Artificial Intelligence: A Short History of Thinking Machines

Artificial intelligence has been “ten years away” for about seventy years, and then, quite suddenly, it was in everyone’s pocket. This guide tells the real story: the question Alan Turing asked, the summer workshop that named the field, the booms and the winters, the machines that beat world champions, and the simple learning loop that sits at the heart of today’s systems. You can even train a tiny neuron yourself on this page.

The question: can machines think

Modern AI starts with a question. In 1950 the British mathematician Alan Turing published a paper, “Computing Machinery and Intelligence”, that opened with: can machines think? Because “think” is hard to define, he proposed a game. A human judge chats by text with two hidden partners, one human and one machine. If the judge cannot reliably tell which is which, the machine has passed what we now call the Turing test.

Turing’s real point was not the test itself but a way of making the question concrete: look at behaviour, not at philosophy. He also predicted machines that could learn from experience, which was remarkably far-sighted for 1950.

1956: a summer workshop names the field

In the summer of 1956 a small group gathered at Dartmouth College in New Hampshire for a workshop proposed by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon. McCarthy coined the term “artificial intelligence” for the proposal. Their optimism was breathtaking: they thought a summer of work might make significant progress on getting machines to use language, form concepts and improve themselves.

Progress was slower than hoped, but the first programs were exciting. Some proved mathematical theorems, some played checkers and learned from their mistakes, and in 1966 Joseph Weizenbaum’s ELIZA imitated a psychotherapist by turning the user’s words into questions. Users were often convinced it understood them, though it did only simple text matching.

Try it: train a tiny neuron

In 1958 Frank Rosenblatt built the perceptron, a simple model of a neuron that learns from examples. It multiplies each input by a weight, adds a bias, and fires if the total is above zero. When it gets an example wrong, it nudges the weights toward the right answer. Pick a task and press “Train one epoch” to watch it learn. An epoch is one pass through all four examples.

AND and OR can be learned: the perceptron quickly finds weights that give every answer correctly. XOR cannot. In 1969 Marvin Minsky and Seymour Papert showed in their book Perceptrons that a single layer of neurons cannot learn functions like XOR, where the right answer is not separable by a straight line. That result helped cool enthusiasm for neural networks for years. The fix, layers of neurons trained together, was known by the 1980s but needed much more data and computing power to shine.

The perceptron in Python

data = [((0, 0), 0), ((0, 1), 0), ((1, 0), 0), ((1, 1), 1)]   # the AND function
w1 = w2 = b = 0.0
lr = 0.1

for epoch in range(1, 11):
    errors = 0
    for (x1, x2), target in data:
        out = 1 if w1 * x1 + w2 * x2 + b > 0 else 0
        err = target - out
        if err:
            w1 += lr * err * x1
            w2 += lr * err * x2
            b += lr * err
            errors += 1
    print(f"epoch {epoch}: errors={errors}  w1={w1:.1f} w2={w2:.1f} b={b:.1f}")
    if errors == 0:
        break

Output:

epoch 1: errors=1  w1=0.1 w2=0.1 b=0.1
epoch 2: errors=3  w1=0.2 w2=0.1 b=0.0
epoch 3: errors=3  w1=0.2 w2=0.1 b=-0.1
epoch 4: errors=2  w1=0.2 w2=0.2 b=-0.1
epoch 5: errors=1  w1=0.2 w2=0.1 b=-0.2
epoch 6: errors=0  w1=0.2 w2=0.1 b=-0.2

After six passes there are no mistakes. This is the same loop used inside the biggest systems today, just with billions of weights instead of three.

Boom, bust and the “winters”

Early AI promised more than it could deliver. Computers of the 1960s and 1970s had tiny memories and slow processors, and problems that looked easy, such as understanding a photograph or a spoken sentence, turned out to be monstrous. A critical British report in 1973 led to funding cuts, and the field entered its first “AI winter” in the mid-1970s.

In the 1980s, expert systems revived interest. These programs stored the rules of specialists, such as how to diagnose a disease or configure computers, and companies spent heavily on them. But they were costly to maintain and brittle when faced with new situations. When the market for specialised AI hardware collapsed, a second winter followed in the late 1980s and early 1990s.

Machines that beat champions

Chess fell first. In May 1997 IBM’s Deep Blue beat the reigning world champion Garry Kasparov in a six-game match. Deep Blue relied on huge numbers of calculations and chess knowledge written by humans. Go, a far larger game, was thought to be out of reach until 2016, when DeepMind’s AlphaGo beat Lee Sedol 4 to 1. AlphaGo combined neural networks with search and learned from both human games and self-play.

How modern AI learns

Three ingredients lined up around 2012. First, huge labelled datasets such as ImageNet. Second, graphics processors (GPUs) that could do the vast number of calculations needed. Third, improved training methods. The result was deep learning, networks with many layers. It transformed image recognition, speech recognition and translation. In 2017 the transformer design made it practical to train very large language models on text, and by 2022 chatbots built on them were used by many millions of people. Chips made of billions of transistors power all of it.

AI in India

Indian-born researchers have shaped the field. Raj Reddy, born in India and working in the United States, shared the 1994 Turing Award, computing’s highest honour, for work on large-scale AI systems. Today India has a huge pool of software engineers and a special challenge and opportunity: 22 scheduled languages and hundreds of dialects.

The government’s IndiaAI Mission, approved in March 2024, aims to build computing capacity and datasets for AI in the country. A related public effort, Bhashini, works on translation and speech technology for Indian languages. Examples in everyday life include voice assistants that understand Hindi and other languages, crop advice for farmers, and tools that help read and translate government documents.

Mind map: AI at a glance

Common myths

• Myth: AI appeared suddenly in 2022. The field is about seventy years old; 2022 was when chatbots reached the public.
• Myth: AI understands like a human. Current systems find patterns in data; whether that amounts to understanding is debated.
• Myth: Passing the Turing test means a machine is conscious. The test measures behaviour only.
• Myth: AI is one technology. It is a family: rule-based programs, search, statistics and neural networks.

Frequently asked questions

Keep exploring

Sources

• History of artificial intelligence, Encyclopaedia Britannica
• Alan Turing, Encyclopaedia Britannica
• The Turing Test, Stanford Encyclopedia of Philosophy
• Attention Is All You Need, arXiv
• The Nobel Prize in Physics 2024, Nobel Prize Outreach
• IndiaAI, Government of India

Featured image: Image: Elliott & Fry, via Wikimedia Commons, Public domain. Dates are as commonly reported.