Eighty years of artificial intelligence.
shikhar.wikiTwo ideas compete for most of this story. One builds minds from symbols and rules. The other builds them from networks that learn. This site follows both, one chapter at a time.
machines
Logicians, neurophysiologists and engineers ask one question from three sides: can a machine think? Money and confidence grow fast. By 1973 both run out.
A neuron as a logic gate.
Warren McCulloch and Walter Pitts describe an ideal neuron. It fires when the sum of its inputs passes a threshold. Networks of these units can compute logical functions. 0.95
“I propose to consider the question, ‘Can machines think?’”— Alan Turing
Turing then replaces the question with a game. A judge talks in text to a person and to a machine, and tries to tell which is which. 0.97
The field gets a name.
John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon propose a summer study at Dartmouth College. Their 1955 proposal uses the words “artificial intelligence”. It claims that every feature of intelligence can, in principle, be described so exactly that a machine can simulate it. 0.95
A program that proves theorems.
Allen Newell, Herbert Simon and Cliff Shaw build the Logic Theorist. It proves theorems from Whitehead and Russell’s Principia Mathematica by searching through the rules of logic. 0.90
The perceptron learns.
Frank Rosenblatt’s perceptron changes its own weights from examples. The US Navy funds a hardware version, the Mark I. 0.92 The press calls it the embryo of a computer that will one day walk, talk and know that it exists. 0.75
A language for symbols.
McCarthy designs LISP. Programs and data share one form, the list. It stays the main language of symbolic AI for thirty years. 0.93
(CONS (QUOTE A)
(QUOTE (B C)))
→ (A B C)A program that gets better by playing.
Arthur Samuel’s checkers program at IBM improves by playing against itself. In his 1959 paper he uses the phrase “machine learning”. 0.85
“Machines will be capable, within twenty years, of doing any work a man can do.”— Herbert Simon
The optimism of the decade, in one sentence. 0.85
The first chatbot.
Joseph Weizenbaum’s ELIZA plays a therapist. It matches patterns and turns the user’s words back into questions. People trust it more than Weizenbaum expects, and this troubles him. 0.92
Men are all alike. IN WHAT WAY? They're always bugging us about something or other. CAN YOU THINK OF A SPECIFIC EXAMPLE?
Translation stalls.
The ALPAC report tells the US government that machine translation is slower, less accurate and more costly than human work. Funding for it falls sharply. 0.90
The limits of one layer.
Minsky and Seymour Papert publish Perceptrons. They prove that a single-layer perceptron cannot compute some simple functions, such as XOR. 0.93 Many accounts say the book stopped neural network research for ten years. Historians dispute how much. 0.55
The first winter begins.
In Britain, the Lighthill Report finds that AI has not kept its promises outside narrow problems. UK support for AI research falls. In the US, DARPA moves away from open research. 0.90
winter
- Funding cuts in the UK and US
- Prolog and logic programming
- Frames and scripts
systems
- XCON at Digital Equipment
- Japan’s Fifth Generation project
- Backpropagation returns, 1986
winter
- The LISP machine market collapses
- Convolutional networks read digits
from data
- Deep Blue beats Kasparov, 1997
- LSTM, 1997
- ImageNet, 2009
learning
- AlexNet, 2012
- ResNet, 2015
- AlphaGo beats Lee Sedol, 2016
- The transformer, 2017
- BERT, 2018
- GPT-3, 2020
- ChatGPT, 2022
The current year moves too fast for chapters. It runs day by day, with arcs and trends, on its own site: the AI Timeline.
timeline.shikhar.wiki →