The idea that a machine might think is older than computers themselves. In the 17th century, Leibniz imagined a calculus of reasoning that could settle any dispute through mechanical calculation. In the 19th century, Ada Lovelace speculated about the capabilities and limits of Charles Babbage’s proposed Analytical Engine. But the dream did not become a practical research program until the mid-20th century, when electronic computers made it possible to build machines that could, for the first time, manipulate symbols fast enough to simulate something resembling thought.
The history of artificial intelligence is a story of enormous ambition, repeated disappointment, and eventual, unexpected success. The thinking machines we have today bear little resemblance to what the pioneers imagined. But they work, and their capabilities are advancing faster than almost anyone predicted.
Colossus: The Secret Origin (1943)
The first electronic computers were not built to think. They were built to win a war. Colossus, designed by Tommy Flowers and built at the Post Office Research Station in London, was operational by December 1943. It was used at Bletchley Park to break the Lorenz cipher, a high-level German encryption system used for strategic military communications.
Colossus was not a general-purpose computer. It was designed for a single task: finding the settings of the Lorenz cipher machine by testing thousands of possible configurations at electronic speed. But it demonstrated that electronic machines could process information faster than any human, and its existence (kept secret until the 1970s) spurred the development of more versatile electronic computers after the war.
Alan Turing, who worked at Bletchley Park on the related problem of breaking the Enigma cipher, was deeply influenced by his wartime experience. The cryptanalytic machines he helped design were not Colossus (that was Flowers’s achievement), but the intellectual environment at Bletchley convinced Turing that electronic machines could do far more than break codes. By 1950, he was asking the question that would define the field of artificial intelligence: can a machine think?
Turing’s Question (1950)
In his 1950 paper “Computing Machinery and Intelligence,” Turing proposed replacing the vague question “Can machines think?” with a concrete test. A human interrogator communicates by text with two hidden respondents: one human and one machine. If the interrogator cannot reliably distinguish the machine from the human, the machine should be considered intelligent. This became known as the Turing test.
Turing also anticipated and rebutted nearly every objection to machine intelligence that would be raised over the following decades: the theological objection (machines have no souls), the mathematical objection (Gödel’s incompleteness theorems limit what machines can prove), the consciousness objection (machines cannot truly understand), and the creativity objection (machines can only do what they are programmed to do). His paper remains one of the most prescient documents in the history of computing.
Turing died in 1954, before the field he envisioned had a name. But his ideas shaped everything that followed.
The Birth of AI (1956)
The term “artificial intelligence” was coined at the Dartmouth Conference in the summer of 1956, organized by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. The proposal for the conference stated, with remarkable optimism, that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.”
The early AI researchers pursued what is now called symbolic AI (or “good old-fashioned AI”): the idea that intelligence consists of manipulating symbols according to rules, and that a sufficiently clever program running on a sufficiently fast computer could reproduce any intelligent behavior. They built programs that could prove mathematical theorems, play chess, solve logic puzzles, and carry on simple conversations.
The results were impressive enough to generate enormous enthusiasm. In 1965, the AI pioneer Herbert Simon predicted that “machines will be capable, within twenty years, of doing any work a man can do.” This prediction proved spectacularly wrong, but the optimism it reflected drove funding and research for years.
The AI Winters (1970s and 1980s)
By the early 1970s, the limitations of symbolic AI were becoming apparent. Programs that worked brilliantly on toy problems failed completely on real-world tasks. Language translation, visual recognition, and common-sense reasoning proved far harder than the pioneers had expected. The gap between what AI could do in a laboratory demonstration and what it could do in practice was enormous.
Funding agencies noticed. The Lighthill Report (1973) in Britain concluded that AI had failed to deliver on its promises and recommended cutting funding. Similar skepticism emerged in the United States. The result was the first “AI winter”: a period of reduced funding, diminished interest, and widespread disillusionment.
A second wave of enthusiasm emerged in the 1980s with expert systems, programs that encoded human expertise as sets of if-then rules. Companies invested millions in expert systems for medical diagnosis, financial analysis, and industrial process control. But expert systems proved brittle: they worked well within their narrow domains but failed unpredictably when confronted with situations their rules did not cover. By the early 1990s, the expert systems boom had collapsed, producing a second AI winter.
The Neural Network Revolution (1980s to 2010s)
While symbolic AI was failing to deliver on its promises, a different approach was quietly developing. Neural networks, inspired by the structure of biological brains, consisted of layers of simple processing units (neurons) connected by adjustable weights. Instead of being programmed with explicit rules, neural networks learned by being shown examples and adjusting their weights to minimize errors.
The basic idea dated back to the 1940s (Warren McCulloch and Walter Pitts published a model of artificial neurons in 1943), but early neural networks were too simple and computers too slow to produce useful results. The development of the backpropagation algorithm in the 1980s (discovered independently by several researchers) made it possible to train networks with multiple layers, and progress accelerated.
The real breakthrough came in the 2010s, when three factors converged: vastly more powerful computers (especially GPUs, originally designed for video games), vastly more training data (thanks to the internet), and improved network architectures. Deep learning, the use of neural networks with many layers, began achieving superhuman performance on specific tasks: image recognition, speech recognition, game playing, and translation.
In 2012, a deep neural network trained by Geoffrey Hinton’s group won the ImageNet competition (a benchmark for visual recognition) by a margin that stunned the field. In 2016, DeepMind’s AlphaGo defeated the world champion at Go, a game so complex that brute-force search (the approach that had conquered chess) was impossible. These achievements demonstrated that AI had crossed a threshold: machines could now outperform humans on tasks that had seemed permanently beyond their reach.
Large Language Models (2017 to Present)
The current era of AI was launched by a 2017 paper from Google titled “Attention Is All You Need,” which introduced the transformer architecture. Transformers process sequences of text by attending to relationships between all elements simultaneously, rather than reading one word at a time. This architecture proved extraordinarily effective for language tasks.
OpenAI’s GPT series (Generative Pre-trained Transformer), beginning with GPT-2 in 2019 and culminating in GPT-4 in 2023, demonstrated that transformers trained on enormous amounts of text could generate coherent, contextually appropriate language at a level that startled even the researchers who built them. ChatGPT, released in November 2022, made this capability available to the public and triggered a global conversation about the capabilities and risks of artificial intelligence.
Large language models do not think in the way that Turing imagined. They are statistical engines that predict the next word in a sequence, trained on billions of pages of text. Whether they “understand” language in any meaningful sense is one of the most debated questions in contemporary science and philosophy. What is not debated is their practical capability: they can write essays, summarize documents, translate languages, generate code, and answer questions with a fluency that would have seemed miraculous to the AI pioneers of the 1950s.
The Foundations
The intellectual foundations of artificial intelligence trace back to the same wartime period that produced Colossus and ENIAC. Alan Turing’s theoretical work on computation (the Turing machine, 1936) established that any computable function can be computed by a sufficiently simple machine. John von Neumann’s design for the stored-program computer (the EDVAC report, 1945) provided the practical architecture on which all AI software runs.
Kronecker Wallis’s edition of the EDVAC Report presents the document that defined the computer architecture underlying every AI system in existence. Printed on blue Fabriano paper in monospace type, it is the blueprint for the machines that now generate text, recognize faces, and play games at superhuman levels.
Turing’s own wartime work is preserved in Kronecker Wallis’s edition of the Treatise on the Enigma, the original Bletchley Park manuscript in which Turing described the cryptanalytic methods that were among the earliest applications of machine intelligence. The document, known as “The Prof’s Book” to Turing’s colleagues, reveals the thinking of the man who asked, before anyone else, whether machines could think.
What the Pioneers Got Right and Wrong
The AI pioneers of the 1950s were right that machines would eventually perform tasks requiring intelligence. They were wrong about almost everything else: how long it would take, what methods would work, and what “intelligence” would look like when machines achieved it.
The symbolic AI approach, building intelligence from logical rules, largely failed. The approach that succeeded, training neural networks on massive datasets, was dismissed by the mainstream AI community for decades. The resulting systems do not reason in the way that humans reason. They do not have goals, beliefs, or understanding in any obvious sense. They are, in many ways, as alien as the intelligence that science fiction writers imagined, just in a completely different way than anyone expected.
The history of thinking machines is a reminder that the hardest problems in science are often not the ones we expect. The pioneers thought that chess would be hard and language would be easy. The opposite turned out to be closer to the truth. They thought that intelligence required explicit rules. It turned out to require enormous amounts of data and processing power. The path from Colossus to ChatGPT is not a straight line. It is a winding road through decades of wrong turns, dead ends, and unexpected breakthroughs, a road that is still being traveled.
No products in the cart. 