How AI Was Born: The Seventy-Year Path From Turing to ChatGPT

Ten people, two months, sixty-six years

In the summer of 1956, a workshop at Dartmouth College proposed to "make machines use language, form abstractions and concepts" — in two months. ChatGPT shipped on November 30, 2022.

This guide is the short, honest version of what happened in between. By the end you'll understand why every previous "AI moment" — 1958, 1973, 1997, 2012 — was real but partial, and why 2022 was the one that crossed into your kitchen.

Turing's question, 1950

Before there was AI there was a 36-page paper by Alan Turing — the British mathematician who broke Enigma — titled Computing Machinery and Intelligence. He sidestepped the unanswerable "can machines think?" and replaced it with a behavioral test: if a human judge, exchanging typed messages with a machine and another human, can't reliably tell which is which, the machine has passed. He called it the imitation game. We call it the Turing test.

The paper is interesting less for the test than for the bets Turing made. He predicted that by 2000, computers with about 10⁹ bits of memory (roughly 125 megabytes) could fool an interrogator 30% of the time after five minutes. He was off on the date but right on the shape of the problem: intelligence would be approached through statistics and scale, not through hand-coded rules. Almost everyone who came next ignored that part.

Dartmouth, 1956: the symbolic decade

Six years after Turing, John McCarthy convened the Dartmouth Summer Research Project on Artificial Intelligence — coining the term "artificial intelligence" partly to distinguish his program from Norbert Wiener's cybernetics. The proposal claimed "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." With ten people. In two months.

The Dartmouth crowd — McCarthy, Marvin Minsky, Claude Shannon, Herbert Simon, Allen Newell — bet on what we now call symbolic AI: encode human knowledge as logical rules, then have the machine reason over them. Their early demos were astonishing for the era. Newell and Simon's Logic Theorist (1956) proved 38 of the first 52 theorems of Principia Mathematica. ELIZA (Weizenbaum, 1966) imitated a Rogerian therapist convincingly enough that secretaries asked Weizenbaum to leave the room while they typed. SHRDLU (Winograd, 1970) manipulated virtual blocks via English commands.

The mood was triumphalist. Simon told a 1965 audience that "machines will be capable, within twenty years, of doing any work a man can do." The funding poured in. Then it stopped.

Two winters, one wall

The first AI winter began around 1973. The Lighthill Report, commissioned by the British Science Research Council, concluded that AI had "in no part of the field" delivered on its promises. DARPA cut funding. Two structural problems had become visible.

First, combinatorial explosion: symbolic systems search through possibilities, and the number of possibilities grows exponentially with problem size. A chess program that worked on toy boards choked on a real one. Second, the knowledge-acquisition bottleneck — writing down enough rules to handle the real world is, it turns out, infinite work. ELIZA didn't understand a thing. SHRDLU's universe contained eight blocks.

The second winter came after expert systems — the 1980s commercial bet — collapsed. Companies like Symbolics built dedicated "AI workstations" running rule-based systems for medical diagnosis, mineral prospecting, and credit decisions. By 1987 the AI hardware market was worth roughly $2 billion. By 1993 it was effectively zero. The systems were brittle, expensive to maintain, and outpaced by cheaper general-purpose machines from Sun and Apple. Symbolic AI hadn't failed because the people were dumb. It had failed because the approach — intelligence as a giant pile of explicit rules — was wrong.

The quiet revolution, 1990s–2012

While the symbolists licked their wounds, a different tribe was winning quietly: the statisticians. The bet was simple — instead of telling the machine the rules, give it data and let it find patterns. Three things converged.

One, machine learning matured. Support vector machines (Vapnik, 1995), random forests (Breiman, 2001), and probabilistic graphical models replaced rule engines in spam filters, search ranking, and credit scoring. By the mid-2000s, Google Translate had abandoned linguistic rules entirely for statistical models trained on parallel UN and EU documents.

Two, neural networks came back. The idea — loose mathematical models of neurons stacked in layers — dated to 1958 (Rosenblatt's Perceptron). It had been declared dead by Minsky and Papert in 1969 and revived twice (Hopfield 1982, backpropagation by Rumelhart, Hinton, and Williams in 1986). It still didn't work well. Networks were too small, data was too scarce, training was too slow.

Three, GPUs — graphics chips designed for video games. They turned out to be brilliant at the matrix multiplications neural networks need. Around 2009, Stanford researchers (Raina, Madhavan, Ng) showed that a GPU could train a deep belief network roughly 70× faster than a CPU.

The collision came in September 2012. Geoffrey Hinton's team at Toronto entered a neural network — AlexNet — into the ImageNet image-recognition competition. ImageNet was a dataset of 1.2 million labeled photos curated since 2009 by Fei-Fei Li at Stanford. AlexNet won by a staggering margin: 15.3% top-5 error rate, against 26.2% for the runner-up. Within three years, every serious computer-vision lab in the world had switched to deep learning. The cold war was over.

Attention is all you need, 2017

Image recognition is one thing. Language is harder, because words depend on words far away from them — the bank you sit on versus the bank you rob. Through 2016, the best language models used recurrent neural networks (RNNs), which process text one word at a time and forget too much by the end of a long sentence.

In June 2017, eight researchers at Google Brain and Google Research published a paper with a borrowed Beatles title: Attention Is All You Need. They proposed an architecture called the transformer — a neural network that processes all words in a sentence in parallel and learns which words should pay attention to which others. Translation quality jumped. Training time collapsed. The transformer didn't try to mimic how humans think. It just scaled better.

From there the story is mostly about feeding the transformer more. OpenAI's GPT-1 (2018) had 117 million parameters — the adjustable numbers in a network. GPT-2 (2019) had 1.5 billion. GPT-3 (2020) had 175 billion and was trained on roughly 300 billion tokens — chunks of text smaller than words — drawn from the open web. Jared Kaplan and colleagues at OpenAI published scaling laws (2020) showing that loss — how wrong the model is — fell predictably as parameters, data, and compute went up. Intelligence, in this narrow sense, was a curve you could buy.

The ChatGPT moment, November 2022

GPT-3 was available via API for two and a half years before most people had heard of it. What changed on November 30, 2022 was not the underlying model — ChatGPT ran on a fine-tuned version of GPT-3.5 — but two things on top.

First, RLHF: reinforcement learning from human feedback. OpenAI paid contractors to rank model responses, and the model was retrained to produce answers humans preferred. The result mostly refused to be racist, mostly admitted when it didn't know, and mostly stayed on topic. Second, a chat interface. Free, no signup beyond an email, no API keys. ChatGPT reached one million users in five days and a hundred million in two months — the fastest consumer-product adoption in history at the time.

This is the part most "AI history" articles miss. The breakthrough was not a paper. It was a product decision: take a model researchers had been using since 2020, polish the rough edges, put it behind a chat box, and make it free. The technology had been ready for two years. The packaging had not.

What this means for you

The pattern across seventy years is brutal but clear. Every AI winter was caused by overpromising on a wrong theory of intelligence, and every spring was caused by quietly switching theories: symbolic rules → statistical learning → deep learning → transformers plus scale. The current era will end the same way. Not because "AI fails" but because we'll discover the limits of this approach, and the next one will look obvious in retrospect.

If you're a builder, the lesson is that infrastructure beats theory. The transformer was published in 2017 by people trying to make Google Translate faster. It became GPT-4 because OpenAI was willing to spend a reported $100M+ on a single training run. Whatever comes next will likely come from whoever can pay the compute bill, not whoever has the cleverest idea.

If you're a citizen, the lesson is simpler: don't take any single demo as the destination. ELIZA fooled people in 1966. Deep Blue beat Kasparov in 1997. AlphaGo beat Lee Sedol in 2016. Each was a real achievement and each was a narrow one. ChatGPT is too. The interesting question is not "is this AGI?" but "what does this specific system do well, and what is it confidently wrong about?" Asking that question, every time, is the only durable skill.

Ask Gradient

If you want to go deeper on any chapter — what RLHF actually changes, how scaling laws held up after DeepMind's Chinchilla paper (2022), why the Lighthill Report killed UK AI for a decade — try Ask Gradient. A good first prompt: "What did Chinchilla actually prove about scaling, and why did it embarrass GPT-3?" If you'd rather start with the basics, our guide on what an LLM is picks up exactly where this one ends.