A Brief History of AI

In November 2022, a chat box appeared on the internet and roughly a hundred million people tried it within eight weeks. If you were paying attention, it felt like AI arrived out of nowhere. It did not. The research behind that chat box had been running for seventy-two years, through two funding collapses, several rounds of public ridicule, and a long stretch where “AI researcher” was a slightly embarrassing thing to put on a business card.

This matters to you for one practical reason. Every week now, some company announces something and calls it a breakthrough. Most of those announcements are not breakthroughs. They are small improvements dressed up in big language, and if you cannot tell the difference, you will spend your attention badly. You will chase the loud thing and miss the important one.

By the end of this lesson you will know the actual shape of AI’s history, the three specific things that broke it open after decades of stalling, and a working test you can apply to any new announcement: does this change what is possible, or just what is faster. That single question will save you a lot of wasted enthusiasm.

1950 to 1956: The Question and the Promise

In 1950, the British mathematician Alan Turing published a paper that opened with a question he immediately admitted was badly formed: “Can machines think?” The word “think” was too slippery to argue about, so he replaced it with a game. Put a human judge at a keyboard. Have them exchange typed messages with two hidden participants, one human and one machine. If the judge cannot reliably tell which is which, Turing argued, arguing about whether the machine “really” thinks is a waste of breath. He called it the imitation game. We call it the Turing test.

Six years later, a small group of researchers spent the summer at Dartmouth College working on the problem. Their funding proposal contained a sentence that has aged into legend: they believed a significant advance could be made if a carefully selected group of scientists worked on it together “for a summer.” That workshop is where the term artificial intelligence was coined. The leading figures walked away predicting that a machine matching human general intelligence was perhaps twenty years out.

Here is the part people get wrong when they laugh at that prediction. It was not stupid. It was reasonable given what they had just seen. In under a decade, computers had gone from nothing to proving theorems in symbolic logic, some of them more elegantly than the textbook proofs. A program called the Logic Theorist re-derived proofs from Russell and Whitehead’s Principia Mathematica. If a machine could do formal mathematics, which most people consider the hardest thinking humans do, then surely the easy stuff, recognizing a chair, understanding a sentence, walking across a room, was just a matter of more programming.

That assumption was exactly backwards. The things humans find hard, like chess and logic, turned out to be relatively easy for computers, because they have clean rules and a fixed board. The things a four-year-old does without effort, like recognizing their grandmother’s face at a strange angle in bad light, turned out to be brutally hard. Researchers spent the next forty years learning this the expensive way.

The Two AI Winters: When the Money Left

Funding for AI collapsed twice. The first collapse came in the mid-1970s, after government reviews on both sides of the Atlantic concluded the field had produced impressive demos and very little that worked outside a laboratory. The second came in the late 1980s and ran into the mid-90s, when the commercial “expert system” industry, which had grown into a real business, essentially evaporated. Researchers call these the AI winters, and the name stuck because it captures the mood: the work continued, quietly, underfunded, unfashionable.

The bottleneck was not computer speed, though that was a factor. The bottleneck was the method. Every serious AI system of that era worked by hand-written rules. Humans sat down with domain experts and tried to write out intelligence as a list of if-then statements.

Try it yourself with something trivially easy. Write the rules for recognizing a cup. Rule one: it has a handle. Except many cups do not. Rule two: it holds liquid. So does a bathtub. Rule three: it is roughly cylindrical and fits in a hand. Now you have excluded a wide flat teacup and included a soup can. Every rule you add creates two exceptions, and the exceptions need their own rules, and those need exceptions. This is the actual reason the field stalled. Not a lack of cleverness. The real world contains more exceptions than any team can write down, and each one has to be anticipated in advance by a person.

1997, 2011, 2012: Three Famous Wins, Only One That Mattered

In May 1997, IBM’s Deep Blue beat world chess champion Garry Kasparov in a six-game match. The photographs of Kasparov with his head in his hands ran on front pages worldwide, framed as the moment machines surpassed us.

Be precise about what happened, because the popular version is misleading. Deep Blue evaluated around 200 million chess positions per second and picked the move that led to the best outcome. It had chess-specific knowledge tuned by grandmasters written directly into it. It did not learn. It could not be asked about anything but chess. It had no idea it was playing chess, or that a person sat across from it, or what a person is. It was a very fast, very narrow search machine, and it proved something real but limited: on a problem with fixed rules and a countable set of moves, raw computation beats human intuition. That is a genuine result. It is not thinking.

In February 2011, IBM’s Watson won Jeopardy against champions Ken Jennings and Brad Rutter, which was harder, because Jeopardy clues are full of puns and misdirection. Watson still ran on a large hand-engineered pipeline of specialized components. Impressive engineering, narrow system.

The turning point was quieter and got almost no press. In 2012, a neural network called AlexNet entered the ImageNet competition, an annual contest to correctly label photographs across a thousand categories. The best systems the previous year had an error rate around 26 percent. AlexNet came in at roughly 15 percent. In a field where a one-point improvement was a good year’s work, it cut the error by more than a third in a single jump.

Two things made it work. It was a deep neural network trained on 1.2 million labeled images, and it was trained on GPUs, graphics chips designed for video games. That combination was the proof of concept. Within three years essentially every serious computer vision team on earth had switched to neural networks. If you want a single date where modern AI actually starts, it is 2012, not 2022.

2017: The Transformer, the Engine Under Everything You Use

In 2017, a team of researchers at Google published a paper with the confident title “Attention Is All You Need.” It introduced an architecture called the Transformer, and it is the single most consequential piece of AI engineering of the last twenty years. The “T” in GPT stands for Transformer. Claude and Gemini are built on the same foundation.

Before it, language models read text the way you read a ticker tape: one word, then the next, each step depending on the one before. That created two problems. By the end of a long paragraph, the beginning had faded. And because each step waited on the previous step, you could not spread the work across many processors. Training was slow, and slow training caps how much data you can learn from.

The Transformer let the model take in a whole passage at once and, for every word, calculate how much every other word should influence its meaning. That mechanism is called attention. Consider the sentence: “The trophy would not fit in the suitcase because it was too big.” What does “it” refer to? You know instantly it means the trophy, because a thing that does not fit is the larger thing. Change one word to “too small” and “it” flips to the suitcase. A ticker-tape reader struggles with this. A Transformer looks at the entire sentence simultaneously and assigns “it” a strong connection to “trophy,” a weaker one to “suitcase,” and near-zero to “the.” That weighting is learned from data, not written by hand.

Now finish the analogy, because the mechanism only half explains the impact. Reading the whole passage at once did not just improve comprehension. It meant training could run in parallel, thousands of chips working on different pieces at the same time instead of standing in a queue. That is what made scale possible. Feed it ten times the text on ten times the hardware and it gets meaningfully better rather than just slower. Every large model since is a version of that bet.

Use that on anything you need to get oriented in quickly: your own industry, a client’s sector, a topic your kid is studying. The final instruction is the valuable part. Forcing a defense of one pivotal entry turns a flat list into an argument you can agree or disagree with, which is how you actually learn a field.

November 2022: The Interface, Not the Model

ChatGPT launched on 30 November 2022 and reached an estimated 100 million monthly users within about two months, the fastest consumer adoption recorded at the time. For comparison, Instagram took roughly two and a half years to reach the same number.

Here is the detail almost everyone misses. The underlying model was not new. GPT-3 had been publicly available to developers since 2020, two full years earlier. The capability sat there, largely unused by the general public, because using it meant writing code, managing an API key, and phrasing your request as a text-completion problem rather than a question.

What changed in November 2022 was a text box on a web page and a model tuned to behave like something you could talk to. That is it. The technical work behind that tuning was real and difficult, but the leap in adoption came from removing the requirement to be a programmer. It is the same pattern as the web: the internet existed for years before the graphical browser, and almost nobody used it. Access, not capability, was the constraint. Remember that the next time someone tells you a product is revolutionary. Ask whether the underlying ability is new, or whether they just made an existing ability easier to reach. Both are valuable. They are not the same claim.

2023 to 2026: The Era You Are Standing In

Three shifts define the period since ChatGPT launched, and you will meet all three in this course.

Multimodal models. Early ChatGPT handled text only. Current systems take in images, audio, documents and video in the same conversation. Practically, that means you can photograph a confusing insurance letter and ask what it obligates you to do, or hold up a broken appliance part and ask what it is called so you can order a replacement.

Large context windows. The context window is how much the model can hold in mind at once. Early versions managed a few thousand words. Current systems handle hundreds of thousands, which is the difference between pasting a paragraph and uploading an entire 200-page contract, then asking which clauses expose you.

Agents. Instead of answering and stopping, an agent takes multiple steps toward a goal: searching the web, reading results, writing a file, checking its own work. This is the least mature of the three and the one most oversold right now. Agents fail in ways that are hard to notice, because a confident wrong answer looks the same as a right one.

One quiet consequence matters to you directly. Competition drove capability into the free tiers. The tools you will use in this course cost nothing, and they are stronger than what enterprises paid heavily for in 2023. That is why this course is free to take.

The Three Things That Broke the Logjam

If you remember one thing from this lesson, make it this. Modern AI required three ingredients to arrive together, and none of them alone was enough.

Test the claim by removing one. The Transformer with 1980s data and 1980s chips would have been an interesting paper nobody could run. GPUs and the whole internet without the Transformer gets you 2015-era systems: good at labeling photos, poor at holding a conversation. All the data and the architecture without hardware to train on leaves you with a design that would take centuries to finish. Three legs. Remove any one and it falls over.

This is also your filter for announcements. When you read that a new model is out, ask which of the three legs it advances. Usually the answer is “none, it is a bigger version of the last one,” and that is fine, it is just not a breakthrough.

Common Mistakes to Avoid

• Thinking AI was invented in 2022. The research is over seventy years old and the technical foundations of what you use today were laid in 2012 and 2017. November 2022 is when the interface got good enough for ordinary people. Getting this wrong leads you to treat AI as a fad that might pass, when it is a seventy-year research programme that finally reached usable form.
• Assuming every new model release is a breakthrough. The overwhelming majority are incremental: a few percentage points on a benchmark, faster responses, slightly cheaper. Real breakthroughs are rare, maybe one every few years, and you recognize them because they change what is possible, not what is faster. AlexNet made image recognition work at all. The Transformer made scaling worthwhile. Most of what you will read about this month does neither.
• Assuming the tool knows what happened last week. Models are trained up to a cutoff date and then frozen. Unless a tool is actively searching the web, its knowledge of recent events may be months out of date, and it will often answer confidently anyway.

Where This Leaves You

The honest position is that we are early. A useful comparison is the internet in 1998. It was obviously important, the money was pouring in, and most of what was actually being built was bad. Corporate websites were brochures nobody read. Plenty of serious people said the whole thing was overhyped, and in the short term they were right, the crash came in 2000. They were wrong about the long term.

The people who quietly learned how it worked in 1998, not the ones who invested in pet food delivery, spent the next twenty years with an advantage that compounded. They were not smarter. They just started earlier and stayed practical. You are standing at the same point with this. That is the whole reason to take a course like this now rather than in five years.

Key Takeaways

• AI stalled for forty years because researchers tried to hand-write the rules of intelligence, and the real world contains more exceptions than anyone can write down.
• The decisive shift was moving from systems that are told rules to systems that are shown examples and find the patterns themselves. AlexNet in 2012 proved it worked.
• Modern AI needed three things at once: internet-scale data, GPUs, and the Transformer architecture. Any two without the third produces nothing you would recognize as ChatGPT.
• The Transformer mattered because it read whole passages in context and trained in parallel, which is what made throwing more data and hardware at the problem actually pay off.
• ChatGPT’s launch was an interface breakthrough, not a capability breakthrough. The model had been available to developers for two years and almost nobody used it.
• Judge every announcement by one question: does it change what is possible, or only what is faster and cheaper? Nearly all of them are the second kind.