History of Artificial Intelligence

AI had a theory before it had a name. In 1943 Warren McCulloch and Walter Pitts published 'A Logical Calculus of Ideas Immanent in Nervous Activity,' the first mathematical model of an artificial neuron. They showed that simple binary neurons wired into networks could, in principle, compute any function a Turing machine could compute. Every modern neural network traces its conceptual lineage to that paper. In 1950 Alan Turing reframed the central question in 'Computing Machinery and Intelligence,' arguing that 'Can machines think?' was the wrong thing to ask. What mattered was whether a machine could convincingly imitate human conversation, an idea now known as the Turing Test.

These two contributions set the two poles that the field still works between: a bottom-up account of intelligence as something networks of simple units learn, and a behavioral test that judges intelligence by what a system does rather than how it is built. Neither author had the computing power to test their ideas at scale, so for more than a decade the work stayed mostly on paper. That gap between a clear idea and the hardware to run it is a pattern that recurs throughout the history of artificial intelligence.

Artificial intelligence became a named discipline at a two-month workshop at Dartmouth College in the summer of 1956. John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon organized it, and the proposal stated 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 gathering coined the term 'artificial intelligence' and brought together most of the people who would lead the field for the next two decades. A year later, in 1957, Frank Rosenblatt built the Mark I Perceptron at Cornell, the first hardware implementation of a neural network, able to learn to classify simple visual patterns.

The optimism of this period was real and, in retrospect, overstated. The New York Times reported the Perceptron as an 'Electronic Brain' that the Navy expected would one day walk, talk, see, and be conscious of its existence. The ambition set at Dartmouth drove decades of funded research, but it also created promises the technology could not keep. The overpromise-and-underdeliver cycle that started here would repeat, and it is the direct cause of the funding collapses that followed.

The first reversal came from inside the field. In 1969 Marvin Minsky and Seymour Papert published 'Perceptrons,' proving mathematically that a single-layer perceptron could not solve the XOR problem or other tasks that are not linearly separable. The proof was correct but widely overgeneralized: multi-layer networks could solve those problems, yet the book was read as proof that neural networks were fundamentally limited. Funding and researchers moved elsewhere, and neural network research stalled for over a decade. In 1973 the Lighthill Report delivered a broader verdict in the United Kingdom, concluding that 'in no part of the field have the discoveries made so far produced the major impact that was then promised,' which triggered deep funding cuts.

What went wrong was not the underlying ideas but the gap between narrow successes and the general intelligence that had been promised. When a field sells human-level capability and delivers pattern classifiers, a backlash follows. The damage was real: important work, including the backpropagation algorithm, sat largely ignored through the anti-neural-network climate of the 1970s. The lesson that inflated expectations invite funding collapse is the throughline of both AI winters, and it is worth studying on its own.

The recovery began with a training breakthrough. In 1986 David Rumelhart, Geoffrey Hinton, and Ronald Williams published 'Learning Representations by Back-propagating Errors' in Nature, showing that backpropagation could train multi-layer networks effectively and reviving the line of research the 1969 critique had buried. Twenty years later, in 2006, Hinton's work on deep belief networks showed that deep networks could be trained layer by layer, often treated as the starting gun for modern deep learning. The decisive proof arrived in 2012, when AlexNet, built by Alex Krizhevsky, Ilya Sutskever, and Hinton, won the ImageNet competition by a large margin and cut the error rate sharply while training on two consumer GPUs.

AlexNet settled the question of whether deep learning worked at scale, and the field reorganized around it almost overnight. In 2016 DeepMind's AlphaGo beat Lee Sedol four games to one at Go, a game long considered decades away from being solved, and its Move 37 showed the system could be creative rather than merely fast. Earlier, in 1997, IBM's Deep Blue had beaten world chess champion Garry Kasparov, but that was brute-force search rather than learning, a distinction that mattered. The 2006 to 2017 span is where AI moved from promising research to systems that outperformed the best humans in hard domains.

In June 2017 eight researchers at Google published 'Attention Is All You Need,' introducing the Transformer architecture. It replaced recurrence with self-attention, which let models process whole sequences in parallel and scale far beyond what came before. The Transformer is the architecture behind GPT, BERT, Claude, Gemini, and nearly every frontier system since, which makes it the most consequential AI paper of the 2010s. In 2020 OpenAI released GPT-3 with 175 billion parameters, a hundred times larger than its predecessor, and without fine-tuning it could write essays, code, and poetry from a few examples in the prompt. GPT-3 gave the scaling hypothesis, the idea that bigger models would simply get more capable, real credibility.

The public arrived in November 2022. ChatGPT, built on GPT-3.5 and tuned with reinforcement learning from human feedback, reached one million users in five days and a hundred million in two months, the fastest-growing consumer application on record. GPT-4 followed in March 2023 as a multimodal model that read text and images, passed the bar exam near the 90th percentile, and forced every industry to reckon with what these systems could do. This is the stretch where AI stopped being expert knowledge and became something most people had used.

By 2025 the frontier shifted from better answers to systems that could act. Frontier models were wrapped in software that could browse the web, call tools, edit files, execute code, manage state, and carry multi-step tasks forward with limited supervision. Claude Code, OpenAI's Operator, Google's Project Mariner, and a wave of agent frameworks turned 'AI agent' from a research label into a product category. The change was less about conversation quality and more about systems that observe, plan, and act across real software environments.

That shift raised sharper questions than the chatbot era did: permissions, oversight, reliability, and what happens to knowledge work when software can execute rather than only suggest. Read against the full arc, the agentic era is a stronger answer to the same questions Turing and the Dartmouth founders posed, namely whether machines can reason, learn, and act usefully in the world. The history of artificial intelligence is still being written, and the open problems now are about trust and control as much as raw capability.