History & Timeline of AI

Before the term existed#

Formal ideas about "thinking machines" predate the field's name by several years. In 1950, mathematician Alan Turing published "Computing Machinery and Intelligence," proposing what's now called the Turing Test: rather than trying to define "thinking" philosophically, judge a machine by whether a human conversing with it (via text, blind to which party is which) can reliably tell it apart from another human. Turing's paper didn't build anything, but it framed the field's central question — can behavior alone stand in for intelligence — in a way that still shapes debates about modern chatbots passing (or gaming) exactly this kind of test.

1956: Dartmouth and the birth of "artificial intelligence"#

The field got its name and its founding moment at a summer workshop at Dartmouth College in 1956, organized by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. McCarthy coined the term "artificial intelligence" in the workshop's funding proposal, and the attendees — a small group that would go on to found or lead most of the field's major labs for decades — proposed, with startling optimism, that "every aspect of learning... can in principle be so precisely described that a machine can be made to simulate it." That same year, Allen Newell and Herbert Simon's Logic Theorist program proved theorems from Whitehead and Russell's Principia Mathematica using symbolic search — often cited as the first genuine AI program.

Symbolic AI and expert systems (1950s–1980s)#

The field's first three decades were dominated by symbolic AI: the working assumption that intelligence could be captured as explicit logic — rules and symbols manipulated by a program, not patterns learned from data. Joseph Weizenbaum's ELIZA (1966), a simple pattern-matching chatbot simulating a psychotherapist, showed how convincingly a system with no real understanding could still produce compelling conversation — and reportedly unsettled Weizenbaum himself when people treated it as genuinely empathetic.

Through the 1970s this matured into expert systems: programs that encoded a human specialist's knowledge as a large set of hand-written if-then rules. DENDRAL (Stanford, mid-1960s) inferred chemical structures from mass spectrometry data; MYCIN (Stanford, 1970s) diagnosed bacterial infections and recommended antibiotics, reportedly matching or exceeding human specialists on narrow test cases, though it was never deployed clinically, partly over liability concerns. XCON (Digital Equipment Corporation, 1980) configured orders for VAX computer systems and was estimated to save DEC tens of millions of dollars a year, making it one of the first commercially successful AI systems.

These systems worked well in narrow, well-understood domains, but hand-writing and maintaining thousands of rules didn't scale to messier problems like vision or open-ended language — and funding followed the disappointment. A 1973 UK government report (the "Lighthill report") triggered the first AI winter (roughly 1974–1980), a sharp cut in funding after inflated expectations went unmet. A second AI winter followed around 1987–1993, as the market for specialized "Lisp machine" hardware collapsed and expert systems proved expensive to build and maintain in production.

Statistical machine learning (1990s–2000s)#

Rather than hand-coding rules, researchers increasingly built systems that learned patterns statistically from labeled data. Backpropagation — the algorithm that lets a neural network learn from its own errors — had been described earlier, but a 1986 paper by Rumelhart, Hinton, and Williams popularized it as a practical way to train multi-layer networks, though networks of this era stayed fairly shallow due to limited data and compute. More of the era's progress came from other statistical methods: support vector machines (formalized by Vapnik and Cortes in 1995), decision trees and random forests, and statistical natural language processing that replaced hand-written grammar rules with probabilities learned from text corpora. IBM's Deep Blue defeated world chess champion Garry Kasparov in 1997 — a landmark result, though it came from brute-force search over huge numbers of positions plus a hand-tuned evaluation function, not learning, illustrating that "AI beats a human" and "AI learns like a human" were (and remain) two different claims.

Deep learning takes off (2006–2016)#

Neural networks moved back to the center of the field once three things converged: much larger labeled datasets, much cheaper parallel compute (GPUs originally built for video games), and refinements that made deep networks easier to train. Geoffrey Hinton's 2006 work on "deep belief networks" helped revive interest in networks with many layers. Fei-Fei Li's ImageNet dataset (2009), with millions of hand-labeled images across thousands of categories, gave the field a shared, large-scale benchmark. The turning point most historians point to is 2012: Krizhevsky, Sutskever, and Hinton's AlexNet, a deep convolutional neural network trained on GPUs, won the ImageNet competition by a huge margin over the next-best (non-deep-learning) entrant — the moment that convinced much of the field deep learning wasn't a niche technique. Word2vec (2013) showed that word meaning could be captured as vector arithmetic learned from raw text. DeepMind's AlphaGo beat Go champion Lee Sedol in 2016, pairing deep neural networks with search and reinforcement learning to master a game long considered far harder for computers than chess, because Go's search space is too vast for Deep-Blue-style brute force alone.

Transformers and the LLM era (2017–present)#

In 2017, a team at Google published "Attention Is All You Need," introducing the transformer architecture (see this site's LLM tutorials for how self-attention actually works). It replaced the sequential processing of recurrent neural networks with a mechanism that could process an entire sequence in parallel, which turned out to scale dramatically better with data and compute. Google's BERT (2018) applied transformers to language understanding tasks; OpenAI's GPT series applied them to text generation, scaling from GPT-1 (2018) to GPT-2 (2019, notable enough at the time that OpenAI initially withheld the full model over misuse concerns) to GPT-3 (2020, 175 billion parameters), demonstrating that a single model, given the right prompt, could perform tasks it was never explicitly trained on. ChatGPT's public release in November 2022 was the moment this technology reached mainstream, non-technical audiences at scale, followed by rapid iteration across the industry — larger and more capable models, competing model families from multiple labs, and expansion into multimodal models that handle images and audio alongside text.

A quick-reference timeline#

Era	Roughly	Defining idea
Foundations	1950–1956	Turing Test proposed; the field is named at Dartmouth
Symbolic AI / expert systems	1950s–1980s	Intelligence as hand-written logic and rules
AI winters	1974–1980, 1987–1993	Funding collapses after expectations outrun results
Statistical ML	1990s–2000s	Learning patterns from labeled data instead of hand-coded rules
Deep learning	2006–2016	Multi-layer neural networks learn features automatically from raw data
Transformers / LLMs	2017–present	Self-attention enables large, general-purpose language models

Common mistakes#

• Treating "AI" as a recent invention — the field is over 70 years old, and several of its core ideas (search, logic, learning from data) were worked out decades before deep learning made them practical at scale.
• Assuming progress was a steady, uninterrupted climb — the field went through two well-documented funding collapses ("AI winters") when hyped expectations outran what the technology of the time could actually deliver, a pattern worth remembering when evaluating bold capability claims today too.
• Conflating "beats humans at a task" with "learns/thinks like a human" — Deep Blue's 1997 chess win came from brute-force search and a hand-tuned evaluation function, not learning; it's a fundamentally different kind of achievement from AlphaGo's 2016 win, which combined learning with search.