This is Part 2 of a four-part series on the history of artificial intelligence.
Part 1 — Thinking automata and the dawn of computation (1920s–1960s) Part 2 — AI winter, expert systems, and the rise of machine learning (1970s–2000s) Part 3 — Deep learning, AlphaGo, and the language renaissance (2010s) Part 4 — ChatGPT, generative AI, and the Web4 horizon (2020s and beyond)
The First Cold Shower: AI Winter
By the early 1970s, the optimism began to cool.
Machines were supposed to start teaching us. Instead they stalled at puzzles we thought they'd already solved. AI was clever inside narrow boxes — chessboards, medical labs, single-purpose domains. Outside those boxes, it floundered. Common sense, contextual reasoning, the kind of fluid generalization any five-year-old does effortlessly — all unreachable.
In 1973, a UK government report by Sir James Lighthill delivered a cutting verdict: artificial intelligence had failed to meet its promises. Funding was pulled. Research programs shut down across Britain. A wave of skepticism rolled through the U.S. and Europe.
The term AI itself became a kind of taboo — a synonym for overpromising and underdelivering.
Why the First Approach Failed
Part of the problem was conceptual. The perceptron — once heralded as a breakthrough in machine learning — couldn't even solve basic logic problems like XOR. The book Perceptrons by Marvin Minsky and Seymour Papert made the limitations painfully clear and effectively buried neural networks for a generation.
The other problems were physical. Computers were slow. Memory was scarce. The ambitious attempts at automatic translation and speech recognition collapsed under their own complexity. One Cold War-era project aimed to translate Russian to English for military use. After years of work, the output was better suited for comedy than combat.
Yet AI didn't die. It changed faces.
Researchers turned from building general minds to solving narrow, specific problems. Quietly, the field began to rebuild — under a different name.
A Second Wind: Expert Systems (1980s)
The 1980s brought AI back from the brink. Not through magic. Through expertise.
The vision of machines that could think like humans gave way to something more useful: systems that could mimic a specialist in a narrow field.
MYCIN, XCON, and the Commercial Awakening
One of the early pioneers was MYCIN, developed at Stanford. It diagnosed infectious diseases and recommended antibiotics. MYCIN was never deployed clinically — liability concerns kept it sidelined — but it matched the performance of average physicians. The lesson was significant: AI didn't need to think like a human to be useful.
Then came XCON, built by Digital Equipment Corporation. It automatically configured VAX computers for customers, saving the company millions of dollars in operations. Suddenly, businesses took notice. AI wasn't a pipe dream. It was cutting costs.
This narrow utility sparked a wave of excitement. In 1982, Japan launched its Fifth Generation Computer Systems project — aiming to build machines that could reason, process natural language, and transform global computing. The West, wary of falling behind, doubled down on funding. Startups proliferated. New languages emerged (CLIPS, OPS5). Expert-system shells made it easier to encode knowledge bases.
By 1985, AI was "hot" again. Fortune featured it on the cover: Why AI Is Back in Style. It looked like a renaissance.
The Cracks Underneath
The systems had flaws.
They were expensive to maintain. Their knowledge had to be hand-coded by domain experts — a slow, expert-dependent process that didn't scale. And when the world changed — markets, regulations, scientific consensus — the systems became obsolete fast. Keeping them relevant required constant, costly updates.
Japan's ambitious Fifth Generation project quietly fell short. Despite real progress on parallel computing and logic programming, the revolution it promised never quite arrived. By the end of the 1980s, disillusionment returned. Industry backed away. Universities kept digging.
What they found in the 1990s would change everything.
The Paradigm Shift: Machines That Learn (1990s)
In the early 1990s, something began to stir — slowly, mostly behind closed doors.
The new idea was deceptively simple: don't teach the machine. Let the machine learn.
From Rules to Data
The shift from hard-coded logic to data-driven learning was profound.
Statistical models. Bayesian inference. Probabilistic reasoning. These approaches replaced the brittle, handcrafted rule sets of expert systems. In 1993, IBM published a new approach to machine translation: instead of grammar rules, it used massive datasets of paired sentences. Meaning came from pattern, not syntax.
Meanwhile, the internet changed everything. Suddenly, data was everywhere. Texts, images, clicks, transactions — an infinite stream of training material. Hardware caught up. Cheaper computers, early GPU acceleration, and access to massive corpora let researchers run experiments that were previously impossible.
Deep Blue Beats Kasparov
Then came the victories.
In 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov. It wasn't a thinking machine in the human sense — more a highly specialized chess engine combining brute-force search with hand-tuned evaluation. But the outcome rewrote public opinion overnight. For the first time in decades, people believed in AI again.
Behind the scenes, a second wave was building. Speech recognition kept improving. Neural networks quietly came back from the dead, augmented by larger datasets and better optimization. Data mining became its own discipline. Labs were humming. Algorithms were evolving.
AIBO and the First Domestic Hint
In 1999, Sony released AIBO — a robotic dog that could learn tricks. It was a toy. But it pointed toward a future where machines wouldn't just respond. They would adapt.
The vocabulary of AI changed. Words like training, model, dataset, validation replaced rules, expert, inference engine. A new generation of researchers grew up speaking the new language.
Algorithms That Learn (2000s)
As the new millennium dawned, the rules changed for good.
Computers became ubiquitous. The internet stopped being a novelty and became infrastructure. The world began drowning in data: billions of emails, clicks, purchases, photos, GPS pings, financial transactions. All of it — fuel for the next generation of AI.
The old dream of "hard-coding intelligence" finally died. A new mantra replaced it: give the machine enough examples, and it will learn. Machine learning wasn't just a method — it became a mindset. No longer theory-driven, it was unapologetically empirical.
Search, Recommendations, and the Crowd's Implicit Wisdom
Search engines like Google began to train ranking algorithms on user behavior. What did people click? Which links did they linger on? Which words did they search next? Relevance was no longer defined by experts — it was learned from the crowd.
In 2006, Netflix launched the famous Netflix Prize: a million dollars to anyone who could improve their recommendation algorithm. Thousands of teams joined. New techniques were born. The lesson: the best technology often doesn't come from rules. It comes from listening to data.
Banks adopted machine learning to detect fraud. Doctors began using it to analyze X-rays. Retailers used it to forecast demand. Human intuition became optional — if you had the right datasets.
The Era of Benchmarks
Research advanced in parallel. Open datasets became the new gold: MNIST for handwritten digits. CIFAR for image recognition. ImageNet for real-world photos. These benchmarks sparked global competitions: who could train the most accurate, most efficient, most elegant model? The annual leaderboards became scoreboards for an entire scientific community.
AI was no longer a spectacle. It was an assistant.
In 2011, IBM Watson competed on Jeopardy! — the legendary American quiz show — and beat its best human champions. Watson didn't just retrieve facts. It analyzed puns, double meanings, metaphors, ambiguity. And it did so at speed.
That same year, Apple released Siri. Not at a conference. Not in a lab. In your pocket. Suddenly you could speak to your phone and it would answer. AI stopped being a theory and became a service.
Machines Take the Wheel
While Siri got smarter, machines started to drive.
In 2004 and 2005, DARPA — the U.S. military's research agency — hosted the Grand Challenge in the Nevada desert: build a car that can drive 200 kilometers across a desert without a human inside. In 2004, no team finished. In 2005, five cars made it to the end. Stanford's robot vehicle Stanley won.
This was no longer science fiction. It was code, sensors, lidar, computer vision. AI behind the wheel wasn't a metaphor anymore.
In 2007, DARPA raised the bar with the Urban Challenge: now the cars had to navigate a simulated city with traffic rules and dynamic obstacles. Google took notice. In 2009, it launched the self-driving car project that would later become Waymo. By the early 2010s, test cars were quietly weaving through California traffic. Inside: neural networks. Behind the wheel: no one.
GPUs — The Quiet Hero
None of this would have happened without one supporting actor: the GPU.
Graphics processing units, originally built to render video games, turned out to be a near-perfect fit for the parallel math that neural networks demand. Tasks that once took weeks now took hours. Then minutes.
Soon came specialized chips: Google's TPU. Apple's Neural Engine. Dozens more from startups and incumbents. AI was accelerating — and becoming less artificial, more real.
The Doorstep of a Revolution
By 2010, the world was ready. Ready for the next step. For deep learning.
Neural networks — the same idea that had been shelved in the 1970s — were coming back. This time deeper, faster, and fed with orders of magnitude more data. And most importantly: they worked.
At first the change was quiet. Deep networks crept into text recognition, financial forecasting, simple recommendation engines. Nothing yet to call a "revolution."
Then, in 2012, that changed.
A deep convolutional neural network called AlexNet entered the ImageNet competition — a global challenge to identify objects in photos. It crushed the field. Its error rate was dramatically lower than anything seen before.
AlexNet had eight layers of neurons. It was trained on a million images. And it was powered by a single GPU. What used to take weeks took days. The results were astonishing.
Overnight, deep learning became the mainstream. Recognition. Prediction. Classification. Everything began to change.
The next decade would be defined by what AlexNet started. That's where Part 3 picks up.
Continue to Part 3 — Deep learning, AlphaGo, and the language renaissance (2010s).
