When Sen. Bernie Sanders' briefing convenes on September 16, one name on the witness list carries more history than any headline about it will mention. Geoffrey Hinton, the researcher press coverage calls the "Godfather of AI," spent decades on an idea the rest of the field had declared dead. He was still working on it when almost nobody else was funded to. Understanding why the field abandoned that idea twice, and what actually brought it back, explains something the first three parts of this series could not: why so much of today's alarm is aimed at a technology whose comeback is barely fifteen years old.
The First Winter
Artificial intelligence research had already crashed once by the time most people alive today were born. In 1966, a National Research Council report concluded that machine translation, one of the field's first big promises, was not competitive with human translation on cost, speed, or quality, and saw no near-term prospect of becoming useful, after something on the order of twenty million dollars in government funding had gone into the effort over the prior decade. Funding for that line of work ended almost immediately. Three years later, a book by Marvin Minsky and Seymour Papert laid out the mathematical limits of a simple neural network design called the perceptron, a narrow critique the field nonetheless treated as a verdict on the whole approach, and mainstream research into neural networks stalled for most of a decade on the strength of it. Then in 1973, a UK government report by Sir James Lighthill concluded that AI research had failed to deliver on its own stated goals and identified a problem, combinatorial explosion, that made many of the field's algorithms unworkable outside of toy examples. British funding dried up immediately, and the freeze held across the field for roughly six years, until a new approach called expert systems revived interest in the early 1980s. In the United States, a 1969 law requiring defense research to serve a specific military purpose gave the Pentagon a reason to cut its own AI funding the following year, including several speech-research programs, among them work underway at Carnegie Mellon.
The Second, Longer Winter
The field clawed back into fashion in the early 1980s around a new idea called expert systems, programs meant to capture a human specialist's knowledge as a set of rules. Companies built an entire hardware industry, specialized computers called Lisp machines, to run them. Then in 1987 that specialized hardware market collapsed within about a year, once ordinary workstations from companies like Sun and Apple started outperforming the specialized machines for a fraction of the price. The expert systems themselves turned out to have a deeper problem: they were expensive to maintain, could not learn from new information, and broke in brittle ways outside the narrow scenarios they were built for. A year later, the officials then overseeing DARPA's Strategic Computing Initiative cut its funding sharply, having grown skeptical that expert systems amounted to much more than clever programming dressed up as intelligence. Japan's own attempt to leapfrog the field, a multi-hundred-million-dollar national effort called the Fifth Generation Computer project aimed at building machines that could reason and hold conversations, wound down in 1992 without meeting its own goals. This second winter stretched into the early 2000s, more than a decade during which building a career around neural networks was, by most accounts, a good way to not get funded.
The Comeback Nobody Called
Geoffrey Hinton kept working on neural networks through most of that stretch anyway, refining the backpropagation training method he had helped popularize back in 1986, and running a lab at the University of Toronto that trained a small number of researchers who would go on to define the field. In 2012, two of them, Alex Krizhevsky and Ilya Sutskever, built a network under Hinton's supervision called AlexNet and entered it in an annual computer vision contest called ImageNet, where systems compete to correctly label objects across 1.2 million photographs sorted into a thousand categories. AlexNet won by a margin nobody had seen before: an error rate of 15.3 percent against a second-place finisher at 26.2 percent, a gap of almost eleven points in a competition where progress usually moved by fractions of a point. The result rested on an idea the field had circled for years without being able to prove: that a large enough neural network, trained on enough data using the raw processing power of ordinary graphics cards instead of specialized hardware, could outperform every hand-built alternative. Sutskever later described it as the moment that vindicated the bet that scale, more data and more compute, would keep paying off. Neural networks, the exact approach Minsky and Papert had helped bury in 1969, were credible again.
The Second Piece
AlexNet proved neural networks could see. It took another five years for a second breakthrough to give them something closer to a working memory for language. In 2017, eight researchers at Google published a paper called "Attention Is All You Need," introducing the transformer architecture that Part 2 of this series already walked through: the attention mechanism that lets a model weigh which earlier words in a passage matter most to the word it is working on right now. The older approach, a family of models that read text one word at a time in strict sequence, could not be parallelized across modern chips, which capped how large and how fast those models could ever get. Attention removed that ceiling. Every major language model built since, GPT, Claude, and everything that followed, descends directly from that one paper. All eight of its authors have since left Google.