From boom to bust: the AI winter
Start with the broad timeline
Read this overview of AI's changing expectations and approaches, then try the short practice on historical claims.
Start with: From boom to bust: the AI winter
RedLectern topic
History of AI
Major approaches and milestones · changing expectations · successes and setbacks.
For: Everyone
After this page you can: Place current claims within the field's history.
Page reviewed
Part of: Artificial Intelligence
About History of AI
Artificial intelligence did not begin with a single invention. In 1950, Alan Turing proposed the imitation game as a way to reframe the question “Can machines think?” The 1956 Dartmouth Summer Research Project gave the emerging field a name and a research gathering place. Its proposal expressed a bold expectation: that features of intelligence could be described precisely enough for machines to simulate them.
AI research developed along several paths, including rule-based systems, neural networks, machine learning and robotics. Expectations sometimes outran results: the perceptron’s early promise was followed by reduced funding and a slowdown in AI research in the early 1970s. Rule-based systems later proved useful in specific domains, while another period of declining funding and interest followed unmet expectations in the 1980s. Later milestones, including a chess-playing computer’s 1997 match victory and neural-network results on image classification, showed success on bounded tasks. They did not settle what intelligence means or establish that the same methods work everywhere.
To place a current claim in context, ask what task was demonstrated, which approach was used, what evidence supports the result, and how broad the claim is compared with the historical example.
Start here
Learn the foundations
1.Milestones and testsThe 1956 Dartmouth Summer Research Project on Artificial Intelligence helped establish AI as a named research field.
The 1956 Dartmouth Summer Research Project on Artificial Intelligence helped establish AI as a named research field. Its proposal voiced the expectation that aspects of learning and intelligence could be described precisely enough for machines to simulate them. Alan Turing’s 1950 “Computing Machinery and Intelligence” framed a different kind of question through the imitation game: could an interrogator tell a machine from a person in conversation?
A later, bounded milestone came in 1997, when the chess-playing computer Deep Blue defeated reigning world champion Garry Kasparov in a six-game match. A result on chess is evidence about performance in that task and match setting; by itself, it does not show general ability across other tasks.
- Computing Machinery and Intelligence
Read Turing’s original framing of the imitation game.
- Garry Kasparov vs. Deep Blue
Use a documented chess match to think about what a task-specific result can establish.
Turing test
A historical test based on whether conversation can distinguish a machine from a person.
AI winter
A period of reduced confidence, funding or activity following unmet AI expectations.
Try it
In two or three sentences, explain what Deep Blue’s 1997 match demonstrated and name one conclusion it did not establish.
Deep Blue defeated reigning world chess champion Garry Kasparov in a six-game match in May 1997. That demonstrated success at chess in that match; it did not establish that a chess system could perform every task associated with intelligence.
2.Approaches, expectations and setbacksEarly AI research included perceptrons and neural networks as well as programs using rules to represent and apply knowledge.
Early AI research included perceptrons and neural networks as well as programs using rules to represent and apply knowledge. Those approaches had different strengths and limitations. The Open University’s history describes how high expectations for perceptrons were followed by funding reductions and a slowdown in AI research in the early 1970s. It also describes rule-based systems that made progress on narrower tasks, such as disease diagnosis and analysis of scientific data.
Expectations rose again in the 1980s around advanced rule-based systems. The Open University account says those systems could be hard to maintain and failure-prone, and describes funding and interest dwindling when large programs did not meet expectations. The term AI winter captures a downturn in confidence, funding or activity; it does not mean all AI research stopped or every system failed. The history is better understood as changing expectations and different lines of work than as a simple sequence in which one approach replaced all others.
- A Brief History of AI
A second overview for comparing symbolic, neural-network and learning approaches.
Try it
Name one early AI approach and one reason confidence or funding fell, using the historical overview.
Perceptrons were one early approach. The Open University account says they did not meet high expectations, and that the 1969 book Perceptrons is often credited with helping end their generous funding.
3.Learning from data and changing methodsNeural networks returned as an active line of work after the earlier setbacks around perceptrons.
Neural networks returned as an active line of work after the earlier setbacks around perceptrons. A 2012 paper reported training a deep convolutional neural network to classify images from the ImageNet LSVRC-2010 contest. That result belongs to a particular image-classification task and dataset, rather than serving as a measure of every AI capability.
In 2017, “Attention Is All You Need” proposed the Transformer architecture and reported experiments on machine-translation tasks. These examples help locate shifts in methods: research has included rule-based systems, neural networks, statistical learning and architectures designed for sequence tasks. A milestone can show that an approach worked for a stated task under stated conditions. It does not, by itself, support a broader claim about intelligence or future performance. When reading a current claim, identify the task, evidence and approach before deciding how much the historical comparison tells you.
- Attention Is All You Need
Read the paper that proposed the Transformer architecture and its translation experiments.
Try it
Compare the ImageNet paper and “Attention Is All You Need.” What task did each paper report work on?
The ImageNet paper reported image classification using a deep convolutional neural network. “Attention Is All You Need” proposed the Transformer and reported experiments on machine translation tasks.
Latest
No recent updates are available.
Books
- The Quest for Artificial Intelligence: A History of Ideas and Achievements
The Quest for Artificial Intelligence: A History of Ideas and Achievements
Nils J. Nilsson · 2010
Nilsson’s full text is available from Stanford. It traces AI’s history chronologically and discusses what the field has tried, achieved and left unresolved.
Machines Who Think: A Personal Inquiry into the History and Prospects of Artificial Intelligence
Pamela McCorduck · 2004
The 2004 edition brings McCorduck’s account up to date. It follows the field through the scientists she met and interviewed.
AI: The Tumultuous History of the Search for Artificial Intelligence
Daniel Crevier · 1993
Crevier combines intellectual and business history, tracing AI’s projects, notable successes and failures through interviews with leading researchers.