A Graduate Discussion Primer
The Climate of Machine Intelligence
AI history does not move in a straight line — it moves like weather. Warm springs of funding and optimism give way to cold winters of disillusionment, then boil over into hot booms of rapid deployment. This timeline reads 40 milestones from 1943 to 2025 through that lens, and argues that the winters were failures of hardware and funding, not of ideas.
Scroll, filter by theme, or click a band in the barometer below to jump to an era.
The Barometer — click a band to jump there
Foundations
Warm1943–1955
The theoretical seeds of AI are planted before electronic computers can even run them at scale: formal neuron models, Hebbian learning, and Turing's operational test for intelligence.
A Logical Calculus of the Ideas Immanent in Nervous Activity
Warren McCulloch, Walter Pitts
McCulloch and Pitts propose a mathematical model of an artificial neuron, showing that networks of simple binary threshold units can in principle compute any logical function — the first formal theory of neural computation.
The Organization of Behavior
Donald Hebb
Hebb proposes that when one neuron repeatedly helps fire another, the connection between them strengthens — later summarized as 'cells that fire together, wire together.' This principle underlies most learning rules used in neural networks today.
Computing Machinery and Intelligence
Alan Turing
Turing proposes the 'imitation game' — later known as the Turing Test — as a practical way to sidestep the philosophical question 'can machines think?' in favor of an operational, behavioral test.
The Golden Years
Warm1956–1973
A wave of optimism follows the Dartmouth workshop's naming of the field, but the era closes with Minsky and Papert's mathematical critique and Lighthill's damning government review — a golden age that ends in its own indictment.
The Dartmouth Summer Research Project
Pivot MomentJohn McCarthy, Marvin Minsky, Nathaniel Rochester, Claude Shannon
An eight-week workshop at Dartmouth College brings together researchers to explore the conjecture that every aspect of learning or intelligence can in principle be so precisely described that a machine can simulate it. McCarthy coins the term 'artificial intelligence' in the funding proposal, giving the field its name and its founding optimism.
Consequence
Names and launches AI as a distinct research discipline, setting a decades-long agenda around symbolic reasoning and search.
The Perceptron
Frank Rosenblatt
Rosenblatt publishes the perceptron — a single-layer neural network that learns to classify simple patterns, first demonstrated in software and soon after as the custom-built Mark I hardware — and claims it as the seed of a machine that could eventually walk, talk, and reproduce itself.
Samuel's Checkers Player
Arthur Samuel, IBM
Samuel writes a checkers-playing program that improves through self-play, coining the term 'machine learning' along the way and demonstrating that a program could get better at a task without being explicitly reprogrammed.
ELIZA
Joseph Weizenbaum, MIT
Weizenbaum builds a simple pattern-matching chatbot that imitates a Rogerian psychotherapist. To his alarm, many users attribute genuine understanding and empathy to the program, an early demonstration of how easily people anthropomorphize machines.
Perceptrons
Pivot MomentMarvin Minsky, Seymour Papert
Minsky and Papert publish a rigorous mathematical analysis showing that single-layer perceptrons cannot compute the XOR function, a basic non-linear problem. Though multi-layer networks could solve it in principle, the book is widely read as a verdict against neural networks altogether.
Consequence
Funding for neural network research collapses for over a decade; the field pivots hard toward symbolic AI, helping set up the First AI Winter.
The Lighthill Report
Pivot MomentSir James Lighthill, UK Science Research Council
Commissioned by the British government, mathematician James Lighthill's report concludes that AI research has failed to achieve its grandiose objectives and that the 'combinatorial explosion' problem makes many approaches fundamentally intractable at scale.
Consequence
UK funding for AI research outside a few universities is dismantled almost overnight, and the report's pessimism ripples across US funders too — the opening act of the First AI Winter.
The First AI Winter
Cold1974–1980
Funding, not ideas, freezes first. Even during the drought, Fukushima's Neocognitron quietly lays groundwork for the convolutional networks that won't matter for another thirty years.
DARPA Cuts Speech and AI Funding
U.S. Defense Advanced Research Projects Agency
Following disappointing results from speech-understanding research and mounting Congressional skepticism, DARPA sharply cuts undirected AI research funding in the United States, echoing the Lighthill Report's conclusions in the UK.
The Neocognitron
Kunihiko Fukushima
Working in relative obscurity during the funding drought, Fukushima designs a hierarchical, multi-layered neural network for visual pattern recognition, with alternating layers directly foreshadowing the convolutional and pooling layers of modern CNNs.
Expert Systems & the Second Boom
Warm1980–1987
Rule-based expert systems briefly deliver real commercial value, and Rumelhart, Hinton, and Williams answer the decade-old XOR objection to neural networks — but the industry that grows up around it is built on fragile, specialized hardware.
XCON deployed at Digital Equipment Corporation
John McDermott, Carnegie Mellon / DEC
DEC deploys XCON, a rule-based expert system that configures orders for VAX computer systems, saving the company an estimated $25 million a year. Its commercial success helps ignite a boom in expert-system startups.
The Fifth Generation Computer Systems Project
Japan's Ministry of International Trade and Industry
Japan launches a massive, decade-long national initiative to leapfrog Western computing by building parallel, logic-programming based machines for AI. The project spurs matching government investment in the US and UK, fearful of falling behind.
Learning Representations by Back-Propagating Errors
Pivot MomentDavid Rumelhart, Geoffrey Hinton, Ronald Williams
The trio publish a landmark paper popularizing backpropagation as a practical way to train multi-layer neural networks, directly answering the XOR objection Minsky and Papert raised seventeen years earlier.
Consequence
Revives neural network research technically, but the idea sits mostly dormant for 26 years — waiting on data and compute that don't yet exist. The convergence doesn't happen until AlexNet in 2012.
The Second AI Winter
Cold1987–1993
The LISP machine hardware market's collapse — not a failure of the underlying ideas — triggers the second winter. LeCun's LeNet, built and deployed during these very years, proves the thesis: winters starve funding and hardware, not theory.
The LISP Machine Market Collapse
Pivot MomentSymbolics, Lisp Machines Inc., and the expert-systems industry
Specialized LISP machine hardware, once the backbone of the expert-systems boom, is wiped out almost overnight by cheaper general-purpose workstations from Sun and PCs running competing software. Venture capital and corporate AI budgets evaporate within a year.
Consequence
Demonstrates that Winter II is a commercial failure, not a scientific one — the underlying expert-system ideas still worked, but the business model didn't survive contact with cheaper hardware.
LeNet
Yann LeCun, Bell Labs
LeCun applies backpropagation to a convolutional neural network trained to read handwritten zip codes for the US Postal Service, one of the first real commercial deployments of a deep neural network — built in the depths of Winter II.
Quiet Progress — Statistics Wins
Warm1993–2011
Away from the AI label entirely, statistical and connectionist methods quietly win: LSTM (1997) solves recurrent learning, cloud computing and CUDA remove infrastructure barriers, and Fei-Fei Li's ImageNet reframes the whole bottleneck as data, not algorithms — not unlike LeNet before it, LSTM is another idea that outlives its funding winter.
Long Short-Term Memory
Sepp Hochreiter, Jürgen Schmidhuber
Hochreiter and Schmidhuber introduce a recurrent network architecture with gated memory cells that solves the vanishing-gradient problem plaguing earlier recurrent networks, enabling learning over much longer sequences.
Deep Blue Defeats Kasparov
IBM
IBM's Deep Blue, a specialized chess-playing supercomputer relying on brute-force search rather than learning, defeats reigning world champion Garry Kasparov in a six-game match — a landmark for engineered search, not machine learning.
Amazon Web Services Launches EC2
Amazon
Amazon opens Elastic Compute Cloud, letting anyone rent computing power by the hour instead of buying servers outright. Cloud infrastructure quietly removes one of the biggest barriers to running large-scale experiments.
Hinton Popularizes 'Deep Learning'
Geoffrey Hinton, Simon Osindero, Yee-Whye Teh
Hinton and colleagues show that deep belief networks can be trained layer-by-layer using an unsupervised pretraining trick, reviving serious academic interest in deep neural networks and giving the approach its modern name.
NVIDIA Releases CUDA
NVIDIA
NVIDIA opens up its graphics processors for general-purpose parallel computing with the CUDA toolkit, turning gaming hardware built for rendering pixels into an accessible platform for the matrix multiplications neural networks depend on.
ImageNet
Pivot MomentFei-Fei Li and collaborators, Princeton/Stanford
Li's team release a labeled image dataset assembled largely through crowdsourced annotation on Mechanical Turk — growing to over 14 million images across more than 20,000 categories — and in 2010 launch the annual ILSVRC challenge to benchmark visual recognition algorithms against it.
Consequence
Reframes the bottleneck in AI progress as data, not algorithms — the neural network architectures needed already existed; what had been missing was a dataset large enough to train them well.
Watson Wins Jeopardy!
IBM
IBM's Watson system defeats the show's two greatest champions using a combination of natural language processing, information retrieval, and statistical confidence scoring — showcasing how far statistical methods had come.
The Deep Learning Boom
Hot2012–2019
Backpropagation (1986), ImageNet-scale data (2009), and GPU compute (CUDA, 2007) finally converge in AlexNet, and deep learning becomes the default approach across vision, games, and eventually language.
AlexNet Wins ImageNet
Pivot MomentAlex Krizhevsky, Ilya Sutskever, Geoffrey Hinton
A deep convolutional network trained on two consumer NVIDIA GPUs cuts the ImageNet error rate nearly in half compared to the previous year's best, stunning the computer vision community.
Consequence
The convergence point: 1986 backpropagation, 2009 ImageNet-scale data, and 2007 CUDA-enabled GPU compute finally meet at once. Deep learning goes from academic curiosity to the dominant paradigm within two years.
Google Acquires DeepMind
Google buys the four-year-old London AI lab DeepMind for roughly $500 million, betting on deep reinforcement learning research well before it has an obvious product, signaling how seriously industry now takes deep learning talent.
Generative Adversarial Networks
Ian Goodfellow et al.
Goodfellow proposes training two networks against each other — a generator trying to fool a discriminator — as a new way to train models that produce realistic synthetic images, opening the door to modern generative modeling.
OpenAI Is Founded
Sam Altman, Elon Musk, Ilya Sutskever, Greg Brockman, and others
A group of researchers and technologists launch OpenAI as a non-profit lab explicitly dedicated to ensuring artificial general intelligence benefits humanity, backed by $1 billion in pledged funding.
AlphaGo Defeats Lee Sedol
DeepMind
DeepMind's AlphaGo, combining deep neural networks with Monte Carlo tree search and reinforcement learning, defeats world champion Lee Sedol 4-1 at Go, a game long considered far too complex for brute-force computer play.
Attention Is All You Need
Pivot MomentAshish Vaswani et al., Google Brain
Google researchers introduce the Transformer architecture, discarding recurrence entirely in favor of self-attention mechanisms that can be parallelized across massive amounts of hardware and data.
Consequence
Becomes the architecture behind essentially every major large language model that follows, from BERT and GPT to Claude and Llama.
BERT and GPT-1
Google AI; OpenAI
Within months of each other, Google and OpenAI release large pretrained Transformer language models — BERT and GPT-1 — showing that pretraining on unlabeled text and fine-tuning for specific tasks beats task-specific architectures across the board.
GPT-2
OpenAI
OpenAI trains a 1.5-billion-parameter language model and initially withholds the full model, citing concerns about malicious use of its fluent generated text — an early flashpoint in the debate over responsible AI release.
The Generative Era
Hot2020–2025
Scale itself becomes the strategy: GPT-3 shows what happens when Transformers meet internet-scale data, and ChatGPT turns a research artifact into the fastest-adopted consumer product in history.
GPT-3
Pivot MomentOpenAI
OpenAI releases a 175-billion-parameter language model demonstrating that scaling Transformer models on internet-scale text yields broad few-shot capabilities across tasks it was never explicitly trained on.
Consequence
Establishes scale itself as a research strategy, kicking off an industry-wide race to train ever-larger models and setting the stage for the chat-based products of 2022.
AlphaFold2
DeepMind
DeepMind's AlphaFold2 predicts 3D protein structures from amino acid sequences with accuracy rivaling experimental methods, cracking a 50-year-old grand challenge in biology and demonstrating deep learning's reach beyond language and vision.
Anthropic Is Founded
Dario Amodei, Daniela Amodei, and former OpenAI researchers
A group of former OpenAI safety researchers found Anthropic, an AI safety company focused on building steerable, interpretable, and reliable large models — which later releases the Claude family of assistants.
Stable Diffusion
Stability AI, CompVis, Runway
An openly released, freely runnable text-to-image diffusion model puts high-quality image generation on consumer hardware, rapidly spawning a large ecosystem of tools, fine-tunes, and creative applications.
ChatGPT Launches
Pivot MomentOpenAI
OpenAI packages a GPT-3.5-class model behind a free conversational interface. It reaches an estimated 100 million users within two months, among the fastest consumer product adoptions on record.
Consequence
Moves large language models from research demos into daily consumer and enterprise use overnight, triggering the current wave of investment, competition, and public debate about AI.
GPT-4, Claude, and Llama
OpenAI, Anthropic, Meta
Multiple labs release next-generation large language models within months of each other — GPT-4's multimodal reasoning, Anthropic's Claude, and Meta's open-weight Llama — turning what had been a single-vendor story into genuine platform competition.
NVIDIA H100 GPUs Ship at Scale
NVIDIA
NVIDIA's Hopper-architecture H100 GPUs, purpose-built for transformer workloads, become the most sought-after hardware in the industry, with demand so high that GPU access becomes a genuine strategic bottleneck for AI labs and nations alike.
Reasoning Models Emerge
OpenAI (o1), and others
OpenAI releases o1, a model trained to spend extra inference-time computation deliberating step-by-step before answering, trading latency for higher accuracy on hard math, coding, and logic problems — a new axis for scaling beyond just parameter count.
Agentic AI and DeepSeek
Multiple labs; DeepSeek
Models increasingly act as agents — planning, calling tools, and executing multi-step tasks with less human supervision — while Chinese lab DeepSeek releases highly capable, low-cost open-weight models, intensifying global competition and dispelling assumptions about the cost of frontier training.
Sources
- Putchuon. (n.d.). The entire history of artificial intelligence [Video]. YouTube.
- IEEE Computer Society. (n.d.). The evolution of AI. IEEE Computer Society.
- Roser, M., Hasell, J., Herre, B., & Mathieu, E. (2023). A brief history of artificial intelligence. Our World in Data.
- Lighthill, J. (1973). Artificial intelligence: A general survey. Science Research Council.
- Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323(6088), 533–536.
- Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25.
- Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.
- Russell, S., & Norvig, P. (2020). Artificial intelligence: A modern approach (4th ed.). Pearson.
AI Tools Used — Disclosure
Claude was used to compile milestones and draft copy for this timeline. Dates, attributions, and the stated causes of each AI winter were verified against the sources listed above. The climate framing, the selection of which moments count as pivots, and the thesis that both AI winters were failures of hardware and funding rather than of ideas are the author's own editorial choices.