Summary
This is Part 2 of 6 in the series Complete History of AI: From Turing's Machines to Generative Models. The period between 1974 and 1993 is the most dramatic in the field's history: it includes a genuine commercial renaissance, a $2 billion market boom, and a second collapse more brutal than the first — this time caused not by unfulfilled promises but by platform shift. At the same time, on the field's periphery, a handful of researchers kept the neural network paradigm alive that the world would rediscover decades later.
Context
The First AI Winter (1974-1980) was not death — it was selective hibernation. Funding for speculative general AI research dried up, but researchers working on specific, verifiable applications continued. The context that enabled the renaissance was not a new theoretical breakthrough — it was a change of question. Instead of "how do we make machines think in general," researchers began asking "how do we capture a human expert's knowledge in a narrow domain and make it accessible via computer." That question had a viable answer with 1980s technology. And it had an immediate market: companies that needed consistent, fast, scalable specialized decisions in domains like medical diagnosis, equipment configuration, and financial analysis.
— 1980-1987: The Expert Systems Boom
The first National Conference on Artificial Intelligence (AAAI) at Stanford in August 1980 gathered about 300 researchers and signaled that the field was finding focused practical applications. What emerged as the dominant paradigm were expert systems — programs that encoded human expert knowledge into rule bases and used inference mechanisms to solve problems in specific domains. MYCIN, developed at Stanford in the 1970s, had proven the concept: it diagnosed bacterial blood infections with accuracy comparable to specialist physicians. XCON, developed by Digital Equipment Corporation in 1980, automatically configured VAX computer systems, saving the company tens of millions of dollars annually. These two cases were convincing enough to trigger a boom. Companies like DuPont, IBM, and General Electric created dedicated expert systems departments. In Japan, the government launched the ambitious Fifth Generation Computer Systems Project (1982-1992), investing billions of yen in AI and logic-based computers using PROLOG. This triggered competitive responses in the US and Europe. The global expert systems market grew from practically nothing in 1980 to $2 billion in 1988. Companies like Symbolics and Lisp Machines Inc. built specialized hardware for running AI programs — Lisp Machines cost ten times more than conventional workstations but were the state of the art for expert systems.
— 1986-1987: Hinton and the Neural Network Renaissance
While expert systems dominated commercial applications, a quiet revolution was occurring in academic research. David Rumelhart, Geoffrey Hinton, and Ronald Williams, working at different institutions in the US and Canada, published seminal works on backpropagation in 1986 — the algorithm for training multilayer neural networks. The basic idea of backpropagation had been independently developed by several people in previous decades, but these researchers demonstrated its practical effectiveness and popularized it. The algorithm worked by propagating errors backward through the network, adjusting connection weights to minimize the difference between predicted and desired outputs. This solved the fundamental problem Minsky and Papert had identified in 1969: multilayer networks could now learn complex hierarchical representations of data, overcoming single-layer Perceptron limitations. Hinton had spent the years of the first winter studying cognitive psychology and neuroscience at Cambridge, convinced that the path to artificial intelligence lay in understanding how the biological brain learns — not in encoding explicit rules. His persistence through years of mainstream skepticism produced the algorithm that, combined with sufficient hardware decades later, would make deep learning possible. In 1987, Hinton participated in the first modern workshop on neural networks — a field that had been considered dead a decade earlier.
— 1987-1993: The Second Collapse
The second AI winter arrived not from lack of technical progress but from platform shift — a pattern that technology history would repeat with microcomputers, smartphones, and cloud computing. In 1987, the specialized AI hardware market collapsed in under two years. The cause was simple and brutal: conventional workstations from Apple and Sun Microsystems became sufficiently powerful and dramatically cheaper than Lisp Machines. The specialized hardware that had cost fortunes lost its reason to exist overnight. Symbolics, which had been considered the future of computing in 1985, collapsed. The expert systems market shrank dramatically. DARPA again cut AI funding. The Japanese Fifth Generation project ended without meeting its most ambitious goals. The second winter was more psychologically devastating than the first because it hit a field that had experienced genuine commercial success — not just unfulfilled promises. The structural problem of expert systems became clear in retrospect: knowledge engineering — the process of extracting expert knowledge and encoding it into rules — was expensive, slow, and did not scale. Each new domain required years of manual work. The promise that it could scale to all forms of human knowledge was never verified.
Comparative Analysis
The second winter illuminates a fundamental distinction that AI history would continue to teach: the difference between AI that works because someone encoded explicit knowledge, and AI that works because it learned patterns from data. Expert systems were of the first kind — impressive within domains where knowledge could be articulated as rules, brittle in any situation the knowledge engineer had not foreseen. Hinton's neural networks were of the second kind — too slow to be practical at the time, but fundamentally scalable in a way expert systems never would be. The historical irony is that while the mainstream field bet on the paradigm that would collapse, Hinton and his collaborators were quietly developing the paradigm that would dominate. In 1989, Yann LeCun, then at Bell Labs, demonstrated convolutional neural networks recognizing handwritten digits with precision sufficient for commercial use — one of the first deep learning systems in real production. It was a seed planted in the middle of winter.
Analysis
The 1974-1993 period establishes two critical patterns for understanding AI in 2026. First: AI paradigms that depend on manually encoded knowledge have a low scaling ceiling — they work well in narrow, controlled domains, but the expansion cost grows linearly with scope, making them economically unviable for general knowledge. Second: hardware platform shifts destroy specialized software ecosystems faster than any technical failure — Symbolics was not defeated by a better expert system, it was defeated by general-purpose hardware that became good enough. In 2026, with specialized AI chips (GPUs, TPUs, Huawei Ascend) creating ecosystems similar to the Lisp Machines of 1985, the historical pattern deserves attention.
Synthesis
The most important legacy of the 1974-1993 period is not the boom and collapse of expert systems — it is the survival of the connectionist paradigm through two consecutive winters. Hinton, LeCun, and their collaborators kept alive a paradigm the mainstream had declared dead twice. What sustained them was not institutional funding — it was the conviction that learning from data was fundamentally more powerful than encoding rules. That conviction would be validated decades later with hardware they could not have imagined. The lesson for 2026 is for those evaluating AI paradigms today: what seems impractical for lack of infrastructure may become dominant when the infrastructure arrives. The question is not whether the paradigm works in theory — it is whether the infrastructure necessary for it to work in practice is on the way. In 1987, it was. In 2026, the relevant question is which current paradigm is waiting for infrastructure that does not yet exist. Next: Part 3 — The Productive Silence (1993–2011)
