The AI Winters
Two boom-and-bust cycles, and the people who kept working through the frost
Twice now, artificial intelligence has gone from boundless funding to near-total collapse. The pattern repeats: bold promises, government money, a damning report, a freeze. This trail follows both winters end to end - not as cautionary trivia, but because every era of AI optimism, including this one, gets measured against them.
The workshop that named artificial intelligence
The 1955 Dartmouth proposal by McCarthy, Minsky, Rochester, and Shannon coined 'artificial intelligence' and launched the field.
The field is born already optimistic: the Dartmouth proposal suggests a significant advance can be made in one summer if a carefully selected group works on it together.
Simon and Newell's 1957 ten-year predictions
In 1957 Herbert Simon and Allen Newell predicted that within ten years a computer would be world chess champion and prove a new theorem.
Within two years the field's leaders are making dated, falsifiable promises. Some eventually come true - decades late. The pattern of overpromising starts here.
The perceptron: a machine that learns from examples
Frank Rosenblatt's perceptron was an early trainable neural network that adjusted its own connections to classify patterns.
The perceptron arrives with Navy-funded fanfare and newspaper stories about machines that will walk, talk, and be conscious. The hype writes checks the hardware cannot cash.
1966 ALPAC report ends a decade of machine-translation funding
The 1966 ALPAC report found no near-term prospect of useful machine translation and urged cutting funding, collapsing US support for roughly two decades.
The first bill comes due in machine translation: a government committee concludes there is no near-term prospect of useful MT, and funding collapses for twenty years.
Perceptrons and the first neural network winter
Minsky and Papert's book exposed the limits of single-layer perceptrons, helping freeze neural network funding for years.
Then the mathematics: Minsky and Papert prove what single-layer perceptrons can never compute. Fair as written, devastating as read - neural networks become a career dead end.
The 17-year neural network freeze
Minsky and Papert 1969 book Perceptrons exposed what single-layer networks cannot do, helping freeze neural-network research until backprop revived it in 1986.
The freeze lasts seventeen years. The lesson is not that the critics were wrong, but that a true negative result about a narrow case can take down an entire research direction.
The Lighthill Report triggers the first AI winter in the UK
Sir James Lighthill's survey for the UK Science Research Council judged AI a disappointment, cutting British AI funding for a decade.
The UK runs the same play with a single author: Lighthill's report judges AI a disappointment and British funding is cut for a decade. The phrase 'AI winter' now has a textbook case.
XCON/R1, the first big commercial expert system
R1, later called XCON, was a rule-based expert system that configured DEC's VAX computer orders and became one of the first to pay off in business.
The thaw comes from an unexpected direction - not learning, but rules. XCON configures DEC's computer orders and proves AI can pay its own way. The expert-systems boom ignites.
Japan launches the Fifth Generation Computer Systems project
Japan MITI launched a ten-year national project in 1982, run by the ICOT institute, to build knowledge-processing computers based on logic programming.
Japan goes all-in with a ten-year national project, and Western governments panic-fund their own programs in response. The second boom is now geopolitical.
The rise and fall of the Lisp machine industry
Around 1980 the MIT AI Lab spun out Symbolics and LMI to sell purpose-built Lisp computers; both collapsed within a decade as cheap workstations caught up.
An entire hardware industry grows up to run AI workloads - and is then undercut by ordinary workstations that got fast enough. Specialized AI hardware would have to wait for GPUs.
The expert systems boom and bust
In the early 1980s 'knowledge engineering' was sold as AI's future, sparking a wave of expert-systems firms; by the decade's end the market had collapsed.
Expert systems hit their ceiling: brittle rules, expensive maintenance, and knowledge that walks out the door when the expert does. The market collapses by decade's end.
The second AI winter
In the late 1980s the commercial AI boom collapsed as the Lisp-machine market crashed and expert systems failed their hype, starting a long downturn near 1987.
Boom number two ends like boom number one. 'AI' becomes a word grant writers avoid; survivors rebrand as machine learning, informatics, or just software.
CIFAR Funds Deep Learning Through the Winter (2004)
In 2004 Canada's CIFAR launched a program directed by Geoffrey Hinton that sustained Hinton, Bengio, and LeCun through the neural-network winter.
What ends a winter is patient money: a Canadian institute quietly funds Hinton, Bengio, and LeCun to keep working on the unfashionable idea of deep neural networks.
AlexNet wins ImageNet 2012
A deep convolutional neural network crushed the ImageNet contest, proving deep learning could outperform hand-built computer vision.
Eight years later the unfashionable idea wins ImageNet by a margin nobody can ignore. The thaw is instant, and the field has not cooled since - so far.