chapter 00 / history · prologue · estimated time 45-60 min
Artificial intelligence did not march in a straight line to where we are today. It lived through two fevered springs, two bone-chilling winters, several wars over methodology (symbolism vs. connectionism, rules vs. learning), and a handful of moments when "the entire field was rewritten by a single paper." Understanding this history is not just nostalgia—the cause of every winter (overpromising) and the recipe for every breakthrough (algorithm × data × compute arriving at the same time) are still playing out today.
Below is a canvas you can pan and zoom without limit. Seven lanes correspond to AI's seven main threads, and the larger a dot is, the more important the event. Use the scroll wheel to zoom into the 2012-2026 stretch—and you'll get a visceral sense of what "more has happened in the last decade than in the previous seventy years combined" really means.
AI's starting point was not computer science but a marriage of neuroscience and logic: in 1943, McCulloch and Pitts proved that networks of neurons can implement arbitrary logical operations. In 1950, Turing gave an operational criterion for "what counts as intelligence." In 1956, at the Dartmouth workshop, "Artificial Intelligence" was formally named and established as a discipline.
Over the next twenty years, two routes sprouted at the same time but met very different fates: symbolism (logic, search, rules) became the mainstream orthodoxy—chess programs, theorem provers, ELIZA; connectionism (neural networks), after the perceptron frenzy of 1957, was cast into the cold by Minsky's 1969 book Perceptrons, which proved that "a single-layer network can't even learn XOR."
In the 1980s, AI regained its commercial ticket through expert systems: encoding human experts' knowledge into thousands of if-then rules. XCON saved DEC tens of millions of dollars a year, Japan launched its "Fifth Generation Computer" program, and the US responded with matching investment—a second spring. But the maintenance cost of rule bases exploded with scale, the bubble burst after 1987, and the second winter set in.
The real spark was lit inside the cold cell: in 1986 backpropagation made multilayer networks trainable (a direct answer to the 1969 critique), in 1989 LeCun's CNN began reading handwritten checks, and in 1997 the LSTM solved the vanishing-gradient problem in recurrent networks. The connectionists laid the infrastructure brick by brick in years when no one was cheering—Hinton, LeCun, and Bengio would later share the Turing Award, a reward precisely for those years spent riding the bench.
Two winters taught the field humility: don't talk about "intelligence," solve concrete problems first. The SVM (1995) dominated academia with its elegant convex-optimization theory and powerful performance on small data; random forests and boosting engineered "the wisdom of the crowd"; machine learning quietly entered spam filtering, credit scoring, and search ranking—AI was everywhere, just nobody called it AI. The methodology distilled in this period (train/test splits, cross-validation, regularization) is exactly the fundamentals you just learned in Chapter 1.
Meanwhile, three seemingly unrelated events were storing up energy for the detonation: the internet accumulated massive amounts of data; Fei-Fei Li released ImageNet in 2009; and NVIDIA's gaming GPUs were found to be remarkably well suited to matrix operations (CUDA was released in 2007). Algorithm, data, compute—the three ingredients were about to be complete.
On September 30, 2012, AlexNet cut the error rate in the ImageNet competition from 26.2% to 15.3%. The shock of that number lay in its cliff-like drop: previous champions had improved by a toothpaste-squeeze of 1-2 percentage points each year. Within three years, every subfield of computer vision was rewritten by CNNs; in 2015 ResNet surpassed human-level performance on ImageNet; and in 2016 AlphaGo's move 37 let the whole world—especially China and Korea—see machine "intuition" with their own eyes.
The 2017 paper Attention Is All You Need was meant to be just a machine-translation paper, yet it accidentally provided a general architecture that was a perfect fit for GPU parallelism and whose performance grew smoothly with scale. The GPT series began a high-stakes bet: scale up the same architecture tenfold, a hundredfold, generation after generation. In 2020, GPT-3 (175B) proved the bet right—scale brought not only quantitative change but the qualitative, emergent ability to "learn a new task from just a few examples." That same year, the Scaling Laws paper turned "buying intelligence with compute" into a computable engineering decision.
On November 30, 2022, ChatGPT launched. One hundred million users in two months. The pace that followed you've probably lived through yourself: GPT-4 and the move to closed source, LLaMA and the open-source ecosystem, multimodality, the reasoning models o1/R1 and the new dimension of "test-time compute," MCP and the tool ecosystem, agentic coding rewriting software engineering... The timeline gets crowded here, and that's not an illusion—it's the real acceleration of this field. Pull the timeline all the way to the right: the frontier questions of 2026—long-horizon agents, multi-agent collaboration, RL with verifiable rewards—are exactly the technical details that Chapters 8 and 11 of this course will dig into.
Having brought the history this far, this is the perfect moment to watch a documentary. DeepMind's official AlphaGo film documents in full the 2016 Seoul matches—it's both a visual archive of Act Four's climax and the best material for understanding the impact of "machine intuition":