1943 — McCulloch & Pitts publish “A Logical Calculus of the Ideas Immanent in Nervous Activity,” the first mathematical model of an artificial neuron. They prove networks of simple binary units can compute any logical function.

1950 — Alan Turing publishes “Computing Machinery and Intelligence,” proposing the Imitation Game (Turing Test) as a practical measure of machine intelligence.

1956 — Dartmouth Workshop formally founds AI as a field. McCarthy coins the term “Artificial Intelligence.”

General Problem Solver (1957): Newell and Simon created a program that could solve logic puzzles using means-ends analysis — the first program to separate problem-solving strategy from domain knowledge.

ELIZA (1966): Joseph Weizenbaum’s chatbot at MIT simulated a Rogerian psychotherapist using simple pattern matching. Some users became emotionally attached, foreshadowing today’s debates about AI relationships.

Perceptron (1958): Rosenblatt’s learning machine could classify patterns by adjusting weights from examples — the first algorithm that learned from data.

Government funding flowed freely. DARPA (then ARPA) invested heavily in AI research at MIT, Stanford, and Carnegie Mellon. Researchers made bold predictions about imminent breakthroughs in language understanding, vision, and general reasoning.

1969 — Minsky & Papert publish Perceptrons, mathematically proving that single-layer perceptrons cannot solve XOR or any non-linearly-separable problem. This devastated neural network research for over a decade.

1973 — Lighthill Report (UK): Sir James Lighthill concluded that AI had failed to achieve its “grandiose objectives” and recommended drastic funding cuts. The British government largely withdrew AI funding.

1974 — DARPA cuts: The U.S. Congress pressured DARPA to fund only “mission-oriented” research, ending broad AI funding.

AI winters follow a predictable cycle: hype → overpromise → underdelivery → disillusionment → funding collapse. The gap between what researchers promised and what they delivered eroded trust. This pattern would repeat in the late 1980s and remains a cautionary tale today.

AI rebounded through expert systems — programs that encoded human expert knowledge as if/then rules. Companies paid millions for systems that could diagnose diseases, configure computers, or analyze chemical compounds.

R1/XCON (1980): DEC’s computer configuration system saved an estimated $40 million per year. By 1985, companies were spending over $1 billion annually on expert systems.

Japan’s Fifth Generation (1982): A $400 million government initiative to build intelligent computers spurred competitive investment from the US and UK.

Expert systems proved brittle, expensive to maintain, and unable to learn. Every new scenario required manual rule authoring by scarce domain experts. The specialized Lisp machine market collapsed as cheaper general-purpose workstations caught up.

Japan’s Fifth Generation project ended in 1992 without achieving its goals. DARPA’s Strategic Computing Initiative was cancelled. The AI industry contracted sharply.

Quietly, important work continued:
1989: Yann LeCun demonstrates CNNs for handwriting recognition (LeNet)
1992: Gerald Tesauro’s TD-Gammon masters backgammon via reinforcement learning
1997: IBM’s Deep Blue defeats Garry Kasparov in chess

Three forces converged:

1. Data: The internet produced massive datasets. ImageNet (2009, Fei-Fei Li) provided 14 million labeled images across 20,000 categories — the fuel deep learning needed.

2. Compute: GPUs, originally designed for gaming, turned out to be ideal for the parallel matrix operations neural networks require. Training that took weeks on CPUs took hours on GPUs.

3. Algorithms: Hinton’s 2006 deep belief networks showed that deep networks could be trained effectively. ReLU activation, dropout, and batch normalization followed.

In 2012, Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton entered AlexNet in the ImageNet competition (ILSVRC). It achieved a top-5 error rate of 15.3% — crushing the second-place entry at 26.2%. The gap was so large it shocked the computer vision community. Deep learning had arrived.

2017: Google researchers publish “Attention Is All You Need” (Vaswani et al.), introducing the transformer architecture. By replacing recurrence with self-attention, transformers process entire sequences in parallel — enabling massive scaling.

2018: GPT-1 (117M parameters) and BERT demonstrate that pretraining on large text corpora produces powerful language representations.

2020: GPT-3 (175B parameters) shows that scaling transforms quantity into quality — the model can write code, translate, and reason without task-specific training.

1. Hype cycles are real: Every era of AI has been marked by inflated expectations followed by disappointment. The current excitement about LLMs is unprecedented, but the pattern warns us to be realistic.

2. Compute + data > clever algorithms: The biggest breakthroughs (AlexNet, GPT-3) came from scaling existing ideas with more data and compute, not from entirely new algorithms.

3. The “bitter lesson”: Rich Sutton’s 2019 essay argues that methods leveraging computation (search, learning) always eventually beat methods leveraging human knowledge (hand-coded rules).