> History of AI_

McCulloch and Pitts published a mathematical model of an artificial neuron — showing that the logic of thought could be represented as on/off switches. Computers didn't even exist yet. But in a wartime paper that almost nobody read, the theoretical foundation for "thinking machines" was quietly laid. Every neural network alive today traces its lineage to this moment.

Alan Turing published "Computing Machinery and Intelligence" and posed the question that would haunt the next century: "Can machines think?" His proposal was elegant — if a machine could hold a text conversation indistinguishable from a human, it should be considered intelligent. Over 70 years later, the question remains unanswered. But the machines keep getting better at faking it.

Arthur Samuel at IBM built a checkers program that learned from experience — and eventually beat its own creator. It studied thousands of board positions, developed its own evaluation strategies, and improved without being explicitly programmed for each move. Samuel later coined the term "Machine Learning" in 1959, giving a name to the idea that machines could teach themselves. The concept that would take over the world started with a board game.

John McCarthy gathered the brightest minds to a summer workshop and coined the term "Artificial Intelligence." An era of wild optimism — participants genuinely believed AGI was 20 years away. McCarthy, Minsky, Shannon, Rochester — legends in a room, dreaming big. Reality would prove far more stubborn than any of them imagined.

Frank Rosenblatt invented the Perceptron — the first machine that could genuinely "learn" from examples. The New York Times ran the headline: "Navy reveals embryo of computer designed to read and grow wiser." The hype was enormous. Researchers declared it the dawn of machine intelligence. But a devastating critique was already brewing —

At Stanford, Edward Feigenbaum, Bruce Buchanan, and Nobel laureate Joshua Lederberg built DENDRAL — the first "expert system." It could identify organic compounds from mass spectrometry data, matching and sometimes exceeding the performance of human chemists. The idea was revolutionary: encode human expertise as rules, and the machine becomes the expert. This approach would dominate AI for two decades — and then spectacularly fail.

Joseph Weizenbaum created ELIZA at MIT. A simple rule-based program that mimicked a Rogerian therapist — it mostly just rephrased your own words back at you. Yet people became deeply captivated. Weizenbaum's own secretary asked him to leave the room so she could talk to it privately. The "ELIZA effect" revealed something unsettling: humans will project emotion and intelligence onto machines at the slightest provocation.

Marvin Minsky and Seymour Papert published "Perceptrons" — mathematically proving that single-layer perceptrons couldn't solve even simple problems like XOR. The book was devastating. AI funding evaporated almost overnight. Researchers scattered to other fields. The dream of thinking machines, barely a decade old, froze solid. The First AI Winter had begun.

Terry Winograd at MIT built SHRDLU — a program that could hold full English conversations about a virtual world of colored blocks. "Put the red block on the blue one." "Now move it." SHRDLU understood what "it" referred to. It handled context, pronouns, and ambiguity. The AI community was electrified. Human-level language understanding seemed imminent. But SHRDLU only understood its tiny blocks world. The real world — with its infinite ambiguity, sarcasm, and common sense — remained completely beyond reach. The optimism SHRDLU generated would contribute directly to the coming crash.

The UK government commissioned mathematician James Lighthill to evaluate the state of AI research. His verdict was merciless: "In no part of the field have the discoveries made so far produced the major impact that was then promised." British AI funding was nearly completely eliminated. DARPA followed with similar cuts across the Atlantic. AI researchers became academic pariahs.

Expert systems roared back with a vengeance. DEC's XCON configured computer orders using over 10,000 IF-THEN rules, saving $25 million per year. Stanford's MYCIN diagnosed bacterial infections with 69% accuracy — better than most doctors — but was never deployed due to ethical and legal concerns. Corporations poured billions into AI. For a brief, shining moment, it looked like the winter was over. It wasn't.

Japan's MITI launched the Fifth Generation Computer Project through ICOT, spending 54 billion yen (~$400 million) over ten years. The goal: build a parallel inference machine running at 1 Giga-LIPS that would leapfrog Western computing. The US and UK panicked and launched rival programs. The project failed commercially — parallel Prolog never took off. But it trained an entire generation of Japanese computer scientists and proved that national AI ambition could reshape global research priorities.

Hinton, Rumelhart, and Williams published the paper that made backpropagation practical for multi-layer neural networks. Finally, a way to train networks with hidden layers — to teach machines complex, layered representations of the world. The algorithm that would eventually power the deep learning revolution now existed. But the world wasn't ready. Data was scarce, computers were slow, and nobody was listening. The seeds were planted in frozen ground.

Expert systems crumbled under their own weight — brittle, expensive, impossible to maintain. The LISP machine market collapsed overnight, sending companies into bankruptcy. "AI" became a toxic label that repelled investors. Researchers learned to survive by camouflage, deliberately renaming their work "machine learning," "informatics," or "computational intelligence." The dream went underground again.

Yann LeCun at Bell Labs built LeNet — a convolutional neural network (CNN) that could read handwritten zip codes for the US Postal Service. Trained with backpropagation, it learned to recognize digits with remarkable accuracy. By the late 1990s, LeNet was processing millions of checks per day in ATMs across America. In the depths of the AI winter, while the mainstream had abandoned neural networks, LeCun's CNN quietly proved they had real-world value. If you deposited a check at an ATM in the 2000s, a neural network designed in 1989 was reading your handwriting. Some banks reportedly still run his code.

Gerald Tesauro at IBM created TD-Gammon — a program given nothing but the rules of backgammon. It played itself 1.5 million times and reached human world-champion level. No training data. No expert games to study. Pure self-play and reinforcement learning. More remarkably, TD-Gammon discovered strategies that humans had never considered in thousands of years of play. Professional backgammon players studied the program and changed how they played. Twenty-four years later, AlphaGo's legendary Move 37 — the move no human had ever played — was a direct descendant of what TD-Gammon proved possible.

Richard Wallace created A.L.I.C.E. and invented AIML — an open XML language for writing conversation rules. Released as open source, developers worldwide contributed tens of thousands of patterns. It won the Loebner Prize three times. The inspiration behind Spike Jonze's film "Her." Not intelligent — but proof that conversation AI still captivated the human imagination.

IBM's Deep Blue defeated world chess champion Garry Kasparov in a six-game match. But this wasn't neural networks or learning — it was brute-force search, evaluating 200 million positions per second. Kasparov said afterward that he sensed "a new kind of intelligence" across the table. He was facing a calculator. A phenomenally fast one. The world celebrated, but AI researchers knew: raw computation isn't thought.

Sepp Hochreiter and Jurgen Schmidhuber solved one of neural networks' most stubborn problems: the vanishing gradient. Their Long Short-Term Memory networks could finally remember information over long sequences — hundreds or thousands of steps back. LSTM was largely ignored for years. Then it quietly powered everything: Siri's voice recognition, Alexa's language understanding, Google Translate's breakthroughs. The architecture that taught machines to remember would dominate sequence modeling until the Transformer arrived.

Eleven autonomous vehicles navigated 60 miles of urban traffic — merging, turning, parking, and obeying traffic laws — with no human intervention. Carnegie Mellon's "Boss" won. Stanford's "Junior" came second. The technology was rough, the speeds were slow, and safety drivers were on standby. But it proved autonomous driving was possible. Within a decade, Google would spin out Waymo, Tesla would ship Autopilot, and self-driving would become a trillion-dollar race.

IBM's Watson crushed human champions Ken Jennings and Brad Rutter on the quiz show Jeopardy!, winning $1 million. The first time an AI "understood" natural language questions — puns, wordplay, obscure references — and answered them live on national television. Watson wasn't connected to the internet. It reasoned from 200 million pages of pre-loaded text. Jennings wrote on his answer screen: "I for one welcome our new computer overlords."

Google's Tomas Mikolov released Word2Vec, and a stunning equation emerged: "King - Man + Woman = Queen." Words could be represented as vectors in geometric space, and their relationships could be computed with simple arithmetic. The same year, Google acquired DeepMind for roughly $500 million — a startup with no products, only research papers and ambition. The AI arms race between tech giants had officially started. Talent became the scarcest resource on earth.

Ian Goodfellow, after an argument with friends at a bar in Montreal, went home and coded a new architecture in a single night. By morning, it worked. He called them Generative Adversarial Networks — two neural networks locked in competition: one forges, one detects. The forger gets better at fooling the detector. The detector gets better at catching fakes. Both improve endlessly. AI could now create photos, art, and music — not just classify. The era of generative AI had quietly begun.

Kaiming He at Microsoft Research introduced residual connections — a deceptively simple idea that let neural networks grow to 152 layers deep without collapsing during training. Previous networks topped out at around 20 layers. ResNet achieved 3.57% error on ImageNet. Human performance: 5.1%. For the first time in history, a machine could identify objects in photographs more accurately than the people who took them. The depth barrier was shattered.

OpenAI released GPT-1 in June, Google released BERT in October. A new paradigm crystallized: pre-train on massive text, then fine-tune on small task-specific data. No more building separate models for each problem. One foundation model, infinite applications. GPT-1 had 117 million parameters — modest by today's standards. But the architecture was right, and the insight was profound: language models trained on enough text develop emergent capabilities nobody explicitly programmed.

OpenAI refused to fully release GPT-2, citing risks of fake news generation. The first time AI researchers publicly declared their own creation "too dangerous to release." The AI community was split — half called it responsible caution, half called it marketing genius. The generated text was indistinguishable from human writing. Coherent paragraphs, consistent style, factual-sounding claims that were entirely fabricated. A preview of a world where you can't trust what you read.

OpenAI unveiled DALL-E: type a description, get an image. "An astronaut riding a horse in the style of Andy Warhol" — and there it was. GitHub launched Copilot, an AI that writes code alongside you, trained on billions of lines from open-source repositories. AI crossed a critical threshold: it moved from research labs into creative tools that ordinary people could use. Artists, writers, and programmers suddenly had AI collaborators. The question shifted from "can AI create?" to "who owns what AI creates?"

In August 2022, Stability AI released Stable Diffusion with fully open weights and code. Unlike DALL-E 2 or Midjourney, locked behind corporate APIs and waitlists, Stable Diffusion could run on any consumer GPU with just 2.4 GB of VRAM. The power to generate photorealistic images from text was suddenly free and unrestricted. The internet was flooded with AI-generated art overnight. Artists filed class-action lawsuits alleging copyright infringement. An AI-generated painting won a fine art competition, igniting a global debate about creativity and authorship. The question of whether AI should be open or controlled became the defining schism of the field.