80+ years of breakthroughs — from McCulloch–Pitts neurons to GPT-4 and beyond
Warren McCulloch & Walter Pitts publish the first mathematical model of a neuron — laying the computational foundation for all future neural networks.
Donald Hebb publishes The Organization of Behavior, introducing Hebbian learning: "Neurons that fire together, wire together" — still core to modern DL.
Alan Turing publishes Computing Machinery and Intelligence — "Can machines think?" He proposes the Imitation Game (Turing Test) as a criterion for machine intelligence.
Marvin Minsky & Dean Edmonds build SNARC using 3,000 vacuum tubes to simulate 40 neurons — the first physical ANN ever constructed.
Arthur Samuel builds a checkers-playing program that improves by playing itself — the world's first practical demonstration of machine learning.
John McCarthy organizes a summer workshop at Dartmouth College and coins the term "Artificial Intelligence" — officially founding the field.
Frank Rosenblatt develops the Perceptron at Cornell — a single-layer ANN that learns from labeled examples. The first algorithm to automatically acquire knowledge.
Arthur Samuel uses the phrase "machine learning" in a paper on checkers — the first documented use of the term that now defines an entire discipline.
Joseph Weizenbaum at MIT creates ELIZA, a natural language processing program that simulates a psychotherapist. Users couldn't tell it was a machine — an early Turing-test moment.
Edward Feigenbaum's Dendral at Stanford identifies chemical compounds from mass spectrometer data — the first AI system to encode expert knowledge in a domain.
SRI International builds Shakey — the first mobile robot capable of reasoning about its own actions using logic, vision, and path planning. A landmark in embodied AI.
Minsky & Papert publish Perceptrons, proving mathematically that single-layer ANNs cannot solve XOR. Research funding collapses across the US and UK.
Sir James Lighthill's review concludes AI research has failed to achieve its goals. The UK Science Research Council cuts nearly all AI funding. The First AI Winter begins.
Paul Werbos introduces backpropagation in his PhD thesis — a method to train multi-layer networks. Largely ignored for a decade, it would later transform the field.
Digital Equipment Corporation deploys XCON (R1) to configure computer orders — saving $40M/year. AI enters the commercial world. Expert system market booms to $2B+ by 1988.
Rumelhart, Hinton & Williams publish the definitive backprop paper — enabling multi-layer neural networks to learn. The modern deep learning era begins from this moment.
John Hopfield introduces recurrent neural networks that can store and recall patterns — inspiring associative memory research and Boltzmann machines.
Yann LeCun applies backpropagation to convolutional neural networks at Bell Labs — successfully reading handwritten ZIP codes. The birth of practical CNNs.
Cortes & Vapnik publish SVMs — a powerful supervised learning algorithm based on statistical theory. Dominated ML benchmarks for over a decade before deep learning.
IBM's Deep Blue defeats world chess champion Garry Kasparov in a 6-game match — the first time a computer beats a reigning world champion under tournament conditions.
Hochreiter & Schmidhuber introduce Long Short-Term Memory networks — solving the vanishing gradient problem. LSTMs power speech recognition, translation & text for 20+ years.
LeCun's LeNet-5 is deployed by US banks to read millions of handwritten checks per day — the first wide-scale commercial application of convolutional neural networks.
iRobot launches the Roomba, using sensor fusion and ML-based navigation. Becomes the first widely-adopted consumer robot — over 40M sold to date.
Geoffrey Hinton publishes a method to efficiently train deep networks using Restricted Boltzmann Machines — reigniting deep learning research and launching the modern era.
Fei-Fei Li's team at Stanford releases ImageNet — 14 million labeled images across 20,000+ categories. The dataset that would fuel the CNN revolution of 2012 and beyond.
IBM's Watson defeats Jeopardy! champions Brad Rutter & Ken Jennings — demonstrating NLP, knowledge retrieval, and probabilistic reasoning at superhuman level in natural language.
Krizhevsky, Sutskever & Hinton's AlexNet wins ImageNet by a 10-point margin using a GPU-trained deep CNN. The moment that convinced the world deep learning had arrived.
Google's Mikolov introduces Word2Vec — encoding words as dense vectors where "king − man + woman ≈ queen". Transforms NLP and is still used in many systems today.
Ian Goodfellow invents Generative Adversarial Networks — two competing networks (generator vs discriminator) that produce photorealistic synthetic images. The foundation of modern generative AI.
DeepMind's AlphaGo beats 18-time world champion Lee Sedol 4–1 at Go — a game with more positions than atoms in the universe. Reinforcement learning + deep networks triumph.
Google Brain's paper introduces the Transformer architecture. Self-attention replaces RNNs, enabling parallel sequence processing. Every modern LLM (GPT, BERT, Claude, Gemini) is built on this.
Google releases BERT (Bidirectional Encoder Representations from Transformers) — contextual understanding of language in both directions. Transforms search engines, Q&A, and summarization.
OpenAI releases GPT-2 (1.5B params) — initially withheld due to fears of misuse. Its ability to generate coherent, fluent text previews the disruption to come with GPT-3.
OpenAI releases GPT-3 — 100× larger than GPT-2. Few-shot learning without fine-tuning. Can write essays, code, poetry, and answer questions across nearly any domain.
OpenAI's DALL·E generates images from text descriptions. CLIP connects vision and language in a single model. Midjourney, Stable Diffusion, and Adobe Firefly follow.
DeepMind's AlphaFold 2 accurately predicts 3D protein structures from amino acid sequences — solving a 50-year grand challenge in biology. Awarded the 2024 Nobel Prize in Chemistry.
OpenAI launches ChatGPT — the fastest product to 100 million users in history. LLMs enter daily life for writing, coding, tutoring, and beyond. The generative AI consumer era begins.
OpenAI (GPT-4), Anthropic (Claude), Google (Gemini), and Meta (Llama) release multimodal foundation models — capable of understanding text, images, audio, and code simultaneously.
OpenAI o1 introduces chain-of-thought reasoning. Sora generates photorealistic video. Autonomous AI agents complete multi-step tasks using tools, APIs, and memory without human input.
AI compute doubles every ~6 months. Governments worldwide introduce AI regulation frameworks. The debate on Artificial General Intelligence (AGI) timelines intensifies across labs and policy circles.