A Brief History of AI Winters and Springs
You'll learn to
- -Understand why AI progress has historically come in boom-bust cycles
- -Recognize the major eras from the 1958 Perceptron to modern large language models
- -Explain with the XOR example why a single-layer perceptron hit a wall, and what a second layer changes
- -Name the three ingredients behind each breakthrough, and use them to judge a new AI claim
Tip: the highlighter is on - just select any text below to mark it. Use the highlighter button up top to change color or turn it off. Highlights are saved on this device only. Log in to save them to your account instead.
AI has not progressed steadily. It has moved in cycles of inflated hope, hard technical walls, and funding collapse, called "winters," followed by breakthroughs that restart the cycle, called "springs." Knowing this history helps calibrate expectations. The field has looked solved before and turned out not to be, more than once.
An Analogy: Seasons
Seasons are a good way to hold this history in your head. In a spring, a new idea works better than anyone expected, money and talent pour in, and promises grow. In summer, the hype peaks. Then autumn: the idea hits a hard limit the early demos hid, and the promises are not kept. In winter, funding dries up and the field goes quiet, while a few researchers keep working until the next idea arrives.
The First Spring and Winter (1950s-1970s)
Frank Rosenblatt's 1958 Perceptron, a single artificial neuron that could learn simple classification tasks, triggered enormous optimism. Then in 1969, Marvin Minsky and Seymour Papert proved that a single-layer perceptron mathematically could not solve certain simple problems, like the XOR logic gate. The result was a heavy blow to neural-network research, and funding dried up over the following years. The first AI Winter had begun.
Why One Neuron Was Not Enough: XOR
The wall the perceptron hit is easy to see. XOR means exclusive or: the answer is 1 when exactly one of the two inputs is 1. So (0, 0) gives 0, (0, 1) gives 1, (1, 0) gives 1, and (1, 1) gives 0. Draw those four points on paper: the two points that should give 1 sit on opposite corners, and so do the two points that should give 0.
A single perceptron can only draw one straight line to separate the points. No straight line puts both 1s on one side and both 0s on the other. That is what Minsky and Papert proved in 1969. The fix, a second layer of neurons, was already suspected: with two lines combined, the problem is easy. What was missing was a way to train more than one layer.
Backpropagation and a Second Spring (1980s-1990s)
The missing way to train more than one layer arrived with backpropagation, popularized in 1986: an efficient method to work out how much each weight in every layer should change. With it, multi-layer networks could solve XOR and much more, and interest returned. But on the computers and datasets of the time it could only be shown on small problems, so the excitement faded again and the field cooled through the 1990s.
The Deep Learning Era (2012-Present)
In 2012, a deep neural network called AlexNet dramatically outperformed every other approach on a major image recognition competition, made possible by GPUs finally providing enough compute and the internet providing enough labeled data. This kicked off the deep learning boom that produced modern computer vision, then large language models later in the 2010s and 2020s.
AlexNet's margin was large. On the 2012 ImageNet competition its top-five error (a miss counts only when the right answer is not among the model's five best guesses) was roughly 15%, against roughly 26% for the runner-up. The next big step came in 2017 with the transformer architecture, which scaled to far more data and compute than earlier designs and led to the large language models of the 2020s.
The Three Ingredients
Look across all the eras and every real breakthrough needed three things at once. The early waves each had one or two and lacked the rest:
- -Algorithms: an idea that works, such as the perceptron, backpropagation, convolutional networks or the transformer.
- -Data: enough examples to learn from. In 1986 there were no internet-scale labeled datasets.
- -Compute: hardware fast enough to train the model. GPUs made 2012 possible.
When you hear a bold new claim today, ask three questions. What exactly can it do, measured on what? What data and compute does it need? And what happens when it fails? An honest answer to all three is a sign of real capability; a vague answer to any of them is a sign of hype.
The pattern across every era is the same. A genuine breakthrough creates real capability, hype outruns that capability, reality catches up, and the field resets around whatever survived contact with reality. It is worth keeping this pattern in mind for any new claim you hear today.
The origin story below goes deeper into the people and papers behind these eras:
From Perceptrons to Reasoning Models
The papers, breakthroughs, and origin stories behind every idea in this course - from the 1958 Perceptron to modern LLM agents.
When machines first tried to think
Scale + GPUs = the deep learning revolution
The paper that changed everything
From 117M parameters to 100M users
Grounding AI in facts, not fiction
From text generators to autonomous workers
From notebook to 10 million users
Your company's leadership wants to make a large, multi-year bet based on a hot new AI capability. What does the history of AI winters suggest you should do before committing?
"Nothing, the current wave is different. It is already proven, so history does not really apply."
"History suggests distinguishing genuine, demonstrated capability from the hype layered on top of it before committing multi-year resources. Every past cycle had a real breakthrough underneath the hype, AlexNet's results were real, backpropagation's math was real, but the hype consistently outran what the technology could reliably do in production, and that gap is where expensive bets went wrong. I would want the capability demonstrated directly on our actual use case, not just an impressive general demo, before betting multiple years on it."
What caused the first "AI Winter" following the initial excitement around the Perceptron?
Discussion0
Join the Discussion
Sign in to leave comments, reply to others, or like insights.
No comments yet. Be the first to start the thread!