Skip to content
learn
  1. Learn
  2. Artificial Intelligence
  3. Orientation

AI AI-00.02

A short history of AI

From Turing's question and the Dartmouth workshop through expert systems, two AI winters and deep learning to transformers and large language models.

Beginner 14 min Last reviewed

The big idea

AI is not new. People have been building "thinking machines" since the 1950s, and the field has swung between wild optimism and deep disappointment several times. Periods of hype brought money and bold promises; when the promises were not kept, funding dried up in what became known as AI wintersA period when interest and funding for AI research collapsed after expectations were not met. The main ones were in the mid 1970s and the late 1980s to early 1990s.. Each comeback was powered by a different idea: logic and search, then expert systems, then statistics, then deep neural networks.

Knowing this history is practical, not nostalgic. Many of today's techniques are old ideas that finally got enough data and computing power. And the pattern of over-promising followed by a correction is worth recognizing when you read headlines about AI today.

How it works

Step 1: The shape of seventy years

Waves of excitement in AI from 1950 to today An illustrative wave of excitement and funding: a first peak in the 1960s around logic and search, a first winter in the 1970s, a second peak in the mid 1980s around expert systems, a second winter around 1990, a modest rise with statistical machine learning in the 2000s, then a steep climb from 2012 with deep learning and large language models. 19501960197019801990200020102020yearexcitement and funding (illustrative) logic and search first winter expert systems second winter statistical ML deep learning LLMs Figure 1. A qualitative sketch, not measured data. Each boom was driven by a new idea and big promises; each winter followed when the promises outran what the technology could do.

Step 2: The symbolic era (1950s to 1980s)

In 1950 Alan Turing asked "Can machines think?" and proposed the imitation game, now called the Turing testTuring's proposal to judge machine intelligence by whether a person conversing in text can tell the machine from a human.. In the summer of 1956, a workshop at Dartmouth College gave the field its name, and its organizers believed key problems could be solved quickly.

Early AI was symbolic: intelligence as manipulating symbols with logic and search. Programs proved theorems, played checkers and solved puzzles. At the same time, Frank Rosenblatt built the perceptron, a machine that learned from examples, an ancestor of today's neural networks:

A diagram from the Mark I Perceptron operator's manual showing the machine's sensory units connected through association units to response units.
The Mark I Perceptron, from its 1960 operator's manual: light sensors feed a layer of randomly wired units whose connection weights are adjusted as it learns. It is the first hardware neural network. Image: Mark I Perceptron Operators' Manual (Hay and Murray, 1960), US federal government work, public domain, via Wikimedia Commons

Progress was real but much slower than promised. A single-layer perceptron provably cannot learn some simple patterns, critical reports questioned what had been delivered, and in the 1970s funding fell sharply: the first winter. In the 1980s expert systemsPrograms that encode a human expert's knowledge as many if-then rules to give advice in a narrow domain, such as diagnosing infections or configuring computers. brought a second boom. Companies built rule-based advisors for narrow jobs, but the rules were costly to write, brittle and hard to maintain. By the late 1980s the market collapsed: the second winter.

Step 3: The statistical turn (1990s to 2000s)

Researchers increasingly let data do the work. Probabilistic models, decision trees and support vector machines learned from examples and handled uncertainty better than hand-written rules. In 1997 IBM's Deep Blue beat world chess champion Garry Kasparov, using massive search rather than learning. Through the 2000s, machine learning quietly took over spam filtering, search ranking, recommendations and fraud detection.

Step 4: Deep learning and the LLM era (2012 onward)

Neural networks with many layers had existed for decades, but needed more data and computing power than anyone had. In 2012 a deep network called AlexNet, trained on graphics cards, won the ImageNet image recognition challenge by a wide margin, and the deep learning boom began. In 2016 DeepMind's AlphaGo beat Lee Sedol, one of the world's best Go players, a game long considered too intuitive for computers. In 2017 the paper "Attention Is All You Need" introduced the transformerA neural network architecture built on attention, which lets every token look at every other token. It is the basis of modern large language models., which turned out to scale remarkably well. Larger and larger language models followed, and the public release of ChatGPT in late 2022 brought generative AI to hundreds of millions of people.

Try it

  1. 1950 Turing's question

    Alan Turing's paper Computing Machinery and Intelligence asks whether machines can think and proposes the imitation game as a practical test.

  2. 1956 Dartmouth workshop

    A summer workshop at Dartmouth College, proposed by John McCarthy and colleagues, names the field artificial intelligence and sets out an ambitious research program.

  3. 1958 The perceptron

    Frank Rosenblatt introduces the perceptron, a learning machine inspired by neurons. The Mark I Perceptron hardware learned to recognize simple shapes from a camera of light sensors.

  4. 1966 ELIZA

    Joseph Weizenbaum's ELIZA imitates a therapist with simple pattern matching, and people confide in it anyway: an early lesson in how readily we attribute understanding to fluent text.

  5. 1970s First AI winter

    Results fall far short of predictions, a single-layer perceptron's limits are proved, and critical government reviews lead to deep funding cuts.

  6. 1980s Expert systems boom

    Businesses deploy rule-based expert systems for narrow tasks such as configuring computer orders. Neural network training with backpropagation is popularized in 1986.

  7. Late 1980s Second AI winter

    Expert systems prove expensive and brittle, the specialized AI hardware market collapses, and funding retreats again.

  8. 1997 Deep Blue

    IBM's Deep Blue defeats world chess champion Garry Kasparov in a six-game match, using specialized hardware to search hundreds of millions of positions per second.

  9. 2012 AlexNet and ImageNet

    A deep convolutional network trained on GPUs wins the ImageNet challenge by a large margin, convincing the field that deep learning works at scale.

  10. 2016 AlphaGo

    DeepMind's AlphaGo, combining deep neural networks with tree search, defeats Lee Sedol 4 games to 1 at Go.

  11. 2017 The transformer

    Attention Is All You Need introduces the transformer architecture, which becomes the foundation of large language models.

  12. 2020s Large language models

    Models trained on huge text collections write, summarize, translate and code. ChatGPT's release in November 2022 brings them to a mass audience, followed by multimodal models that handle images, audio and video.

Scroll sideways and open any card.

Try this:

  1. Open the two winter cards. What did each boom promise that it could not deliver?
  2. Find the milestones that rely on search and those that rely on learning. Which one combines both?
  3. Pick a claim about AI from this week's news. Which era's promises does it sound like?

Going deeper

Old ideas, new scale

Neural networks, backpropagation and even attention-like mechanisms existed long before they dominated. What changed after 2012 was scale: far more data from the internet, far more computing power from GPUs, and engineering tricks that made deep networks trainable. That is a recurring theme: an idea can be right decades before the resources exist to prove it. The ML track tells the neural network side of the story in The perceptron and the artificial neuron .

Search never went away

Symbolic AI's tools are not obsolete. Search, logic and planning run map routing, game engines, compilers, schedulers and the verification of chips and software. Modern agents often combine a language model with classical search and planning. The first half of this track, starting at Agents and environments , covers these ideas.

Trade-offs

OptionWins whenLoses when
Symbolic AI (rules, logic, search) Problems with clear rules: exact, explainable, needs no training data Perception and language, where the rules cannot be written down
Statistical machine learning Structured data with many labelled examples: fraud, ranking, forecasting Raw images, audio and text, where hand-made features fall short
Deep learning and LLMs Perception, language and generation at scale Cost, data needs, guaranteed correctness and explainability

Common mistakes

  • Believing AI started with ChatGPT. Most core ideas are decades old.
  • Reading history as a straight line of progress. It went through booms and collapses driven by over-promising.
  • Assuming the newest approach replaced everything before it. Classical methods still do much of the work in real systems.
  • Treating a milestone in one game or benchmark as general intelligence. Each was impressive and narrow.

Check yourself

  1. What was the Dartmouth workshop of 1956 known for?

    Show answer

    B. Naming the field of artificial intelligence The 1956 workshop gave the field its name and launched it as a research area.

  2. What caused the AI winters?

    Show answer

    B. Results fell far short of promises, so funding collapsed Over-optimistic predictions met hard limits, and funders pulled back.

  3. Which 2012 event started the deep learning boom?

    Show answer

    B. AlexNet winning the ImageNet challenge A deep convolutional network trained on GPUs beat other methods by a wide margin on ImageNet.

  4. Which milestone combined deep learning with tree search?

    Show answer

    C. AlphaGo AlphaGo used neural networks to evaluate positions and guide a tree search.

Recap

  • AI began in the 1950s with Turing's question and the 1956 Dartmouth workshop.
  • Symbolic AI and expert systems each brought a boom, followed by an AI winter.
  • Statistical machine learning took over practical tasks in the 1990s and 2000s.
  • Deep learning broke through in 2012; transformers (2017) led to today's large language models.
  • Many modern techniques are old ideas that finally got enough data and compute.

Get new lessons by email

Get new posts on backend and AI internals. No spam, unsubscribe anytime.

Your email is only used to send new posts. Privacy

Next up