The AI Winters

Alan Turing (1912–1954) stands as the conceptual progenitor, though the exhibit honors the unglamorous researchers who persisted through winter: Marvin Minsky and John McCarthy (founders of MIT AI Lab, 1956), who survived the first winter by pivoting to expert systems; Geoffrey Hinton, Yann LeCun, and Yoshua Bengio, whose deep learning revival (2006 onward) ended the second winter. The true hero is institutional patience—the Defense Advanced Research Projects Agency (DARPA) and academic departments that maintained skeleton crews when venture capital fled.

The first winter (1974–1980) exposed the brittleness of early symbolic systems: LISP machines required hand-coded knowledge, could not learn from data, and demanded computational resources that 1970s hardware could not supply. The Lighthill Report (1973, UK) and subsequent U.S. congressional skepticism cut funding because systems like ELIZA and early expert systems overpromised. The second winter (1987–1993) followed the collapse of the expert-systems market—companies like Symbolics and Lisp Machines Inc. failed when rule-based systems proved unmaintainable and inflexible. Both winters ended when new architectures emerged: connectionist models (neural networks) in the 1980s–90s, and crucially, the availability of large datasets and GPU computing (2006 onward). The transformer architecture (2017) unified attention mechanisms with scalable training, finally delivering on the promise of learned representations rather than hand-coded rules.

The Age of AI Revolutions spans 1950–present, but the two winters—1974–1980 and 1987–1993—are the exhibit's temporal anchors because they reveal the cyclical nature of technological promise and disillusionment. The first winter followed the Dartmouth Summer Research Project on Artificial Intelligence (1956), which had promised 'significant progress' within a generation. By 1974, DARPA withdrew funding after symbolic systems (MYCIN, DENDRAL) proved narrow, expensive, and unable to scale. Researchers retreated to 'AI in the basement'—small labs at MIT, Stanford, and CMU kept the field alive through the doldrums. The second winter was triggered by the collapse of the Lisp machine market (1987–1990) and the failure of the Fifth Generation Computer Project (Japan's ambitious 1982–1992 initiative). The winter ended not with a single breakthrough but with a convergence: the availability of large datasets (ImageNet, 2009; Wikipedia), the realization that neural networks could be trained on GPUs (Hinton's 2006 deep belief networks), and the gradual shift from symbolic to statistical reasoning. The transformer (Vaswani et al., 2017) and the subsequent scaling laws (Kaplan et al., 2020) finally vindicated the connectionist vision—but only after decades of winter.

The AI Winters existed because the field had made promises it could not keep. The early symbolic approach—encoding human knowledge as logical rules—hit a fundamental wall: the 'knowledge representation problem' (how do you encode common sense?) proved intractable, and the computational costs of search-based reasoning grew exponentially. Funding agencies and corporations expected rapid returns; when ELIZA (Weizenbaum, 1966) failed to cure depression and early machine translation proved inadequate, confidence evaporated. The winters also reflected deeper theoretical confusion: the field lacked a unifying principle. Symbolic AI and connectionism were treated as rival paradigms rather than complementary approaches. The second winter was particularly brutal because the expert-systems bubble had created unrealistic expectations in the business world. When Symbolics and Lisp Machines Inc. collapsed, the entire field was tainted by association with failed startups. The winters were necessary—they forced a reckoning with what was actually possible and what was hype. They also created space for patient, foundational work: the theoretical development of backpropagation, the study of learning theory, and the slow accumulation of computational resources that would eventually enable the deep learning era.

During the first winter (1974–1980), AI researchers worked in relative obscurity. Graduate students at MIT and Stanford pursued connectionist models despite the prevailing skepticism toward neural networks. DARPA-funded projects shifted toward narrower, more defensible goals: robotics, natural language processing for military intelligence, and specialized expert systems for medical diagnosis. In industry, the few companies that survived (Teknowledge, Inference Corporation) sold expensive expert-system shells to large corporations, where domain experts spent months encoding rules. The second winter (1987–1993) was more visible but equally grim: AI researchers rebranded themselves as 'knowledge engineers' or 'machine learning' researchers to distance themselves from the tainted 'AI' label. Academic departments maintained small groups; funding came from NSF, ONR (Office of Naval Research), and a few far-sighted corporations like Bell Labs and Xerox PARC. The daily work was unglamorous: implementing algorithms from papers, running experiments on workstations, publishing in specialized conferences (ICML, founded 1988; NIPS, founded 1987) that few outside the field read. The turning point came around 2006, when Hinton's deep belief networks showed that neural networks could learn useful representations from unlabeled data—suddenly, the daily work of AI researchers became visible again, and funding returned.

The AI Winters were not constructed but rather emerged from the collision of technological limits and market expectations. The first winter (1974) was triggered by a specific policy decision: the U.S. Congress, responding to the Lighthill Report (a damning 1973 assessment of AI progress in the UK), cut DARPA's AI budget from ~$20 million annually to ~$2 million. This forced the closure of AI labs at universities and corporations, leaving only the most committed researchers. The second winter was more organic: the expert-systems market collapsed when companies realized that maintaining rule-based systems was economically unsustainable. The 'AI Winter' concept itself was not formalized until the late 1980s, when researchers began to speak openly about the field's crisis. The construction of the recovery (2006 onward) involved: (1) the availability of large labeled datasets (ImageNet, 2009); (2) the realization that GPUs could accelerate neural network training (Hinton's 2006 work); (3) the rise of deep learning conferences (ICML, NIPS, ICLR) that created a separate identity from the discredited 'AI' label; (4) the shift from academic funding to corporate investment (Google, Facebook, OpenAI, DeepMind) starting around 2010. The transformer architecture (2017) was not a response to the winters but rather the culmination of decades of patient theoretical work that had continued despite them.

The winters were not uniform across all AI subfields. Machine learning and probabilistic reasoning (Judea Pearl's Bayesian networks, 1980s) were less affected because they made more modest claims and delivered measurable results. Robotics, funded separately by DARPA and NSA, continued to advance (though slowly) through both winters. Natural language processing suffered most during the first winter but recovered earlier during the second, as statistical methods (n-grams, hidden Markov models) proved more practical than symbolic parsing. Computer vision was nearly dormant during both winters, revived only with the availability of large datasets and deep learning (2012 onward). The geographic variation was significant: the U.S. experienced both winters acutely; Japan's Fifth Generation Computer Project (1982–1992) was a separate phenomenon, driven by government ambition rather than market forces, and its failure was also a winter. The UK and Europe experienced the first winter most severely, with the Lighthill Report effectively ending government funding for AI research for a decade. China and the Soviet Union had separate AI programs with different funding cycles, less visible to Western observers.