From the Golem to GPT: The History of Artificial Intelligence
At the end of 2022, something strange happened. For the first time in history, millions of people around the world began having free-flowing conversations with a non-human intelligence. It was called ChatGPT, and within a few weeks it became a global phenomenon—the fastest-adopted technology in human history. For many, it was a shock—the sudden appearance of something that had until then belonged to the realm of science fiction. Yet what looked like an overnight revolution was, in reality, the culmination of a dream as old as civilization itself: the dream of creating an artificial mind.
The history of artificial intelligence (AI) is not a simple, linear story of technological progress. Rather, it is a turbulent saga full of cycles, in which philosophical dreams collide with engineering reality, while a handful of key breakthroughs repeatedly redefine what is considered impossible. It is a journey that takes us from the clay giants of ancient myths, through philosophers’ studies and dusty university laboratories, to city-sized server farms that power our digital world today.
To understand how we reached this point, we must travel back in time—not decades, but millennia. We will trace the ancient echoes of this idea, meet the founding fathers who, in an act of almost audacious optimism, christened a new field of science, endure the painful “AI winters,” when enthusiasm froze and funding dried up, and finally examine the three great waves of the 21st-century revolution that led to the current explosion. This is the story of how humanity tried to imprison the spirit in a machine.
Table of contents
-
Antiquity and the Middle Ages: Myths, Logic, and the First Automata
-
Founding Fathers – Turing and Dartmouth. The Turbulent History of Artificial Intelligence
-
Long Winters and Short Summers – Cycles of Hype and Collapse
-
The Revolution of the 21st Century – Three Eras That Changed Everything
Period | Milestone | Significance |
Antiquity | Myths of the Golem and Talos; Aristotle’s logic | Establishing the desire to create artificial life and the first attempts to formalize thought. |
1843 | Ada Lovelace’s Work | The prediction that analytical engines could manipulate symbols, not just numbers – the birth of the concept of programming. |
1950 | Turing Test | Alan Turing proposes a practical test of machine intelligence, bypassing the philosophical debate about consciousness. |
1956 | Dartmouth Conference | The official birth of the field and the coining of the term “artificial intelligence” by John McCarthy. |
1966 | ELIZA | Creation of the first chatbot, which revealed how easily people anthropomorphize machines. |
1974–1980 | The First “AI Winter” | Unfulfilled promises and funding cuts lead to a collapse in AI research. |
1997 | Deep Blue defeats Kasparov | IBM’s supercomputer defeats the world chess champion, symbolizing the triumph of computing power over human intuition. |
2012 | AlexNet’s Breakthrough in the ImageNet Competition | A deep learning neural network outperforms its rivals in image recognition, launching the deep learning revolution. |
2017 | Publication of “Attention Is All You Need” | Introduction of the Transformer architecture, which became the foundation for modern large language models. |
2022 | Public release of DALL-E 2 and ChatGPT | The explosion of generative AI, reaching ordinary users on a mass scale for the first time in history. |
Antiquity and the Middle Ages: Myths, Logic, and the First Automata
Talos and Golem
Before the first transistor was created, and even before humanity mastered electricity, the idea of artificial intelligence was already alive in stories. The ancient Greeks spun myths about Talos – a gigantic bronze automaton that guarded the shores of Crete, hurling boulders at the ships of invaders. In Jewish folklore, the legend of the Golem was born, a being molded from clay and brought to life by a mystical word to protect the Prague community. These stories, though fantastical, reveal a fundamental human desire: the wish to bring into existence a being in our likeness, endowed with strength and purpose, yet created rather than born.
However, the history of AI unfolded along two parallel tracks that did not intersect for centuries. While myths and legends explored the idea of an artificial body, ancient philosophy began laying the foundations for a mechanical mind. Aristotle, strolling through the Athenian Lyceum, was the first to attempt to formalize human reasoning. His system of syllogisms was a revolutionary concept: a set of strict rules that made it possible to draw logical conclusions from premises, regardless of who applied them. It was the first attempt to separate the process of thinking from the thinker—to create an algorithm of deduction.
Charles Babbage
First Calculators
For more than two millennia, these two currents—the idea of an artificial being and the formalization of thought—flowed separately. Only in the 17th century did they slowly begin to merge. Blaise Pascal and Gottfried Wilhelm Leibniz, two brilliant mathematicians, built the first mechanical calculators. Leibniz’s machine was particularly groundbreaking because it was based on the binary system—the language of zeros and ones that, more than 250 years later, became the foundation of the digital age. These early automata gave physical form to logical processes for the first time.
The culmination of this early mechanical era was Charles Babbage’s unfinished work in the 19th century. His Analytical Engine, powered by steam and programmed using punched cards, was effectively the design for the first computer. Yet it was his collaborator, Ada Lovelace, who recognized the device’s true, revolutionary potential. She understood that Babbage’s machine could manipulate not only numbers, but any symbols—letters, musical notes, and images. In her 1843 notes, she predicted that such a device would be able to “compose elaborate and scientific pieces of music of any degree of complexity.” This was the first vision in history of a general-purpose computer—a machine capable of simulating mental processes. The dream of the Golem and Aristotle’s logic had finally found a common language.
Photo: Geni: https://en.wikipedia.org/wiki/Difference_engine
Founding Fathers – Turing and Dartmouth. The Turbulent History of Artificial Intelligence
Only the upheaval of World War II and the emergence of the electronic computer transformed these early ideas into a real scientific discipline. At the center of this transformation stood one man: Alan Turing, the British mathematician whose brilliant mind shaped the twentieth century. His work in breaking the German Enigma code not only helped the Allies win the war but also gave him unique insight into the capabilities of computing machines. As early as 1936, long before the first computer was built, Turing described in theory a “universal computing machine”—an abstract device that, given the appropriate program, could perform any computational task. This was theoretical proof that it was possible to create a machine capable of solving any problem.
Artistic Vision of the Turing Machine
In 1950, in an article entitled “Computing Machinery and Intelligence,” Turing posed the question that defined the entire field: “Can machines think?” Realizing that the definitions of “thinking” and “intelligence” were philosophically treacherous ground, he proposed a brilliantly simple way around the problem: the “imitation game,” known today as the Turing test. In the test, a judge engages in a text-based conversation with two hidden participants—a human and a machine. If, after five minutes, the judge is unable to say with certainty which of them is the machine, then the machine should be considered intelligent. This approach is also known as the Turing test. Turing did not ask about consciousness or the soul; he asked about the ability to engage in convincing conversation. He was also an optimist, predicting that by the year 2000, computers would be able to fool judges in 30% of cases.
Although Turing provided the philosophical and theoretical framework, artificial intelligence as a field of study was born six years later, during the hot summer of 1956 on the campus of Dartmouth College. It was not a typical conference with lectures and presentations, but rather a weeks-long, informal “brainstorming session” convened by four visionaries: John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon.
Photo: Raynaudmarc
The very creation of the field was a deliberate act of invention, and even of marketing. John McCarthy, a young mathematician, deliberately coined the term “artificial intelligence” to distinguish his endeavor from cybernetics, an already existing field dominated by the figure of Norbert Wiener. Cybernetics focused on feedback loops and control mechanisms, whereas McCarthy and his colleagues had a far more ambitious goal: to recreate higher functions of the human mind, such as creativity, abstraction, and language. The name “artificial intelligence” was catchy, futuristic, and perfectly suited to attracting funding.
In their proposal to the Rockefeller Foundation, the group formulated a sentence that became both the creed and the curse of the entire field for decades to come: “The study will proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be described so precisely that a machine can be built to simulate it.” This was the height of optimism. The founding fathers believed that, over the course of a single summer and with ten scientists, they would make significant progress toward solving the problem of intelligence. They could not have been more mistaken about the scale of the challenge, but their audacious vision ignited a spark that still burns today.
The Golden Years of Artificial Intelligence and the First Disappointments – ELIZA, Shakey, and the Collision with Reality
After the Dartmouth conference, AI entered a golden age. Fueled by optimism and generous funding, primarily from the U.S. Department of Defense, scientists began creating programs that seemed to confirm the founders’ bold claims. Allen Newell and Herbert Simon developed the Logic Theorist, which independently proved mathematical theorems, astounding even its creators. John McCarthy created LISP, an elegant programming language that became AI researchers’ primary tool for decades. It seemed that machines were indeed beginning to think.
In 1966, a program appeared on the scene that forever changed the public perception of AI. Joseph Weizenbaum of MIT created a chatbot named ELIZA. Its mechanism was deceptively simple. The program pretended to be a Rogerian psychotherapist, using active listening techniques. It identified keywords in the user’s sentences (e.g., “father,” “problem”) and transformed them into open-ended questions (“Tell me more about your family”) or repeated fragments of the speaker’s statements. ELIZA did not understand a single word of the conversation.
The reaction surpassed Weizenbaum’s wildest expectations—and terrified him. People, including educated MIT employees, began treating ELIZA as a real confidant. They shared their deepest fears and secrets with her, spending hours in “therapeutic” sessions. A famous anecdote tells of Weizenbaum’s secretary, who, after a brief conversation with the program, asked him to leave the room because she wanted to be alone with ELIZA. Weizenbaum created ELIZA to expose the superficiality of human-machine communication, but instead he discovered the “ELIZA effect”: a deeply rooted human tendency to attribute understanding and empathy to machines, even when there is no basis for doing so. Shocked by his creation, he became one of AI’s most outspoken critics, arguing that there are spheres of human experience, such as therapy, that should never be automated.
At the same time that ELIZA easily simulated human conversation in the sterile world of text, another project showed how monumentally difficult it is to transfer intelligence into the physical world. It was Shakey, the first mobile robot capable of perceiving and reasoning about its surroundings, built at Stanford Research Institute (SRI) between 1966 and 1972. Shakey, whose name came from its shaky, uncertain movement (“it shakes like hell”), was equipped with a television camera and tactile sensors. It could analyze its surroundings, create an internal map of a room containing blocks and ramps, and then, using planning software called STRIPS, execute complex, multi-step commands such as “move the block off the platform.”
These two parallel projects exposed a fundamental problem with early AI. It turned out that simulating intelligent conversation was relatively simple if one could limit oneself to manipulating symbols. In contrast, performing even the simplest tasks in the real world required solving extraordinarily complex problems of perception, planning, and coping with unpredictability. People were easily deceived by the illusion of a mind in ELIZA, while Shakey’s slow, clumsy movements showed just how vast the gap was between the “toy worlds” of early programs and the complexity of reality. This gap between promises and actual capabilities would soon lead to the first major downturn in the field of AI.
Long Winters and Short Summers – Cycles of Hype and Downfall
By the end of the 1960s, the initial, almost boundless optimism began to give way to disappointment. It turned out that methods that worked excellently for solving simple, logical problems failed completely when confronted with more complex challenges. The main culprit was the so-called “combinatorial explosion”
Combinatorial Explosion
As the complexity of the problem increased, the number of paths that could be analyzed grew at an astronomical rate, exceeding the capabilities of computers at the time. Promises to create a fully intelligent machine within a decade proved to be a pipe dream.
We can illustrate this using the example of a game of chess:
- In the early stages of the game, the number of possible moves is relatively small.
- With each successive move, the number of possible continuations increases at an enormous rate. After just a few moves, the number of potential paths is already so large that even today’s supercomputers have difficulty analyzing them fully.
In a nutshell, early AI was based on decision paths.
As a result, attempts to create artificial intelligence based on exhaustively searching through all possible solutions proved infeasible for computers at the time. This phenomenon was one of the main reasons why the promises of rapidly creating fully intelligent machines in the 1950s and 1960s were not fulfilled. It was only the development of heuristic algorithms (which do not analyze all options but select the most promising ones) and the increase in computing power that made it possible to address this problem in some more limited areas.
In 1973, concerned about the lack of practical results, the British government commissioned mathematician Sir James Lighthill to prepare a report on the state of AI research. His conclusions were devastating. Lighthill stated that AI had failed to live up to its grand promises and had no chance of solving real-world problems in the foreseeable future. The report, combined with similar pressure from the U.S. Congress, led to drastic cuts in research funding. The first “AI winter” began—a period of stagnation in which enthusiasm waned and research funding almost completely disappeared.
Expert systems based on artificial intelligence.
The thaw arrived in the early 1980s with the commercial success of “expert systems.” This was a pragmatic departure from dreams of general, human-like intelligence in favor of creating programs specialized in narrow fields. These systems, such as MYCIN for diagnosing bacterial infections and XCON for configuring computer systems, were based on hundreds or even thousands of “if-then” rules that encoded the knowledge of human experts. For the first time, AI became a profitable business, and by the middle of the decade, the market was worth more than one billion dollars.
However, this brief period of prosperity also rested on fragile foundations. Expert systems proved to be “brittle” – they could make absurd mistakes when faced with situations outside their strict rules. They were also extremely expensive to create and update, requiring the painstaking work of “knowledge engineers.” At the same time, powerful and relatively inexpensive workstations from companies such as Sun Microsystems entered the market, making specialized, expensive AI computers—so-called LISP machines—obsolete almost overnight. In 1987, the LISP machine market collapsed, bringing the entire industry down with it. The second, though less severe, AI winter had arrived.
These cycles of boom and bust were not merely the result of technological or financial problems. They were a symptom of the failure of the entire dominant paradigm, known as “Good Old-Fashioned AI” (GOFAI). This approach, originating directly from the Dartmouth conference, assumed that intelligence could be achieved by manually programming symbolic representations of the world and logical rules. The AI winters were painful proof that this model had fundamental limitations. Manually encoding knowledge about the world is an impossible task, and systems based on rigid rules cannot learn, adapt, or cope with the ambiguity of the real world. These failures forced the field to make a painful but necessary shift toward an entirely new idea: instead of programming intelligence, machines could be allowed to learn it themselves.
Turning Point – Kasparov vs. Deep Blue
In May 1997, the world held its breath. In New York, Garry Kasparov, the reigning world chess champion and considered by many to be the greatest player in history, sat down for a rematch against his opponent—the IBM supercomputer known as Deep Blue. This was not an ordinary game of chess. It was a carefully orchestrated spectacle, a clash between man and machine, a symbolic duel between human intuition, creativity, and strategic depth and the cold, relentless logic of silicon.
Deep Blue was not intelligent in the human sense of the word. It did not “understand” chess, feel the beauty of a combination, or possess intuition. It was, however, a computational monster, a machine built for one purpose: analyzing chess positions. Thanks to its massively parallel architecture, it was able to evaluate up to 200 billion positions per second. Its “intelligence” lay in brute computational force—in its ability to look far enough into the future of a game to choose the move with the highest statistical value.
A year earlier, in 1996, Kasparov won the first match by leveraging his unpredictability and changing his playing style to confuse the machine. However, IBM had significantly improved its program for the rematch. The match was fiercely contested and highly dramatic. Kasparov was under enormous psychological pressure. At one point, he accused the IBM team of cheating when the computer made a move that seemed too “human” to him, too subtle for a machine. The climax came in the sixth and decisive game. After just 19 moves, in a position that many analysts still considered defensible, Kasparov, clearly demoralized, resigned. For the first time in history, the reigning world chess champion had been defeated by a computer under tournament conditions.
Deep Blue’s victory was a symbolic earthquake. To the general public, it was proof that artificial intelligence was catching up with, and even surpassing, humans. In reality, however, the event meant something else. It symbolically marked the end of the romantic vision of AI as an attempt to create a machine that thinks and reasons like a human. Deep Blue did not win because it was smarter than Kasparov in the human sense. It won because it was unimaginably faster at calculating.
This moment forever changed the metric of success in AI. The question was no longer, “Does the machine think like a human?” but instead, “Is the machine capable of defeating a human at a given task?” This pragmatic shift in focus from process to outcome opened the door to an entirely new approach that would come to dominate the 21st century. The future of AI no longer lay in elegant logical systems, but in the power of data and computing capacity. After the fall of chess, the last bastion of human intellect, researchers set themselves a new goal—the ancient Chinese game of Go, orders of magnitude more complex than chess, considered impossible for machines to crack. They did not yet know that conquering it would take them less than two decades.
The Revolution of the 21st Century – Three Eras That Changed Everything
After years of stagnation and false dawns, conditions for a true revolution began to mature at the dawn of the 21st century. Three powerful forces converged, creating a perfect storm that brought the AI winters to an end forever. First, Big Data: the explosion of the internet, social media, and mobile devices generated previously unimaginable amounts of digital data—text, images, sounds—that could serve as fuel for algorithms. Second,
Enormous Computing Power: graphics processing units (GPUs), originally designed with computer gaming in mind, turned out to be ideal for the parallel processing of mathematical operations underlying neural networks, giving researchers access to inexpensive supercomputing power. Third,
Better Algorithms: concepts in neural networks that had lain fallow for years were refined and ready to scale.
Era I: Deep Learning (approx. 2010–2017)
The breakthrough came in 2012 and had a very specific name: ImageNet. It was an annual competition in object recognition on millions of images, a kind of world championship in the field of computer vision. That year, a team from the University of Toronto, led by one of the “godfathers” of AI, Geoffrey Hinton, entered a neural network called AlexNet in the competition. Its result was a shock. AlexNet achieved an error rate of 15.3%, while the second-best team, using traditional methods, achieved 26.2%. This was not a minor improvement; it was a crushing victory. In an instant, the entire field of computer vision, and soon AI as a whole, changed course.
What set AlexNet apart was a technique called deep learning. Unlike symbolic AI, where rules were manually programmed by humans, deep learning relies on multilayered neural networks that learn to recognize patterns on their own from a huge number of examples. This can be compared to the way a child learns to recognize a cat: not by learning definitions (“it has fur, four legs, whiskers”), but by looking at thousands of different cats until their brain independently extracts the key features. Neural networks work in a similar way, automatically learning a hierarchy of features—from simple edges and colors in the first layers to complex shapes, such as eyes or ears, in deeper layers.
The deep learning revolution was not, in essence, a philosophical breakthrough in our understanding of intelligence. It was primarily an engineering and resource breakthrough. It demonstrated that, given sufficient data and computing power, relatively simple, scalable architectures can outperform decades of work by human experts attempting to manually encode knowledge about the world. It was the ultimate confirmation of the path indicated by Deep Blue: quantity of data has its own unique quality. In the new era of AI, the most important skill ceased to be the creation of clever rules and became the accumulation of gigantic datasets and the construction of powerful computing clusters.
Era II: Transformers (2017–2022)
Deep learning has revolutionized perception-related tasks, such as image and speech recognition. However, natural language processing (NLP) remained a challenge. Older models, such as recurrent neural networks (RNNs), processed text sequentially, word by word, which made it difficult for them to handle long-range dependencies and prevented them from efficiently leveraging the power of GPUs.
In 2017, a group of researchers from Google published a paper with the prophetic title: “Attention Is All You Need.” In it, they presented a new neural network architecture called
Transformer. At its heart was the “self-attention” mechanism, which allowed the model, while processing one word, to simultaneously analyze and weigh the meaning of all the other words in the sentence or even the entire document. This gave the model a deep, contextual understanding of language. In the sentence “The chicken did not cross the street because he was too tired,” the attention mechanism allowed the model to understand that “he” refers to “the chicken,” not “the street.”
The Transformer proved to be not just another model, but a new, fundamental Lego block for modern AI. Its scalability and efficiency in processing relationships within data were so revolutionary that it became the default architecture not only for language, but also for images, audio, and other types of data. It was precisely this architectural breakthrough that enabled the construction of increasingly larger models and led directly to the next era.
Two main development paths emerged from the Transformer architecture, dominating the NLP landscape:
- BERT (Google, 2018): A model based exclusively on the Transformer encoder, designed to understand language. Its key feature was bidirectionality—it analyzed the entire sentence at once, considering both the words preceding and following a given word. This made it extremely effective at tasks such as information retrieval, sentiment analysis, and question answering.
- GPT (OpenAI, 2018 and later): A model based exclusively on the decoder, designed to generate language. It operated autoregressively, meaning it predicted the next word in a sequence based on the preceding words. This made it a powerful tool for writing text, from simple sentences to complex essays and code snippets.
The discovery of deep learning provided the method (learning from data), while the Transformer architecture provided the tool—a highly efficient and scalable structure that made it possible to apply this method on a previously unimaginable scale.
Era III: The Generative Explosion (2022–Present)
The technology that drove today’s revolution had been maturing for several years (the GPT-3 model was created as early as 2020), but it was primarily available to researchers and developers. The year 2022 changed everything, as AI entered people’s homes for the first time in history. This happened thanks to two groundbreaking products: DALL-E 2, an image generator capable of creating photorealistic and artistic graphics based on simple text prompts, and ChatGPT, a chatbot based on the GPT-3.5 language model.
ChatGPT’s success was not merely a triumph of technology, but above all a triumph of user interface design. OpenAI packaged the powerful yet complicated technology of large language models (LLMs) into the most natural and intuitive interface imaginable: conversation. It was the “iPhone moment” for artificial intelligence. Just as the iPhone did not invent the touchscreen or mobile internet, but combined them into a product that anyone could use, ChatGPT reduced the barrier to entry for advanced AI to zero. Anyone who knew how to type on a keyboard could now use one of the most powerful AI models in the world.
The result was a social and economic shockwave. Within a few months, ChatGPT became a work tool for millions of people, assisting with writing emails, creating marketing content, programming, and education. This triggered an unprecedented investment boom and sparked a corporate arms race among technology giants such as Microsoft, Google, and Meta.
At the same time, this sudden democratization of AI forced society to confront its darker side. A global debate erupted over ethics and threats, including mass unemployment in creative and knowledge-based professions, disinformation driven by realistic deepfakes, copyright issues surrounding training data, the perpetuation of biases embedded in that data, and privacy concerns. Artificial intelligence ceased to be an abstract concept confined to research laboratories. It became a powerful, ubiquitous, and deeply controversial tool shaping our reality.
Summary
The history of artificial intelligence has come full circle. Ancient dreams of thinking automata and the bold visions of the founding fathers at Dartmouth have finally, in a strange and unexpected way, begun to materialize. Yet this has not happened through elegant logic and precise modeling of the human mind, as originally assumed, but through statistical approximations trained on continent-sized datasets and processed by server farms consuming as much energy as small cities.
Another troubling motif also recurs in this story: the creator’s remorse. In the 1960s, Joseph Weizenbaum turned away from his creation, ELIZA, in horror as he saw how easily people were deceived by the illusion of understanding. Half a century later, his concerns have returned in a far more powerful form. Deep-learning pioneers such as Geoffrey Hinton publicly express regret and warn of the existential threat their work may pose to humanity. It turns out that AI’s greatest prophets are often its deepest skeptics.
The paradox of modern AI is that as machines become increasingly capable, our understanding of their inner workings becomes increasingly murky. Early symbolic systems were transparent; their logical reasoning could be traced step by step. Today’s giant neural networks are “black boxes.” Even their creators cannot fully explain why a model made one decision rather than another. We are creating systems that surpass us in an ever-growing number of fields without fully understanding how they do it.
This lack of transparency is leading to a deep divide among experts themselves. The optimism of the founders in the 1950s was almost unanimous. Today, the field is split—some foretell a technological utopia, while others, like Hinton, warn of extinction. Despite incredible progress, fundamental questions remain unanswered. We are closer than ever to passing the Turing test, but are we any closer to understanding what intelligence and consciousness are, and what it means to be human? The long, behind-the-scenes history of artificial intelligence has come to an end. The public chapter of this story is only just beginning.