Chapter 5: AI Did Not Happen Overnight
The illusion of sudden arrival
For most people, artificial intelligence arrived on November 30, 2022, the day ChatGPT was released to the public. Within two months it had reached an audience faster than any consumer product before it, and the conversation shifted almost overnight from does this exist to what does this mean.
That date is real, and it matters. But it marks the arrival of an interface, not a technology. The science underneath it had been developing for seventy years, in public, with published papers and known researchers and a long record of predictions that did and did not come true.
I want you to know that history for a practical reason, not an academic one. This field has been through several cycles of enormous excitement followed by collapse, and the people who navigated those cycles well were the ones who understood which claims rested on demonstrated results and which rested on extrapolation. You are living through the largest such cycle yet. The history is your calibration.
1950s: The question is posed
In 1950, Alan Turing published a paper asking whether machines could think, and (recognizing that the question was hopelessly vague) replaced it with a test. If a person conversing with a machine through text could not reliably tell it from a human, on what grounds would we deny it thought?
Six years later, a summer workshop at Dartmouth College gathered the small group of researchers who would define the field. They coined the term artificial intelligence, and they proposed that a significant advance could be made if a carefully selected group worked on it together for a summer.
That estimate was off by roughly seventy years. It also established a tradition of confident timelines that the field has never quite broken, including at present.
The early results were nonetheless real. Programs proved mathematical theorems, solved algebra problems, played checkers competently. The prevailing belief was that intelligence was fundamentally a matter of symbol manipulation and logical rules, that if you could write down enough rules precisely enough, thinking would follow.
1960s–1980s: Rules, expert systems, and the first winters
That belief produced decades of work on what is now called symbolic AI: encoding human knowledge as explicit rules a machine could follow.
The high point was the expert system. Interview a specialist (a physician, a geologist, a chemist), extract their decision-making into thousands of if-then rules, and encode it. Some of these worked well enough to be commercially deployed, and by the 1980s there was a real industry around them.
After finishing my master’s degree in artificial intelligence, I developed an expert system for an auto-parts company.
An expert system is an early form of artificial intelligence designed to imitate the decision-making and problem-solving abilities of a human specialist. Instead of learning statistically from massive datasets, as today’s AI systems do, it relies on explicitly programmed logical rules and an organized collection of facts about a particular field.
What I created may sound ordinary today, but at the time it was quite innovative. The application helped every employee at the auto-parts company recommend the additional components a customer might need when purchasing a particular part.
With the system’s guidance, even a newly hired employee—someone who might have been working at McDonald’s the week before—could make recommendations like an experienced mechanic. The computer would explain that if the customer replaced Part A, they would probably also need Parts B and C.
This was good for the customer, good for sales, and especially good for the new employee, who no longer had to pretend that staring thoughtfully at Part A would somehow reveal the existence of Parts B and C.
Some forms of expertise, however, could not be translated into computer rules. Expert systems eventually hit a wall, deflating many of the high hopes that had surrounded them.
Human expertise turns out not to be made all of rules. Ask a doctor how they knew something was wrong and you often get “the patient just looked unwell”: a judgment drawn from thousands of unarticulated pattern matches. You cannot interview that out of someone. And the systems were brittle in a distinctive way: superb inside their narrow domain, catastrophically stupid one inch outside it, with no ability to notice they had crossed the line.
Funding collapsed twice, once in the mid-1970s and again in the late 1980s. These periods are called AI winters, and they followed the same pattern both times: striking demonstrations, extrapolation from those demonstrations to imminent general capability, failure to deliver, withdrawal of money and credibility.
Anyone telling you today that AI progress is guaranteed to continue at its current rate should be asked about the winters. Anyone telling you it is certain to stall should be asked why this time resembles those times.
The quiet parallel track: learning instead of rules
Alongside the rule-writers, a smaller group pursued a different idea: instead of programming the knowledge in, build a system that learns it from examples.
The inspiration was loosely biological. A brain is a network of simple units connected to each other, and it learns by adjusting the strength of those connections. Frank Rosenblatt built an early version, the perceptron, in the late 1950s.
In 1969, Marvin Minsky and Seymour Papert published a rigorous analysis showing what simple perceptrons fundamentally could not do. The critique was mathematically sound, and its practical effect was to nearly kill the approach for fifteen years.
The revival came in the 1980s with a training technique called backpropagation, which made it possible to train networks with multiple layers: to work out how much each connection in a deep stack contributed to the error, and adjust accordingly. The math worked. The results were modest, for a reason that would not be obvious for another twenty years: the method was sound, but the field had neither the data nor the computing power to feed it.
1997 and 2011: Two famous machines that mattered less than people thought
In 1997, IBM’s Deep Blue defeated world chess champion Garry Kasparov. In 2011, IBM’s Watson won at Jeopardy! against the game’s strongest human players.
Both were genuine engineering achievements and both were widely misread. Deep Blue did not think about chess; it evaluated positions at enormous speed using rules written by human experts. Watson was a sophisticated information-retrieval system. Neither could do anything else at all.
The lesson is one you will use constantly when reading AI news: a system that performs superbly at one visible task tells you almost nothing about its general capability. That was true of Deep Blue and it remains a live question about the systems you are using today.
From Business Intelligence to AI-Assisted Decisions
Business intelligence gathers information from sales, finance, operations, customers, and other sources, then transforms it into reports, dashboards, charts, and performance indicators. It gives executives something they could not easily obtain before: a clear view of the business based on evidence rather than intuition alone.
Decision support systems took this idea a step further. Instead of merely showing managers what had happened, they helped them evaluate what might happen next and what they could do about it. A decision support system could compare alternatives, test assumptions, project possible outcomes, and answer questions such as: What happens to profits if material costs rise? Which location is best for a new store? How much inventory should we order? The computer supplied the analysis, but the manager remained responsible for the decision. That principle continues in modern AI: technology can expand human judgment without replacing human accountability.
Big data was another important development in this evolution, although some would call it a parallel movement rather than a direct step toward AI. Either way, it gave organizations access to a vast digital record of their operations, customers, competitors, and the wider world. Business intelligence organized that information. Decision support systems used it to compare choices. Modern AI can now search it, summarize it, identify patterns, make predictions, and explain possible courses of action in ordinary language. Each generation has made the computer a more capable participant in decision-making, while leaving the final responsibility where it belongs: with the human being.
2012: The turn
In 2012, a neural network called AlexNet entered an image-recognition competition and won by a margin so large that the field reorganized around it within about a year.
Nothing conceptually new had been invented. The architecture was decades old. What had changed was the arrival of two missing ingredients at once: the internet had produced enormous labeled datasets, and graphics processors (built for video games) turned out to be exceptionally good at the specific mathematics neural networks require.
Old idea, plus data, plus compute. That combination is the engine of everything that followed, and it is the reason progress since 2012 has correlated so closely with how much money is spent.
The results came quickly. Speech recognition became usable. Machine translation went from a joke to a tool. Image recognition surpassed human accuracy on narrow benchmarks. In 2016, AlphaGo defeated one of the world’s best Go players: a game long considered resistant to computational brute force because of its intuitive character.
2017: The architecture that made everything else possible
In 2017, a group of Google researchers published a paper introducing an architecture called the transformer.
Its key idea, called attention, gave the model a way to weigh how much every part of a text relates to every other part: to work out, in the sentence the trophy didn’t fit in the suitcase because it was too large, that “it” refers to the trophy. Earlier approaches processed text in sequence and struggled to hold long-range relationships. Transformers considered everything at once.
The consequence that mattered most was practical: transformers parallelize well. They can be trained on enormous quantities of text using enormous quantities of hardware, which earlier architectures could not. That opened the door to scale.
Every major AI system discussed in this book is a descendant of that paper.
2018–2022: Scale, and a surprise
What followed was less a series of breakthroughs than a series of enlargements. Bigger models, more data, more computation: GPT-1 in 2018, GPT-2 in 2019, GPT-3 in 2020, each an order of magnitude beyond the last.
The surprise was that scaling kept working, and that it produced capabilities nobody had specifically engineered. A model trained only to predict the next word in text began doing arithmetic, translating between languages it was never explicitly taught to translate, writing functioning code, summarizing, and answering questions in formats it had never been shown.
This is worth sitting with, because it is the genuinely strange fact at the center of this technology. Nobody wrote a translation module. Translation fell out of prediction at sufficient scale, because to predict text well across the whole of human writing, you have to internalize a great deal about what that writing describes.
Researchers still argue about how deep that internalization goes. That argument is unresolved and you should be suspicious of anyone on either side who tells you it is settled.
Then, in November 2022, someone wrapped a model in a chat box and made it free.
The technology had existed for two years. The interface is what changed the world — the same lesson the personal computer taught in 1984 and the smartphone taught in 2007. Capability sitting behind a difficult interface reaches thousands of people. The same capability behind an easy one reaches hundreds of millions.
2023–2026: Four directions at once
The years since have been busy rather than revolutionary. Four developments matter for how you will use these tools.
Multimodality. Models stopped being text-only. They now take in images, audio, documents, and video, and produce images and audio as well as text. Practically, this is why you can photograph an error message or a rash or a spreadsheet and ask about it.
Reasoning. A category of models emerged that work through problems in extended internal steps before answering, rather than responding immediately. They are slower and more expensive, and substantially better at multi-step problems: math, logic, code, complex analysis. This is the origin of the “thinking” or “reasoning” toggle and of the tiered model menus now found in many tools.
Tools and retrieval. Models gained the ability to reach outside themselves: to search the web, run code, read files you provide, and connect to other software. This is the most underappreciated development of the period, because it directly addresses the failure mode of a system answering from frozen memory.
Agents. Systems that carry out multi-step tasks with reduced supervision, chaining actions together toward a goal. This is the current frontier, the current locus of overclaiming, and a subject that deserves careful treatment.
What the history teaches
Four things.
Progress is lumpy. Long plateaus, then sudden jumps when a missing ingredient arrives. This makes both “it will keep improving at this rate” and “it has hit a wall” unreliable predictions. Neither is knowable now.
The field has a chronic timeline problem. From the Dartmouth summer onward, serious people have consistently underestimated how long things take. Weigh predictions about the next five years against that record.
Demonstrations are not capabilities. Deep Blue, Watson, and a hundred impressive demos since. Ask what a system does when conditions change, not what it does in the video.
Interfaces matter as much as engines. The most consequential AI event of the last decade was a chat box. Watch for the next such moment in how people access the technology, not only in how the models improve.
Theory Versus Actual Results
Around the year 2000, I participated in a highly ambitious project. It was particularly interesting because the founders intended to compete in the home décor and remodeling market with two of the largest companies in the United States: Home Depot and Lowe’s, both Fortune 50 companies. At the heart of the project was technology that promised unprecedented capabilities in home-remodeling design. The company was also being positioned for a rapid initial public offering (IPO), all during the Internet boom and the bust that followed.
The founders planned to launch a new chain of stores. I was responsible for providing the company’s conventional information systems: the proven, standard technology needed to run the business at that time.
Another technology company was participating in the project. Its role was to provide a highly intelligent application that would allow customers to design their remodeling projects from home. Having studied artificial intelligence, I was highly skeptical that the state of the art could deliver the capabilities being promised, but it was not my place to challenge those claims.
The company appeared to have advanced systems capable of managing complex bills of materials, but it did not have the sophisticated graphical user interface that had been promised. The reality was that the technology and the processing power it required simply were not ready. Only now, approximately twenty-five years later, are systems with those capabilities becoming widely available.
Exercise 5.1: Find the winter in a current claim
Locate an article predicting where AI will be in five years. Identify precisely what its prediction rests on: a demonstrated result, an extrapolated trend line, or an assumption that current progress continues.
Then find the sentence where extrapolation begins. It is almost always identifiable, and it is almost always earlier than the article implies.
Exercise 5.2: Ask the machine about itself
Explain the history of AI from 1950 to today in about 500 words, for someone with no technical background.
Compare it against this chapter. Note anything it emphasized differently, anything it left out, and anything it stated that you cannot verify. Keep the list: Chapter 8 will explain what you found.
Learn More About Artificial Intelligence
Learn more about the CLEAR, TRUST and LEARN methods for Artificial Intelligence