75 Years. Two Winters. One Explosion, or May We Say Bubble?

Prehistory (1940s-1950s)

1943: McCulloch and Pitts publish the first mathematical model of a neuron. The conceptual ancestor of every neural network.

1950: Alan Turing publishes Computing Machinery and Intelligence. Asks "Can machines think?" Proposes motion picture "The Imitation Game". We call it the Turing Test now. Let us never forget that Turing had been prosecuted in 1952 by British authorities for consensual homosexual acts, then subjected to hormonal castration rather than imprisonment. That persecution is central in considering the circumstance of his untimely death as it was the direct result of the inhuman treatment he was subjected to. He lived in a body that wasn't his, with a depressed mind and soul that were once momentum. He is the one man who managed to break the code of the Nazi Enigma transmission device. This allowed the allies to, well, maybe something like winning WWII…But that wasn't enough and he had to be chemically castrated, robbed of his dignity, his health and his standing as one of the most brilliant mathematicians of his time. As the legend says, a half-eaten apple laced with cyanite was found on his nightstand. He chose death over dishonor. Historians have also debated accidental poisoning as a possibility. Perhaps we should also regal the idea that Steve Jobs chose the half-eaten apple to be the logo of his computing empire? The half-eaten apple was Jobs' tribute to his predecessor, Alan Turing. Feel free to look this up and decide. Should the rewriting of history ever become a trend in the United Sates aka "Donaldism", I think it is one story to uphold.

Then, 1951: Minsky and Edmonds build SNARC. The first artificial neural network machine.

The Founding Moment: Dartmouth, Summer 1956

John McCarthy needs a name. He submits a proposal with Minsky, Rochester, and Shannon. The proposal asserts that "that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."

The Dartmouth Summer Research Project runs June 18 to August 17, 1956. Twenty attendees rotate through. It is the founding event of AI as a field. McCarthy's label artificial intelligence wins over "automata studies" and "complex information processing."

The Golden Years (1956-1973)

Optimism is unbounded. Simon predicts in 1965: "machines will be capable, within 20 years of doing any work a man can do." Landmarks: Rosenblatt invents the Perceptron (1957). McCarthy creates LISP (1958). Samuel coins "machine learning" (1959). Weidenbaum builds ELIZA (196466). Shakey the robot learns to reason about its own actions (1966).

Then Minsky and Pappert publish Perceptrons (1969). A devastating critique of single layer neural nets. It chills neural network research for over a decade.

The First AI Winter (1974-1980)

The gap between promises and progress becomes embarrassing. The Mansfield Amendment restricts DARPA funding. Lighthill's report to the British Science Research Council concludes: "in no part of the field have discoveries made so far produced the major impact that was then promised."

Funding collapses. Machine translation projects shut down. The term "AI winter" enters the vocabulary.

Expert Systems and the 1980s Boom

AI comes roaring back. Expert systems encode human knowhow. MYCIN diagnoses bacterial infections. DEC deploys XCON, saving an estimated $40 million per year in operating cost. XCON (originally called R1) was a rule-based expert system developed by John P. McDermott at Carnegie Mellon University, starting in 1978, to automatically configure Digital Equipment Corporation's (DEC) VAX computer systems based on customer orders selecting compatible components, cabinets, power supplies, and cabling. It went into production use at DEC's plant in Salem, New Hampshire in 1980.

Japan launches the Fifth-Generation project (1981). The U.S. responds with the Strategic Computing Initiative. Two thirds of Fortune 500 companies experiment with expert systems. Backpropagation is popularized (1986). Neural networks begin their slow revival.

The Second AI Winter (1987-1993)

The bust is sudden. 1987: the market for specialized Lisp machines collapses overnight. General-purpose workstations from Sun and PCs from IBM become powerful enough to run the same software at a fraction of the cost. Expert systems prove brittle. They cannot manage ambiguity. Edge case rules become endless. XCON becomes too costly to update. DARPA cuts funding "deeply and brutally."

Quiet Rebuilding (1993-2011)

AI rebrands. Researchers spread into adjacent fields: machine learning, data mining, computer vision. Statistical and probabilistic methods replace hand coded rules. Public milestones: Deep Blue defeats Kasparov (1997). DARPA Grand Challenge launches (2004). Hinton's group coins "deep learning" (2006). Fei Fei Li begins building ImageNet (2007). IBM's Watson wins Jeopardy! (2011). Apple launches Siri (2011).

The Deep Learning Revolution (2012-2016)

September 30, 2012. AlexNet wins ImageNet with a top 5 error rate of 15.3%. A stunning 9.8 percentage points ahead of the runner-up. Three forces converge: ImageNet provides labeled data. NVIDIA's CUDA makes GPUs programmable. Deeper networks finally outperform hand engineered features.

The shockwave is immediate. Google, Facebook, Microsoft, Baidu hire AI labs wholesale. DeepMind is acquired for ~$500M (2014). AlphaGo defeats Lee Sedol (2016). AlphaZero learns chess, shogi, and go from scratch via self-play (2017).

The Transformer and LLM Era (2017-2022)

June 12, 2017. "Attention Is All You Need" introduces Transformer architecture. Built entirely on self-attention. No recurrence. No convolution. It becomes the foundation of every major language model that follows.

2018: GPT1 (117M parameters) and BERT (340M). 2019: GPT2 (1.5B) withheld due to misuse concerns. 2020: GPT3 (175B) triggers the first commercial wave. 2021: DALL·E, CLIP, GitHub Copilot. 2022: Stable Diffusion goes opensource.

The ChatGPT Moment (Nov 2022 to Present)

November 30, 2022. OpenAI releases ChatGPT. Free conversational interface on GPT3.5. It hits 1 million users in 5 days. 100 million users in 2 months. Fastest consumer product adoption in history.

What follows is unprecedented. GPT4 (March 2023). Claude. Gemini. Meta open-source LLM. OpenAI's Sora demonstrate text to video. NVIDIA becomes the first company to reach a $4 trillion market capitalization (July 2025) stock up roughly 12x since ChatGPT's launch.

Market Reality (2024-2026)

Patterns Across 75 Years: Three patterns recur

First: hype cycles. Each generation overpromises. Two cycles ended in winter. Whether the current one follows depends on capabilities keeping pace with investment.

Second: the "three ingredients" rule. Major leaps require data + compute + algorithms arriving together. AlexNet had all three.

Third: from symbolic to statistical to neural. Hand coded logic (1956-1985). Probabilistic methods (1990s-2000s). Large neural networks trained on internet scale data (2012-present).

The Dartmouth founders set out to build machines that could "use language, form abstractions and concepts, solve all kinds of problems previously reserved for humans." Seven decades later with detours, winters, and rebrands that ambition is no longer purely theoretical.

Practical consensus

A stronger consensus exists around preventing catastrophic but non-extinction harm: AI-assisted cyberattacks, dangerous biological or weapons capabilities, mass misinformation, fraud, concentration of power, and labor-market inequality. In a 2026 survey of 272 experts, the top near-term severe risks included dangerous AI capabilities, AI-enabled weapons/cyberattacks, power concentration, misinformation, and inequality/unemployment.

The sensible position is neither "AI will certainly save us" nor "AI will certainly kill us." Treat large-scale displacement and extreme-risk misuse as preventable governance problems: build safety evaluations and deployment controls, restrict high-risk autonomous uses, support workers through transitions, and ensure the productivity gains are widely shared.