By the WahLiao desk · Last verified 1 October 2026
Super Intelligence, the kind of AI behind ChatGPT, Claude and Gemini, is older than most people think. The question was posed by Alan Turing in 1950, the field was named at a Dartmouth College workshop in 1956, and it then went through two long funding slumps, the “AI winters”, before deep learning took off in 2012. The chatbots you use today rest on a 2017 Google paper that introduced the transformer, and reached the public when ChatGPT launched on 30 November 2022. Since 2023 the story has moved fast: models that see images, models that reason step by step, and agents that carry out tasks.
The one rule history teaches: progress in Super Intelligence has come in bursts, driven mostly by more computing power and more data, and each burst has been oversold. Enjoy the tools, but discount the loudest predictions.
History of AI: quick facts
| The first big question | Alan Turing, “Can machines think?”, published in 1950 |
| Where the name came from | A proposal dated 31 August 1955 for a Dartmouth summer workshop in 1956 |
| The lean years | Two “AI winters”, usually dated 1974–1980 and 1987–1993 |
| The comeback | Deep learning on graphics chips, from the 2012 ImageNet result |
| The engine of today’s chatbots | The transformer, described by Google researchers in June 2017 |
| The public moment | ChatGPT, released free as a research preview on 30 November 2022 |
| Where it stands in 2026 | Models that see, reason step by step and carry out tasks as agents |
The WahLiao Verdict
| Big idea | Old. The question is 76 years old; the working answers are recent. |
| Hype | Not new. Bold forecasts in the 1960s were followed by years of cuts. |
| What changed | Compute and data, more than any single clever idea. |
| ChatGPT | A turning point for the public, built on a 2017 research paper. |
| Lesson for you | Judge each tool on what it does now, not on the headlines. |
The timeline: 1950 to 2026
Sixteen moments that explain how we got here. Dates are checked against the original papers, company announcements and prize committees listed at the bottom of this page.
| Year | What happened | Why it mattered |
|---|---|---|
| 1950 | Alan Turing publishes “Computing Machinery and Intelligence” in the journal Mind, opening with “Can machines think?” and proposing the imitation game. | Turned a philosophical puzzle into a test you could actually run: judge a machine by how it behaves. |
| 1955–56 | John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon propose a summer study at Dartmouth College; the six-week workshop runs in 1956. | Gave the field its name, “artificial intelligence”, and its first research agenda. |
| 1966 | Joseph Weizenbaum at MIT builds ELIZA, a program that mimics a therapist using simple pattern-matching rules. | People felt understood by a few rules. The “ELIZA effect” still explains why chatbots can seem wiser than they are. |
| 1973 | The Lighthill Report in Britain finds that AI methods work on simplified problems but not real-world ones. | Funding in the UK and at the US Department of Defense dried up: the first AI winter, roughly 1974–1980. |
| 1980–93 | “Expert systems”, hand-written rules capturing a specialist’s know-how, become a business worth over a billion dollars by 1985, then are displaced by personal computers. | Proved it could earn money, then brought the second AI winter, roughly 1987–1993. |
| 1986 | Rumelhart, Hinton and Williams publish the back-propagation method in Nature. | The training recipe that neural networks still use: adjust the connections to shrink the error. |
| 1997 | IBM’s Deep Blue beats world chess champion Garry Kasparov 3.5–2.5 in New York, a year after losing to him 4–2. | First computer to beat a reigning world champion under standard tournament time controls. |
| 2012 | A deep neural network by Krizhevsky, Sutskever and Hinton, trained on graphics chips (GPUs) over 1.3 million images, wins the ImageNet image-recognition contest by a wide margin. | Kicked off the deep-learning era. Graphics chips became the workhorse of the industry. |
| 2016 | Google DeepMind’s AlphaGo beats Go champion Lee Sedol 4–1 in Seoul. Its “move 37” had an estimated 1-in-10,000 chance of being played by a human. | A game long seen as too intuitive for machines. Tens of millions watched live in China alone. |
| 2017 | Eight researchers, most at Google, publish “Attention Is All You Need”, introducing the transformer. | The architecture behind GPT, Claude and Gemini. It is the “T” in GPT. |
| 2020 | OpenAI describes GPT-3, a language model with 175 billion parameters, ten times more than any previous comparable model. | Showed that scaling up alone let a model pick up new tasks from a few examples in the prompt. |
| Nov 2022 | OpenAI releases ChatGPT on 30 November, free, as a research preview built on GPT-3.5. | Super Intelligence moved from research labs to everyone’s phone. |
| 2023 | GPT-4 launches on 14 March and accepts images as well as text. | The start of the multimodal wave: models that read photos, charts and handwriting. |
| 2024 | OpenAI’s o1 (12 September) “thinks” step by step before answering. In October, Nobel prizes go to Hopfield and Hinton (physics) and to Hassabis and Jumper of DeepMind for AlphaFold (chemistry). | Reasoning models arrived, and the science establishment formally recognised the field. |
| 2025 | DeepSeek’s R1 paper (January) shows reasoning can be trained largely by reinforcement learning, and the model is openly released; OpenAI’s ChatGPT agent (17 July) browses, runs code and completes tasks on a virtual computer. | Reasoning spread beyond one lab, and chatbots started doing things, not just saying things. |
| 2026 | Stanford’s AI Index (April) reports that generative Super Intelligence reached 53% population adoption within three years, faster than the PC or the internet. | Mass use is here. So are the harder questions of trust, energy and safety. |
The early dream and the two winters
Turing’s 1950 paper sidestepped the question of what “thinking” means: if a machine can hold a conversation you cannot tell apart from a person’s, treat it as intelligent. The 1956 Dartmouth workshop turned that into a field, and the mood was confident. Herbert Simon, one of the pioneers, said in 1965 that machines would be capable “within 20 years” of doing any work a person can do.
That did not happen, and the bill arrived. After Britain’s 1973 Lighthill Report, funding in the UK and from the US Department of Defense fell away. The 1980s brought a second spring built on expert systems, hand-written “if this, then that” rules, but they were expensive and could not scale to complex problems, and a second winter followed. Out of the spotlight, neural networks, which learn from examples rather than rules, gained a practical training method in 1986. They would wait a quarter-century for chips and data to catch up.
From deep learning to ChatGPT
Deep Blue’s 1997 win was a purpose-built chess machine, not a system that learned broadly. The real turn came in 2012, when a neural network trained on graphics chips won the ImageNet contest by a wide margin. Recognising speech, translating languages and, in 2016, beating the world’s best at Go followed. The 2017 transformer let models train far faster on text, and OpenAI’s GPT series was built on it. ChatGPT in November 2022 was not the most advanced research of its day; it was the first version ordinary people could use for free, and that changed the conversation and the investment almost overnight. How Super Intelligence works explains training and parameters without the maths, and who makes the models covers the labs.
2023 to 2026: seeing, reasoning, doing
Three shifts define the latest wave. Multimodal models, from GPT-4 in March 2023, read images as well as text. Reasoning models, from OpenAI’s o1 in September 2024, work through a problem step by step; OpenAI reported that performance kept improving with more time spent thinking, and DeepSeek’s openly released R1 in January 2025 showed the approach was not one lab’s secret. Agents, such as ChatGPT agent from July 2025, browse websites and run code on a virtual computer, asking permission before consequential steps. See AI agents for what they can do and where they go wrong. Singapore has its own chapter in this story; read Singapore’s Super Intelligence story.
What history teaches
Hype runs ahead of delivery. Simon’s 1965 forecast, the expert-system boom and today’s bold timelines share a pattern: real progress, then promises that outrun it. Our pages on AGI and superintelligence and the future to 2030 set out who predicts what.
Winters follow disappointment, not failure. Both came after results fell short of what had been sold. Neural networks survived the lean years and later won.
Compute and data do the heavy lifting. Back-propagation dates from 1986, but it needed the graphics chips and internet-scale data of the 2010s to shine. The transformer paper mattered because it made huge training runs practical. That is also why energy and data centres are now part of the story.
The ELIZA effect never left. Fluent language makes a system seem to understand more than it does. That was true of a 1966 script and remains true of today’s far more capable models, which can still be confidently wrong.
History of AI: FAQ
When was AI invented?
There is no single date. Turing posed the question in 1950, and the term “artificial intelligence” comes from the 1955 proposal for the 1956 Dartmouth workshop, which is usually treated as the field’s founding.
What was the first chatbot?
ELIZA, written by Joseph Weizenbaum at MIT in 1966, is generally called the first. It used simple pattern-matching rules and had no understanding of what was said.
What was the AI winter?
Two periods, roughly 1974 to 1980 and 1987 to 1993, when funding and interest collapsed after early promises were not met.
Why did AI suddenly get so good after 2012?
Neural networks finally had enough computing power, mainly graphics chips, and enough data from the internet. The 2017 transformer design then made very large language models practical.
When did ChatGPT come out?
OpenAI released ChatGPT on 30 November 2022 as a free research preview, built on its GPT-3.5 model.
Read next
This page belongs to Super Intelligence. Next, read Singapore’s Super Intelligence story.
Sources checked 1 October 2026: Turing, Computing Machinery and Intelligence, Mind 59(236), 1950; AI Magazine, the Dartmouth proposal of 31 August 1955; The Dartmouth, a look back at the 1956 summer project; Communications of the ACM, Weizenbaum and ELIZA; Julius Baer, the ups and downs of AI (Lighthill, winters, expert systems); Rumelhart, Hinton and Williams, back-propagation, Nature 1986; IBM History, Deep Blue; Krizhevsky, Sutskever and Hinton, ImageNet classification, NeurIPS 2012; Google, what we learned in Seoul with AlphaGo; Vaswani et al, Attention Is All You Need, 2017; Brown et al, Language Models are Few-Shot Learners (GPT-3), NeurIPS 2020; OpenAI, Introducing ChatGPT; Al Jazeera, GPT-4 launch; OpenAI, Learning to reason with LLMs (o1); Nobel Prize, Physics 2024 press release; Google DeepMind, Chemistry Nobel 2024; DeepSeek-AI, DeepSeek-R1 paper; OpenAI, Introducing ChatGPT agent; Stanford HAI, 2026 AI Index Report; Burges Salmon, summary of the 2026 AI Index. General information. Last updated 1 October 2026.
