the gender and racial prejudices we have described.

Recognizing erased contributions is not merely an exercise in historical justice. It is a condition for understanding what we are building. If we do not know where our technology comes from, we cannot know where it is taking us.

The Maya understood something essential: time is not a line, but a cycle. What has been destroyed can return in other forms. What has been forgotten can be rediscovered. What has been made invisible can become visible again.

The quipus of the Andes teach us that a knotted cord can be a memory. The Maya codices remind us that predicting the future requires understanding the past. The ENIAC programmers show us that yesterday's invisible ones can become tomorrow's pioneers.

The history of artificial intelligence is not an American history in the narrow sense — that of MIT laboratories and Bell Labs. It is an American history in the broad sense — that of a continent where memories were lost and found, destroyed and rebuilt, erased and rewritten.

The Americas did not only invent punch cards and logic circuits. They also invented zero, positional notation, portable databases, precision astronomical calculation. And they produced generations of calculators, programmers, and mathematicians whose names are only beginning to emerge from the shadows.

The artificial intelligence of tomorrow will be what we make of it. It will carry the memories we choose to transmit to it. Will we choose to perpetuate the erasures of the past? Or will we choose to finally tie the threads of all memories?

The cords are still there. The knots are waiting to be made.