Artificial Intelligence does not come from nowhere. It is the result of centuries of ideas, discoveries, and breakthroughs in mathematics, logic, philosophy, and computing. Its roots can be traced back hundreds of years, with important foundations emerging as early as the 17th century. What we call AI today is the culmination of a long journey from early attempts to understand and mechanize human reasoning to the powerful models that can learn, generate, and solve complex problems today. Let’s take a journey through this history to better understand the technology we use every day and the ideas that made modern AI possible.
1. Thinking is arithmetic
In 1651, in Leviathan, the English philosopher Thomas Hobbes proposed a striking idea about the nature of human reasoning. In Chapter 5, Of Reason and Science, he wrote:
“For REASON, in this sense, is nothing but Reckoning, that is, Adding and Subtracting, of the Consequences of generall names agreed upon for the marking and signifying of our thoughts.”
Hobbes was suggesting that reasoning might not be a mysterious or uniquely human faculty. If reasoning could be reduced to a process of adding and subtracting, then perhaps it could be described by rules—and, in principle, carried out mechanically.
The idea raised a profound question: if reasoning is merely calculation, could a machine reason too? Hobbes did not have anything resembling a modern computer, of course, and his “reckoning” was still a philosophical description of how the mind manipulates concepts and language. More importantly, he never provided a practical account of what exactly should be added and subtracted, or how such reasoning could actually be mechanized.
That gap would remain for centuries. Philosophers continued to debate the nature of reason, while mathematicians and logicians gradually developed formal systems for representing and manipulating symbols. Only much later would these ideas converge with mechanical computation, laying some of the intellectual foundations for artificial intelligence.
If Hobbes was right, then anything capable of performing the right kind of calculation could, in principle, reason. The difficult question was not whether reasoning involved calculation, but what exactly was being calculated—and how could we make a machine do it?
2. Algebra
In 1854, George Boole published An Investigation of the Laws of Thought, a work that would eventually transform the way humans understood logic. Boole’s path to mathematics was remarkable. The son of a shoemaker, he received little formal education and largely taught himself Latin and advanced mathematics from books. In 1849, he was appointed Professor of Mathematics at Queen’s College, Cork, Ireland—despite never having attended a university.
Boole’s central idea was radical in its simplicity: logical reasoning could be reduced to mathematics. Instead of treating statements as vague expressions of thought, he represented them using two possible values: true or false. He then developed an algebra for combining these values through logical operations such as AND, OR, and NOT.
For example, the statement “It is raining AND I have an umbrella” is true only when both parts are true. Change either part to false, and the entire statement becomes false. What had traditionally been debated through words and philosophical argument could now be expressed as a formal system of symbols and rules.
Boole had effectively turned logic into something that could be calculated. The ideas of reasoning that philosophers had debated since Aristotle could now be manipulated according to mathematical rules—a crucial step toward the possibility of machines performing logical operations.
Boole died only ten years later, in 1864, after becoming ill following a walk to give a lecture in heavy rain. His algebra attracted relatively little attention at first. Decades later, however, his system of Boolean logic would become one of the fundamental building blocks of digital computers—and, eventually, artificial intelligence.
3. Electricity
In 1937, at just 21 years old, MIT graduate student Claude Shannon noticed a connection that would change the course of computing. An electrical switch has two basic states: open or closed. Boolean logic has two corresponding values: false or true.
Shannon realized that the two systems could be mapped onto each other. Connect switches in series, and they behave like AND: the circuit is active only when every switch is closed. Connect them in parallel, and they behave like OR: the circuit is active when any one of the switches is closed.
In his 1938 master’s thesis, A Symbolic Analysis of Relay and Switching Circuits, Shannon showed that Boolean algebra could be used to design and analyze electrical switching circuits. An abstract mathematical system developed by Boole could now be physically implemented with wires, switches, and relays. His thesis is widely regarded as one of the most influential master’s theses of the 20th century, and its ideas became a foundation of modern digital circuit design.
Machines could now compute logic.
But there was still a fundamental gap. A machine could follow logical rules and manipulate symbols, yet that did not mean it could understand, learn, or reason in the way humans do.
Computing was not the same as thinking.
4. The Switch
In 1943, in Chicago, neurophysiologist Warren McCulloch teamed up with Walter Pitts, a brilliant 20-year-old with no university degree and, by many accounts, an unstable home life. A famous story—though never definitively confirmed—claims that when Pitts was 12, he ran from bullies into a library and spent three days reading Bertrand Russell and Alfred North Whitehead’s Principia Mathematica.
Together, McCulloch and Pitts published a radically simplified mathematical model of a biological neuron. A neuron receives signals from other neurons, adds them together, and fires only when the combined signal reaches a certain threshold.
The idea was remarkably simple, but its implications were profound. By connecting a few of these artificial neurons together, they showed that the network could reproduce basic Boolean operations such as AND, OR, and NOT.
A network of neurons could therefore perform the same kind of logical computation as Shannon’s electrical circuits. Two seemingly different systems—the brain and an electrical machine—could be described using the same underlying language of logic.
But there was a crucial limitation: the network could not learn. Its connections were fixed. If you wanted the network to perform a new task, you had to change its connections manually.
The challenge was no longer simply to build a machine that could compute. It was to build one that could change itself through experience.
5. The Learning
In 1949, Canadian psychologist Donald Hebb proposed a simple but powerful idea about how the brain learns. When one neuron repeatedly helps activate another, the connection between them becomes stronger. In other words, connections that are repeatedly used become stronger, while those that are rarely used become weaker.
This idea, later known as Hebbian learning, introduced something the McCulloch–Pitts model was missing: a mechanism for changing the connections between neurons based on experience.
The pieces were now beginning to fall into place. There was a mathematical model of a neuron, a way to connect neurons into a network, and now a basic rule describing how those connections could change. Someone just had to put them together.
6. The Thinking
In 1950, Alan Turing—the mathematician who had played a crucial role in breaking Germany’s Enigma code during World War II—opened his paper Computing Machinery and Intelligence with a deceptively simple question: “Can machines think?”
Turing immediately recognized a problem: nobody could agree on what thinking actually meant. Rather than getting trapped in a philosophical debate over the definition, he proposed replacing the question with a practical test.
Imagine a person and a machine in separate rooms. A human judge communicates with both through written questions, without seeing either one. If the judge cannot reliably determine which participant is the machine, Turing argued that we could treat the machine’s behavior as evidence of intelligence. This became known as the Turing Test.
But the most forward-looking idea in the paper was not the test itself. Turing suggested that instead of attempting to program a fully developed adult mind, we could build a “child machine” and teach it.
The distinction was profound. Rather than telling a machine exactly what to know, perhaps we could create a machine capable of learning—and let experience do the rest.
The idea that intelligence could be grown rather than simply programmed would become one of the central questions in the development of artificial intelligence.
7. The Name
In 1955, four researchers—including Claude Shannon and a young Marvin Minsky—wrote a funding proposal for a summer research project at Dartmouth College. Their plan was ambitious: bring together ten researchers for two months in the summer of 1956 to study a field they gave a new name: artificial intelligence.
It was the first known use of the term.
The proposal was remarkably confident, almost to the point of bravado. The researchers argued that every aspect of learning and intelligence could, in principle, be described so precisely that a machine could be built to simulate it.
Their goal was equally bold: make significant progress on understanding and reproducing machine intelligence in just one summer.
The Dartmouth workshop would not solve the problem of intelligence in two months. But it gave an emerging field a name, a research agenda, and a shared ambition.
Artificial intelligence had officially entered the vocabulary of science.
8. The Perceptron
In 1958, psychologist Frank Rosenblatt brought together two ideas that had been developed separately: the mathematical neuron of McCulloch and Pitts, and Hebb’s idea that neural connections could change through learning. He called the result the perceptron—and he didn’t leave it as a mathematical model. He built one as physical hardware: the Mark I Perceptron.
This was not software running on a general-purpose computer. It was a large cabinet with 400 light sensors arranged in a 20 × 20 grid, forming its artificial “eye.” Its memory consisted of a bank of adjustable connections controlled by electric motors. Learning literally meant changing those connections; when the machine made a mistake, the connections were adjusted again.
And remarkably, it worked.
Rosenblatt showed the perceptron punched cards containing patterns marked on one side or the other. After only a few dozen training examples, it learned to distinguish left from right without anyone explicitly programming the rule.
The achievement attracted enormous attention. The press quickly imagined a future in which machines might see, speak, walk, write, reproduce themselves, and even become conscious. The New York Times reported predictions from the U.S. Navy along these lines.
The reality was much more modest.
The machine could tell left from right.
But for the first time, a machine was not simply following a fixed set of instructions. It was learning from examples.
9. The First AI Winter
Not everyone was convinced by the excitement surrounding the perceptron. Marvin Minsky, one of the organizers of the Dartmouth workshop, had been a year behind Rosenblatt at the Bronx High School of Science and had debated the capabilities of perceptrons with him for years.
In 1969, Minsky and Seymour Papert published Perceptrons, a detailed analysis of what these early neural networks could—and could not—do. They showed that a single-layer perceptron has a fundamental limitation: it can only separate patterns that can be divided by a single straight boundary.
Some surprisingly simple problems cannot be solved this way. The classic example is XOR—“one or the other, but not both.” Plot its four possible combinations on a grid, and the cases that produce a true result sit diagonally opposite each other. No single straight line can separate them from the false cases.
The problem was not that neural networks could never solve XOR. Add a hidden layer, and the network can represent it. But this introduced a much deeper problem.
If the network produces the wrong answer, we can measure the error at the output. But how do we determine which hidden connections caused that error—and how much should each connection change?
That question would remain a major obstacle for years. As enthusiasm faded and research funding declined, neural networks entered what became known as the first AI winter.
In 1971, just two years after Perceptrons was published, Frank Rosenblatt died in a boating accident on his 43rd birthday.
The idea of learning machines had not disappeared. But researchers still needed a way to teach a multi-layer network how to correct its own mistakes.
10. Backpropagation
The breakthrough came in 1986, when David Rumelhart, Geoffrey Hinton, and Ronald Williams published a landmark paper in Nature describing how to effectively train multi-layer neural networks using backpropagation.
The idea was deceptively simple. Instead of only sending information forward through the network, take the error at the output and propagate it backwards through the layers. Each connection receives a share of the blame based on how much it contributed to the final error. That information is then used to adjust the connections, layer by layer, all the way back to the input.
For the first time, a practical method existed for teaching a multi-layer network how to correct its own mistakes.
But backpropagation was not entirely new. The underlying mathematical technique had already been described by Finnish researcher Seppo Linnainmaa in his 1970 master’s thesis, and Paul Werbos had proposed applying the method to neural networks in 1974. What changed in the 1980s was the demonstration that the technique could be used effectively to train neural networks with multiple layers.
By 1986, two crucial pieces were finally in place: multi-layer neural networks and a practical way to train them.
So why didn’t AI suddenly take off?
The missing ingredients were not ideas. They were data and computing power.
The networks could learn. We simply didn’t yet have enough examples to teach them—or enough machines powerful enough to train them.
11. The Data
In 2007, Princeton professor Fei-Fei Li challenged the prevailing direction of AI research. While many researchers were focused on designing better algorithms, Li argued that something more fundamental was missing: data.
At the time, computers were being trained to recognize objects from only a few thousand carefully selected images. Humans, by contrast, learn to recognize objects after seeing countless examples in the real world. Li believed that if machines were going to learn to see, they needed to be exposed to vastly more visual information.
She turned to WordNet, an existing database of English words and concepts, and selected roughly 22,000 categories of objects, ranging from ambulances to pumpkins. The challenge was enormous: every image had to be collected and labelled. Doing the work manually with a small research team was estimated to take around 18 years.
Then came an unexpected solution.
Amazon had recently launched Mechanical Turk, a platform that allowed researchers to distribute small online tasks to thousands of workers around the world. By using it to crowdsource image labelling, Li’s team reduced the estimated effort from nearly two decades to roughly two years.
The result, released in 2009, was ImageNet: a dataset containing more than 14 million labelled images across thousands of object categories.
Li then launched an annual competition, challenging researchers to build systems that could recognize these images as accurately as possible. At first, progress was modest. Some researchers even questioned whether the dataset was simply too large and difficult for existing algorithms to handle.
But ImageNet had created something AI had never had before at this scale: a massive supply of labelled examples with which to learn.
The algorithms were about to catch up.
12. The Power
The computing problem was solved by an unlikely industry: video games.
Rendering a 3D game scene requires performing millions of relatively simple mathematical operations simultaneously—calculating the position, color, lighting, and movement of countless pixels. To handle this workload efficiently, chipmakers developed graphics processing units, or GPUs, designed to perform many calculations in parallel.
Then researchers realized something remarkable: training a neural network requires much of the same kind of mathematics. Neural networks repeatedly perform enormous numbers of matrix and vector calculations, and many of those operations can be performed simultaneously.
Hardware designed to make video games look realistic turned out to be remarkably well suited to teaching machines.
The GPU had been built to render virtual worlds.
Now it could help machines learn from data.
13. AlexNet
In 2012, Alex Krizhevsky, a student of Geoffrey Hinton, teamed up with fellow student Ilya Sutskever to put the pieces together: a large neural network, a massive dataset, and the computing power of GPUs.
Their network had around 60 million parameters and was trained on two gaming graphics cards for roughly five or six days—reportedly in Krizhevsky’s parents’ house. They entered it in Fei-Fei Li’s ImageNet competition.
The result was astonishing.
Their network, later known as AlexNet, achieved a top-5 error rate of 15.3%. The second-place system had an error rate of 26.2%. In a competition where improvements of just one or two percentage points per year were considered significant, the gap was extraordinary.
AlexNet demonstrated something the field had been waiting decades to see: given enough data, enough computing power, and a sufficiently deep neural network, machines could learn to recognize complex visual patterns far better than previous approaches.
The implications were immediate. Within months, Google acquired the small company Hinton had formed with Krizhevsky and Sutskever. It had no product, no revenue, and only three employees.
But it had something much more valuable: a working demonstration of where AI was heading.
14. The Attention
In 2014, researchers were beginning to push neural networks beyond image recognition and into language. But machine translation had a fundamental problem.
The prevailing approach would read an entire sentence, compress its meaning into a single fixed-size internal representation, and then generate the translation from that representation. Short sentences could work reasonably well. But as sentences became longer and more complicated, the system struggled. Too much information had to be squeezed into too little space.
Yoshua Bengio compared the process to reading an entire book in French, closing it, and only then trying to write the English translation. You would inevitably forget important details.
Bengio’s intern, Dzmitry Bahdanau, asked a simple question: How do human translators actually work?
They don’t read the entire sentence once and then rely entirely on memory. As they write each word, they look back at the original and focus on the part that matters.
So Bahdanau and his collaborators designed a neural network that could do something similar. Instead of forcing the entire source sentence into one fixed representation, the model could look back at different parts of the sentence at each step, assigning more weight to the words that were most relevant to the word it was producing.
They called this mechanism attention.
It was a deceptively simple change, but it introduced a powerful new idea: instead of forcing a neural network to remember everything equally, let it decide what to focus on.
15. The Transformer
Earlier attention-based models still processed words sequentially, one after another. That made them difficult to parallelize: a model had to finish processing one word before it could move on to the next. GPUs are extremely good at performing many calculations simultaneously, so this sequential bottleneck limited how efficiently they could be used.
In 2017, eight researchers at Google proposed a radical simplification: remove sequential processing entirely and build the architecture around attention.
Their paper was titled Attention Is All You Need—a playful nod to the Beatles’ song All You Need Is Love. The architecture they introduced became known as the Transformer.
Instead of reading a sentence strictly from left to right, a Transformer can process the tokens in parallel while using attention to determine how they relate to one another. This made much better use of GPU hardware and allowed researchers to train substantially larger models.
The breakthrough changed the direction of AI research. With the sequential bottleneck removed, researchers could scale up data, computing power, and model size far more effectively.
The era of increasingly large neural networks had begun.
16. The GPT
OpenAI, where Ilya Sutskever had gone on to become chief scientist, took the Transformer architecture and gave it an almost comically simple task: predict the next word.
The process was straightforward. Show the model enormous amounts of text, hide the next word, ask it to predict what comes next, and use backpropagation—the same fundamental technique developed decades earlier—to adjust the model when it gets the answer wrong. Then repeat the process billions of times.
But why would predicting words produce anything resembling intelligence?
Consider what it takes to make a good prediction. To complete “The capital of France is…”, the model needs to know something about geography. To predict how a story will end, it needs to understand the characters, their goals, and the relationships between them. To predict what comes next in a technical explanation, it needs to capture something about the underlying subject.
The better the model becomes at predicting language, the more information about the world behind that language it needs to capture.
In 2020, OpenAI introduced GPT-3, with 175 billion parameters. It could generate remarkably fluent text, but there was a problem: fluency was not the same as following instructions. The model could produce convincing answers while ignoring what the user actually wanted—or confidently generating something incorrect.
So OpenAI added another stage of training.
People were shown multiple answers to the same prompt and asked to compare them, indicating which responses were more helpful, accurate, and aligned with what was requested. Those human preferences were then used to train the model further through a technique known as reinforcement learning from human feedback (RLHF).
The model had first learned from the patterns in enormous amounts of text.
Now it could also learn from people correcting its behavior.
That combination—pretraining on massive datasets followed by human-guided refinement—became a key part of the path from language models that could predict text to systems designed to interact with people.
17. OpenAI
In November 2022, OpenAI released the result as a relatively low-key “research preview.” After decades of ambitious promises and disappointing chatbots, there was little reason to assume this one would be different.
And the, it was!
The product reached one million users in about five days and roughly 100 million monthly active users within two months, according to widely cited estimates. It became one of the fastest-growing consumer internet products ever. The name was ChatGPT. For the first time, millions of people could interact directly with a powerful language model through a simple conversation—and AI moved from something most people encountered indirectly to something they could use themselves.