AI Timeline: From Turing to Large Language Models

By Victor Hoang · Updated Oct 21, 2026

In July 1972, Sir James Lighthill, a Cambridge mathematician, sat down to survey artificial intelligence. The UK's Science Research Council had asked for it. Grant applications kept arriving in a field it couldn't judge. Years earlier, a small group had promised that a summer of work could teach machines to learn.

The AI timeline starts in 1950, when Alan Turing asked whether machines can think. It runs through Dartmouth in 1956 and a long funding winter in the 1970s. Then come the deep learning breakthrough of 2012, the Transformer of 2017, ChatGPT in 2022 and AI agents today. The throughline is a rise, a fall and a return. Bold promises hit a hard limit, then a new method made them work. This page is the dated chronology. For the manager's view of the same story, see The History of AI.

Timeline at a Glance

The Paper Years (1936-1955)

Logic, neurons and a test, all on paper

Picture the machines of 1943: relay switches, vacuum tubes and punched tape filling rooms, while the people who would define AI worked mostly with pencils. In 1936 Alan Turing had described an abstract machine that could carry out any step-by-step procedure, which gave later researchers a precise idea of what "computing" meant (the physical story is in the history of computers). Seven years later, in December 1943, Warren McCulloch and Walter Pitts published a logical calculus of the ideas immanent in nervous activity. They argued that because nerve cells fire in an all-or-none way, neural events can be treated with propositional logic. The Computer History Museum calls it a foundational paper in artificial neural networks.

Then came the question itself. In October 1950, Turing published "Computing Machinery and Intelligence" in the journal Mind. It begins by proposing to consider the question, "Can machines think?" He then swapped that vague wording for a game anyone could run, now called the Turing Test. A machine passes if, from written answers alone, you can't tell it from a person.

By August 31, 1955, the question had a plan. John McCarthy (Dartmouth), Marvin Minsky (Harvard), Nathaniel Rochester (IBM) and Claude Shannon (Bell Labs) signed a proposal for a summer research project on artificial intelligence. It rested on the conjecture that every aspect of learning or any other feature of intelligence could in principle be described precisely enough for a machine to simulate it. Its list of topics included automatic computers, using language, neuron nets, self-improvement and abstractions. The same year, Allen Newell, Herbert Simon and J.C. Shaw began work on Logic Theorist, a program that would eventually prove 38 theorems from Principia Mathematica.

The problem this era solved was vocabulary: it gave the field a question, a test and a target. The problem it created was scale. Everything was proposed on paper, and the machines of the day were far too small to try most of it.

For teams today: a clear test beats a clear definition. Write down what "working" looks like before you evaluate any AI tool. The older roots of these ideas, from automata to ancient myths of made minds, are covered in What Is the Origin of Artificial Intelligence?.

The Symbolic Years (1956-1971)

Rule-based programs and the first overpromises

The summer of 1956 at Dartmouth put the name "artificial intelligence" on a field, and the next fifteen years belonged to symbols. Researchers wrote programs that manipulated lists of facts and rules, and they believed that was how reasoning worked. In 1958, McCarthy invented LISP, a language in which programs and data were both simply lists, which made it easy for a program to treat another program as data. It stayed in use in AI work for decades.

The programs got more impressive. In 1965, a Stanford team led by Ed Feigenbaum, Joshua Lederberg and Carl Djerassi built DENDRAL, which the Computer History Museum describes as the first "expert system": if-then rules from specialists, used to identify the structure of organic compounds. In 1966, Joseph Weizenbaum finished ELIZA at MIT, whose DOCTOR script answered like a psychotherapist by substituting key words into predetermined phrases. It fooled some users for a while. In 1968, Terry Winograd began the thesis behind SHRDLU, which could follow typed commands about a small world of blocks, cones and pyramids. In 1970, SRI's Shakey became the first mobile robot controlled by AI, though it moved at about 2 meters per hour.

Each success was real, and each lived in a tiny world. ELIZA understood nothing, and SHRDLU understood only blocks. The pattern of the era was a demo that worked in a narrow room and a prediction that it would soon work everywhere. That gap is what the next era had to answer for.

For teams today: a demo proves the interface, not the accuracy. Ask for results on your own data, in your own mess.

The First Winter (1972-1979)

Funding dries up after the Lighthill survey

Imagine the Science Research Council's position in 1972: a growing pile of AI grant requests, a field full of confident claims and no outside judge. Lighthill's survey, dated July 1972, divided the work into three categories. Category A was advanced automation, replacing people with machines for specific purposes. Category C was computer-based study of the central nervous system in humans and animals. Category B was the bridge between them, and Lighthill said progress there had been even more disappointing than elsewhere.

His central diagnosis was a single idea. Earlier hopes had failed, he argued, because researchers hadn't reckoned with the "combinatorial explosion": the number of possible ways to combine elements grows so fast that a program searching a large body of knowledge drowns in options. The most notorious disappointment, he wrote, was machine translation, where enormous sums had produced very little useful result, a verdict a US National Academy of Sciences review had reached in 1966.

Historians call what followed the first AI winter. The label matters less than the shift in behavior. Funders asked harder questions, and researchers stopped promising general intelligence. The work didn't stop. In 1972, a team at BBN completed LUNAR, a natural language system that let geologists query chemical data on Apollo 11 moon rocks. In 1979, the Stanford Cart crossed a room on its own while steering around a chair. Both were narrow, both had clear goals and both worked.

The winter solved a credibility problem by forcing the field to measure. It created a quieter one: for years, the most ambitious ideas, neural networks among them, struggled to get attention or funding.

For teams today: a winter is a correction, not an ending. When a technology disappoints your team, narrow the use case and keep the parts that worked.

The Learning Turn (1980s-2011)

Programs stop being written and start being trained

By the 1980s, the symbolic approach had a business-friendly cousin. Expert systems, descended from DENDRAL, packaged a specialist's rules into software that companies could deploy. Their weakness was that every exception required another hand-written rule.

The turn came from neural networks. In October 1986, David Rumelhart, Geoffrey Hinton and Ronald Williams published back-propagation in Nature. Their procedure repeatedly adjusts the weights of a network's connections to shrink the difference between actual and desired output, and internal "hidden" units come to represent important features of the task. The abstract distinguishes it from earlier methods like the perceptron-convergence procedure precisely because it can create useful new features. That is the technical heart of what we now call machine learning: learn the pattern from examples instead of writing the rule.

Two public contests showed what specialized systems could do. In February 1996 in Philadelphia, Garry Kasparov beat IBM's Deep Blue 4-2, although the computer won Game 1. In May 1997 in New York, an upgraded Deep Blue won the rematch 3.5 to 2.5. It used 32 processors in parallel and could evaluate 200 million chess positions per second, which is brute force, not learning. In February 2011, IBM's Watson, built by a team led by David Ferrucci, defeated Brad Rutter and Ken Jennings on Jeopardy!, answering questions posed in natural language without an internet connection.

Neither contest was a path to general intelligence, and both were the same kind of win: a narrow task, a clear score, a huge amount of engineering. Meanwhile, unglamorous learning systems quietly took on scoring, ranking and forecasting inside businesses. The era solved the rule-writing bottleneck. It created a data bottleneck, because learning needs examples, and the examples were hard to come by until the web and cheap graphics hardware arrived.

For teams today: the cheapest AI win is often a prediction on data you already own, such as which leads convert or which invoices are late. You don't need a chatbot to start.

The Deep Learning Breakthrough (2012-2016)

Data and graphics chips make neural networks win

Picture a large academic competition in 2012: teams submitting software that has to name the object in each of a million photographs, and the entries are scored on how often the right answer appears in the top five guesses. That was the ImageNet challenge. In 2012, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton entered a large deep convolutional neural network and won with a top-5 error of 15.3%, against 26.2% for the second-best entry. The network had 60 million parameters and 500,000 neurons, and the authors credit an efficient GPU implementation for making the training practical.

The ingredients came from outside AI: the web supplied labeled images (see the history of the internet) and gaming chips supplied cheap parallel math. The result was a reliable deep learning recipe that spread from images to speech and translation.

The public moment came from the board game Go. In October 2015, DeepMind's AlphaGo beat the three-time European champion Fan Hui 5-0. In March 2016 in Seoul, it beat Lee Sedol 4-1. DeepMind's page notes that in game two, AlphaGo played Move 37, which it estimated had a 1 in 10,000 chance of being played by a human, and the move helped win the game. Go has far too many positions to search exhaustively, so a system that won there had to learn something that looked like judgment.

This era solved perception: reading images, sound and text characters became dependable. It left language itself unsolved, because the best language models of the time read text one word after another and were slow to train.

For teams today: perception tasks such as reading invoices, transcribing calls and sorting photos have been solved for a decade. If your team still keys these in by hand, that's adoption debt, not a technology gap.

The Transformer Age (2017-present)

One architecture scales into chat, reasoning and agents

In June 2017, eight researchers posted "Attention Is All You Need". It proposed the Transformer, a network built solely on attention mechanisms, dispensing with recurrence and convolutions entirely. On English-to-German translation it scored 28.4 BLEU, beating the best published results by more than 2 points, while being more parallelizable and needing significantly less time to train. Our explainer on the transformer architecture goes through how it works.

Training faster meant training bigger, and the pace of papers shows it. In October 2018, Jacob Devlin and three co-authors posted BERT, a model for pre-training deep bidirectional Transformers on language understanding. In May 2020, OpenAI researchers submitted GPT-3, an autoregressive language model with 175 billion parameters, ten times more than any previous non-sparse language model. Its unusual trait was that it handled tasks specified purely through text, with no gradient updates or fine-tuning. A person could describe a job in plain English and get a first attempt. Models like these are what we now call large language models.

On November 30, 2022, OpenAI introduced ChatGPT, a model that interacts in a conversational way and can answer follow-up questions, admit its mistakes and reject inappropriate requests. It was a free research preview, trained with reinforcement learning from human feedback. The technology wasn't brand new. The chat box was. OpenAI also said plainly that it sometimes writes plausible-sounding but incorrect or nonsensical answers. In March 2023 came GPT-4, a large multimodal model accepting image and text inputs, which scored around the top 10% of test takers on a simulated bar exam, where GPT-3.5 scored around the bottom 10%. The wave of generative AI products that followed turned drafting text, code and images into a routine task.

The newest turns are about thinking and acting. On September 12, 2024, OpenAI released o1-preview, a series of models designed to spend more time thinking before they respond. In January 2025, DeepSeek posted a paper on incentivizing reasoning capability in LLMs via reinforcement learning. Both belong to the family of reasoning models. Models also learned to use software. In October 2024, Anthropic released computer use in public beta, letting developers direct Claude to look at a screen, move a cursor and type, and called it experimental and at times cumbersome and error-prone. In November 2024, Anthropic open-sourced the Model Context Protocol, a standard for connecting AI assistants to the systems where data lives. On December 9, 2025, it donated the protocol to a Linux Foundation fund, reporting more than 10,000 active public MCP servers. Systems that call tools and take actions are the core of agentic AI.

This era solved language and, increasingly, action. It created the problems you can see in the headlines: confident wrong answers, unclear ownership of what an agent does, and questions of who reviews its work. For the business side of this stretch, from lab to daily tool, read How AI Moved from Research Labs to Everyday Work.

For teams today: treat the model as a general layer you configure with instructions and company data, and keep a person accountable for anything that leaves the building.

Key People

Key Facts

• The Dartmouth proposal is dated August 31, 1955, a year before the summer project it describes.
• Lighthill's survey is dated July 1972 and names the combinatorial explosion as a general cause of the field's disappointments.
• Deep Blue evaluated 200 million chess positions per second and beat Kasparov 3.5 to 2.5 in May 1997.
• AlexNet cut the top-5 error on ImageNet to 15.3%, against 26.2% for the runner-up.
• GPT-3 had 175 billion parameters, ten times more than any previous non-sparse language model.
• The Transformer paper was submitted to arXiv on June 12, 2017.

The Pattern Behind the History

Across the six eras, the same four moves keep repeating. We call the sequence the Overreach-Limit-Reset Loop. It's our synthesis of the milestones above, not a research finding.

1. A narrow win gets read as a general one. ELIZA, SHRDLU and Deep Blue each worked brilliantly in a small world.
2. A hard limit appears. Lighthill's combinatorial explosion, the cost of hand-written rules and the shortage of training data each capped the previous method.
3. Funding and attention fall back. The winter is the field's way of repricing its promises.
4. A new ingredient resets the game. Back-propagation, GPUs plus ImageNet and the Transformer each unlocked what the last method couldn't.

The loop isn't a law, but it's a useful checklist. When a new capability is announced, ask which step you're in. If you're at step one, the demo is real and the generalization is not yet proven. If you're at step four, the early adopters are about to learn what the new limit is.

What It Teaches Teams Today

• Test on your own data. Every era's first claims were made on clean samples.
• Start narrow. The wins that lasted, from DENDRAL to scoring models, picked a small domain with a clear score.
• Expect a limit. Each method hit one. Ask what happens when your volume doubles or the inputs get messy.
• Budget for upkeep. Rules, prompts and training data all need an owner once the business changes.
• Match review to autonomy. A drafting tool needs a reader. An agent that acts needs approvals and an undo.